Pathology-Aware Multi-View Contrastive Learning for Patient-Independent ECG Reconstruction

1 P athology-A ware Multi-V ie w Contrasti v e Learning for Patient-Independent ECG Reconstruction Y oussef Y oussef and Jitin Singla Abstract —Reconstructing a 12-lead electrocardiogram (ECG) from a r educed lead set is an ill-posed in verse pr oblem due to anatomical variability . Standard deep learning methods often ignore underlying cardiac pathology losing vital morphology in precordial leads. W e propose Pathology-A ware Multi-V iew Contrastive Learning, a framework that regularizes the latent space through a pathological manif old. Our architecture in- tegrates high-fidelity time-domain wavef orms with pathology- aware embeddings learned via supervised contrastive alignment. By maximizing mutual inf ormation between latent r epresenta- tions and clinical labels, the framework learns to filter anatomical ”nuisance” variables. On the PTB-XL dataset, our method achieves approx. 76% reduction in RMSE compared to state- of-the-art model in patient-independent setting. Cross-dataset evaluation on the PTB Diagnostic Database confirms superior generalization, bridging the gap between hardwar e portability and diagnostic-grade r econstruction. Index T erms —Electrocardiogram (ECG), Lead Reconstruc- tion, W earable ECG, Supervised Contrastive Learning, Deep Learning, Latent Representation. I . I N T R O D U C T I ON T HE standard 12-lead electrocardiogram (ECG) is the clinical gold standard for assessing cardiac electrical activity [1]–[3], yet its utility is constrained by the requirement for ten electrodes and precise anatomical placement, making it challenging for continuous monitoring in amb ulatory and home settings [3]–[6]. T o bridge the gap between clinical interpretability and hardware portability , there is a growing need to reconstruct high-fidelity 12-lead wa veforms, especially precordial V 1 − V 6 leads, from a reduced set of measured leads (e.g., Leads I, II, and V 2 ) [4], [7]. The feasibility of ECG lead reconstruction is theoretically grounded in the spatial redundancy of cardiac dipoles [4], [8], [9]. Ho wev er , the relationship between the cardiac source v ( t ) ∈ R 3 and body surface potentials Φ is go verned by a non-linear subject- specific lead-field matrix A ( s ) , where s represents latent anatomical v ariables such as torso geometry , heart orientation and tissue conductivity [10], [11]: Φ ( t ) = A ( s ) v ( t ) + ϵ (1) In a patient-independent setting, s is unobserv ed. Conse- quently , standard models attempting a direct mapping f : X → Y . Y oussef is with the Department of Computer Science and Engineering, Indian Institute of T echnology Roorkee, Roorkee 247667, India (e-mail: youssef yy@cs.iitr .ac.in). J. Singla is with the Department of Biosciences and Bioengineering, Indian Institute of T echnology Roorkee, Roorkee 247667, India (e-mail: jitin.singla@bt.iitr .ac.in). *Corresponding author: Jitin Singla. Y are mathematically forced to minimize error by learning the marginalized transfer function E [ A ( s )] . This is expressed as: p ( Y |X ) = Z p ( Y |X , s ) p ( s |X ) d s (2) Because the influence of s is inte grated (averaged) across the population, the model suffers from a ”regression to the mean” effect, where high-frequency diagnostic details are lost to satisfy the population av erage. This loss is particularly high in the precordial ”transitional zone” ( V 3 – V 4 ) where anatomical variance is maximal [12]. T o counteract this marginalization, we propose Pathology- A ware Multi-V iew Contrastiv e Learning, a frame work that reformulates lead reconstruction by regularizing the latent space through a pathological manifold. W e argue that while the anatomical matrix A ( s ) varies stochastically between patients, the cardiac signal can be constrained to lower -dimensional manifold dictated by the