A foundation model for electrodermal activity data

A f ound a tion model f or electr odermal a ctivity d a t a Leonardo Alc hieri Univ ersità della Svizzera Italiana (USI) Via la San ta 1 6962 Lugano, Switzerland leonardo.alchieri@usi.ch Matteo Garzon Univ ersità della Svizzera Italiana (USI) Via la San ta 1 6962 Lugano, Switzerland Lidia Alecci Univ ersità della Svizzera Italiana (USI) Via la San ta 1 6962 Lugano, Switzerland F rancesco Bombassei De Bona Univ ersità della Svizzera Italiana (USI) Via la San ta 1 6962 Lugano, Switzerland Martin Gjoreski Univ ersità della Svizzera Italiana (USI) Via la San ta 1 6962 Lugano, Switzerland Gio v anni De F elice Univ ersità della Svizzera Italiana (USI) Via la San ta 1 6962 Lugano, Switzerland Silvia San tini Univ ersità della Svizzera Italiana (USI) Via la San ta 1 6962 Lugano, Switzerland Marc h 19, 2026 Abstract F oundation mo dels hav e recen tly extended b eyond natural language and vision to timeseries domains, including physiological signals. Ho wev er, progress in electro dermal activity (ED A) mo deling is hindered b y the absence of largescale, curated, and openly accessible datasets. ED A reﬂects sympathetic nervous system activity and is widely used to infer cognitive load, stress, and engagemen t. Y et v ery few wearable devices provide contin uous, unobtrusiv e sensing, and the only largescale archiv e to date is proprietary . T o address this gap, we compile EDAMAME, a collection of ED A traces from 24 public datasets, comprising more than 25,000 hours from 634 users. Using this resource, w e train UME, the ﬁrst dedicated foundation model for ED A. In eight out of ten scenarios, UME outp erforms baselines and matc hes generalist timeseries foundation mo dels while using 20 × fewer computational re- sources. Our ﬁndings, how ever, also highlight the intrinsic challenges of ED A mo deling, motiv ating further researc h to unlo ck its full p otential. All datasets, model weigh ts, and co de are released to supp ort further research. 1 In tro duction Thanks to their ability to learn general patterns from broad, diverse data and to adapt them to a wide range of downstream tasks, foundation models hav e attracted atten tion b eyond their traditional applications in natural language pro cessing and computer vision. In particular, generalist foundation mo dels for time series , lik e Mantis [35] or Chronos [ 8 , 9 ], hav e b een prop osed recently . T rained on cross-domain dataset A f ounda tion model f or electr odermal activity da t a collections, these mo dels achiev e remarkable p erformance on a diverse range of do wnstream tasks. Emerging foundation mo dels for physiolo gic al time series data – e.g., for photoplethysmogram (PPG) [ 1 , 72, 81, 32, 62], electro cardiogram (ECG) [ 1 , 66, 59], and electro encephalogram (EEG) data [84] – show comp elling p erformance, too. Y et their training remains constrained by the absence of largescale, curated, and op en datasets for physiological signals. As highlighted b y Abbasp ourazad et al. [ 1 ] with resp ect to time series data in the medical domain, “ datasets ar e usual ly smal l in c omp arison to other domains, which is an obstacle for developing neur al network mo dels for biosignals ” . This paucity of data is particularly acute for electrodermal activity (EDA) signals. EDA refers to changes in the skins electrical conductance caused by v ariations in sweat gland activity , whic h is itself regulated b y the sympathetic branc h of the autonomic nerv ous system. Alongside more commonly used physiological sig- nals, EDA has numerous applications in p ersonal informatics systems. Sp eciﬁcally , because ED A increases with physiological arousal, it is commonly used to assess cognitiv e load [52, 79], stress [39, 73], and engage- men t [36, 38, 19]. Y et ED A sensors are not (y et) commonplace on w earable devices and only few (such as, e.g., the Empatica E4 wristband) feature the unobtrusive, contin uous measurement mo dality necessary to collect longitudinal traces of EDA in real-w orld settings. The only large-scale arc hive of ED A data is a proprietary dataset collected using the Fitbit Sense 2 [65]. This lack of (op en) data has, to date, hamp ered the developmen t of foundation mo dels for EDA. T o cop e with this challenge, w e ﬁrst constructed a large-scale archiv e of EDA data that in tegrates records from 24 diﬀeren t publicly av ailable datasets and a total of more than 25’000 hours of data of 634 diﬀerent users. W e curated this collection of datasets, to which we refer to as EDAMAME ( ED A M ulti-dataset A rc hive for M o del training and E v aluation), to train a foundation mo del for EDA data. The mo del, called UME (fo U ndation M o del for E lectro dermal activit y data), has b een trained on approximately 275 million of 60-second windows of EDA data and it is, to b est of our knowledge, the ﬁrst ED Asp eciﬁc foundation mo del rep orted in the literature. W e ev aluated UME on several downstream tasks and found that it surpasses baseline mo dels trained on b oth generic and EDA-speciﬁc handcrafted features in 8 out of 10 tests and matc hes the p erformance of generalist time series foundation mo dels, while requiring at least 20 × less computational resources. Our results, how ev er, also highlight the intrinsic diﬃcult y of w orking with EDA signals: balanced accuracy scores rarely exceed 0.7 and exhibit substantial v ariability . F urther researc h is therefore needed to fully harness the potential of EDA for b oth unimo dal and multimodal ubiquitous sensing. T o enable further research, we make all artifacts pro duced as part of this w ork publicly av ailable. The ED AMAME dataset collection can b e obtained up on signing a single data sharing agreement, in accordance with the provisions of the individual datasets. UME’s weigh ts and the entire co debase used to pro duce the mo del is av ailable for download: [Link to the rep ository anonymized] 1 . 2 Bac kground & related w ork In this section, we provide an o verview of the background and relev an t related work. W e discuss the relev ant background associated with w earable devices and EDA data. Then, we describ e existing researc h on foundation models for ph ysiological data, with a focus on signals collected from wearable devices. W e also present work on the use of self-sup ervised learning applied to EDA data, since self-sup ervised learning is one of the main building blo cks for training foundation mo dels. Bac kground F oundation mo dels require large scale datasets to b e trained [14]. Researchers ha ve obtained state-of-the-art results in the NLP domain thanks in part to the a v ailability of large scale textual corp ora, e.g., [10, 17]. Researc hers create textual corp ora by , for example, cra wling websites and gathering all av ail- able text, e.g., as done by the Common Crawl dataset 2 . On the other hand, collecting physiological signals requires the use of sp ecialized equipment as well as signiﬁcant human and time resources [60, 83]. Researchers ha ve explored foundation mo dels for ph ysiological data through the use of large scale clinical corpora [58]. Only recen t work, e.g., [ 1 ], explores the use of foundation mo dels trained from real-w orld w earable ph ysio- logical signals. This is due to signiﬁcantly higher op erational diﬃculties in gathering high-quality data when collecting wearable data, compared to clinical studies [13]. Additionally , EDA, even if used in the healthcare domain to detect seizures [21], is not as commonly collected in clinical settings as PPG or ECG [25, 74]. 1 The link to the rep ository will b e made av ailable up on completion of the reviewing pro cess 2 https://commoncrawl.org 2 A f ounda tion model f or electr odermal activity da t a T able 1: Overview of selected foundation mo dels for physiological data and the datasets used to train them. F oundation mo del train data Ref. Signal(s) P arams. A v ail. Dataset(s) Data type Hours (k) Users A v ail. Clinical data [72] PPG ∼ 30 - 150 M Op en source VitalDB [57] Clinical ∼ 17 ∼ 6k 3 MIMIC-I I I [50] Clinical ∼ 20 ∼ 6k 3 MESA [94] Clinical ∼ 20 ∼ 2k 3 [32] PPG ∼ 1-8 M Code only [32] Clinical ∼ 300 ∼ 29k 3 [66] ECG ∼ 300 M Open source MIMIC IV [51] Clinical ∼ 140 ∼ 160k 3 Ph ysioNet2021 [78] Clinical ∼ 14 N/A 3 UHN-ECG [66] Clinical ∼ 100 ∼ 180k 3 [59] ECG ∼ 30M Op en source HEEDB [54] Clinical ∼ 1,000 ∼ 2M 3 [84] EEG not stated Op en source curated collection Clinical ∼ 350 ∼ 9 7 3 W earable data [ 1 ] PPG ∼ 3M Priv ate AHMS [88] W earable ∼ 300 ∼ 141k 7 ECG ∼ 3M Priv ate ∼ 30 ∼ 106k 7 [68] multi-modal ∼ 1-100M Priv ate [68] W earable ∼ 40,000 ∼ 165k 7 [81] PPG ∼ 30M Op en source MOODS [69] W earable ∼ 40 ∼ 120 7 Other [62] multi-modal ∼ 140M Op en source curated collection Mixed ∼ 15 N/A 7 3 Ours Ours ED A ∼ 1M Op en source EDAMAME W earable ∼ 25 ∼ 630 3 F oundation mo dels for physiological data In T able 1 we provide an ov erview of a selected set of existing foundation models for physiological data, their size, their a v ailabilit y and information about the data used to train them. Current research fo cuses on PPG, ECG, or m ulti-mo dal approac hes. A ma jority of the selected foundation mo dels relies on clinical data, given its abundance. T o the b est of our knowledge, Saha et al. [81]’s work is the only one using exclusively w earable physiological data to train an op en source foundation mo del. Researc hers hav e explored foundation mo dels for physiological data using b oth clinical and wearable data. Abbasp ourazad et al. [ 1 ] w ere the ﬁrst to use a large scale, priv ate, dataset of wearable data to train foundation mo dels on PPG and ECG data. Their results sho w that features from foundation mo dels can b e used to p erform multiple downstream tasks, with p erformance on par to existing approaches. F ollowing these results, other researc hers, e.g., [81, 72, 32, 66, 59], ha v e trained foundation mo dels for either PPG or ECG data. How ev er, most of the existing approaches rely on clinical PPG and ECG signals, with limited w ork on real-world physiological data collected from wearable devices [81], and, esp ecially , no work on EDA data. Multi-mo dal approac hes hav e also recently been prop osed. Nara y answam y et al. [68] trained a family of m ultimo dal foundation models on data aggregated o v er 1 min ute. They compute aggregated features from PPG (e.g., mean heart rate), EDA (e.g., mean ED A v alues), and other physiological and b ehavioral signals. They ﬁnd that their foundation mo del signiﬁcan tly outp erforms (up to 50 % improv emen t) ov er baseline metho ds. How ev er, their approach relies on proprietary data and the authors released neither the co de nor the weigh ts for their mo del. Using publicly av ailable data, Luo et al. [62] trained a multi-modal foundation mo del for physiological data, e.g., PPG, ECG, EDA, from a curated collection of datasets with ab out 15’000 hours of data. In their exp erimental setup, they show promising zero-shot do wnstream p erformance for their foundation model. Ho wev er, the authors’ fo cus w as not on EDA data, which means that their collection of datasets only contained a limited amount of ED A, compared to other physiological signals. Overall, a curated EDA dataset collection has not y et b een made av ailable to the wider research communit y . Self-sup ervised learning for EDA data Researc hers ha ve also explored the use of self-sup ervised learn- ing on wearable data, including ED A. Dissanay ak e et al. [33] trained a self-sup ervised learning mo del, through con trastive learning, for emotion estimation using PPG, ED A and skin temperature. Saeed et al. [80] sim- ilarly used a multi-modal self-supervised approac h in a federated learning settings. Recently , Matton et al. 3 Individual datasets av ailable, but not the curated collection. 3 A f ounda tion model f or electr odermal activity da t a [64] trained a self-sup ervised learning model on ED A data to p erform stress classiﬁcation. They used con- trastiv e learning and data augmentation techniques to achiev e state-of-the-art p erformance on in-distribution do wnstream tasks. Ov erall, existing work suggests that foundation mo dels for physiological data can b e trained and used on a div erse set of downstream tasks when using PPG, ECG or m ulti-mo dal approaches. How ev er, a large part of publicly a v ailable physiological foundation mo dels rely on clinical data, and there is limited evidence that features from these mo dels can be used with data from wearable devices and real-life settings. Recen t work on m ulti-mo dal [62] and PPG [81] foundation mo dels shows that, even without large-scale proprietary datasets, researc hers are able to train op en source foundation mo dels. Ho wev er, recen t work do es not provide a readily av ailable collection of datasets for researchers to build up on their results, hindering the dev elopment of w earable foundation mo dels. Given promising w ork on self-supervised learning for EDA data [64], we create a collection of datasets, EDAMAME, which we make av ailable to the research communit y and, with it, we train an op en source foundation mo del for EDA data, UME, whose co de and weigh ts w e make av ailable to the research communit y . 