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

Conference: AAAI 2026 arXiv: 2512.15762 Code: GitHub Area: Medical Imaging Keywords: intraoperative hypotension prediction, test-time adaptation, cross-sample retrieval, time series forecasting, personalized medicine

TL;DR

This paper proposes the CSA-TTA framework, which enhances personalized intraoperative hypotension prediction at test time by constructing a cross-sample bank, performing coarse-to-fine retrieval, and applying multi-task optimization to retrieve hypotension event signals from other patients' data.

Background & Motivation

• Intraoperative Hypotension (IOH): MAP < 65 mmHg lasting ≥ 1 minute, which can lead to acute kidney injury, myocardial infarction, stroke, or death.
• Accurate IOH prediction is critical for early intraoperative intervention, yet inter-patient physiological variability is substantial.
• Limitations of existing methods:
  • Population-level models such as CMA (attention mechanism) and HMF (multi-feature fusion) fail to capture individual differences.
  • Clinical interventions (anesthesia, drug administration) introduce covert distribution shifts, degrading generalization of population models.
• Test-Time Adaptation (TTA) is a promising direction: TTT and TTT++ fine-tune models at inference time via self-supervised auxiliary tasks.
• TTA challenges in IOH: hypotensive events are extremely sparse.
  • In the VitalDB dataset, hypotension accounts for only 12.6% of samples.
  • Most patients experience hypotension during less than 10% of surgery time.
  • Standard TTA relies on single-sample adaptation, failing to capture sudden blood pressure drops and producing overly smooth predictions.
• Core Insight: Patients with similar physiological characteristics exhibit similar intraoperative responses → hypotensive events from other patients can be retrieved to enrich the adaptation signal.

Method

Overall Architecture

CSA-TTA consists of three core steps (as shown in Figure 2):

1. Cross-Sample Bank Construction
2. Coarse-to-Fine Retrieval
3. Multi-task Optimization

Two operational modes are supported: fine-tuning mode (offline fine-tuning followed by TTA) and zero-shot mode (direct TTA without prior fine-tuning).

Key Designs

1. Cross-Sample Bank Construction

Physiological time series from all patients in the historical dataset are segmented into fixed-length clips to form the cross-sample bank \(\mathcal{B}\):

\[\mathcal{B} = \mathcal{B}_{\text{hypo}} \cup \mathcal{B}_{\text{non-hypo}}\]

• Clips are partitioned into two subsets based on whether they contain hypotensive events.
• Hypotension is defined as MAP < 65 mmHg lasting ≥ 1 minute.
• This partition enables targeted retrieval of hypotensive samples during adaptation.

2. Coarse-to-Fine Retrieval Strategy

An adaptive context window is used to process streaming patient data. At time step \(t\), the historical window \(\mathcal{W}_{t-m:t}^{\text{hist}}\) is defined.

Coarse-grained retrieval: - K-Shape clustering is applied separately to the hypotensive subset \(\mathcal{B}_{\text{hypo}}\) and the non-hypotensive subset. - K-Shape relies on shape-based matching, making it well suited to physiological time series (no explicit temporal alignment or amplitude normalization required). - A query sample is first assigned to a category, then matched to the most similar cluster centroid, rapidly narrowing the search space.

Fine-grained retrieval: - Within the cluster identified by coarse retrieval, Dynamic Time Warping (DTW) is used to compute semantic similarity. - The top-\(K\) most similar samples are selected to form the candidate set \(\mathcal{D}_{\text{retrieval}}\). - Retrieved samples are augmented with perturbations (Gaussian noise, time scaling) to increase diversity.

Constructing the adaptation dataset:

\[\mathcal{D}_t^{\text{CSA-TTA}} = \mathcal{W}_{t-m:t}^{\text{hist}} \cup \text{Aug}(\mathcal{D}_{\text{retrieval}})\]

3. Multi-task Optimization

The model \(F_\theta = (f_\theta, h_\theta, g_\theta)\) comprises: - A shared feature encoder \(f_\theta\) - A prediction branch \(h_\theta\) (primary task: sequence prediction) - A self-supervised branch \(g_\theta\) (auxiliary task: masked reconstruction)

Total loss:

\[\min_{f_\theta, h_\theta, g_\theta} \frac{1}{N} \sum_{n=1}^{N} \mathcal{L}_{\text{Pred}}(X_n, Y_n; f_\theta, h_\theta) + \mathcal{L}_{\text{Recon}}(X_n; f_\theta, g_\theta)\]

• Masked reconstruction enhances time series representation learning and helps capture subtle signal variations.
• Retrospective sequence prediction uses known historical data for self-supervised training.

Loss & Training

• Partial fine-tuning strategy: only input layers, output layers, and LayerNorm parameters are updated, balancing adaptability and generalization.
• Fine-tuning mode: 1 epoch of TTA updates; zero-shot mode: 3 epochs.
• Offline fine-tuning: 10 epochs, lr = 1e-4, batch size = 64, dropout = 0.01.
• Hybrid trigger mechanism for IOH prediction:
  • Hard trigger: detects sustained hypotensive episodes.
  • Soft trigger: average risk assessment within a sliding window.
  • The combination produces the final probability estimate.