underlying pathology l . Since s is inaccessible for ne w patients, we introduce a pathology-aw are representation h = f ϕ ( X ) to act as a structural anchor and conditioning v ariable, decomposing the generative process as: p ( Y |X , h ) ∝ p ( Y |X , l ) p ( h |X , l ) (3) W e maximize the Mutual Information I ( h ; l ) = H ( l ) − H ( l | h ) between latent representation h and pathology label l by optimizing a supervised contrastive loss. This partitions latent space by clinical condition rather than anatomical noise, providing a prior that restricts the solution space for Y to the ”pathological subspace”. Mathematically , this condition- ing reduces the conditional entropy of the reconstruction as H ( Y |X , h ) < H ( Y |X ) , providing the reconstruction decoder with a semantic ”anchor” to recov er features typically lost in unconditioned models. Our contributions are two-fold: 1) W e propose a patient- independent multi-view architecture that fuses high-fidelity wa veforms with contrastive embeddings; 2) W e achieve state- of-the-art results on PTB-XL Dataset, including 76% reduction in RMSE over existing benchmarks, and demonstrate robust cross-dataset generalization on the PTB Diagnostic Database. I I . R E L A T E D W O R K Early ECG reconstruction methods used linear transforma- tion techniques [13] (e.g., the in verse Dower matrix [14], regression-based models [15]) to model spatial relations be- tween leads. These mathematical models were interpretable; howe ver , their precision is highly af fected by variations in human anatomy and electrode placement [7], [9]. T o overcome these limitations, recent research has changed direction to 2 using deep learning (DL) methods to model the non-linear relationship between leads from the data directly . There are sev eral common architectures reported in the literature, in- cluding: • U-Net architectures: These ha ve been used to reconstruct a full 12-lead ECG from a single input lead, achie ving correlations up to 0.9 for specific leads [16]. • Hybrid CNN-LSTM Models: One-dimensional Con volu- tional Neural Netw orks (CNN) combined with Bidirec- tional Long Short-T erm Memory (LSTM) models pro- vided better temporal modeling of the 12-lead ECG signal [17], [18]. • Attention-based architectures: Attention-modified U-Net models hav e shown a verage Pearson correlations of around 0.80 and R 2 values of approximately 0.64 [19]. • Lightweight architectures: Frequenc y-based decomposi- tion and parameter-ef ficient 12 lead networks hav e been dev eloped for use in resource-constrained environments (such as wearables) [20]. Recently , the ECG reconstruction have mo ved away from direct signal-to-signal mapping to training on channel-agnostic latent representation. These methods in volv e joint alignment and reconstruction across multiple leads or multimodal con- trastiv e alignment such as integrating cardiac MRI and ECG data. These representations aim to preserve channel-specific information as well as facilitate information transfer across different lead subsets [21], [22]. Most existing studies are con- ducted on PTB Diagnostic Database [4], [7], [16], [17], [23], [24], which, despite its historical significance, is relativ ely small in terms of subject count and pathological di versity . In contrast, the more recent and extensiv e PTB-XL dataset remains underutilized [25]–[27], leaving a gap in ev aluating how these models scale to larger , more heterogeneous clinical populations. Furthermore, performance is typically inflated by subject-dependent partitioning as intra-patient data is mixed across training and test sets [19], [23], [25], [28]. This leads to subject-specific memorization rather than clinical generaliza- tion. Only fe w studies hav e been done in patient-independent settings [26], [27]. I I I . P R O P O S E D M E T H O D O L O G Y The proposed methodology