3 ED AMAME: a collection of electro dermal activit y datasets In this section, we describ e how we address the av ailability of large scale datasets containing wearable ED A data through EDAMAME ( ED A M ulti-dataset A rc hive for M o del training and E v aluation). EDAMAME is a collection of existing, smaller scale, datasets prepared and pre-pro cessed in a uniﬁed manner. The goal of EDAMAME is to enable researchers to train foundation mo dels for wearable EDA data. 3.1 Description of the collection of datasets Our goal is to create a large-scale and div erse corpus to address the scarcity of EDA-speciﬁc collections. T o this end, w e ﬁrst iden tify a set of scenarios that are relev an t for EDA data through relev ant literature [47]. The scenarios are: sle ep monitoring , str ess/emotion induction , engagement , r e al-world str ess , workplac e analysis , daily living . Then, we select and search datasets from the literature using the following criteria: 1. datasets hav e to contain r aw ED A data from wearable devices; 2. the individual datasets hav e to b e either op en source or av ailable to researchers up on signing a data sharing agreement; 3. datasets hav e to contain data collected during tasks or moments that inﬂuence EDA signals; 4. the collection has to con tain datasets collected using diﬀerent proto cols. e.g., lab environmen t or in-the-wild collection; 5. the collection needs to con tain at least one dataset from each one of the six scenarios deﬁned; 6. in order to limit the searc h size, we also add a saturation criterion: once we reach the threshold of 100 users and 1000 hours for one of the six scenarios, we stop searching for additional data; 7. the total size of ED AMAME has to b e more than 15’000 hours and more than 100 users, which is a size similar to that of datasets used to train existing op en source foundation mo dels for wearable data [81, 62]. In order to satisfy the ﬁrst criterion, we collect data only from datasets using Empatica E4 devices 4 . W e c ho ose to use only Empatica E4 data since it is one of the few research-grade w earable devices that allow to con tinuously collect wearable EDA. W e searc h datasets through op en source databases con taining physiological data, sp eciﬁcally Ph ysioNet 5 , Zotero 6 , Kaggle 7 . W e also search for dataset information using databases of scien tiﬁc articles, e.g., Go ogle Sc holar 8 , ACM Digital Library 9 . Finally , we also searc h through the online databases of scientiﬁc journals 4 https://www.empatica.com/research/e4/ 5 https://physionet.org/ 6 https://zenodo.org/ 7 https://www.kaggle.com 8 https://scholar.google.com 9 https://dl.acm.org 4 A f ounda tion model f or electr odermal activity da t a T able 2: Summary of the datasets included in the training split. The collection spans div erse physiological scenarios and environmen ts. Dataset Name Duration (h) # Users Scenario En vironmen t APSync [38] 168 27 Real-world Stress Wild BIG IDEAS [11] 2607 16 Daily Living Wild BiHeartS [ 3 ] 2911 11 Sleep Monitoring Wild DREAMT [92] 882 100 Daily Living Wild Dynamics in the w orkplace [61] 1813 42 W orkplace Analysis Wild EmpaticaE4Stress [20] 5 29 Stress/Emotion Induction Lab EPM-E4 [37] 22 47 Engagement Wild HeartS [ 2 ] 886 5 Sleep Monitoring Wild HHISS [42] 3166 46 Real-world Stress Wild LA UREA TE [55] 1406 46 Engagement Wild M2Sleep [40] 8403 16 Sleep Monitoring Wild MEF AR [28] 27 23 Stress/Emotion Induction Lab Nurses’ Stress [48] 800 18 W orkplace Analysis Wild PPG-Dalia [77] 38 15 Daily Living Wild SEED [29] 340 31 Engagement Wild SEED-I I-Lab [29] 29 25 Engagement Lab SEED-I I-Wild [29] 156 6 Engagement Wild Stress Predict [49] 32 35 Stress/Emotion Induction Lab T oadStool [85] 10 10 Stress/Emotion Induction Lab USILaughs [30] 1.5 30 Stress/Emotion Induction Lab WEEE [41] 17 17 Daily Living Lab WESAD [82] 29 15 Stress/Emotion Induction Lab WESD [75] 123 10 Real-world Stress Wild W orkplace [31] 830 14 W orkplace Analysis Wild T otal 24’735 634 that publish datasets containing physiological data, e.g., Nature Scientiﬁc Data 10 , IMWUT 11 . W e identify a p oten tial 37 datasets. Through the aforementioned criteria, we select a total of 24 datasets for EDAMAME. W e rep ort in T able 2 an o verview of the 24 datasets. In total, EDAMAME contains approximately 25’000 hours of EDA data from 634 users. The size of EDAMAME is in line, as outlined by our sev en th selection criterion, with collections of datasets used by Saha et al. [81], Luo et al. [62] to train op en source foundation mo dels for wearable physiological data. The EDA signal in all 24 datasets is sampled at 4 Hz , since this is the default sampling rate for ED A from the Empatica E4 12 . All timestamps are con v erted to UTC time. Whenev er timestamps are missing, w e assign unix-time 0 to the start of a timeseries of EDA data. 3.2 Data v ariability in EDAMAME In this section, we sho w the data v ariability in the EDAMAME collection of datasets. W e highlight the div ersity of EDA data in EDAMAME to highlight its feasibility in training foundation mo dels for EDA data. In particular, we discuss the EDA data distribution and the distribution of users and data p er user across the datasets. Figure 1 shows six example EDA signals from diﬀerent datasets, represen ting six distinct scenarios: sle ep monitoring , str ess/emotion induction , engagement , r e al-world str ess , workplac e analysis , daily living . Data distribution Figure 2 shows the distribution of raw ED A v alues across the 24 datasets in ED AMAME. W e highlight ho w data distribution is similar across datasets that are collected in similar scenarios. W e highlight that higher EDA v alues are asso ciated with stronger ANS arousal activ ation [15]. F or example, datasets in the Str ess/Emotion Induction scenario hav e longer tails, in their distribution, com- pared to others. On the other hand, datasets in the R e al-world Str ess scenario ha v e low er tails and more 10 https://www.nature.com/sdata/ 11 https://dl.acm.org/journal/imwut 12 https://www.empatica.com/blog/decoding- wearable- sensor- signals- what- to- expect- from- your- e4- data 5 A f ounda tion model f or electr odermal activity da t a 0 10 20 30 40 50 T ime (s) 0.72 0.74 0.76 0.78 0.80 0.82 0.84 0.86 E D A ( S ) Daily Living (WEEE) 0 10 20 30 40 50 T ime (s) 0.08 0.10 0.12 0.14 0.16 0.18 E D A ( S ) Engagement (SEED) 0 10 20 30 40 50 T ime (s) 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 E D A ( S ) Real-world Str ess (APSYNC) 0 10 20 30 40 50 T ime (s) 1.45 1.46 1.47 1.48 1.49 1.50 1.51 E D A ( S ) Sleep Monitoring (BiHeartS) 0 10 20 30 40 50 T ime (s) 0.86 0.88 0.90 0.92 0.94 0.96 0.98 E D A ( S ) Str ess/Emotion Induction (CLUES) 0 10 20 30 40 50 T ime (s) 8.50 8.75 9.00 9.25 9.50 9.75 10.00 E D A ( S ) W orkplace Analysis (W orkplaces) Figure 1 : Example of six EDA signals (randomly drawn). Each signal is from a diﬀeren t dataset, each for one of the six scenarios. ED A v alues close to 0 µ S . These tw o distinct patterns highligh t the challenges of real-w orld EDA data: a lo wer distribution of EDA v alues in the R e al-world Str ess scenario suggests that during daily life less arousal- asso ciated moments are present. Ov erall, ED AMAME contains div erse data, across b oth scenarios and EDA v alues. This diversit y is imp ortant when training foundation mo dels, since it exp oses the model to a wide range of EDA data. Num b er of users and data p er user Figure 3 shows the num b er of users compared to the data density p er user, for each dataset in EDAMAME. The plot shows the structural top ology of the collection of datasets. Sp eciﬁcally , the collection contains tw o types of datasets: datasets that contain less users but more data p er user (longitudinal depth); and datasets that contain more users, but less data p er user (high p opulation div ersity). The ﬁrst set of datasets in which multiple days of data for eac h user are present contains more in tra-p ersonal v ariability information, i.e., how p eople’s data c hanges ov er time, compared to the second set. The second set, on the other hand, contains more in ter-p ersonal v ariability information, i.e., how data diﬀers b et ween diﬀerent participants, compared to the ﬁrst one. ED AMAME contains data from a diverse range of scenarios and users. The collection also balances datasets with longitudinal depth and datasets with high p opulation diversit y . This diversit y and v ariabilit y is in line with recent literature on foundation mo del train data [23, 18]. 4 UME: op en source foundation mo del for EDA data In this section, we describ e UME (A foUndation Mo del for Electro dermal activit y data), the ﬁrst open source foundation mo del trained on wearable EDA data. In particular, w e explain the data pre-pro cessing, the training criteria, the mo del architecture and additional information ab out training the mo del. With UME, our ob jective is to obtain a mo del whose internal representations enco de k ey c haracteristics of the input EDA signal, suc h that these represen tations can b e extracted and used as substitutes for domain- informed features in downstream tasks. T o ac hieve this objective, as common for foundation mo dels for ph ysiological data, e.g., [ 1 , 72, 81], we rely on self-sup ervised learning with a con trastiv e learning ob jective. T o train UME we use a subset of the EDAMAME collection of datasets, consisting of around 19’000 hours of EDA data and approximately 400 users. W e op en source the co de and the weigh ts for the mo del, and the ED AMAME collection of datasets is av ailable to researc hers (see section 8 for additional details). W e sho w in Figure 4 an ov erview of the pip eline used to train and ev aluate UME. 4.1 Data pre-pro cessing and preparation Data split W e divide EDAMAME in to tw o parts: a train and a downstream ev aluation parts. F or simplicit y , we call the ﬁrst ED AMAME-train and the second ED AMAME-test. W e use ED AMAME-train to train our foundation mo del for wearable EDA data, while EDAMAME-test to p erform ev aluation on 6 A f ounda tion model f or electr odermal activity da t a APSYNC Datasets (A-H) BIG IDEAS BiHeartS CLUES Dreamt Dynamics EPM EmpaticaE4Stress 0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00 EDA V alue ( S) HHISS HeartS Datasets (H-T) LAUREA TE M2Sleep MEF AR Nurses' Stress Dataset PPG_FieldStudy SEED SEED-II-Lab 0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00 EDA V alue ( S) SEED-II-W ild Stress Predict Dataset Datasets (T -W) T oadStool USILaughs WEEE WESAD WESD 0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00 EDA V alue ( S) W orkplaces Scenarios Daily Living Engagement Real-world Stress Sleep Monitoring Stress/Emotion Induction W orkplace Analysis Figure 2: Ridgeline plot with distribution of ED A v alues from eac h dataset. The colors represen t the six scenarios each dataset w as collected in. The plot has b een truncated, for visualization’s sake, in the range 0 − 2 . 5 µ S . The original range was 0 − 40 µ S (the E4’s max theoretical v alue is 100 µ S ). 1 2 5 10 20 50 70 Number of Users 1 0 0 1 0 1 1 0 2 A verage Duration per User (hours) Scenarios Daily Living Engagement Real-world Stress Sleep Monitoring Stress/Emotion Induction W orkplace Analysis T otal Dataset Size 10h 100h 500h 1000h Figure 3: Number of users and data p er user across the EDAMAME collection of datasets. Mark er size is prop ortional to the size of the dataset (in terms of hours of EDA data). a set of downstream classiﬁcation tasks. W e select 17 datasets for EDAMAME-train and 7 datasets for ED AMAME-test. W e p erform this split to allow models to b e tested on data never seen during training, con taining new users, lo cations, and protocols as w ell. W e select the datasets in the downstream task in order to hav e a diverse set of lab els relev an t for EDA data. W e rep ort in T able 3 the binary task from eac h dataset in the do wnstream ev aluation part. Pre-pro cessing W e pre-pro cess the data from b oth EDAMAME-train and EDAMAME-test parts follow- ing related work on EDA data [ 4 , 30, 39, 46]. First, we apply a Butterworth low-pass ﬁlter (cutoﬀ 0 . 4 Hz ) to remo ve high-frequency noise. Then, we decomp ose the ED A signals into phasic and tonic comp onents, using the cvxEDA metho d from Greco et al. [45]. This is a common pro cedure applied when working with EDA data. The phasic component is asso ciated with short-term c hanges, e.g., momentary stress, while the tonic comp onen t with longer term v ariations [15]. T o train the UME foundation mo del, we use b oth the phasic and tonic comp onents, as well as the non-decomp osed EDA signal. Using all three signals, i.e., phasic, tonic 7 A f ounda tion model f or electr odermal activity da t a Low-pass ﬁlter (0.4 Hz cutoff) cvxEDA Decomposition Raw EDA Signal (60s window @ 4 Hz) Phasic Component T onic Component Original EDA 3-Channel EDA Input Augmentation 1 Augmentation 2 EfﬁcientNet EfﬁcientNet InfoNCE Loss Maximize Similarity EfﬁcientNet Shared weights Logistic Regression e.g., Stress / Relax = trainable parameters = frozen parameters Conv1D Stacked MBConv1DBlocks Conv1D Pooling Linear Projection conv . operations x L EfﬁcientNet T raining (Contrastive Learning) Evaluation (Linear Probing) Figure 4: Overview of the train and downstream ev aluation pro cess we used for the UME foundation mo del. F or reference, w e rep ort a sketc h of the EﬃcientNet architecture, which we use as backbone for UME. and original EDA signal, is a common pro cedure in classiﬁcation-based tasks using EDA data, since it allows mo dels to learn b oth short and long-term eﬀects in the data [ 4 , 5 ]. Data segmen tation After pre-pro cessing the data, we segment the EDA signals into ﬁxed-length windows of 60 seconds. W e use the same window length as Matton et al. [64] and Schmidt et al. [82], who highlight ho w this length allows to capture b oth short and long-term changes in the EDA data. T o train UME, on ED AMAME-train we select maximum ov erlapping windows, i.e., with an ov erlap step of 0 . 25 s (the sampling rate). W e implement this approach as done by Matton et al. [64], on self-sup ervised learning for EDA data, and Ansari et al. [ 8 ], Gosw ami et al. [43], F eofano v et al. [35], on generalist time series foundation mo dels for other physiological signals. With this approach, we obtain a train set consisting of appro ximately 275 mil lion windows of EDA data. W e also apply the same 60-second segmentation on the data used for ev aluation, i.e., EDAMAME-test. F or ev aluation, w e assign to eac h window a binary lab el, corresp onding to the asso ciated task: w e refer to T able 3 for an o v erview of the binary do wnstream tasks used. Whenever a dataset contains EDA collected sim ultaneously from b oth sides of the b o dy , for ev aluation we use only the data from the b o dy side most asso ciated with the task, follo wing guidelines by Alchieri et al. [ 4 ], e.g., right-side EDA signals for cognitiv e load classiﬁcation. On EDAMAME-test, we use non-o v erlapping windows, since the literature sho ws how testing on either o verlapping or non-ov erlapping windows leads to similar results [26, 87]. Discussion on rescaling the data W e do not apply any normalization or standardization, e.g., min- max normalization, on the prepared windo ws of ED A data. Researchers working with ED A data collected in lab-controlled environmen ts apply p er-user min-max normalization to reduce in ter-p ersonal v ariabilit y and improv e classiﬁcation p erformance signiﬁcantly , e.g., [67, 82, 7 ]. Ho w ever, ED A data is aﬀected by b oth in ter-p ersonal v ariability , which min-max normalization addresses, and intra-personal v ariabilit y , i.e., a user’s data changes ov er time. A ﬁx calibration, like the one applied by min-max normalization, is rendered inv alid o ver time if a user’s signal morphology changes [91, 12]. Moreov er, “cold-start” approaches, i.e., a machine learning mo del is applied directly on a new user, are the preferred approach for wearable physiological data [93]. Normalization approaches cannot b e implemented directly in “cold-start” scenarios. 