Key Experimental Results

Main Results (Zero-shot & Fine-tuning)

Dataset: VitalDB (2,150 non-cardiac surgeries, 2s/30s sampling) + in-house dataset (130 test cases).

Zero-shot setting (VitalDB 30S):

Model	F1↑	Recall↑	MAE↓	MSE↓
TimesFM	64.17	58.87	6.49	92.77
TimesFM + CSA-TTA	64.90	59.27	6.28	85.28
UniTS	52.23	43.24	7.32	99.96
UniTS + CSA-TTA	57.30	50.70	7.19	95.84

UniTS + CSA-TTA: Recall +7.46%, F1 +5.07%.

Fine-tuning setting (VitalDB 2S):

Model	F1↑	Recall↑	MAE↓	MSE↓
TimesFM	64.20	64.93	6.03	77.87
TTT	64.00	64.77	6.02	77.70
TTT++	64.10	64.80	6.02	77.68
CSA-TTA	64.83	65.99	5.94	76.19

In-house dataset (zero-shot):

Model	F1↑	Recall↑	MAE↓	MSE↓
UniTS	56.10	43.77	6.45	91.97
UniTS + CSA-TTA	63.80	53.33	6.30	88.69

Recall +9.56%, F1 +7.70%.

Ablation Study

Multi-task optimization (TimesFM fine-tuned, 5-minute prediction):

Pred	Recon	F1	MAE	MSE
✗	✓	70.00	4.82	55.81
✓	✗	70.60	4.79	54.08
✓	✓	70.60	4.77	53.17

Data augmentation strategy (fine-tuning setting):

Bank	Aug	F1	MAE	MSE
✗	✗	65.90	5.87	74.79
✓	✗	66.03	5.85	74.19
✓	✓	66.07	5.82	72.93

Top-K selection: \(K=3\) yields F1 = 64.90 and MSE = 85.28, balancing relevance and diversity.

Key Findings

• CSA-TTA yields larger gains on weaker models (UniTS zero-shot F1 +5.07%), suggesting that cross-sample signals are particularly valuable for underfitting models.
• The combination of a cross-sample bank and perturbation augmentation outperforms either component used alone.
• Case studies show that CSA-TTA captures abrupt blood pressure drops and rebounds that standard models smooth over.
• Computational overhead is manageable: only 1.06% of TimesFM parameters are updated, with approximately 6.6 seconds per epoch.

Highlights & Insights

1. First application of TTA to personalized IOH prediction: addresses the failure of standard TTA under sparse event conditions.
2. Elegant cross-sample retrieval design: the two-stage coarse-to-fine pipeline balances efficiency and precision; the K-Shape → DTW combination is well suited to physiological signals.
3. Plug-and-play compatibility: applicable to arbitrary time series foundation models (TimesFM, UniTS), effective in both zero-shot and fine-tuning settings.
4. Clinical relevance: improvements in Recall directly reduce the miss rate of hypotensive events.

Limitations & Future Work

• Zero-shot Precision occasionally degrades (e.g., UniTS on VitalDB 30S: Precision −4.58%), indicating that recall gains come at a partial cost to precision.
• The cross-sample bank requires sufficiently large historical data, which may be unavailable for newly established clinical units.
• DTW computational cost scales rapidly with data volume, potentially requiring approximate algorithms.
• Validation is limited to MAP prediction and has not been extended to other vital signs.
• The retrieval strategy assumes that similar physiological features imply similar responses, without accounting for specific surgical type or medication differences.

Related Work & Insights

• TTT/TTT++ provide the foundational TTA paradigm; CSA-TTA extends this by incorporating external information sources.
• The application of K-Shape clustering to time series retrieval is worth adopting, as it is far more efficient than exhaustive DTW search.
• The cross-sample bank concept parallels Retrieval-Augmented Generation (RAG), adapted here to the time series adaptation setting.
• The partial parameter update strategy is generalizable to other online adaptation scenarios.

Rating

• Novelty: ⭐⭐⭐⭐ — The cross-sample TTA concept is original and addresses a genuine clinical pain point.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ — Two datasets, two backbone models, multiple ablations, and case analysis.
• Writing Quality: ⭐⭐⭐⭐ — Motivation is clear and method description is complete.
• Value: ⭐⭐⭐⭐ — High practical value with direct applicability to clinical decision support.

Related Papers

• [AAAI 2026] MIRAGE: Scaling Test-Time Inference with Parallel Graph-Retrieval-Augmented Reasoning Chains
• [ICCV 2025] Progressive Test Time Energy Adaptation for Medical Image Segmentation
• [AAAI 2026] GP-MoLFormer-Sim: Test Time Molecular Optimization through Contextual Similarity Guidance
• [CVPR 2026] SPEGC: Continual Test-Time Adaptation via Semantic-Prompt-Enhanced Graph Clustering for Medical Image Segmentation
• [AAAI 2026] Provably Minimum-Length Conformal Prediction Sets for Ordinal Classification