reformulates ECG lead recon- struction as a multi-view integration task, combining morpho- logical fidelity (time-domain wa veforms) with a pathology- aware latent space. The system comprises three stages: signal preprocessing, dual-representation construction, and stacked latent decoding. A. Signal Cleaning and Se gmentation Each 12-lead ECG is processed through a narrowband notch filter , 0.5–45Hz Butterworth band-pass filter and Median- based filter eliminating powerline interference, baseline drift and residual low-frequency wander , respectively . Cleaned sig- nals are segmented into temporal windo ws of T = 256 samples with a 64-sample hop size. Segments are qual- ity controlled based on lead-wise amplitude and root-mean- square criteria using empirical percentile bounds (0.1–99.9%). Artifact-free segments x ∈ R 3 × 256 (leads I, II, and V 2 ) serve as inputs for reconstructing leads V 1 and V 3 – V 6 . B. P athology-aware Latent Repr esentation T o capture clinically meaningful in variants, we train a network f ϕ : x → h to structure a latent space where similar pathologies cluster together . Clean segments are transformed into augmented morphology-centered (R-peak aligned) and signal-augmented vie ws, then processed by a 1-D conv olu- tional encoder into an embedding h ∈ R 128 and projected into an ℓ 2 -normalized embedding z ∈ S d − 1 . W e employ a supervised contrastiv e loss L sc with temperature τ = 0 . 07 to maximize mutual information between the latent space and high-confidence diagnostic labels l (certainty ≥ 80 ): L sc = 1 |B | X i ∈B   − 1 |P ( i ) | X p ∈P ( i ) log e z ⊤ i z p /τ P j ∈B\ i e z ⊤ i z j /τ   (4) where P ( i ) contains indices of samples in batch B sharing at least one label with anchor i . L sc shapes the latent rep- resentation with high intra-class compactness and inter-class separation, pro viding a robust, pathology-aware anchor that can guide the do wnstream reconstruction decoder . All clean segments are projected into h after training f ϕ . C. Normalization and Fusion The time-domain wa veform x and learned representation h are normalized to stabilize the reconstruction process. The segment x is normalized lead-wise to zero mean and unit variance: ˆ x c,i = ( x c,i − µ c ) / ( σ c + ϵ ) , where µ c and σ c are calculated across the temporal dimension T = 256 . Similarly , h is normalized as ˆ h = ( h − µ ) / ( σ + ϵ ) using the mean vector and standard deviation of h . D. Stack ed Latent Decoder The final reconstruction network f θ : ˆ x , ˆ h → y synthesizes target lead by integrating ˆ x ∈ R 3 × 256 and ˆ h ∈ R 128 . Dedicated projection modules map both inputs to a shared 128-channel temporal dimensionality . The features are then concatenated into a stacked latent tensor ∈ R 256 × 256 and fused via a 1-D con volutional layer . A compact temporal decoder then recovers the tar get wav eform y ∈ R 256 . Independent decoders are trained for each output lead. The ov erall workflo w of the methodology is illustrated in Fig.1. The Decoder loss function combines Mean Squared Error for sensiti vity to large deviations and Mean Absolute Error to provide a steady gradient against outliers (CITE P APER HERE). I V . E X P E R I M E N T S A. Data P artitioning and Pr otocol W e utilized PTB-XL dataset [29] with a patient-wise parti- tioning strategy to prev ent inter-subject leakage. First 8 folds were used for training, 9th for validation, and last for testing. For cross-dataset e v aluation, we used PTB Dataset [30]. All signals were resampled to 100 Hz for consistency . 