4.2 Mo del training task & architecture T raining ob jectiv e W e adopt con trastive learning to train UME. Contrastiv e learning is used by multiple researc hers to train foundation mo dels for wearable physiological data, e.g., [ 1 , 72]. This approach has also b een used for generalist time series foundation mo dels sp ecialized in classiﬁcation tasks, e.g., [35]. Con trastive learning is a self-sup ervised learning metho d that consists in training an enco der-only mo del to learn a latent em b edding space where representations of similar data pairs — typically created via augmentations of the 8 A f ounda tion model f or electr odermal activity da t a T able 3: Summary of Datasets and Asso ciated Binary T asks Dataset Binary T asks APSync Lo w/High Engagemen t HeartS Sleep/W ak e USILaughs Cog. Load/Relaxation WESAD High/Lo w Arousal, High/Low V alence (self-rep ort) Nurses’ Stress Low/High Stress DREAMT Sleep/W ak e HHISS Lo w/High Stress same signal — are attracted to each other, while representations of dissimilar pairs are rep elled. The latent em b eddings often contain information which allows them to b e used eﬀectively in downstream tasks [56]. W e also exp eriment with an additional self-supervised learning method: masked reconstruction. This metho d consists in training an enco der-deco der to reconstruct the whole signal from a mask ed version of it, i.e., a signal which is missing some parts. W e rep ort in App endix B details ab out this additional exp eriment which, ho wev er, failed to learn useful represen tations of our EDA data. Mo del arc hitecture W e c ho ose the architecture for our foundation mo del from similar w ork in the lit- erature [1, 68, 72, 35]. Our c hoice reﬂects the decision to implement a generalist time series foundation mo del on ED A data. This choice is in line with research on the ﬁrst PPG and ECG-sp eciﬁc foundation mo dels [ 1 , 68]: our ob jective is to sho w case how features computed from ED A-speciﬁc foundation mo dels p erform, as well to pro vide weigh ts and code to the researc h comm unit y . Each physiological signal, from PPG to ECG and ED A, has morphological elemen ts sp eciﬁc to them. It is p ossible to deﬁne arc hitectures for foundation mo dels that adapt to these morphological elements, as recent work from Saha et al. [81] show. Ho wev er, this implementation go es b eyond the scop e of the current work. W e c hose an Eﬃcien tNet [86] architecture, as Abbasp ourazad et al. [ 1 ], since it is computationally less exp ensiv e than traditional CNN backbones. W e consider this choice also in light of the fact that foundation mo dels for w earable data ha v e the potential to be used on w earable devices themselv es [ 1 ]. W e adapt the Eﬃcien tNet for our input data, i.e., 1-dimensional time series data of 240 v alues ( 60 s at 4 Hz ) with 3 c hannels (tonic, phasic and original ED A signal). W e report in App endix B ablation studies on the mo del size and hyperparameters. W e implement a version of EﬃcientNet with approximately 1M parameters and a latent representation of d = 64 . 4.3 T raining pro cess T rain loss Using the contrastiv e learning training ob jective and Eﬃcien tNet architecture deﬁned in sub- section 4.2, we train UME using the InfoNCE loss [89]. F ormally , for a given pair of embeddings ( z i , z j ) , the loss is deﬁned as: L i,j = − log exp ( sim ( z i , z j )/ τ ) P 2 N k =1 ⊮ [ k  = i ] exp ( sim ( z i , z k )/ τ ) (1) where sim ( · , · ) denotes the cosine similarit y , τ = 0 . 1 is a temp erature h yperparameter, N is the batch size, and ⊮ is the indicator function. This ob jective maximizes the similarity b et ween representations of the same underlying signal while minimizing agreement with unrelated samples. P airs of EDA data W e generate positive pairs of ED A signals by applying tw o distinct sto chastic aug- men tations to the same EDA signal segment (an anchor segment). W e employ the set of data augmentations optimized for EDA signals prop osed b y Matton et al. [64]. Negative pairs consist of comparisons b etw een the anc hor segment and all other segments in the mini-batch. This set combines b oth standard data augmenta- tion tec hniques (e.g., signal warping) and augmen tations speciﬁc to EDA data (e.g., lo ose sensor artifact). W e rep ort the complete list of augmentations and additional details ab out the training in App endix A (T able A.5). 9 A f ounda tion model f or electr odermal activity da t a 5 Ev aluation of UME In this section, we rep ort the results of the ev aluation procedure of UME on the selected downstream tasks from ED AMAME-test. W e use line ar pr obing with frozen weigh ts to ev aluate the p erformance on the selected classiﬁcation tasks, as frequently done to ev aluate p erformance of foundation mo dels [ 1 , 81, 72, 95]. W e compare the UME feature set to v arious baseline features, including generic handcrafted features, a set of ED A-sp eciﬁc features, and features computed from generalist time series foundation mo dels. Finally , we also ev aluate the computation complexity of the UME foundation mo del in extracting features, computed to the other baseline metho ds. 5.1 Exp erimen tal ev aluation setup Ev aluation through linear probing W e ev aluate the features computed from the UME foundation mo del using line ar pr obing on the downstream tasks from ED AMAME-test. F or each dataset, we freeze the trained weigh ts from the foundation mo del to compute features on the EDA data. Then, we train a logistic regression classiﬁer on the computed feature set. The usage of frozen weigh ts and linear probing is commonly done in similar work on foundation mo dels for physiological data [ 1 , 72, 81]. W e use the same ev aluation pro cedure for UME and the other baseline features sets. W e use a linear model for all approac hes since we are gauging the representativ eness of the features sets on classiﬁcation tasks, and not that of the downstream mo del itself. Linear probing also allows to estimate the linear separability of the diﬀerent feature sets. Cross-v alidation proto cols for the do wnstream tasks W e ev aluate the linear probing using tw o distinct cross-v alidation metho ds. The ﬁrst metho d is Leav e-One-P articipant-Out (LOPO) cross-v alidation. LOPO cross-v alidation is used b y researc hers to ev aluate how mac hine learning mo dels generalize to new users [76]. With this metho d, we test robustness to inter-personal v ariability . The second v alidation metho d is Time-A ware (T A) cross-v alidation [ 6 ], which we use to ev aluate the mo del’s ability to generalize to data from users already seen in the train set, and this we test robustness to intra-personal v ariabilit y . W e partition users into N folds: the mo del trains on all external groups plus the ﬁrst chronological 2/3 of the target fold, reserving the ﬁnal 1/3 of data strictly for testing. This approach sim ulates a realistic scenario where a mo del uses a sp eciﬁc participant’s historical data to predict their future states. In our exp erimental setup, we use N = 5 . In b oth v alidation metho ds, we p erform hyperparameter tuning for the logistic regression at train set, i.e., at eac h cross-v alidation iteration. W e p erform hyperparameter selection using a 3-fold “inner” cross-v alidation with grid search. W e rep ort in App endix A information ab out the hyperparameter grid used. Baseline feature sets W e compare the p erformance using linear probing on the features computed from UME and using other baseline metho ds. Speciﬁcally , to follow the same exp erimental setup of similar work on foundation mo dels for physiological data [ 1 , 72, 68], we deﬁne a set of baseline, generic , handcrafted features. These features are: the mean, the standard deviation, the minimum and the maximum of a 60 s ED A signal. W e compute these features for all three EDA comp onents, i.e., phasic, tonic and original signal. In total, the dimensionalit y of this feature set is d = 12 . Ho wev er, the aforemen tioned generic handcrafted feature set do es not represent the state-of-the-art approac h for ED A-based classiﬁcation tasks [ 4 , 39, 63]. T o this end, we also implement a second handcrafted feature set, which we call EDA-sp e ciﬁc handcrafted features. This feature set includes b oth statistics, e.g., av erage of the ﬁrst deriv ative, as well as feature sp eciﬁc to EDA data, e.g., num b er of EDA p eaks and their av erage amplitude. As with the generic handcrafted feature set, we compute these features on 60 s windows and for all three EDA comp onents. In total, the EDA-speciﬁc handcrafted features are d = 45 . W e also compare the performance using linear pro ving of the feature set from UME with other generalist time series foundation mo dels. W e compare with the follo wing: Chronos [ 8 ], MOMENT [43] and Mantis [35]. W e select these three foundation mo dels since Alchieri et al. [ 5 ] show how they achiev e p erformance similar to the EDA-speciﬁc handcrafted features when using EDA data. Saha et al. [81] also show that Chronos [8] and MOMENT [43] achiev e, on av erage, p erformance similar to their PPG-sp eciﬁc foundation mo del on the do wnstream tasks selected b y the authors. W e also select the recen t foundation model Man tis [35] since it is trained sp eciﬁcally for classiﬁcation tasks on time series data. Man tis takes time series of length 512 as input: we ov ersample our time series, whic h hav e length of 240, to matc h this desired length. F rom a single time series, Mantis computes a set of embeddings, similarly to our foundation model. Mantis has a feature set size of d = 768 (embedding size of 256 across three channels). 10 A f ounda tion model f or electr odermal activity da t a 0.5 0.6 0.7 0.8 0.9 1.0 Balanced A cc. Generic Handcraf ted F eatur es 0.5 0.6 0.7 0.8 0.9 1.0 Balanced A cc. E D A M A M E F e a t u r e s V alidation method T A cv L OPO Dataset + task (mark er) APSYNC: L ow/high engagement Dr eamt: Deep Sleep/REM Dr eamt: Sleep/W ak e HeartS: Sleep/W ak e WES AD: L ow/High V alence USILaughs: Cognitive load/r elaxation WES AD: L ow/High (Self -R eported) Ar ousal HHISS: Str ess/calm Figure 5: P airplot with comparison b etw een UME and the baseline features for the selected downstream tasks. Colors represent the tw o v alidation metho ds, T A and LOPO. The embedding size of MOMENT is d = 1024 . How ever, Chronos computes embeddings ( d = 512 ) for each timep oin t in the series. This leads, with our input, to high dimensionalit y of the data. T o address the high dimensionality problem, we av erage across the time axis the em b eddings extracted from eac h timep oin t. Ov erall, the feature set from Chronos is of size d = 1536 (embedding size of 512 across three channels). Rep orting and task selection information W e rep ort all results in terms of b alanc e d ac cur acy . W e use this metric since the downstream ev aluation datasets w e use contain binary lab els which, in a subset of cases, are imbalanced. Balanced accuracy accounts for class imbalance, rep orting results which are more represen- tativ e of the real p erformance on a speciﬁc task [16, 70]. A t the same time, compared to other metrics for im balanced data, e.g., Matthew’s correlation co eﬃcient (MCC), balanced accuracy is more interpretable [44]. In addition to the results from the iden tiﬁed feature sets, we also rep ort results from a dumm y classiﬁer. The dummy classiﬁer rep orted is the one achieving the highest balanced accuracy in each giv en task, among the following: most fr e quent , which predicts alwa ys the most frequen t class set; uniform , which predicts the binary lab els by drawing randomly from a uniform distribution; and prior , which predicts the binary lab els b y dra wing randomly from the distribution of lab els in the train set. W e rep ort in this section results from tasks that are solvable , i.e., at least one feature set achiev es balanced accuracy higher than the dummy classiﬁer. If no mo del ac hieves balanced accuracy higher than, for example, random chance, we conclude that it is not due to issues with the feature sets, but with the task itself, e.g., the task is not solv able with the giv en constrain ts. 5.2 Comparison with generic handcrafted feature set W e presen t in Figure 5 the comparison when p erforming linear probing on the selected downstream tasks, b et ween the feature set from our UME foundation mo del and the generic handcrafted features. W e show results for b oth v alidation metho ds, i.e., LOPO and T A. The results show that the features from our foun- dation mo del outp erform the generic handcrafted features in 9 out of 10 tasks. The improv emen t holds true regardless of the scenario and the task selected. W e conclude from these ﬁndings that the UME foundation mo del learns features useful for EDA data in performing do wnstream predictions. Our ﬁndings are in line with works on foundation mo dels for PPG, when comparing to generic handcrafted feature sets [ 1 , 72, 68]. 