Cross-Sample Augmented Test-Time Adaptation for Personalized Intraoperative Hypotension Prediction

📝 Abstract

Intraoperative hypotension (IOH) poses significant surgical risks, but accurate prediction remains challenging due to patient-specific variability. While test-time adaptation (TTA) offers a promising approach for personalized prediction, the rarity of IOH events often leads to unreliable test-time training. To address this, we propose CSA-TTA, a novel Cross-Sample Augmented Test-Time Adaptation framework that enhances training by incorporating hypotension events from other individuals. Specifically, we first construct a crosssample bank by segmenting historical data into hypotensive and non-hypotensive samples. Then, we introduce a coarseto-fine retrieval strategy for building test-time training data: we initially apply K-Shape clustering to identify representative cluster centers and subsequently retrieve the top-K semantically similar samples based on the current patient signal. Additionally, we integrate both self-supervised masked reconstruction and retrospective sequence forecasting signals during training to enhance model adaptability to rapid and subtle intraoperative dynamics. We evaluate the proposed CSA-TTA on both the VitalDB dataset and a realworld in-hospital dataset by integrating it with state-of-theart time series forecasting models, including TimesFM and UniTS. CSA-TTA consistently enhances performance across settings-for instance, on VitalDB, it improves Recall and F1 scores by +1.33% and +1.13%, respectively, under finetuning, and by +7.46% and +5.07% in zero-shot scenarios-demonstrating strong robustness and generalization.

💡 Analysis

Intraoperative hypotension (IOH) poses significant surgical risks, but accurate prediction remains challenging due to patient-specific variability. While test-time adaptation (TTA) offers a promising approach for personalized prediction, the rarity of IOH events often leads to unreliable test-time training. To address this, we propose CSA-TTA, a novel Cross-Sample Augmented Test-Time Adaptation framework that enhances training by incorporating hypotension events from other individuals. Specifically, we first construct a crosssample bank by segmenting historical data into hypotensive and non-hypotensive samples. Then, we introduce a coarseto-fine retrieval strategy for building test-time training data: we initially apply K-Shape clustering to identify representative cluster centers and subsequently retrieve the top-K semantically similar samples based on the current patient signal. Additionally, we integrate both self-supervised masked reconstruction and retrospective sequence forecasting signals during training to enhance model adaptability to rapid and subtle intraoperative dynamics. We evaluate the proposed CSA-TTA on both the VitalDB dataset and a realworld in-hospital dataset by integrating it with state-of-theart time series forecasting models, including TimesFM and UniTS. CSA-TTA consistently enhances performance across settings-for instance, on VitalDB, it improves Recall and F1 scores by +1.33% and +1.13%, respectively, under finetuning, and by +7.46% and +5.07% in zero-shot scenarios-demonstrating strong robustness and generalization.

📄 Content

Cross-Sample Augmented Test-Time Adaptation for Personalized Intraoperative Hypotension Prediction Kanxue Li1, Yibing Zhan1*, Hua Jin2*, Chongchong Qi3, Xu Lin3, Baosheng Yu4 1School of Computer Science, Wuhan University 2First People’s Hospital of Yunnan Province 3Yunnan United Vision Technology Company Limited 4Nanyang Technological University likanxue@whu.edu.cn Abstract Intraoperative hypotension (IOH) poses significant surgical risks, but accurate prediction remains challenging due to patient-specific variability. While test-time adaptation (TTA) offers a promising approach for personalized prediction, the rarity of IOH events often leads to unreliable test-time train- ing. To address this, we propose CSA-TTA, a novel Cross- Sample Augmented Test-Time Adaptation framework that enhances training by incorporating hypotension events from other individuals. Specifically, we first construct a cross- sample bank by segmenting historical data into hypotensive and non-hypotensive samples. Then, we introduce a coarse- to-fine retrieval strategy for building test-time training data: we initially apply K-Shape clustering to identify representa- tive cluster centers and subsequently retrieve the top-K se- mantically similar samples based on the current patient sig- nal. Additionally, we integrate both self-supervised masked reconstruction and retrospective sequence forecasting sig- nals during training to enhance model adaptability to rapid and subtle intraoperative dynamics. We evaluate the pro- posed CSA-TTA on both the VitalDB dataset and a real- world in-hospital dataset by integrating it with state-of-the- art time series forecasting models, including TimesFM and UniTS. CSA-TTA consistently enhances performance across settings—for instance, on VitalDB, it improves Recall and F1 scores by +1.33% and +1.13%, respectively, under fine- tuning, and by +7.46% and +5.07% in zero-shot scenar- ios—demonstrating strong robustness and generalization. Code — https://github.com/kanxueli/CSA-TTA Introduction Intraoperative hypotension (IOH) — typically defined as blood pressure falling below a critical threshold for a sus- tained period (Dong et al. 2024; Wesselink et al. 2018)—is a common but serious complication during surgery. It is strongly associated with adverse outcomes such as acute kidney injury, myocardial infarction, stroke, and even mor- tality (Jeong et al. 2024; Lee et al. 2021). Accurate and timely prediction of IOH is critical for enabling early in- terventions before blood pressure drops to dangerous lev- *Corresponding authors: Yibing Zhan and Hua Jin. Copyright © 2026, Association for the Advancement of Artificial Intelligence (www.aaai.org ). All rights reserved. Figure 1: An illustrative comparison. Standard TTA, relying on recent stable history, often produces overly smooth pre- dictions and misses sudden changes. CSA-TTA leverages a cross-sample augmented dataset to capture diverse temporal patterns, enabling personalized IOH prediction. els (Hwang et al. 2023; Yoon et al. 2020), thereby re- ducing both the incidence and severity of these adverse events (Mukkamala et al. 2025). However, due to the com- plex, dynamic, and highly patient-specific nature of physi- ological responses during surgery, reliable IOH prediction remains a significant challenge despite advances in intraop- erative monitoring and machine learning. Recent studies have explored various methods to im- prove IOH prediction (Shi et al. 2023a; Sidiropoulou et al. 2022; Lee et al. 2018). For instance, CMA (Lu et al. 2023) employed attention mechanisms to capture temporal and feature-level dependencies, while HMF (Cheng et al. 2024) integrated contextual, physiological, and temporal features. However, the ability of these models to generalize remains limited by individual differences in patients’ physiology and the influence of clinical interventions (e.g., anesthesia or drug administration) (Cai et al. 2025; Mohammadi et al. arXiv:2512.15762v1 [cs.LG] 12 Dec 2025 Ψ hθ gθ Ψ’ Ω Ψ Attention Normalize Forward Normalize 풓풆έζ(Ψ, Ψ) 풑풓풆ή(Ω, Ω) wt wt-m wt+m Figure 2: The main proposed CSA-TTA framework. It comprises three key steps: (1) Cross-sample bank construction, (2) Coarse-to-fine retrieval, and (3) Multi-task optimization. 2024; Li et al. 2022). These factors introduce implicit dis- tribution shifts in real-time signals that are typically poorly captured by population-level models. TTA offers a promising paradigm to tackle such distri- bution shifts by adapting models using test data during in- ference (Liang, He, and Tan 2025). Techniques such as TTT (Sun et al. 2020) and TTT++ (Liu et al. 2021) lever- age self-supervised auxiliary tasks to refine models at infer- ence time and have shown effectiveness in fields like com- puter vision (Karmanov et al. 2024; Lim et al. 2023) and NLP (Shi et al. 2024). In the context of IOH, TTA holds po- tential for personalizing predictions by utilizing recent pa- tient history to dynamically adjus

This content is AI-processed based on ArXiv data.