3 Fig. 1: Proposed ECG reconstruction framework. (a) Prepro- cessing and segmentation, (b) Multi-V ie w Supervised con- trastiv e pretraining to learn pathology-a ware latent represen- tations, and (c) Reconstruction network that integrates con- trastiv e representation with clean ECG signals via a temporal decoder . B. T raining Configur ation and Evaluation Metrics Decoder is trained using AdamW and early stopping on validation loss. For testing, non-ov erlapping segments are utilized to reconstruct the complete 10-second signal, ensuring a fair and realistic estimation of RMSE in terms of clinical viability . Fidelity is quantified using Root Mean Squared Error (RMSE) in mV , the Coefficient of Determination ( R 2 ), and the Pearson Correlation Coef ficient [4], [7], [17], [19], [25], [31]. C. Latent Space Structur e and Clinical Separability T o quantify clinical separability , we analyze the cosine similarity of test data identifying k = 10 nearest neighbors of clean signal ˆ x and latent embeddings ˆ h and compute class-to-class af finity matrix A, where element A i,j is the mean proportion of neighbors for class i that belongs to class j . Fig. 2 clearly sho ws that the af finity for latent h is stronger along the diagonal than the original clean baseline signal x . The diagonal a verage (intra-class consistency) for h increased to 0.837 compared to 0.083 in x , demonstrating that the supervised contrasti ve loss effecti vely partitions the latent space in pathology-aware clusters. D. Lead Reconstruction P erformance The impact of the pathology-aware anchor on signal fidelity is evaluated across all target leads for the test fold. As detailed in T able I, the proposed Clean + h model (C-h) outperforms the baseline model (C) with Clean-only signal across all metrics. Notable improvements are observed in the transitional leads ( V 3 – V 4 ), with R 2 increasing by approximately 11.0% and 13.5%, respectiv ely . This suggests that the latent anchor effecti vely restricts the solution space to a pathological man- ifold, allo wing the decoder to recov er morphological features Fig. 2: k-NN af finity matrices ( k = 10 ) comparing (a) the clean baseline x and (b) the learned latent representation h . T ABLE I: Reconstruction Performance: Baseline (C) vs. Pro- posed (C-h) Lead RMSE (mV) ↓ R 2 ↑ Pearson ↑ C C-h ∆% C C-h ∆% C C-h ∆% V 1 0.063 0.052 +18.2% 0.754 0.839 +11.2% 0.940 0.951 +1.2% V 3 0.115 0.101 +12.7% 0.670 0.744 +11.0% 0.913 0.926 +1.5% V 4 0.118 0.108 +8.7% 0.605 0.687 +13.5% 0.893 0.906 +1.4% V 5 0.094 0.084 +10.3% 0.684 0.717 +4.8% 0.922 0.931 +1.0% V 6 0.071 0.063 +11.4% 0.699 0.760 +8.8% 0.935 0.944 +1.0% Mean 0.092 0.082 +12.2% 0.683 0.749 +9.9% 0.921 0.932 +1.2% typically lost to population-wide a veraging. Quantitative ro- bustness across pathologies is further analyzed via radar plot for four most and least abundant diagnostic classes (Fig. 3). The C-h model maintains consistently lower error rates com- pared to the baseline, even for rare pathologies, indicating that supervised contrastiv e alignment anchors the reconstruction to disease-specific in v ariants rather than merely optimizing for the majority class. Qualitativ e results (Fig. 4) also confirm that C-h model achiev es superior alignment with ground truth wa veforms compared to C model. E. Comparison with Prior W ork V ery fe w studies strictly adhere to a patient-independent ev aluation protocol. Many existing frameworks, such as those by Lu et al. [25] and Garg et al. [19], utilize random record- lev el splits where ECG from same subject may appear in both training and test sets, thereby inflating results. In contrast, we adopt a rigorous patient-wise partition to ensure realistic clinical generalization. W e primarily contextualize our results against Rajotte et al. [26], which utilized patient-wise partition on PTB-XL dataset with input leads I, II, and V 2 . Also, Rajotte et al. pretrained a CNN+BiLSTM model on PTB dataset first before fine-tuning on PTB-XL and reconstructed full 10s signal. T o ensure a fair comparison we adopted a record-lev el ev aluation protocol by reassembling our 2.56- second reconstructed se gments into full 10-second signals. The proposed C-h model achie ves statistically rob ust improv ements across all precordial leads (T able II), achie ving an average RMSE reduction of 76.4% and 6.6% improvement in R 2 compared to Rajotte et al. The proposed model capacity also remains modest, with ∼ 239 K parameters, which is comparable to prior work ( ∼ 252 K parameters) [26]. 