11 A f ounda tion model f or electr odermal activity da t a 5.3 Comparison with other baseline feature sets W e report in T able 4 the results for the ev aluation of our UME foundation mo del, in terms of balanced accuracy and for tw o distinct v alidation metho ds, i.e., Time-A w are (T A) cross-v alidation and Lea v e-One- P articipant-Out (LOPO) cross-v alidation. W e compare the feature set computed using UME to the fol- lo wing feature sets: a set of generic handcrafted features; a set of EDA-speciﬁc handcrafted features; and features computed using generalist time series foundation models, sp eciﬁcally Mantis [35], Chronos [ 8 ] and MOMENT [43]. Results using T A The T A metho d tests the intra-personal generalizability of mo dels trained using the diﬀeren t feature sets, i.e., the abilit y to generalize to new data from users already seen in the train data. Our ﬁrst ﬁnding is that, using linear probing, models trained using features from UME alw ays outp erform non-foundation mo del metho ds, i.e., b oth generic and EDA-speciﬁc handcrafted features. W e conclude that the embeddings from UME capture user-sp eciﬁc dynamics, which allo w to obtain a balanced accuracy on par, or higher, than handcrafted-based metho ds. W e also ﬁnd that mo dels trained using the feature set from Man tis [35], which is trained to solve generic classiﬁcation tasks, p erform similarly to those trained using the feature set from UME. W e notice how the other tw o foundation mo dels, MOMENT [43] and Chronos [ 8 ], ha ve lo wer p erformance, on a ma jority of tasks, than both the t w o handcrafted feature sets and the foundation mo dels (UME and Man tis [35]). Results using LOPO The LOPO cross-v alidation metho d tests the inter-personal generalizability of mo dels trained using the diﬀerent feature sets, i.e., the ability to generalize to data from new users not seen in the train set. W e ﬁnd that models trained from the UME features hav e similar balanced accuracy to models trained from either the EDA-speciﬁc handcrafted features and the embeddings computed using Mantis [35]. F eatures from the other foundation mo dels ha ve similar or low er p erformance as well. Generalization to new users is a known issue when working with EDA data [39]. While features from foundation mo dels, b oth our mo del and others, achiev e performance similar to that of ED A-speciﬁc handcrafted features, they do not solv e the “cold-start” problem. A dditional remarks on standard errors W e highligh t how, regardless of the feature set used, the standard error associated with all results leads to ov erlapping conﬁdence in terv als. In other words, while there are trends, e.g., models trained using features from UME outp erforming mo dels trained using ED A- sp eciﬁc handcrafted features in a ma jority of tasks, there is no statistical diﬀerence b etw een results obtained using the diﬀerent feature sets. How ev er, all results are statistically higher (t-test corrected with Bonferroni) compared to the results obtained using the dummy classiﬁer. Result analysis using F riedman-Nemenyi test W e p erform the F riedman test, follow ed by the post- ho c Nemenyi test, to compare mo del p erformance across all exp eriments [27]. W e consider each com bination of dataset-v alidation metho d as a single sample for the statistical analysis. The F riedman test is used to determine whether any statistical diﬀerence is present across all mo dels and exp eriments. F or the F riedman test, we ﬁnd a p-v alue of approximately 0.0001, whic h is below the reference threshold of α = 0 . 05 . W e conclude from this result that, across all exp eriments, there are statistically signiﬁcant diﬀerences. In particular, we attribute this ﬁnding to the diﬀerence b etw een all metho ds, i.e., b oth handcrafted- and foundation mo del-based, and the dummy classiﬁer baseline. W e use the p ost-ho c Nemen yi test to p erform pair-wise statistical comparisons, from the model rankings pro vided b y the Nemenyi test. W e report the ﬁndings with resp ect to our UME foundation mo del only . W e refer to the App endix B to additional results. First, we ﬁnd that our metho d achiev es p erformance statistically higher than the dummy classiﬁer baseline ( p = 0 . 02 < α = 0 . 05 ). Secondly , we ﬁnd that the p erformance diﬀerence in linear probing b etw een feature sets computed using UME and Man tis [35] is not statistically diﬀeren t ( p ≃ 0 . 9 > α > 0 . 05 ). Finally , w e also ﬁnd no statistical diﬀerence b etw een using our UME and the EDA-speciﬁc handcrafted features ( p ≃ 0 . 9 > α = 0 . 05) . F rom the F riedman-Nemen yi test ﬁndings, we conclude that b oth our UME and Man tis [35] capture time series dynamics which allow to achiev e, on a v erage, p erformance similar to the ED A-sp eciﬁc handcrafted feature sets. Our EDA-trained foundation mo del p erforms on-par with a large-scale generalist time series foundation mo del, Mantis [35], even if trained on a relatively smaller dataset and with few er parameters (1M ours vs 8M Man tis). 12 A f ounda tion model f or electr odermal activity da t a T able 4: Results of binary classiﬁcation exp eriments across datasets and tasks, in terms of b alanc e d ac cu- r acy standard er r or . Rep orted are results for both T A and LOPO cross-v alidation metho ds. Acron yms: Gen. HC stands for generic handcr afte d fe atur es ; EDA HC stands for EDA-sp e ciﬁc handcr afte d fe atur es . Dumm y Handcrafted (HC) Generalist F oundation Ours Dataset Binary T ask Gen. ED A-spec. Man tis MOMENT Chronos UME Balanc e d ac cur acy standard er r or Time-A ware cross-v alidation (T A) APSYNC Lo w/High engagemen t . 45 . 20 . 80 . 10 . 80 . 10 . 86 . 07 . 47 16 . 61 . 20 . 89 . 11 Dream t Deep Sleep/REM . 48 . 02 . 55 . 05 . 59 . 04 . 64 . 03 . 65 01 . 63 02 . 63 . 05 Dream t Sleep/W ak e . 50 . 01 . 65 . 02 . 69 . 01 . 73 . 01 . 69 01 . 73 01 . 70 . 01 HeartS Sleep/W ak e . 49 . 00 . 66 . 04 . 72 . 07 . 75 . 09 . 69 06 . 74 06 . 73 . 09 WESAD Low/High V alence . 51 . 03 . 55 . 10 . 54 . 10 . 61 . 04 . 64 05 . 45 . 06 . 63 . 06 Lea v e-one-participant-out (LOPO) cross-v alidation Dream t Sleep/W ak e . 48 . 00 . 70 . 01 . 74 . 01 . 76 01 . 73 01 . 78 01 . 75 . 01 USILaughs Cog. load/relax . 50 . 00 . 66 . 03 . 70 . 04 . 72 . 05 . 60 . 05 . 67 . 03 . 71 . 05 HeartS Sleep/W ak e . 50 . 00 . 66 . 04 . 70 . 03 . 74 . 03 . 70 . 02 . 73 . 03 . 72 . 03 WESAD Low/High Arousal . 58 . 05 . 61 . 04 . 63 . 04 . 56 . 04 . 66 . 04 . 66 . 05 . 61 . 05 HHISS Stress/calm . 50 . 00 . 56 . 02 . 64 . 01 . 63 . 02 . 55 . 01 . 59 . 01 . 60 . 02 5.4 Computational complexit y analysis W e compare the computational complexity of the diﬀerent feature extraction metho ds, i.e., the handcrafted- based approaches, our UME foundation mo del, and the other baseline foundation mo dels. W e compare the metho ds using b oth CPU exe cution time and FLOPs [22]. W e choose this dual approach since handcrafted features cannot b e compared using FLOPs alone. W e p erform all calculations using an Apple M1 Max CPU. T o get CPU exe cution time , we use a random subset of t wen t y 60 s windows from the downstream part of the UME collection. W e also use additionally three samples p er experiment as w arm-up runs. W e p erform FLOPs computation using the fvcore Python library . W e rep ort in Figure 6 the results. W e ﬁnd that the handcrafted feature extraction metho ds hav e the lo west CPU execution time, as exp ected. W e also ﬁnd that our UME foundation model is signiﬁcan tly faster than all other foundation mo dels, b oth with resp ect to CPU exe cution time and FLOPs. W e conclude that our mo del, which is trained sp eciﬁcally only on ED A data, achiev es performance on-par with larger, general purpose, foundation mo dels on a set of do wnstream tasks using EDA data, as detailed in subsection 5.3. How ev er, it do es so with a fraction of the computation resources needed, while also eliminating the need for exp ert-designed or handcrafted features, unlik e traditional feature-based pip elines 6 Discussion, limitations and future w ork Summary of results Through this pap er, we pro vide the researc h communit y with the EDAMAME collection of datasets, whic h enabled the developmen t of UME, the ﬁrst op en-source foundation mo del for w earable EDA. W e ﬁnd that UME, trained on a subset of the EDAMAME collection using contrastiv e 13 A f ounda tion model f or electr odermal activity da t a 0 1000 2000 3000 4000 5000 CPU T ime (ms) EDA-specific hc E D A M A M E ( 1 M ) Mantis (8M) Chronos (28M) MOMENT (128M) CPU T ime per Sample (mean ± SE) (a) CPU Time 0 10 20 30 40 50 60 GFLOPs (Giga Floating Point Operations) E D A M A M E ( 1 M ) Mantis (8M) Chronos (28M) MOMENT (128M) 0.04 0.85 1 1.27 59.80 Computational Complexity: GFLOPs per Sample (60s) (b) Mo del GFLOPs Figure 6: Computational complexity analysis. (a) A v erage (across 20 samples) p er-sample computation time for a 60-second window of EDA data for all three comp onents (phasic, tonic, and original signal). Note: hc stands for handcr afte d ; generic handcrafted features are not sho wn as they require negligible computation time. (b) GFLOPs for the diﬀerent foundation mo dels. learning, surpasses the p erformance obtainable using both generic and EDA-handcrafted features in 8 out of 10 downstream tasks. F urther, UME matc hes the p erformance of larger and muc h less computationally eﬃcien t generalist foundation models for time series data. By relying on publicly a v ailable data only and b y releasing all artifacts pro duced as part of this work, we aim at fostering further researc h on foundation mo dels for wearable ED A as well as other research related to unimodal and m ultimo dal ED A sensing and mo delling. Discussion Ov erall, we observe that features computed using UME, generalist time-series foundation mo dels, and ED A-sp eciﬁc handcrafted features lead to comparable downstream p erformance. It can be argued that other metho dologies, in particular the use of domain-sp eciﬁc architectures elements, may yield sup erior p erformance. How ev er, we b elieve that our results are not b ound to UME’s sp eciﬁc characteristics, b ecause UME’s p erformances not only matches those of generalist foundation models, but also often surpasses results obtained relying on EDA-speciﬁc handcrafted features. This suggests that other methodological approac hes alone may yield incremental rather than substantial gains. Instead, we b elieve that the rep orted results, obtained by UME and other metho ds, are close to the intrinsic upp er limit reachable for the considered downstream tasks, due to the arguably limited information in the input signal that is predictive of the label. P otential additional gains can how ev er b e obtained by coping with the signiﬁcant lab el noise and signal noise of EDA data. W earable EDA is noisy , strongly aﬀected by motion/con tact artifacts and low signal-to-noise ratio, in particular when collected in the wild [39]. A dditionally , high-lev el constructs (e.g., stress, engagement, v alence) might only b e weakly iden tiﬁable from ED A alone resolution without additional context. W e b elieve that this represents an imp ortant direction for future work, e.g., by quantifying the impact of signal quality ﬁltering, lab el reliability , and more. A further element to highlight is that UME, which is sp ecialized for EDA data, requires far fewer parameters to matc h the p erformance on downstream tasks of larger generalist models. As highligh ted by Abbaspourazad et al. [ 1 ], developing foundation mo dels for w earable data includes considering the usage of these models on actual w earable devices, which are resource constrained. W e sp eculate that our results sho w case how sp ecializing foundation mo dels for EDA data limit their size and, consequently , prosp ect their usage on w earable devices. Limitations & future work W e create the EDAMAME collection leveraging data collected exclusively using the Empatica E4 device. Considering data collected from a speciﬁc device remov es hardware-speciﬁc domain shifts, ensuring that foundation mo dels can learn physiological patterns rather than sensor diﬀerences. Nonetheless, future work could fo cus on expanding EDAMAME with data from other devices, e.g., the newer Empatica Embrace Plus 13 . 13 https://www.empatica.com/embraceplus 14 A f ounda tion model f or electr odermal activity da t a In our pre-pro cessing approach, we apply standard low-pass ﬁltering and decomp osition of the EDA signals using the cvxED A metho d [45]. Alternative ﬁltering or decomp osition techniques [90] can b e explored in future research. Lastly , we ackno wledge that the EDAMAME collection of datasets is aﬀected by the same user and domain bias asso ciated with the individual datasets. 7 Conclusions In conclusion, in this w ork we present EDAMAME, a large scale collection of datasets containing w earable ED A data. EDAMAME, whic h is composed of 24 datasets, contains appro ximately 25’000 hours of EDA data from ab out 630 users. With its diversit y and v ariabilit y , it enables the training of foundation mo dels for ED A data, as we show with the training of UME, the ﬁrst foundation mo del for wearable EDA. W e also mak e ED AMAME a v ailable to other researchers, to spur developmen t of further foundation mo dels. Through a comprehensive set of exp eriments, we show that the latent representations obtained using UME outp erform a set of generic handcrafted features when using linear probing on 9 out of 10 downstream tasks, as done in similar w ork in the literature [ 1 , 72]. W e also ﬁnd that UME’s features achiev e performance on-par with sp ecialized EDA features and large-scale generalist foundation mo dels, in the same set of downstream tasks. How ev er, compared to generalist mo dels lik e Chronos [ 8 ] or Man tis [35], UME is computationally ligh ter, e.g., 20 × less demanding than the smallest generalist foundation mo del tested, enabling its p ossible usage on real-world wearable devices. W e hop e that this work enables researchers to further develop EDA, a sensor not (yet) as common as others lik e PPG on wearable devices, through b oth the EDAMAME collection of datasets and the op en source and w eights UME foundation mo del for wearable EDA data. 8 Dataset and co de av ailability W e make the EDAMAME av ailable to other researchers, up on c ompletion of the r eview pr o c ess underway . 11 datasets hav e an op en source license which allows for re-sharing of the data under the original conditions: w e re-share them, prepared in a uniﬁed format, following the original license. 