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

TL;DR

The paper tackles the challenge of predicting intraoperative hypotension (IOH) by introducing CSA-TTA, a cross-sample augmented test-time adaptation framework that leverages a cross-sample bank, coarse-to-fine retrieval, and multi-task optimization to personalize IOH predictions. By integrating both self-supervised masked reconstruction and retrospective forecasting, CSA-TTA enriches the adaptation signal with patterns from hypotensive and non-hypotensive samples across patients, addressing the scarcity of IOH events. Extensive experiments on VitalDB and an in-hospital dataset show consistent improvements over strong baselines (TimesFM, UniTS) in zero-shot and fine-tuning settings, with notable gains in Recall and F1 and robust regression performance. The work demonstrates practical potential for real-time, personalized IOH monitoring and offers a modular framework that could generalize to other perioperative time-series tasks, while highlighting areas for dynamic bank maintenance and broader robustness evaluation.

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 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 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 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 scenarios-demonstrating strong robustness and generalization.

Figures (9)

• Figure 1: An illustrative comparison. Standard TTA, relying on recent stable history, often produces overly smooth predictions and misses sudden changes. CSA-TTA leverages a cross-sample augmented dataset to capture diverse temporal patterns, enabling personalized IOH prediction.
• 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.
• Figure 3: Illustration of the key steps in applying CSA-TTA for personalized IOH prediction. (1) Domain knowledge adaptation, (2) CSA-TTA, and (3) Sequence prediction and IOH detection.
• Figure 4: Case study visualizations on VitalDB (2-second sampling). TimesFM + CSA-TTA predicts 5-minute (top) and 15-minute (bottom) horizons.
• Figure 5: Distribution of hypotension event ratios per patient case and overall class proportions in the VitalDB dataset. The left panel shows the distribution of hypotension event (considered positive sample) ratios for each case, with arrows marking the proportions of cases below 10%, below 20%, and above 80%. The right panel illustrates the global proportion of positive and negative samples for the VitalDB dataset.