4 Fig. 3: Radar plot showing lower RMSE for C-h model for few most and least abundant diagnostic class. n represents the number of se gments in test fold. The closer the vertex to center , the lo wer is the RMSE. The y-values for each ring correspond to [0.05, 0.075, 0.10, 0.125, 0.15] starting from center . Fig. 4: Qualitati ve heartbeat-level reconstruction examples across different diagnostic classes (NORM, INJ AS, R VH) for leads V1,V3–V6. The C-h configuration more closely follo ws the ground truth compared to the C baseline, particularly in capturing QRS complex amplitude and T -wav e morphology . Further , our method shows superior Pearson correlation (0.932) compared to existing models that often utilize random splits, as contextualized in T able III. These results underscore that anchoring reconstruction to a pathology-aware latent space significantly outperforms unconditioned, direct-mapping models. T ABLE II: Record-level Comparison with Rajotte et al. [26] Lead Proposed (C-h) Rajotte et al. [26] RMSE R 2 RMSE ∆ RM S E % R 2 ∆ R 2 % V 1 0 . 054 ± 0 . 002 0 . 839 ± 0 . 010 0 . 276 ± 0 . 009 80.4% 0 . 796 ± 0 . 011 5.4% V 3 0 . 103 ± 0 . 003 0 . 747 ± 0 . 023 0 . 364 ± 0 . 011 71.7% 0 . 702 ± 0 . 023 6.4% V 4 0 . 110 ± 0 . 004 0 . 692 ± 0 . 026 0 . 420 ± 0 . 014 73.8% 0 . 613 ± 0 . 040 12.9% V 5 0 . 086 ± 0 . 003 0 . 719 ± 0 . 037 0 . 350 ± 0 . 011 75.4% 0 . 694 ± 0 . 023 3.6% V 6 0 . 065 ± 0 . 002 0 . 765 ± 0 . 028 0 . 334 ± 0 . 012 80.5% 0 . 729 ± 0 . 018 4.9% Mean 0.084 0.752 0.349 76.4% 0.707 6.6% T ABLE III: Global Performance Benchmarks on PTB-XL Method Patient Indep. Input Leads RMSE Pearson Garg et al. (2023) [19] No II – 0.805 Lu et al. (2025) [25] No V 2 0.160 0.860 Hebiguchi et al. (2025) [18] No I, II, V 2 - 0.93 Ours Y es I, II, V 2 0.082 0.932 T ABLE IV: Cross-Dataset Performance Comparison on the PTB Database Method Patient Indep. Input Leads RMSE Pearson Smith et al. (2021) [23] No 6 limb leads + V 2 0.173 0.921 Moghaddam et al. (2025) [4] Y es 3-leads – 0.76 Ours Y es I, II, V 2 0.095 0.887 F . Cr oss-Dataset Generalization T o ev aluate the robustness of the pathology-aware represen- tation h under distrib ution shifts, we perform a cross-dataset ev aluation on the independent PTB Diagnostic ECG Database. Unlike prior benchmarks that are typically trained and tested on the same dataset, our model is trained exclusi vely on PTB- XL and e valuated on PTB without subject-specific calibration or fine-tuning. T able IV summarizes the comparison of our method with other patient-independent metrics reported in the literature on PTB dataset. Despite the domain shift and reduced input set, our proposed C-h model achie ves an a verage RMSE of 0 . 095 mV , representing a substantial impro vement ov er the 0 . 173 mV reported by Smith et al. . Furthermore, our framew ork attains an av erage Pearson correlation of 0 . 887 without training on PTB dataset. These results empirically validate that anchoring the reconstruction to a disease-aware latent space allows synthesized wa veforms to maintain clini- cally consistent morphologies that generalize across different recording en vironments and hardware. V . C O N C L U S I O N In conclusion, the proposed framework achie ves a strong balance between accuracy and ef ficiency . Despite its compact size, it attains low reconstruction error with average RMSE of 0 . 082 on PTB-XL under a patient-independent setting and maintains good generalization across datasets with average RMSE of 0 . 095 on the PTB database. The model operates with very lo w inference latency of approximately 0 . 