11 datasets need a data sharing agreemen t to b e accessed: we contacted the original authors, who all agreed to re-sharing under a single data sharing agreement, which includes all of the original limitations. Finally , t wo datasets can only b e accessed through their authors: we provide the co de to pro cess them in the same format as we did. W e rep ort information ab out the original licenses in App endix A (T able A.1). W e will also share the code to pro cess these datasets after the ongoing r eview pr o c ess . W e also make the code to train and ev aluate the UME foundation model a v ailable as op en source to other researc hers, as well as the mo del weigh ts. The link will b e provided after the ongoing r eview pr o c ess . 15 A f ounda tion model f or electr odermal activity da t a References [1] Salar Abbasp ourazad, Oussama Elac hqar, Andrew C Miller, Saba Emrani, Udhy akumar Nallasamy , and Ian Shapiro. Large-scale T raining of F oundation Mo dels for W earable Biosignals. In International Confer enc e on L e arning R epr esentations , Vienna, Austria, 2024. In ternational Conference on Learning Represen tations (ICLR). [2] Nouran Ab dalazim, Joseba Aitzol Arbilla Larraza, Leonardo Alchieri, Lidia Alecci, Silvia San tini, and Shkurta Gashi. Heart Rate During Sleep Measured Using Finger-, W rist- and Chest-W orn Devices: A Comparison Study. In Athanasios T sanas and Andreas T rian tafyllidis, editors, Pervasive Computing T e chnolo gies for He althc ar e , pages 18–32, Cham, 2023. Springer Nature Switzerland. ISBN 978-3-031- 34586-9. doi: 10.1007/978- 3- 031- 34586- 9_ 2. [3] Nouran Ab dalazim, Leonardo Alchieri, Lidia Alecci, Pietro Barbiero, and Silvia Santini. The Impact of Domain Shift on Predicting Perceiv ed Sleep Quality from W earables. Sensors , 25(13):4012, Jan uary 2025. ISSN 1424-8220. doi: 10.3390/s25134012. [4] Leonardo Alchieri, Nouran Ab dalazim, Lidia Alecci, Shkurta Gashi, Martin Gjoreski, and Silvia Santini. Lateralization Eﬀects in Electro dermal Activit y Data Collected Using W earable Devices. Pr o c. A CM Inter act. Mob. W e ar able Ubiquitous T e chnol. , 8(1):2:1–2:30, March 2024. doi: 10.1145/3643541. [5] Leonardo Alc hieri, Lino Candian, Nouran Abdalazim, Lidia Alecci, Giov anni De F elice, and Silvia San tini. Exploring Generalist F oundation Mo dels for Time Series of Electro dermal Activit y Data. In Comp anion of the 2025 A CM International Joint Confer enc e on Pervasive and Ubiquitous Comput- ing , UbiComp Companion ’25, pages 885–892, New Y ork, NY, USA, Decem ber 2025. Asso ciation for Computing Machinery . ISBN 979-8-4007-1477-1. doi: 10.1145/3714394.3756186. [6] Leonardo Alchieri, Vittoria Sco cco, Nouran Abdalazim, Lidia Alecci, and Silvia San tini. Improving Human Beha vior Recognition Using Bilateral Electro dermal Activit y Data Collected F rom W earable Devices. In 2025 International Confer enc e on A ctivity and Behavior Computing (ABC) , pages 1–10, Al Ain, United Arab Emirates, April 2025. IEEE. doi: 10.1109/ABC64332.2025.11118532. [7] Ahmad Almadhor, Gabriel A v elino Sampedro, Mideth Abisado, Sidra Abbas, Y e-Jin Kim, Muham- mad A ttique Khan, Jamel Baili, and Jae-Hyuk Cha. W rist-Based Electrodermal A ctivit y Monitoring for Stress Detection Using F ederated Learning. Sensors , 23(8):3984, January 2023. ISSN 1424-8220. doi: 10.3390/s23083984. [8] Ab dul F atir Ansari, Lorenzo Stella, Ali Caner T urkmen, Xiyuan Zhang, P edro Mercado, Huibin Shen, Oleksandr Shc h ur, Syama Sundar Rangapuram, Sebastian Pineda Arango, Sh ubham Kapo or, Jasp er Zsc hiegner, Danielle C. Maddix, Hao W ang, Michael W. Mahoney , Kari T orkkola, Andrew Gordon Wilson, Michael Bohlke-Sc hneider, and Bernie W ang. Chronos: Learning the Language of Time Series. T r ansactions on Machine L e arning R ese ar ch , May 2024. ISSN 2835-8856. [9] Ab dul F atir Ansari, Oleksandr Shch ur, Jaris Kük en, Andreas Auer, Boran Han, Pedro Mercado, Sy ama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Gosw ami, Shubham Kap o or, Danielle C. Maddix, Pablo Guerron, T ony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao W ang, Huzefa Rangw ala, George Karypis, Y uy ang W ang, and Michael Bohlk e-Sc hneider. Chronos-2: F rom Univ ariate to Universal F orecasting, Octob er 2025. [10] Pro ject Ap ertus, Alejandro Hernández-Cano, Alexander Hägele, Allen Hao Huang, Angelika Romanou, An toni-Joan Solergib ert, Barna Pasztor, Bettina Messmer, Dhia Garbay a, Eduard F rank Ďurec h, Ido Hakimi, Juan García Giraldo, Mete Ismayilzada, Negar F oroutan, Skander Moalla, Tiancheng Chen, Vink o Sab olčec, Yixuan Xu, Michael Aerni, Badr AlKhamissi, Inés Altemir Mariñas, Mohammad Hos- sein Amani, Matin Ansarip our, Ilia Badanin, Harold Benoit, Emanuela Boros, Nicholas Browning, F abian Bösch, Maximilian Böther, Niklas Cano v a, Camille Challier, Clement Charmillot, Jonathan Coles, Jan Deriu, Arnout Devos, Lukas Drescher, Daniil Dzenhaliou, Maud Ehrmann, Dongyang F an, Simin F an, Silin Gao, Miguel Gila, María Grandury , Diba Hashemi, Alexander Hoyle, Jiaming Jiang, Mark Klein, Andrei Kuchara vy , Anastasiia Kucherenk o, F rederike Lüb ec k, Roman Machacek, Theoﬁ- los Manitaras, Andreas Marfurt, Kyle Matoba, Simon Matrenok, Henrique Mendonça, F awzi Rob erto Mohamed, Syrielle Montariol, Luca Mouc hel, Sven Na jem-Meyer, Jingw ei Ni, Gennaro Oliv a, Matteo P agliardini, Elia Palme, Andrei Panfero v, Léo Paoletti, Marco Passerini, Iv an Pa vlov, Auguste Poiroux, Kaustubh P onkshe, Nathan Ranchin, Javi Rando, Mathieu Sauser, Jakhongir Saydaliev, Muhammad Ali Sa yﬁddinov, Marian Schneider, Stefano Sch uppli, Marco Scialanga, Andrei Semenov, Kumar Shridhar, Ragha v Singhal, Anna Sotniko v a, Alexander Sternfeld, A yush Kumar T arun, Paul T eiletc he, Jannis V am v as, Xiaozhe Y ao, Hao Zhao, Alexander Ilic, Ana Klimovic, Andreas Krause, Caglar Gulcehre, 16 A f ounda tion model f or electr odermal activity da t a Da vid Rosenthal, Elliott Ash, Florian T ramèr, Jo ost V ande V ondele, Livio V eraldi, Martin Rajman, Thomas Sch ulthess, T orsten Hoeﬂer, An toine Bosselut, Martin Jaggi, and Imanol Sc hlag. Ap ertus: Demo cratizing Op en and Complian t LLMs for Global Language Environmen ts, December 2025. [11] Brinnae Bent, Peter J. Cho, Maria Henriquez, April Wittmann, Connie Thack er, Mark F einglos, Matthew J. Crowley , and Jessilyn P . Dunn. Engineering digital biomarkers of interstitial glucose from nonin v asiv e smartw atches. npj Digital Me dicine , 4(1):89, June 2021. ISSN 2398-6352. doi: 10.1038/s41746- 021- 00465- w . [12] Andrew Beten, Luna Lo co cco, A yaan Baig, and Thilini Karunarathna. Systematic Analysis of Dis- tribution Shifts in Cross-Sub ject Glucose Prediction Using W earable Physiological Data. Engine ering Pr o c e e dings , 118(1):88, 2025. ISSN 2673-4591. doi: 10.3390/ECSA- 12- 26583. [13] Andrea Bizzego, Giulio Gabrieli, Cesare F urlanello, and Gianluca Esp osito. Comparison of W earable and Clinical Devices for Acquisition of Peripheral Nervous System Signals. Sensors (Basel, Switzerland) , 20(23):6778, Nov em b er 2020. ISSN 1424-8220. doi: 10.3390/s20236778. [14] Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Mic hael S. Bernstein, Jeannette Bohg, An toine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shy amal Buc h, Dallas Card, Rodrigo Castellon, Niladri Chatterji, Annie Chen, Kathleen Creel, Jared Quincy Da vis, Dora Demszky , Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etc hemendy , Ka win Etha y ara jh, Li F ei-F ei, Chelsea Finn, T revor Gale, Lauren Gillespie, Karan Go el, Noah Go o dman, Shelby Grossman, Neel Guha, T atsunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas Icard, Saahil Jain, Dan Jurafsky , Pratyusha Kalluri, Siddharth Karamcheti, Geoﬀ Keeling, F eresh te Khani, Omar Khattab, Pang W ei K oh, Mark Krass, Ranjay Krishna, Rohith Kuditipudi, Anan y a Kumar, F aisal Ladhak, Mina Lee, T ony Lee, Jure Lesk ov ec, Isab elle Leven t, Xiang Lisa Li, Xuechen Li, T engyu Ma, Ali Malik, Christopher D. Manning, Suvir Mirchandani, Eric Mitchell, Zanele Munyikw a, Sura j Nair, A v anika Naray an, Deepak Naray anan, Ben Newman, Allen Nie, Juan Carlos Niebles, Hamed Nilforoshan, Julian Ny arko, Giray Ogut, Lau- rel Orr, Isab el Papadimitriou, Jo on Sung Park, Chris Piec h, Ev a Portelance, Christopher P otts, A diti Ragh unathan, Rob Reich, Hongyu Ren, F rieda Rong, Y usuf Ro ohani, Camilo Ruiz, Jac k Ry an, Christo- pher Ré, Dorsa Sadigh, Shiori Sagaw a, Keshav Santhanam, Andy Shih, Krishnan Sriniv asan, Alex T amkin, Rohan T aori, Armin W. Thomas, Florian T ramèr, Rose E. W ang, William W ang, Bohan W u, Jia jun W u, Y uh uai W u, Sang Michael Xie, Mic hihiro Y asunaga, Jiaxuan Y ou, Matei Zaharia, Michael Zhang, Tianyi Zhang, Xikun Zhang, Y uhui Zhang, Lucia Zheng, Kaitlyn Zhou, and Percy Liang. On the Opp ortunities and Risks of F oundation Mo dels, July 2022. [15] W olfram Boucsein. Ele ctr o dermal A ctivity . Electro dermal Activit y , 2nd Ed. Springer Science + Busi- ness Media, New Y ork, NY, US, 2012. ISBN 978-1-4614-1125-3 978-1-4614-1126-0. doi: 10.1007/ 978- 1- 4614- 1126- 0. [16] Kay Henning Brodersen, Cheng So on Ong, Klaas Enno Stephan, and Joachim M. Buhmann. The Balanced A ccuracy and Its P osterior Distribution. In 2010 20th International Confer enc e on Pattern R e c o gnition , pages 3121–3124, Istanbul, T urkey , A ugust 2010. IEEE. doi: 10.1109/ICPR.2010.764. [17] T om Brown, Benjamin Mann, Nic k Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakan tan, Prana v Sh yam, Girish Sastry , Amanda Askell, Sandhini Agarwal, Ariel Herb ert-V oss, Gretc hen Krueger, T om Henighan, Rewon Child, Adit y a Ramesh, Daniel Ziegler, Jeﬀrey W u, Clemens Win ter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray , Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskev er, and Dario Amo dei. Language Mo dels are F ew-Shot Learners. In A dvanc es in Neur al Information Pr o c essing Systems , v olume 33, pages 1877–1901, Online, 2020. Curran Asso ciates, Inc. [18] Alexander Bukharin, Shiyang Li, Zhengyang W ang, Jingfeng Y ang, Bing Yin, Xian Li, Chao Zhang, T uo Zhao, and Haoming Jiang. Data Div ersity Matters for Robust Instruction T uning. In Y aser Al-Onaizan, Mohit Bansal, and Y un-Nung Chen, editors, Findings of the A sso ciation for Computational Linguistics: EMNLP 2024 , pages 3411–3425, Miami, Florida, USA, Nov ember 2024. Asso ciation for Computational Linguistics. doi: 10.18653/v1/2024.ﬁndings- emnlp.195. [19] Maritza Bustos-Lop ez, Nicandro Cruz-Ramirez, Alejandro Guerra-Hernandez, Laura Nely Sánchez- Morales, Nancy Aracely Cruz-Ramos, and Giner Alor-Hernandez. W earables for engagement detection in learning environmen ts: A review. Biosensors , 12(7):509, 2022. [20] Sara Campanella, A yham Altaleb, Alb erto Belli, Paola Pierleoni, and Lorenzo P alma. A Metho d for Stress Detection Using Empatica E4 Bracelet and Machine-Learning T ec hniques. Sensors , 23(7):3565, Jan uary 2023. ISSN 1424-8220. doi: 10.3390/s23073565. 17 A f ounda tion model f or electr odermal activity da t a [21] Marta Casanov as Ortega, Elisa Bruno, and Mark P . Richardson. Electro dermal activity resp onse during seizures: A systematic review and meta-analysis. Epilepsy & Behavior: E&B , 134:108864, September 2022. ISSN 1525-5069. doi: 10.1016/j.yebeh.2022.108864. [22] Jierun Chen, Shiu-hong Kao, Hao He, W eip eng Zhuo, Song W en, Chul-Ho Lee, and S.-H. Gary Chan. R un, Don’t W alk: Chasing Higher FLOPS for F aster Neural Netw orks. In 2023 IEEE/CVF Confer enc e on Computer V ision and Pattern R e c o gnition (CVPR) , pages 12021–12031, V ancouver, BC, Canada, June 2023. IEEE. ISBN 979-8-3503-0129-8. doi: 10.1109/CVPR52729.2023.01157. [23] Zhengyu Chen, Siqi W ang, T eng Xiao, Y udong W ang, Shiqi Chen, Xunliang Cai, Junxian He, and Jingang W ang. Revisiting Scaling Laws for Language Mo dels: The Role of Data Quality and T raining Strategies. In W anxiang Che, Joyce Nab ende, Ekaterina Shuto v a, and Mohammad T aher Pileh v ar, editors, Pr o c e e dings of the 63r d A nnual Me eting of the A sso ciation for Computational Linguistics (V ol- ume 1: L ong Pap ers) , pages 23881–23899, Vienna, A ustria, July 2025. Asso ciation for Computational Linguistics. ISBN 979-8-89176-251-0. doi: 10.18653/v1/2025.acl- long.1163. [24] Davide Chicco and Giusepp e Jurman. The adv an tages of the Matthews correlation coeﬃcient (MCC) o ver F1 score and accuracy in binary classiﬁcation ev aluation. BMC Genomics , 21(1):6, Jan uary 2020. ISSN 1471-2164. doi: 10.1186/s12864- 019- 6413- 7. [25] Donna L. Coﬀman, Xizhen Cai, Runze Li, and No elle R. Leonard. Challenges and Opportunities in Collecting and Mo deling Ambulatory Electro dermal Activit y Data. JMIR biome dic al engine ering , 5(1): e17106, 2020. ISSN 2561-3278. doi: 10.2196/17106. [26] Akbar Dehghani, Omid Sarbishei, T ristan Glatard, and Emad Shihab. A Quan titative Comparison of Overlapping and Non-Overlapping Sliding Windows for Human A ctivit y Recognition Using Inertial Sensors. Sensors , 19(22):5026, Jan uary 2019. ISSN 1424-8220. doi: 10.3390/s19225026. [27] Janez Demšar. Statistical Comparisons of Classiﬁers ov er Multiple Data Sets. J. Mach. L e arn. R es. , 7: 1–30, December 2006. ISSN 1532-4435. [28] Seyma Derdiyok, F atma Patlar Akbulut, and Cagatay Catal. Neurophysiological and biosignal data for in vestigating o ccupational mental fatigue: MEF AR dataset. Data in Brief , 52:109896, F ebruary 2024. ISSN 2352-3409. doi: 10.1016/j.dib.2023.109896. [29] Elena Di Lascio, Shkurta Gashi, and Silvia Santini. Unobtrusive Assessmen t of Studen ts’ Emotional Engagemen t during Lectures Using Electro dermal Activit y Sensors. Pr o c. A CM Inter act. Mob. W e ar able Ubiquitous T e chnol. , 2(3):103:1–103:21, September 2018. doi: 10.1145/3264913. [30] Elena Di Lascio, Shkurta Gashi, and Silvia Santini. Laughter Recognition Using Non-inv asiv e W earable Devices. In Pr o c e e dings of the 13th EAI International Confer enc e on Pervasive Computing T e chnolo gies for He althc ar e , P erv asiv eHealth’19, pages 262–271, New Y ork, NY, USA, May 2019. Association for Computing Machinery . ISBN 978-1-4503-6126-2. doi: 10.1145/3329189.3329216. [31] Elena Di Lascio, Shkurta Gashi, Maik e E. Debus, and Silvia Santini. A utomatic Recognition of Flow During W ork Activities Using Context and Physiological Signals. In 2021 9th International Confer enc e on Aﬀe ctive Computing and Intel ligent Inter action (A CII) , pages 1–8, Nara, Japan, September 2021. IEEE. doi: 10.1109/ACII52823.2021.9597434. [32] Cheng Ding, Zhicheng Guo, Zhaoliang Chen, Randall J Lee, Cynthia Rudin, and Xiao Hu. SiamQuality: A ConvNet-based foundation mo del for photopleth ysmograph y signals. Physiolo gic al Me asur ement , 45 (8):085004, August 2024. ISSN 0967-3334. doi: 10.1088/1361- 6579/ad6747. [33] Vipula Dissana y ake, Sachith Seneviratne, Ra jib Rana, Elliott W en, Tharindu Kaluarac hc hi, and Suranga Nanay akkara. SigRep: T o ward Robust W earable Emotion Recognition With Contrastiv e Rep- resen tation Learning. IEEE A c c ess , 10:18105–18120, 2022. ISSN 2169-3536. doi: 10.1109/ACCESS. 