035 ms, enabling real-time deployment. By leveraging pathology-aware representations, the pro- posed approach preserves clinically relev ant ECG morphology while remaining computationally ef ficient, making it suitable for practical and portable diagnostic applications. R E F E R E N C E S [1] Federico Mason, Amitabh C. Pandey , Matteo Gadaleta, Eric J. T opol, Evan D. Muse, and Giorgio Quer . Ai-enhanced reconstruction of the 12-lead electrocardiogram via 3-leads with accurate clinical assessment. npj Digital Medicine , 7:201, 2024. 5 [2] Y asar Sattar and Lovely Chhabra. Electr ocar diogram . StatPearls Publishing, Treasure Island, FL, 2023. In: StatPearls [Internet]. Updated June 5, 2023. [3] Brian Campbell, Dave Richley , Catherine Ross, and Christopher J. Eggett. Clinical guidelines by consensus: Recording a standard 12-lead electrocardiogram, 2024. Appro ved clinical guidelines, V ersion 4.1. [4] Dorsa E. P . Moghaddam, Anton Banta, Allison Post, Mehdi Raza vi, and Behnaam Aazhang. Reconstructing 12-lead ecg from reduced lead sets using an encoder–decoder conv olutional neural network. Biomedical Signal Processing and Control , 104:107486, 2025. [5] Noemi Giannetta, Giuseppe Campagna, Fla vio Di Muzio, Emanuele Di Simone, Sara Dionisi, and Marco Di Muzio. Accurac y and kno wledge in 12-lead ecg placement among nursing students and nurses: a web- based italian study . Acta Biomedica , 91(12-S):e2020004, 2020. [6] V elislav N. Batchvaro v , Marek Malik, and A. John Camm. Incorrect electrode cable connection during electrocardiographic recording. Eu- r opace , 9(11):1081–1090, 2007. [7] Emmanuel N. Obianom, G. Andre Ng, and Xia Li. Reconstruction of 12-lead ecg: a review of algorithms. F r ontiers in Physiology , 16, 2025. [8] Utkars Jain, Adam A. Butchy , Michael T . Leasure, V eronica A. Co- valesk y , Daniel McCormick, and Gary S. Mintz. Redundancy and novelty between ecg leads based on linear correlation. In Pr oceedings of the 16th International Joint Conference on Biomedical Engineering Systems and T echnologies (BIOSTEC) , pages 359–365. SCITEPRESS, 2023. [9] Stefan Paul Nelwan. Evaluation of 12-Lead Electrocar diogr am Re- construction Methods for P atient Monitoring . PhD thesis, Erasmus Univ ersity Rotterdam, Rotterdam, The Netherlands, 2005. [10] Martin Buist and Andrew Pullan. T orso coupling techniques for the forward problem of electrocardiography . Annals of Biomedical Engineering , 30(10):1299–1312, 2002. [11] K onstantin Ushenin, V italy Kalinin, Sukaynat Gitinova, Oleg Sopo v , and Olga Solovyo va. Parameter variations in personalized electrophysiolog- ical models of human heart ventricles. PLOS ONE , 16(4):e0249062, 2021. [12] O. Presacan, A. Dorobant ¸ iu, J. L. Isaksen, T . W illi, C. Graff, M. A. Riegler , A. R. Sridhar, J. K. Kanters, and V . Thambawita. Evaluating the feasibility of 12-lead electrocardiogram reconstruction from limited leads using deep learning. Communications Medicine , 5(1):139, 2025. [13] D.D. Finlay , C.D. Nugent, J.G. Kellett, M.P . Donnelly , P .J. McCullagh, and N.D. Black. Synthesising the 12-lead electrocardiogram: T rends and challenges. Eur opean Journal of Internal Medicine , 18(8):566–570, 2007. [14] Lars Edenbrandt and Olle Pahlm. V ectorcardiogram synthesized from a 12-lead ecg: Superiority of the in verse dower matrix. Journal of Electr ocar diology , 21(4):361–367, 1988. [15] J. A. Kors, G. v an Herpen, A. C. Sittig, and J. H. v an Bemmel. Reconstruction of the frank vectorcardiogram from standard electrocar- diographic leads: Diagnostic comparison of different methods. European Heart Journal , 11(12):1083–1092, 1990. [16] Sara C. Beco, Jo ˜ ao R. Pinto, and Jaime S. Cardoso. Electrocardiogram lead con version from single-lead blindly-segmented signals. BMC Medical Informatics and Decision Making , 22(1):314, 2022. [17] V ishnuvardhan