2022.3149509. [34] Alexey Dosovitskiy , Lucas Beyer, Alexander Kolesnik o v, Dirk W eissen b orn, Xiaohua Zhai, Thomas Un- terthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylv ain Gelly , Jakob Uszkoreit, and Neil Houlsby . An Image is W orth 16x16 W ords: T ransformers for Image Recognition at Scale. In Inter- national Confer enc e on L e arning R epr esentations , Addis Ababa, Ethiopia, Octob er 2020. International Conference on Learning Represen tations (ICLR). [35] V asilii F eofanov, Songkang W en, Marius Alonso, Romain Ilbert, Hongb o Guo, Malik Tiomoko, Lujia P an, Jianfeng Zhang, and Ievgen Redk o. Mantis: Light w eigh t Calibrated F oundation Model for User- F riendly Time Series Classiﬁcation, F ebruary 2025. 18 A f ounda tion model f or electr odermal activity da t a [36] Nan Gao, W ei Shao, Mohammad Saiedur Rahaman, and Flora D. Salim. N-Gage: Predicting in-class Emotional, Beha vioural and Cognitiv e Engagemen t in the Wild. Pr o c. A CM Inter act. Mob. W e ar able Ubiquitous T e chnol. , 4(3):79:1–79:26, September 2020. doi: 10.1145/3411813. [37] F rancisco M. Garcia-Moreno and Marta Badenes-Sastre. EmoPulse Momen ts E4 Dataset (EPM-E4): An Exhaustive Collection of Emotion-Related Data from Empatica E4 W earable, September 2020. [38] Shkurta Gashi, Elena Di Lascio, and Silvia San tini. Using Unobtrusive W earable Sensors to Measure the Ph ysiological Sync hrony Betw een Presen ters and A udience Mem bers. Pr o c. A CM Inter act. Mob. W e ar able Ubiquitous T e chnol. , 3(1):13:1–13:19, March 2019. doi: 10.1145/3314400. [39] Shkurta Gashi, Elena Di Lascio, Bianca Stancu, V edan t Das Swain, V arun Mishra, Martin Gjoreski, and Silvia Santini. Detection of Artifacts in Am bulatory Electro dermal Activit y Data. Pr o c. A CM Inter act. Mob. W e ar able Ubiquitous T e chnol. , 4(2):44:1–44:31, June 2020. doi: 10.1145/3397316. [40] Shkurta Gashi, Lidia Alecci, Elena Di Lascio, Maike E. Debus, F rancesca Gasparini, and Silvia Santini. The Role of Model P ersonalization for Sleep Stage and Sleep Quality Recognition Using W earables. IEEE Pervasive Computing , 21(2):69–77, April 2022. ISSN 1558-2590. doi: 10.1109/MPR V.2022.3164334. [41] Shkurta Gashi, Chulhong Min, Alessandro Montanari, Silvia Santini, and F ahim Kawsar. A multidevice and m ultimo dal dataset for h uman energy expenditure estimation using w earable devices. Scientiﬁc Data , 9(1):537, September 2022. ISSN 2052-4463. doi: 10.1038/s41597- 022- 01643- 5. [42] Martin Gjoreski, Tine Kolenik, Timotej Knez, Mitja Luštrek, Matjaž Gams, Hristijan Gjoreski, and V eljk o Pejo vić. Datasets for Cognitive Load Inference Using W earable Sensors and Psychological T raits. A pplie d Scienc es , 10(11):3843, January 2020. ISSN 2076-3417. doi: 10.3390/app10113843. [43] Mononito Goswami, K onrad Szafer, Arjun Choudhry , Yifu Cai, Shuo Li, and Artur Dubrawski. MO- MENT: A family of op en time-series foundation mo dels. In Pr o c e e dings of the 41st International Con- fer enc e on Machine L e arning , volume 235 of ICML’24 , pages 16115–16152, Vienna, Austria, July 2024. JMLR.org. [44] Margherita Grandini, Enrico Bagli, and Giorgio Visani. Metrics for Multi-Class Classiﬁcation: An Ov erview, A ugust 2020. [45] Alb erto Greco, Gaetano V alenza, Antonio Lanata, Enzo Pasquale Scilingo, and Luca Citi. cvxEDA: A Con vex Optimization Approach to Electro dermal Activit y Pro cessing. IEEE T r ansactions on Biome dic al Engine ering , 63(4):797–804, April 2016. ISSN 1558-2531. doi: 10.1109/TBME.2015.2474131. [46] Md-Billal Hossain, Y oungsun Kong, Hugo F. Posada-Quin tero, and Ki H. Chon. Comparison of Electro- dermal Activit y from Multiple Body Lo cations Based on Standard EDA Indices’ Quality and Robustness against Motion Artifact. Sensors , 22(9):3177, Jan uary 2022. ISSN 1424-8220. doi: 10.3390/s22093177. [47] Md. Billal Hossain, Y oungsun Kong, Hugo P osada-Quintero, and Ki Chon. Ele ctr o dermal A ctivity: A pplic ations and Chal lenges , pages 475–496. Cam bridge Univ ersit y Press, Nov ember 2024. ISBN 978- 1-316-51855-7. doi: 10.1017/9781009000796.022. [48] Seyedmajid Hosseini, Ra ju Gottum ukkala, Saty a Katragadda, Ravi T eja Bhupatiraju, Ziad Ashkar, Christoph W. Borst, and Kenneth Co chran. A multimodal sensor dataset for contin uous stress detection of nurses in a hospital. Scientiﬁc Data , 9(1):255, June 2022. ISSN 2052-4463. doi: 10.1038/s41597- 022- 01361- y . [49] T alha Iqbal, Andrew J. Simpkin, Dav o o d Roshan, Nicola Glynn, John Killilea, Jane W alsh, Gerard Mollo y , Sandra Ganly , Hannah R yman, Eileen Co en, Adnan Elahi, William Wijns, and A tif Shahzad. Stress Monitoring Using W earable Sensors: A Pilot Study and Stress-Predict Dataset. Sensors , 22(21): 8135, January 2022. ISSN 1424-8220. doi: 10.3390/s22218135. [50] Alistair E. W. Johnson, T om J. Pollard, Lu Shen, Li-w ei H. Lehman, Mengling F eng, Mohammad Ghassemi, Benjamin Mo o dy , P eter Szolo vits, Leo Anthon y Celi, and Roger G. Mark. MIMIC-I I I, a freely accessible critical care database. Scientiﬁc Data , 3(1):160035, May 2016. ISSN 2052-4463. doi: 10.1038/sdata.2016.35. [51] Alistair E. W. Johnson, Lucas Bulgarelli, Lu Shen, Alvin Gayles, A yad Shammout, Stev en Horng, T om J. P ollard, Sicheng Hao, Benjamin Mo o dy , Brian Gow, Li-wei H. Lehman, Leo A. Celi, and Roger G. Mark. MIMIC-IV, a freely accessible electronic health record dataset. Scientiﬁc Data , 10(1):1, January 2023. ISSN 2052-4463. doi: 10.1038/s41597- 022- 01899- x. [52] Daniel Kahneman, Bernard T ursky , David Shapiro, and Andrew Crider. Pupillary , heart rate, and skin resistance changes during a mental task. Journal of exp erimental psycholo gy , 79(1p1):164, 1969. 19 A f ounda tion model f or electr odermal activity da t a [53] Diederik P . Kingma and Jimmy Ba. A dam: A Metho d for Sto chastic Optimization, 2017. [54] Zuzana Kosco v a, V aldery Moura Junior, Matthew Reyna, Shenda Hong, A dit ya Gupta, Manohar Ghan ta, Reza Sameni, Aaron Aguirre, Qiao Li, Sahar Zafar, Gari Cliﬀord, and M Brandon W esto v er. Harv ard-Emory ECG Database, 2025. [55] Matias Lap orte, Martin Gjoreski, and Marc Langheinrich. LA UREA TE: A Dataset for Supp orting Re- searc h in Aﬀective Computing and Human Memory Augmen tation. Pr o c. A CM Inter act. Mob. W e ar able Ubiquitous T e chnol. , 7(3):106:1–106:41, September 2023. doi: 10.1145/3610892. [56] Phuc H. Le-Khac, Graham Healy , and Alan F. Smeaton. Con trastive Representation Learning: A F ramew ork and Review. IEEE A c c ess , 8:193907–193934, 2020. ISSN 2169-3536. doi: 10.1109/A CCESS. 2020.3031549. [57] Hyung-Chul Lee, Y oonsang Park, So o Bin Y o on, Seong Mi Y ang, Dongny eok Park, and Chul-W o o Jung. VitalDB, a high-ﬁdelity multi-parameter vital signs database in surgical patients. Scientiﬁc Data , 9(1): 279, June 2022. ISSN 2052-4463. doi: 10.1038/s41597- 022- 01411- 5. [58] Simon Lee and Kai Akamatsu. F oundation Mo dels for Ph ysiological Signals: Opp ortunities and Chal- lenges, August 2025. [59] Jun Li, Aaron D. Aguirre, V aldery Moura, Jiarui Jin, Che Liu, Lanhai Zhong, Chenxi Sun, Gari Cliﬀord, M. Brandon W esto ver, and Shenda Hong. An Electro cardiogram F oundation Model Built on o v er 10 Million Recordings. Nejm A i , 2(7):AIoa2401033, July 2025. ISSN 2836-9386. doi: 10.1056/aioa2401033. [60] Y u Liu, María-Itatí P alacio, T aha Bikki, Cesar T oledo, Y u Ouy ang, Zhongzheng Li, Zhengyi W ang, F rancisco T oledo, Hong Zeng, and María-T rinidad Herrero. Machine Learning, Physiological Signals, and Emotional Stress/Anxiety: Pitfalls and Challenges. Applie d Scienc es , 15(21):11777, January 2025. ISSN 2076-3417. doi: 10.3390/app152111777. [61] Junoš Lukan, Martin Gjoreski, Heidi Mauersb erger, Annekatrin Hopp e, Ursula Hess, and Mitja Luštrek. Analysing Physiology of Interpersonal Conﬂicts Using a W rist Device. In A c hilles Kameas and Kostas Stathis, editors, A mbient Intel ligenc e , pages 162–167, Cham, 2018. Springer International Publishing. ISBN 978-3-030-03062-9. doi: 10.1007/978- 3- 030- 03062- 9_13. [62] Y unfei Luo, Y uliang Chen, Asif Salekin, and T auhidur Rahman. T ow ard F oundation Mo del for Multi- v ariate W earable Sensing of Physiological Signals, May 2025. [63] Erika Lutin, Ryuga Hashimoto, W alter De Raedt, and Chris V an Ho of. F eature Extraction for Stress Detection in Electro dermal Activit y:. In Pr o c e e dings of the 14th International Joint Confer enc e on Biome dic al Engine ering Systems and T e chnolo gies , pages 177–185, Online Streaming, — Select a Coun- try —, 2021. SCITEPRESS - Science and T ec hnology Publications. ISBN 978-989-758-490-9. doi: 10.5220/0010244601770185. [64] Katie Matton, Rob ert Lewis, John Guttag, and Rosalind Picard. Contrastiv e Learning of Electro dermal A ctivity Represen tations for Stress Detection. In Pr o c e e dings of the Confer enc e on He alth, Infer enc e, and L e arning , pages 410–426, Cam bridge, MA, USA, June 2023. PMLR. [65] Daniel McDuﬀ, Seamus Thomson, Sam y Ab del-Ghaﬀar, Isaac R. Galatzer-Levy, Ming-Zher Poh, Jake Sunshine, Andrew Barakat, Conor Heneghan, and Lindsey Sunden. What Do es Large-scale Electro der- mal Sensing Reveal?, F ebruary 2024. [66] Kaden McKeen, Sameer Maso o d, Augustin T oma, Barry Rubin, and Bo W ang. ECG-FM: An op en electro cardiogram foundation mo del. JAMIA op en , 8(5):o oaf122, Octob er 2025. ISSN 2574-2531. doi: 10.1093/jamiaop en/o oaf122. [67] V arun Mishra, Sougata Sen, Grace Chen, Tian Hao, Jeﬀrey Rogers, Ching-Hua Chen, and David Kotz. Ev aluating the Reproducibility of Ph ysiological Stress Detection Mo dels. Pr o c. A CM Inter act. Mob. W e ar able Ubiquitous T e chnol. , 4(4):147:1–147:29, December 2020. doi: 10.1145/3432220. [68] Girish Nara yansw amy , Xin Liu, Kumar A yush, Y uzhe Y ang, Xuhai Xu, Sh un Liao, Jake Garrison, Sh yam T ailor, Jak e Sunshine, Y un Liu, Tim Althoﬀ, Shrikanth Naray anan, Pushmeet Kohli, Jiening Zhan, Mark Malhotra, Sh wetak Patel, Samy Ab del-Ghaﬀar, and Daniel McDuﬀ. Scaling W earable F oundation Mo dels, Octob er 2024. [69] Sameer Neupane, Mithun Saha, Nasir Ali, Timothy Hnat, Shahin Alan Samiei, Anandatirtha Nandugudi, Da vid M. Almeida, and Santosh Kumar. Momentary Stressor Logging and Reﬂective Visualizations: Implications for Stress Management with W earables. In Pr o c e e dings of the 2024 CHI Confer enc e on Hu- man F actors in Computing Systems , CHI ’24, pages 1–19, New Y ork, NY, USA, May 2024. Asso ciation for Computing Machinery . ISBN 979-8-4007-0330-0. doi: 10.1145/3613904.3642662. 20 A f ounda tion model f or electr odermal activity da t a [70] Michael Owusu-Adjei, James Ben Hayfron-A cquah, T wum F rimp ong, and Gaddaﬁ Abdul-Salaam. Im- balanced class distribution and p erformance ev aluation metrics: A systematic review of prediction ac- curacy for determining mo del p erformance in healthcare systems. PLOS Digital He alth , 2(11):e0000290, No vem ber 2023. ISSN 2767-3170. doi: 10.1371/journal.p dig.0000290. [71] Adam Paszk e, Sam Gross, F rancisco Massa, A dam Lerer, James Bradbury , Gregory Chanan, T rev or Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Y ang, Zac h De Vito, Martin Raison, Alykhan T ejani, Sasank Chilamkurthy , Benoit Steiner, Lu F ang, Junjie Bai, and Soumith Chin tala. PyT or ch: A n Imp er ative Style, High-Performanc e De ep L e arning Libr ary , pages 8026–8037. Curran Asso ciates Inc., Red Ho ok, NY, USA, 2019. [72] Arvind Pillai, Dimitris Spathis, F ahim Ka wsar, and Mohammad Malekzadeh. PaP aGei: Op en F ounda- tion Mo dels for Optical Physiological Signals, F ebruary 2025. [73] Anuja Pinge, Vina y a Gad, Dheryta Jaisighani, Surjya Ghosh, and Sougata Sen. Detection and moni- toring of stress using wearables: a systematic review. F r ontiers in Computer Scienc e , 6:1478851, 2024. [74] Ming-Zher Poh, T obias Lo ddenkemper, Claus Reinsb erger, Nicholas C. Swenson, Shubhi Goy al, Mangw e C. Sabtala, Joseph R. Madsen, and Rosalind W. Picard. Convulsiv e seizure detection us- ing a wrist-worn electro dermal activity and accelerometry biosensor. Epilepsia , 53(5):e93–97, May 2012. ISSN 1528-1167. doi: 10.1111/j.1528- 1167.2012.03444.x. [75] Md. Raﬁul Amin, Dilranjan S. Wickramasuriy a, and Rose T. F aghih. A W earable Exam Stress Dataset for Predicting Grades using Ph ysiological Signals. In 2022 IEEE He althc ar e Innovations and Point of Car e T e chnolo gies (HI-POCT) , pages 30–36, Houston, TX, USA, March 2022. IEEE. doi: 10.1109/ HI- POCT54491.2022.9744065. [76] Saeed Ur Rehman, Anw ar Ali, Adil Mehmo o d Khan, and Cynthia Okpala. Human A ctivity Recognition: A Comparative Study of V alidation Methods and Impact of F eature Extraction in W earable Sensors. A lgorithms , 17(12):556, December 2024. ISSN 1999-4893. doi: 10.3390/a17120556. [77] Attila Reiss, Ina Indlekofer, Philip Schmidt, and Kristof V an Laerhov en. Deep PPG: Large-Scale Heart Rate Estimation with Conv olutional Neural Netw orks. Sensors , 19(14):3079, January 2019. ISSN 1424- 8220. doi: 10.3390/s19143079. [78] Matthew A Reyna, Nadi Sadr, Eric k A Perez Alday , Annie Gu, Amit J Shah, Chad Robic haux, Ali Bahrami Rad, Andoni Elola, Salman Seyedi, Sardar Ansari, Hamid Ghanbari, Qiao Li, Ashish Sharma, and Gari D Cliﬀord. Will T wo Do? V arying Dimensions in Electro cardiography: The Phys- ioNet/Computing in Cardiology Challenge 2021. In 2021 Computing in Car diolo gy (CinC) , volume 48, pages 1–4, Brno, Czec h Republic, September 2021. IEEE. doi: 10.23919/CinC53138.2021.9662687. [79] William L Romine, Noah L Schroeder, Josephine Graft, F an Y ang, Reza Sadeghi, Mahdieh Zabihi- ma yv an, Dip esh Kadariy a, and T anvi Banerjee. Using machine learning to train a w earable device for measuring students cognitive load during problem-solving activities based on electro dermal activ- it y , bo dy temperature, and heart rate: Developmen t of a cognitiv e load trac ker for b oth personal and classro om use. Sensors , 20(17):4833, 2020. [80] Aaqib Saeed, Flora D. Salim, T anir Ozcelebi, and Johan Lukkien. F ederated Self-Sup ervised Learning of Multisensor Representations for Em b edded In telligence. IEEE Internet of Things Journal , 8(2): 1030–1040, January 2021. ISSN 2327-4662. doi: 10.1109/JIOT.2020.3009358. [81] Mithun Saha, Maxwell A. Xu, W an ting Mao, Sameer Neupane, James M. Rehg, and San tosh Kumar. Pulse-PPG: An Op en-Source Field-T rained PPG F oundation Model for W earable Applications across Lab and Field Settings. Pr o c. A CM Inter act. Mob. W e ar able Ubiquitous T e chnol. , 9(3):126:1–126:35, Septem b er 2025. doi: 10.1145/3749494. [82] Philip Schmidt, A ttila Reiss, Rob ert Dueric hen, Claus Marb erger, and Kristof V an Laerhov en. In- tro ducing WESAD, a Multimo dal Dataset for W earable Stress and Aﬀect Detection. In Pr o c e e dings of the 20th A CM International Confer enc e on Multimo dal Inter action , ICMI ’18, pages 400–408, New Y ork, NY, USA, Octob er 2018. Association for Computing Machinery . ISBN 978-1-4503-5692-3. doi: 10.1145/3242969.3242985. [83] Christopher Slade, Yinan Sun, W ei Cheng Chao, Chih-Chun Chen, Rob erto M. Benzo, and Peter W ashington. Current challenges and opp ortunities in active and passive data collection for mobile health sensing: A scoping review. JAMIA op en , 8(4):o oaf025, A ugust 2025. ISSN 2574-2531. doi: 10.1093/jamiaop en/o oaf025. 