Gundlapalle and Amit Acharyya. A novel single lead to 12-lead ecg reconstruction methodology using con volutional neural networks and lstm. In 2022 IEEE 13th Latin American Symposium on Cir cuits and Systems (LASCAS) , pages 1–4, 2022. [18] Kazuki Hebiguchi, Hiroyoshi T ogo, and Akimasa Hirata. Reducing lead requirements for wearable ecg: Chest lead reconstruction with 1d-cnn and bi-lstm. Informatics in Medicine Unlocked , 53:101624, 2025. [19] Akshit Garg, V ijay V ignesh V enkataramani, and U. Dev a Priyakumar . Single-lead to multi-lead electrocardiogram reconstruction using a mod- ified attention u-net framework. In 2023 International Joint Confer ence on Neural Networks (IJCNN) , pages 1–8, 2023. [20] Junseok Lee, Y eonho Y oo, Jinkyu Kim, Dosun Lim, Gyeongsik Y ang, and Chuck Y oo. Parameter -efficient 12-lead ecg reconstruction from a single lead. In Medical Imag e Computing and Computer Assisted Intervention – MICCAI 2025 , Lecture Notes in Computer Science, pages 431–441. Springer, 2025. [21] Nabil Ibtehaz and Masood Mortazavi. Modally reduced representation learning of multi-lead ecg signals through simultaneous alignment and reconstruction. arXiv pr eprint arXiv:2405.19359 , 2024. [22] Alexander Seliv anov , Philip M ¨ uller , ¨ Ozg ¨ un Tur gut, Nil Stolt-Ans ´ o, and Daniel R ¨ uckert. Global and local contrastiv e learning for joint representations from cardiac mri and ecg. In Proceedings of MICCAI 2025 (Medical Image Computing and Computer Assisted Intervention) , pages 217–227, 2025. [23] Gerard H. Smith, Dawie J. V an den Hee ver , and W ayne Swart. The reconstruction of a 12-lead electrocardiogram from a reduced lead set using a focus time-delay neural network. Acta Car diologica Sinica , 37(1):47–57, 2021. [24] Sidharth Maheshwari, Amit Acharyya, P . Rajalakshmi, Paolo Emilio Puddu, and Michele Schiariti. Accurate and reliable 3-lead to 12-lead ecg reconstruction methodology for remote health monitoring applica- tions. In Pr oceedings of the 2013 IEEE 15th International Confer ence on e-Health Networking, Applications and Services (HealthCom) , pages 233–238, IEEE, 2013. [25] Junchen Lu, Ping Hu, Lei Zhao, and Min Xiang. Lightweight u2-net ecg signal reconstruction model with attention fusion. In 2025 IEEE International Conference on Signal Processing Systems (ICSPS) , pages 816–821, 2025. [26] Kiriaki J. Rajotte, Bashima Islam, Xinming Huang, Da vid D. McManus, and Edward A. Clancy . Ecg statement classification and lead reconstruc- tion using cnn-based models. IEEE Journal of Biomedical and Health Informatics , 30(1):162–172, 2026. [27] Jiarong Chen, W anqing W u, T ong Liu, and Shenda Hong. Multi-channel masked autoencoder and comprehensive e valuations for reconstructing 12-lead ecg from arbitrary single-lead ecg. npj Car diovascular Health , 1(1):34, 2024. [28] Anton Banta, Romain Cosentino, Mathews M. John, Allison Post, Skylar Buchan, Mehdi Razavi, and Behnaam Aazhang. A novel con- volutional neural network for reconstructing surface electrocardiograms from intracardiac electrograms and vice versa. Artificial Intelligence in Medicine , 118:102135, 2021. [29] Patrick W agner , Nils Strodthoff, Ralf-Dieter Bousseljot, W ojciech Samek, and T obias Schaef fter . PTB-XL, a large publicly av ailable electrocardiography dataset. PhysioNet , April 2020. V ersion 1.0.1. [30] Ary L. Goldberger , Luis A. N. Amaral, Leon Glass, Jeffrey M. Haus- dorff, Plamen Ch. Iv anov , Roger G. Mark, Joseph E. Mietus, George B. Moody , Chung-Kang Peng, and H. Eugene Stanley . Physiobank, physiotoolkit, and physionet: Components of a new research resource for complex physiologic signals. Cir culation , 101(23):e215–e220, 2000. [31] Jangjay Sohn, Seungman Y ang, Joonnyong Lee, Y unseo Ku, and Hee Chan Kim. Reconstruction of 12-lead electrocardiogram from a three-lead patch-type de vice using a lstm network. Sensors , 20(11):3278, 2020.