21 A f ounda tion model f or electr odermal activity da t a [84] Jamiyan Sukhbaatar, Satoshi Imamura, Ibuki Inoue, Shoy a Murakami, Kazi Mahmudul Hassan, Seung- w o o Han, Ingon Chanp ornpakdi, and T oshihisa T anaka. SingLEM: Single-Channel Large EEG Model, Septem b er 2025. [85] Henrik Svoren, V a jira Thamba wita, Pål Halvorsen, Petter Jak obsen, Enrique Garcia-Ceja, F arzan Ma- jeed No ori, Hugo L. Hammer, Mathias Lux, Michael Alexander Riegler, and Steven Alexander Hic ks. T oadsto ol: A dataset for training emotional intelligen t mac hines playing Super Mario Bros. In Pr o c e e dings of the 11th A CM Multime dia Systems Confer enc e , MMSys ’20, pages 309–314, New Y ork, NY, USA, May 2020. Asso ciation for Computing Machinery . ISBN 978-1-4503-6845-2. doi: 10.1145/3339825.3394939. [86] Mingxing T an and Quoc Le. Eﬃcientnet: Rethinking model scaling for conv olutional neural net w orks. In International c onfer enc e on machine le arning , pages 6105–6114, Long Beac h, CA, USA, 2019. PMLR, PMLR. [87] Andrés T ello, Victoria Degeler, and Alexander Lazovik. T oo go o d to b e true: A ccuracy ov erestimation in (re) curren t practices for human activity recognition. In 2024 IEEE International Confer enc e on Per- vasive Computing and Communic ations W orkshops and Other Aﬃliate d Events (PerCom W orkshops) , pages 511–517, Biarritz, F rance, 2024. IEEE. [88] James T ruslow, Angela Spillane, Huiming Lin, Katherine Cyr, Adeeti Ullal, Edith Arnold, Ron Huang, Laura Rho des, Jennifer Blo c k, Jamie Stark, James Kretlo w, Alexis L. Beatty , Andreas W erdich, Deepali Bankar, Matt Bianchi, Ian Shapiro, Jaime Villalpando, Sharon Ravindran, Irida Mance, Adam Phillips, John Earl, Rahul C. Deo, Sumbul A. Desai, and Calum A. MacRae. Understanding activity and ph ysiology at scale: The Apple Heart & Mov emen t Study. npj Digital Me dicine , 7(1):242, September 2024. ISSN 2398-6352. doi: 10.1038/s41746- 024- 01187- 5. [89] Aaron v an den Oord, Y azhe Li, and Oriol Viny als. Representation Learning with Con trastive Predictiv e Co ding, January 2019. [90] Y eduk ondala Rao V eeranki, Nagara jan Ganapathy , Ramakrishnan Swaminathan, and Hugo F. Posada- Quin tero. Comparison of Electro dermal Activit y Signal Decomp osition T echniques for Emotion Recog- nition. IEEE A c c ess , 12:19952–19966, 2024. ISSN 2169-3536. doi: 10.1109/A CCESS.2024.3361832. [91] Miguel Viana-Matesanz and Carmen Sánc hez-Ávila. Adaptiv e Normalization and F eature Extraction for Electro dermal Activit y Analysis. Mathematics , 12(2):202, January 2024. ISSN 2227-7390. doi: 10.3390/math12020202. [92] Ke W ang, Jiamu Y ang, A yush Shett y , and Jessilyn Dunn. DREAMT: Dataset for Real-time sleep stage EstimA tion using Multisensor wearable T echnology, 2024. [93] Rong jie Yi, Ting Cao, A o Zhou, Xiao Ma, Shangguang W ang, and Mengwei Xu. Bo osting DNN Cold Inference on Edge Devices. In Pr o c e e dings of the 21st A nnual International Confer enc e on Mobile Systems, A pplic ations and Servic es , MobiSys ’23, pages 516–529, New Y ork, NY, USA, June 2023. Asso ciation for Computing Mac hinery . ISBN 979-8-4007-0110-8. doi: 10.1145/3581791.3596842. [94] Guo-Qiang Zhang, Licong Cui, Remo Mueller, Shiqiang T ao, Matthew Kim, Michael Ruesc hman, Sara Mariani, Daniel Mobley , and Susan Redline. The National Sleep Research Resource: T ow ards a sleep data commons. Journal of the A meric an Me dic al Informatics A sso ciation , 25(10):1351–1358, Octob er 2018. ISSN 1527-974X. doi: 10.1093/jamia/o cy064. [95] Y u wei Zhang, Kumar A yush, Siyuan Qiao, A. Ali Heydari, Girish Nara yansw am y , Maxwell A. Xu, Ahmed A. Metw ally , Shawn Xu, Jake Garrison, Xuhai Xu, Tim Althoﬀ, Y un Liu, Pushmeet Kohli, Jiening Zhan, Mark Malhotra, Shw etak P atel, Cecilia Mascolo, Xin Liu, Daniel McDuﬀ, and Y uzhe Y ang. SensorLM: Learning the Language of W earable Sensors, June 2025. 22 A f ounda tion model f or electr odermal activity da t a T able A.1: Summary of datasets with their corresponding data-sharing agreements. This table includes all datasets, including citations and the "Reshareable" column, for inclusion in the app endix. Dataset Name License BiHeartS [ 3 ] 3 - data share agreemen t BIG IDEAS [11] 3 - with original license (ODC-BY 1) Dynamics in the w orkplace [61] 3 - data share agreement EmpaticaE4Stress [20] 3 - (CCBY 4) with original license EPM-E4 [37] 3 - (CCBY 4) with original license LA UREA TE [55] 3 - data share agreemen t MEF AR [28] 3 - (CCBY 4) with original license M2Sleep [40] 3 - data share agreemen t PPG-Dalia [77] 3 - (CCBY 4) with original license SEED [29] 3 - data share agreemen t SEED-I I-Lab [29] 3 - data share agreemen t SEED-I I-Wild [29] 3 - data share agreemen t Stress Predict [49] 3 - (MIT) with original license T oadStool [85] 3 - (CCBY 4) with original license WEEE [41] 3 - (CCBY 4) with original license WESD [75] 3 - with original license (ODC-BY 1) W orkplace [31] 3 - data share agreemen t APSync [38] 3 - data share agreement HeartS [ 2 ] 3 - data share agreemen t USILaughs [30] 3 - data share agreement WESAD [82] 3 - (CCBY 4) with original license Nurses’ Stress [48] 3 - (ODbL 1) with original license DREAMT [92] 7 - PhysioNet Restricted Health Data Use Agreement 1.5.0 HHISS [42] 7 - no license sp eciﬁed A Metho dology , Ev aluation, T ransparency , and A v ailabilit y (MET A) App endix A.1 A dditional dataset information W e report in T able A.1 information ab out the original license asso ciated with the datasets making up the ED AMAME collection. The datasets that we re-share are all under their original license, if so required. F or the datasets that require a data sharing agreemen t to be signed, we contacted the original authors, which all agreed to for re-sharing under the conditions in the original data sharing agreement. A.2 Baseline feature sets In section 5.1 we explain how w e use tw o handcrafted feature sets as baseline to ev aluate the embeddings from our UME foundation mo del. The ﬁrst feature set, whic h w e call generic handcr afte d fe atur e set , consists of four statistics, i.e., mean, minimum, maxim um and standard deviation, computed for b oth the phasic and tonic ED A comp onents, as well as the original non-decomp osed EDA signal. This feature set emulates baseline used b y related work to ev aluate foundation mo dels for ph ysiological data [ 1 , 72]. The second set of features, which we call EDA-sp e ciﬁc handcr afte d fe atur e set , is a broader set of features commonly used in EDA classiﬁcation pip elines [ 4 , 39]. In addition to the generic statistics, this set incorp orates signal dynamics (e.g., slop es and deriv ativ es), morphological p eak characteristics, and frequency-domain comp onen ts. This set consists of 15 base features, resulting in a total of 45 features p er window (15 features × 3 EDA comp onents). T able A.3 details the exact mathematical formulations for all 15 base features and speciﬁes their inclusion in the resp ective baseline sets. W e rep ort in T able A.2 the feature size of the feature extraction metho ds used in this work. 23 A f ounda tion model f or electr odermal activity da t a T able A.2: Overview of the models and feature sets used in the exp erimental ev aluation. W e rep ort the mo del size (num ber of trainable parameters) and the dimensionality of the feature space ( d ) used for linear probing. Mo del / F eature Set Mo del Size F eature Dim. ( d ) Handcrafted Metho ds Generic HC N/A 12 ED A-sp eciﬁc HC N/A 45 Generalist F oundation Mo dels Man tis [35] ∼ 8 M 768 Chronos [ 8 ] ∼ 200 M 1536 MOMENT [43] ∼ 385 M 1024 Ours UME ∼ 1 M 64 A.3 Hyp erparameter grid for the logistic regression In T able A.4 we rep ort the hyperparameter grid we used with the logistic regression. W e used a 3-fold inner cross-v alidation to search for the b est conﬁguration of hyperparameter during our feature ev aluation through linear probing. A.4 T raining details W e train the UME foundation model using the A dam optimizer [53] (learning rate 0.001 and weigh t deca y 0.01). W e also use a learning rate sc heduler, Reduce On Plateau (factor 0.5), to decrease the learning rate during training. W e train the mo del for a maximum of 400 ep o chs, with early stopping, with a batch size of 512. In total, w e train our model for approximately 5 days, using an Nvidia A6000 GPU 14 . W e implement the foundation mo del in Python, using the Pytorch [ 71] and Pytorch-Ligh tning libraries. A.5 Data augmen tations for EDA data W e rep ort in T able A.5 the list of data augmentations used to train our UME foundation mo del with con trastive learning. W e use the same set prop osed b y Matton et al. [64], which contains EDA-speciﬁc augmen tations. 14 https://www.nvidia.com/en- us/products/workstations/rtx- a6000/ 24 A f ounda tion model f or electr odermal activity da t a T able A.3: Mathematical formulations for the 15 base hand-crafted features extracted from EDA time series signals ( x ) of length N . All features are computed independently for the tonic, phasic, and original EDA comp onen ts. The "F eature Set(s)" column indicates whether the feature w as used in the Generic baseline (12 total features) or the EDA-speciﬁc baseline (45 total features). F eature Mathematical Notation or F ormula F eature Set(s) Time-domain Mean 1 N P N i =1 x i Generic & EDA-speciﬁc Minim um min ( x 1 , x 2 , . . . , x N ) Generic & EDA-speciﬁc Maxim um max ( x 1 , x 2 , . . . , x N ) Generic & EDA-speciﬁc Standard Deviation q 1 N − 1 P N i =1 ( x i − ¯ x ) 2 Generic & EDA-speciﬁc Dynamic Range max ( x 1 , x 2 , . . . , x N ) − min ( x 1 , x 2 , . . . , x N ) ED A-speciﬁc only Slop e x N − x 1 N − 1 ED A-sp eciﬁc only Absolute V alue of Slop e    x N − x 1 N − 1    ED A-sp eciﬁc only Mean of the First Deriv ative 1 N − 1 P N − 1 i =1 ( x i +1 − x i ) ED A-speciﬁc only Std. Dev. of the First Deriv ativ e q 1 N − 2 P N − 1 i =1 (( x i +1 − x i ) − ¯ x ′ ) 2 ED A-sp eciﬁc only Num b er of EDA Peaks Count of lo cal maxima in the window ED A-sp eciﬁc only Amplitude of EDA Peaks Mean amplitude of lo cal maxima in the window EDA-speciﬁc only F requency Domain (F ast F ourier T ransform) Direct Current (DC) X 0 ED A-sp eciﬁc only Sum of F requency Co eﬃcients P N k =1 | X k | ED A-sp eciﬁc only Information Entrop y − P N k =1 P ( X k ) log 2 ( P ( X k )) ED A-sp eciﬁc only Sp ectral Energy P N k =1 | X k | 2 ED A-sp eciﬁc only T able A.4: Hyperparameter Grid for Logistic Regression Hyp erparameter V alues Regularization strength ( C ) 0.01, 0.1, 1, 10 Solv er lbfgs, liblinear P enalty L2 Max iterations 10000 Class weigh t balanced 25 A f ounda tion model f or electr odermal activity da t a T able A.5: Summary of Electro dermal Activit y (EDA) Data Augmen tations used to train the UME founda- tion mo del. The table details b oth EDA-speciﬁc transforms designed to isolate ph ysiological comp onents or sim ulate artifacts, and generic time series transforms. The set is the same prop osed by Matton et al. [64]. A ugmen tation T yp e Description P arameter Range F requency Domain & Comp onen t Isolation Lo w-Pass Filter EDA-Speciﬁc Extracts tonic comp onent; remov es high-freq noise. Cutoﬀ f ∈ [0 . 25 , 1 . 0] Hz High-P ass Filter ED A-Sp eciﬁc Extracts phasic comp onent; remo v es slo w drifts. Cutoﬀ f ∈ [0 . 05 , 0 . 25] Hz Band-P ass Filter ED A-Sp eciﬁc Isolates information-ric h ED A fre- quency bands. P ass f ∈ [0 . 05 , 0 . 25] Hz Band-Stop Filter EDA-Speciﬁc Rejects sp eciﬁc frequency bands. Reject f ∈ [0 . 75 , 1 . 0] Hz High F req. Noise ED A-Speciﬁc Adds Gaussian noise only to frequencies > 1 Hz. Noise σ ∈ [0 , 0 . 5] Artifact Sim ulation Jump Artifact EDA-Speciﬁc Simulates abrupt sensor mov e- men t/displacement. Jump ∈ [0 . 01 , 0 . 2] µS Lo ose Sensor ED A-Sp eciﬁc Simulates electro de contact loss (signal drop). Duration t ∈ [5 , 20] s Thermoregulation Sim ulation T onic Const. Scale ED A-Sp eciﬁc Scales tonic comp onent (simulates con- stan t temp). F actor s ∈ [0 . 25 , 2] T onic Amp. W arp ED A-Speciﬁc Time-v arying scale of tonic comp onent (c hanging conditions). Spline σ ∈ [0 . 01 , 0 . 05] Generic Time Series Amp. Const. Scale Generic Applies constant scaling factor to the ra w signal. F actor s ∈ [0 . 25 , 2] Amplitude W arp Generic Applies smo oth, time-v arying scaling to ra w signal. Spline σ ∈ [0 . 01 , 0 . 05] Gaussian Noise Generic A dds random Gaussian noise to the ra w signal. σ ∈ [0 , 0 . 5] Time Shift Generic Shifts the signal window forward or bac kward. Shift t ∈ [5 , 45] s T emporal Cutout Generic Masks/zero es out a random sub- windo w of the signal. Cutout t ∈ [5 , 15] s Time W arp Generic P erturbs temp oral dimension (lo cal stretc h/compress). Spline σ ∈ [0 . 01 , 0 . 1] P ermutation Generic Slices signal and randomly reorders sub- windo ws. Segmen ts n ∈ [2 , 6] Signal Flip Generic In v erts the signal ov er its amplitude di- mension. N/A 26 A f ounda tion model f or electr odermal activity da t a Figure B.1: Example of EDA signal reconstruction using MAE. B Sensitivit y & additional mo del studies W e p erform sensitivity and additional studies to ﬁnd the conﬁguration of our UME foundation mo d el. In this section, we rep ort details for: implemen tation of an additional NN architecture, i.e., a Masked Autoenco der (MAE); studies on mo del size and h yp erparameters for the EﬃcientNet architecture chosen for the UME foundation mo del, trained using contrastiv e learning. W e p erform all training in this section using a subset (about 15% of the total data) of the train part of the EDAMAME collection of datasets. W e do this to sp eed-up experimentation since, as rep orted in subsection 4.3, the complete mo del training takes ab out 5 days in total with a single A6000 NVIDIA GPU, due to the large scale of the train set. B.1 Exp erimen ts with Masked Autoenco ders (MAEs) W e train a reconstructive-based masked auto enco der (MAE), based on vision transformer (ViT) [34]. Nara yansw am y et al. [68] also used a masked-autoenco der to train a foundation mo del for physiological data. W e train the model using the follo wing loss: let x ∈ R T × C denote the input ED A signal, partitioned in to N non-ov erlapping patches of size P , and let ˆ x i and x i denote the reconstructed and original i -th patch, resp ectiv ely . Given a binary mask m ∈ { 0 , 1 } N , where m i = 1 indicates a masked patch and m i = 0 a visible one, the reconstruction loss is deﬁned as L = α L masked + (1 − α ) L visible , (2) with L masked = 1 P i m i N X i =1 m i ℓ ( ˆ x i , x i ) , (3) L visible = 1 P i (1 − m i ) N X i =1 (1 − m i ) ℓ ( ˆ x i , x i ) , (4) where ℓ ( · , · ) denotes either the mean absolute error (MAE) computed within each patch, and α ∈ [0 , 1] con trols the relative imp ortance of masked versus visible patch reconstruction. W e exp eriment with the following conﬁgurations: m ∈ { 0 , 0 . 1 , 0 . 4 } and α ∈ { 0 . 1 , 0 . 5 } . In Figure B.2 we rep ort the v alidation loss during training curve for the conﬁgurations tested, and in Figure B.1 w e sho w an example of a signal reconstruction at train end, ov er a v alidation sample. W e also p erform do wnstream ev aluation on the BiHeartS dataset: how ev er, no conﬁguration leads to p erformance, in terms of balanced accuracy , ab ov e that of the Dummy classiﬁer. Giv en these ﬁndings, we b elieve that masked auto enco ders are not optimal for the sp eciﬁc EDA data present in the EDAMAME collection of datasets. B.2 Eﬃcien tNet ablation studies W e adopt an EﬃcientNet architecture for our UME foundation mo del, which w e train using contrastiv e learning on ED A data. W e p erform ablation studies using diﬀerent conﬁgurations of h yp erparameters, whic h aﬀect the mo del size. W e compare the p erformance of the diﬀeren t mo dels with resp ect to tw o criteria: v alidation loss and p erformance on a selected set of downstream tasks. W e leav e ab out 10% of the 27 A f ounda tion model f or electr odermal activity da t a Figure B.2: V alidation loss during training for diﬀerent conﬁgurations of masking and masking weigh t. 0 10 20 30 40 50 60 70 Epoch 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 L oss T rain vs V al L oss by K er nel Size T rain (k er nel=9) T rain (k er nel=3) V al (k er nel=9) V al (k er nel=3) Figure B.3: train dataset for v alidation during training, to also stop the training using early stopping (patience set to 30). W e ackno wledge that this selection may introduce bias in rep orting the ﬁnal results. W e do not select all p ossible hyperparameters. In particular, we set as ﬁxed the following: loss temp erature τ = 0 . 1 (see loss deﬁnition in Equation 1); weigh ts drop out at 50%. W e perform the hyperparameter selection through m ultiple experiments. W e list the experiments in order of which we p erform them. Kernel size W e compare the follo wing k ernel sizes: 9 and 3. W e c ho ose 9 since the reaction time asso ciated with EDA changes in approximately 2 seconds (9 samples with a sampling rate of 4 Hz is ab out 2 . 25 s ) [15]; w e c ho ose 3 since the minimum rise time of an EDA signal is b etw een 0 . 25 s and 0 . 5 s [39]. W e rep ort in Figure B.3 the train and v alidation loss. W e notice how the loss depends on the set in which it is computed, hence it cannot b e directly compared b etw een train and v alidation; how ever, given the same set and tw o models, the comparison can b e performed, since w e made sure to use the same splits for all exp erimen ts. Giv en these results, we select a k ernel size of 3. Mo del size The second set of exp eriments are related to the mo del size. In order to determine the size of the mo del, we mo dify the following hyperparameters: the stem channels [64, 32], the num b er of conv olutional c hannels [8, 16, 32, 64, 92], the total num ber of conv olutional blo cks [2, 4, 8, 12, 16]; the head channels [32, 64]; and the ﬁnal embeddings dimensionality [32, 64]. 28 A f ounda tion model f or electr odermal activity da t a (a) T rain loss for diﬀerent mo del sizes (b) V alidation loss for diﬀerent mo del sizes W e rep ort the train and v al losses in Figure B.4a and Figure B.4b resp ectively . F rom the comparison b etw een the train and v al losses, we conclude that there is an impact, in learning representation, from the mo del size. Ho wev er, after ab out 1M parameters, the v alidation loss remains similar. W e select a mo del with ab out 1M parameters for our ﬁnal training. The mo del con tains the following parameters: • stem c hannels: 64 • mbconv channels: 64 • num mbconv blo cks: 16 • head c hannels: 248 • embedding dim: 64 • kernel size: 9 29 A f ounda tion model f or electr odermal activity da t a T able C.1: CPU Performance Comparison of Extractors Extractor CPU Time (mean) [ms] CPU Time (SE) [ms] MOMENT (128M) 4823.53 161.11 Chronos (28M) 721.34 28.91 Man tis (8M) 350.23 39.06 UME (1M) 184.56 3.12 ED A-sp eciﬁc hc 35.39 1.40 generic hc 0.09 0.01 C A dditional results Computational analysis complexit y W e rep ort in T able C.1 the detailed performance metrics for the computation complexity analysis. This table mirrors the results in Figure 6. Complete results for the do wnstream ev aluation W e rep ort in T able C.2 and T able C.3 the complete results for our exp erimental ev aluation. Sp eciﬁcally , we rep ort the results in terms of balanced accuracy , Matthew’s Correlation Co eﬃcient (MCC) [24] and F1-score. W e rep ort the results for b oth v alidation metho ds used, i.e., LOPO and T A cross-v alidations. Results are rep orted for all tasks in both v alidation metho ds. F or the T A cross-v alidation, we do not rep ort the results for c o gnitive-lo ad/r elaxation classiﬁcation on the USILaughs dataset [30], since there was not enough data for each user to p erform the temp oral split. 30 A f ounda tion model f or electr odermal activity da t a T able C.2: Results of binary classiﬁcation exp eriments, using the Leav e One P articipant Out (LOPO) cross- v alidation. Results shown as metric ± standard error . Dataset Binary T ask Mo del Metrics Balanced Acc. MCC F1 Dream t Deep Sleep/REM Dumm y 0 . 73 ± 0 . 05 − 0 . 01 ± 0 . 01 0 . 16 ± 0 . 06 Gen. HC 0 . 64 ± 0 . 07 0 . 00 ± 0 . 03 0 . 10 ± 0 . 06 ED A-sp ec. HC 0 . 54 ± 0 . 05 0 . 02 ± 0 . 03 0 . 19 ± 0 . 07 Man tis 0 . 59 ± 0 . 05 0 . 02 ± 0 . 04 0 . 18 ± 0 . 07 MOMENT 0 . 64 ± 0 . 03 0 . 06 ± 0 . 03 0 . 21 ± 0 . 07 Chronos 0 . 64 ± 0 . 04 0 . 03 ± 0 . 03 0 . 19 ± 0 . 07 UME 0 . 56 ± 0 . 05 0 . 04 ± 0 . 04 0 . 17 ± 0 . 06 Dream t Sleep/W ak e Dumm y 0 . 50 ± 0 . 00 0 . 00 ± 0 . 00 0 . 49 ± 0 . 01 Gen. HC 0 . 70 ± 0 . 03 0 . 40 ± 0 . 05 0 . 68 ± 0 . 04 ED A-sp ec. HC 0 . 74 ± 0 . 02 0 . 48 ± 0 . 04 0 . 72 ± 0 . 03 Man tis 0 . 77 ± 0 . 02 0 . 52 ± 0 . 04 0 . 74 ± 0 . 03 MOMENT 0 . 72 ± 0 . 02 0 . 44 ± 0 . 03 0 . 71 ± 0 . 03 Chronos 0 . 78 ± 0 . 02 0 . 55 ± 0 . 03 0 . 75 ± 0 . 03 UME 0 . 75 ± 0 . 02 0 . 50 ± 0 . 04 0 . 73 ± 0 . 03 HHISS Stress/calm Dumm y 0 . 50 ± 0 . 00 − 0 . 00 ± 0 . 01 0 . 24 ± 0 . 03 Gen. HC 0 . 56 ± 0 . 03 0 . 12 ± 0 . 07 0 . 44 ± 0 . 06 ED A-sp ec. HC 0 . 64 ± 0 . 03 0 . 28 ± 0 . 07 0 . 55 ± 0 . 04 Man tis 0 . 63 ± 0 . 03 0 . 26 ± 0 . 06 0 . 53 ± 0 . 04 MOMENT 0 . 55 ± 0 . 02 0 . 09 ± 0 . 05 0 . 44 ± 0 . 03 Chronos 0 . 59 ± 0 . 03 0 . 17 ± 0 . 06 0 . 48 ± 0 . 04 UME 0 . 59 ± 0 . 03 0 . 19 ± 0 . 06 0 . 49 ± 0 . 05 HeartS Sleep/W ake Dumm y 0 . 50 ± 0 . 00 0 . 00 ± 0 . 00 0 . 08 ± 0 . 02 Gen. HC 0 . 66 ± 0 . 09 0 . 28 ± 0 . 16 0 . 32 ± 0 . 11 ED A-sp ec. HC 0 . 70 ± 0 . 05 0 . 32 ± 0 . 13 0 . 36 ± 0 . 11 Man tis 0 . 74 ± 0 . 06 0 . 37 ± 0 . 14 0 . 41 ± 0 . 13 MOMENT 0 . 70 ± 0 . 04 0 . 27 ± 0 . 06 0 . 33 ± 0 . 09 Chronos 0 . 73 ± 0 . 06 0 . 34 ± 0 . 11 0 . 39 ± 0 . 12 UME 0 . 71 ± 0 . 06 0 . 30 ± 0 . 10 0 . 35 ± 0 . 10 WESAD Low/High Arousal Dumm y 0 . 58 ± 0 . 05 − 0 . 02 ± 0 . 03 0 . 05 ± 0 . 03 Gen. HC 0 . 61 ± 0 . 09 0 . 10 ± 0 . 14 0 . 26 ± 0 . 13 ED A-sp ec. HC 0 . 63 ± 0 . 08 0 . 13 ± 0 . 11 0 . 33 ± 0 . 13 Man tis 0 . 56 ± 0 . 08 0 . 06 ± 0 . 13 0 . 27 ± 0 . 14 MOMENT 0 . 51 ± 0 . 06 0 . 03 ± 0 . 10 0 . 22 ± 0 . 12 Chronos 0 . 56 ± 0 . 09 0 . 07 ± 0 . 14 0 . 28 ± 0 . 10 UME 0 . 58 ± 0 . 09 0 . 03 ± 0 . 13 0 . 26 ± 0 . 12 WESAD Low/High V alence Dumm y 0 . 51 ± 0 . 06 0 . 07 ± 0 . 07 0 . 80 ± 0 . 10 Gen. HC 0 . 55 ± 0 . 15 0 . 01 ± 0 . 17 0 . 66 ± 0 . 18 ED A-sp ec. HC 0 . 58 ± 0 . 13 0 . 10 ± 0 . 13 0 . 72 ± 0 . 17 Man tis 0 . 55 ± 0 . 13 0 . 07 ± 0 . 16 0 . 76 ± 0 . 16 MOMENT 0 . 63 ± 0 . 07 0 . 03 ± 0 . 07 0 . 76 ± 0 . 14 Chronos 0 . 56 ± 0 . 10 0 . 05 ± 0 . 11 0 . 77 ± 0 . 15 UME 0 . 59 ± 0 . 13 0 . 08 ± 0 . 12 0 . 77 ± 0 . 13 APSYNC Low/High engagemen t Dumm y 0 . 46 ± 0 . 06 − 0 . 09 ± 0 . 11 0 . 29 ± 0 . 16 Gen. HC 0 . 61 ± 0 . 13 0 . 14 ± 0 . 25 0 . 53 ± 0 . 26 ED A-sp ec. HC 0 . 63 ± 0 . 16 0 . 18 ± 0 . 31 0 . 52 ± 0 . 27 Man tis 0 . 62 ± 0 . 11 0 . 30 ± 0 . 19 0 . 48 ± 0 . 29 MOMENT 0 . 60 ± 0 . 20 0 . 24 ± 0 . 39 0 . 61 ± 0 . 21 Chronos 0 . 48 ± 0 . 16 − 0 . 02 ± 0 . 31 0 . 43 ± 0 . 25 UME 0 . 54 ± 0 . 13 0 . 10 ± 0 . 29 0 . 49 ± 0 . 30 USILaughs Cog. load/relax Dumm y 0 . 50 ± 0 . 00 0 . 00 ± 0 . 00 0 . 80 ± 0 . 00 Gen. HC 0 . 66 ± 0 . 06 0 . 32 ± 0 . 13 0 . 72 ± 0 . 04 ED A-sp ec. HC 0 . 70 ± 0 . 09 0 . 40 ± 0 . 19 0 . 70 ± 0 . 11 Man tis 0 . 72 ± 0 . 09 0 . 43 ± 0 . 19 0 . 76 ± 0 . 10 MOMENT 0 . 60 ± 0 . 09 0 . 20 ± 0 . 18 0 . 61 ± 0 . 11 Chronos 0 . 68 ± 0 . 07 0 . 35 ± 0 . 14 0 . 57 ± 0 . 12 UME 0 . 71 ± 0 . 11 0 . 42 ± 0 . 22 0 . 76 ± 0 . 10 31 A f ounda tion model f or electr odermal activity da t a T able C.3: Results of binary classiﬁcation exp eriments for Time-A ware (T A) cross-v alidation. Results shown as metric ± standard error . Exp eriments whose result is rep orted as OoM (out of memory) means that the required memory size for the embeddings computation exceeded the computational resources av ailable. Dataset Binary T ask Mo del Metrics Balanced Acc. MCC F1 Dream t Deep Sleep/REM Dumm y 0 . 51 ± 0 . 01 0 . 01 ± 0 . 01 0 . 20 ± 0 . 06 Gen. HC 0 . 55 ± 0 . 10 0 . 09 ± 0 . 20 0 . 28 ± 0 . 13 ED A-sp ec. HC 0 . 59 ± 0 . 08 0 . 16 ± 0 . 13 0 . 37 ± 0 . 13 Man tis 0 . 64 ± 0 . 07 0 . 22 ± 0 . 11 0 . 40 ± 0 . 12 MOMENT 0 . 65 ± 0 . 04 0 . 24 ± 0 . 07 0 . 42 ± 0 . 10 Chronos 0 . 64 ± 0 . 04 0 . 23 ± 0 . 08 0 . 41 ± 0 . 10 UME 0 . 63 ± 0 . 09 0 . 20 ± 0 . 13 0 . 38 ± 0 . 11 Dream t Sleep/W ake Dumm y 0 . 50 ± 0 . 01 − 0 . 01 ± 0 . 01 0 . 58 ± 0 . 04 Gen. HC 0 . 65 ± 0 . 04 0 . 32 ± 0 . 11 0 . 69 ± 0 . 06 ED A-sp ec. HC 0 . 69 ± 0 . 01 0 . 38 ± 0 . 03 0 . 72 ± 0 . 02 Man tis 0 . 73 ± 0 . 02 0 . 46 ± 0 . 03 0 . 76 ± 0 . 02 MOMENT 0 . 69 ± 0 . 01 0 . 39 ± 0 . 02 0 . 72 ± 0 . 01 Chronos 0 . 73 ± 0 . 01 0 . 47 ± 0 . 02 0 . 76 ± 0 . 01 UME 0 . 70 ± 0 . 02 0 . 41 ± 0 . 03 0 . 74 ± 0 . 02 HHISS Stress/calm Dumm y 0 . 50 ± 0 . 00 − 0 . 00 ± 0 . 00 0 . 23 ± 0 . 09 Gen. HC 0 . 45 ± 0 . 04 − 0 . 10 ± 0 . 07 0 . 34 ± 0 . 04 ED A-sp ec. HC 0 . 55 ± 0 . 05 0 . 09 ± 0 . 11 0 . 45 ± 0 . 07 Man tis 0 . 51 ± 0 . 02 0 . 20 ± 0 . 03 0 . 50 ± 0 . 03 MOMENT OoM OoM OoM Chronos OoM OoM OoM UME 0 . 49 ± 0 . 06 − 0 . 02 ± 0 . 12 0 . 38 ± 0 . 07 HeartS Sleep/W ak e Dumm y 0 . 49 ± 0 . 00 − 0 . 01 ± 0 . 01 0 . 08 ± 0 . 02 Gen. HC 0 . 71 ± 0 . 06 0 . 35 ± 0 . 11 0 . 39 ± 0 . 08 ED A-sp ec. HC 0 . 72 ± 0 . 14 0 . 36 ± 0 . 25 0 . 41 ± 0 . 20 Man tis 0 . 75 ± 0 . 16 0 . 40 ± 0 . 28 0 . 45 ± 0 . 23 MOMENT 0 . 69 ± 0 . 12 0 . 24 ± 0 . 16 0 . 31 ± 0 . 12 Chronos 0 . 74 ± 0 . 13 0 . 35 ± 0 . 21 0 . 40 ± 0 . 19 UME 0 . 73 ± 0 . 17 0 . 32 ± 0 . 25 0 . 38 ± 0 . 18 WESAD Low/High Arousal Dumm y 0 . 55 ± 0 . 05 0 . 08 ± 0 . 07 0 . 15 ± 0 . 07 Gen. HC 0 . 56 ± 0 . 15 0 . 08 ± 0 . 22 0 . 26 ± 0 . 11 ED A-sp ec. HC 0 . 59 ± 0 . 20 0 . 08 ± 0 . 31 0 . 32 ± 0 . 14 Man tis 0 . 58 ± 0 . 11 0 . 13 ± 0 . 19 0 . 32 ± 0 . 13 MOMENT OoM OoM OoM Chronos 0 . 56 ± 0 . 18 0 . 07 ± 0 . 28 0 . 28 ± 0 . 20 UME 0 . 56 ± 0 . 12 0 . 08 ± 0 . 18 0 . 28 ± 0 . 15 WESAD Low/High V alence Dumm y 0 . 51 ± 0 . 03 0 . 01 ± 0 . 04 0 . 66 ± 0 . 06 Gen. HC 0 . 54 ± 0 . 20 0 . 08 ± 0 . 33 0 . 76 ± 0 . 12 ED A-sp ec. HC 0 . 54 ± 0 . 19 0 . 04 ± 0 . 30 0 . 62 ± 0 . 22 Man tis 0 . 61 ± 0 . 08 0 . 18 ± 0 . 13 0 . 71 ± 0 . 13 MOMENT OoM OoM OoM Chronos 0 . 45 ± 0 . 12 − 0 . 10 ± 0 . 20 0 . 59 ± 0 . 21 UME 0 . 63 ± 0 . 13 0 . 18 ± 0 . 18 0 . 73 ± 0 . 13 APSYNC Low/High engagement Dumm y 0 . 45 ± 0 . 20 0 . 13 ± 0 . 13 0 . 13 ± 0 . 13 Gen. HC 0 . 81 ± 0 . 20 0 . 32 ± 0 . 32 0 . 72 ± 0 . 29 ED A-sp ec. HC 0 . 81 ± 0 . 20 0 . 32 ± 0 . 32 0 . 72 ± 0 . 29 Man tis 0 . 86 ± 0 . 15 0 . 41 ± 0 . 43 0 . 82 ± 0 . 19 MOMENT 0 . 48 ± 0 . 32 − 0 . 05 ± 0 . 10 0 . 33 ± 0 . 30 Chronos 0 . 61 ± 0 . 40 0 . 18 ± 0 . 86 0 . 70 ± 0 . 32 UME 0 . 89 ± 0 . 22 0 . 49 ± 0 . 57 0 . 83 ± 0 . 33 32

A foundation model for electrodermal activity data

Original Paper

Comments & Academic Discussion

Leave a Comment

Original Paper

Related Papers

Comments & Academic Discussion

Leave a Comment