Can a zero-shot time-series foundation model rival task-trained models for intraoperative hypotension prediction? A two-cohort benchmark and the role of covariate-awareness

Technical Summary: Zero-Shot Time-Series Foundation Models for Intraoperative Hypotension Prediction

Problem Statement
Intraoperative hypotension (defined as mean arterial pressure [MAP] < 65 mmHg) is a frequent complication of general anesthesia linked to postoperative organ injury and mortality. Current predictive solutions face significant portability barriers. Commercial indices (e.g., Hypotension Prediction Index) often require invasive arterial waveforms and proprietary features, while state-of-the-art deep learning models (e.g., Temporal Fusion Transformer [TFT], PatchTST) require task-specific training on large, site-specific labeled cohorts. These bespoke models struggle to transfer across institutions with different data cadences, covariate sets, and hypotension prevalence rates. Furthermore, existing models generally condition on drug interventions only as observed history or omit them entirely, failing to utilize "known-future" signals (the planned drug infusion trajectory) that could improve forecasting accuracy.

Methodology
The study benchmarks four zero-shot time-series foundation models against two task-trained baselines (TFT and PatchTST) across two distinct intraoperative cohorts:

1. Development Cohort (VitalDB): 2,708 cases (2,665 subjects) from Seoul National University Hospital, sampled at 15-second intervals.
2. External Cohort (MOVER): 1,827 cases from a multi-center repository, sampled at 60-second intervals with 2–3× higher hypotension prevalence.

Experimental Design:

• Task: A rolling-origin forecasting task where a 30-minute context window predicts MAP trajectories and impending hypotension (MAP < 65 mmHg) over horizons of 1, 3, 5, 7, 10, and 15 minutes.
• Models Evaluated:
  • Zero-Shot Foundation Models: TiRex-2 (multivariate, covariate-aware), Chronos-Bolt, TimesFM-2.5, and Moirai-1.1-R (all univariate). None were trained or fine-tuned on the target task.
  • Task-Trained Baselines: TFT and PatchTST, trained via 5-fold subject-level cross-validation on the development cohort.
• Covariate Handling: The study specifically investigates "covariate-awareness." TiRex-2 is the only model capable of conditioning on both past observations and known-future inputs (the planned drug-infusion trajectory). The authors compare TiRex-2 with and without this covariate (M1 vs. M0) and compare it against univariate models that cannot ingest covariates.
• Evaluation Metrics: Discrimination (AUROC, AUPRC), probabilistic accuracy (CRPS, MAE), calibration (Expected Calibration Error), and decision-curve analysis.

Key Results

• Foundation Model Performance: Among the zero-shot models, TiRex-2 consistently outperformed the others (Chronos-Bolt, TimesFM-2.5, Moirai-1.1-R) across all horizons. Its lead was statistically significant from 3 minutes onward.
• Comparison to State-of-the-Art (SOTA):
  • At short-to-mid horizons (1–7 minutes), zero-shot TiRex-2 matched the performance of the task-trained TFT (e.g., AUROC 0.905 vs. 0.903 at 7 min; p = 0.34).
  • At longer horizons (>7 min), TiRex-2 trailed the stronger supervised baseline, PatchTST (e.g., AUROC 0.849 vs. 0.875 at 15 min).
  • On the external MOVER cohort, TiRex-2 maintained AUROC ≥ 0.85 across all horizons, matching in-domain trained models and outperforming models transferred from VitalDB.
• The Role of Covariate-Awareness:
  • Conditioning on the known-future drug infusion provided a small but significant improvement for TiRex-2 (approx. +1.15% CRPS reduction) specifically in "transition windows" where the infusion plan changed.
  • Crucially, TiRex-2's base forecasting quality (even without the covariate) was sufficient to lead the zero-shot field. The covariate awareness distinguished the best exemplar but was not the sole driver of its superiority over other foundation models.
  • In contrast, task-trained models showed a much larger dependence on the covariate channel (+7% to +12% improvement), suggesting they rely more heavily on explicit intervention signals than the pre-trained foundation model.
• Robustness and Portability: The zero-shot paradigm demonstrated high portability. A single pretrained model applied without labels achieved discrimination comparable to cohort-specific supervised training and superior to supervised transfer learning across the distribution shift between VitalDB and MOVER.
• Internal Representations: Analysis of penultimate-layer activations revealed that zero-shot forecasting skill is decoupled from linear representational structure. Unlike task-trained models, zero-shot models do not encode hypotension risk as a linearly separable feature in their internal states; their value lies in the full encoder-decoder mapping for multi-horizon prediction, not as feature extractors.

Significance and Claims
The paper claims that zero-shot, covariate-aware time-series foundation models represent a competitive and portable alternative for intraoperative hypotension prediction.

• Portability: The primary contribution is demonstrating that a single pretrained model can rival task-trained SOTA (specifically TFT) at clinically urgent horizons (1–7 min) without requiring site-specific labels or retraining. This addresses the "data silo" problem in medical AI.
• Known-Future Conditioning: The study establishes the feasibility of applying known-future conditioning (planned drug infusions) in a zero-shot setting. While the performance gain was modest compared to the base model's capability, it validates a paradigm where foundation models can exploit future intervention plans—a capability previously restricted to task-specific architectures.
• Clinical Utility: The models provide actionable early warnings (1–15 min ahead) with discrimination comparable to supervised baselines, particularly in the "middle horizon" (3–7 min) where naive persistence alarms degrade and bespoke models are hardest to deploy.
• Limitations Acknowledged: The authors note that the "known-future" covariate used was the realized infusion (retrospective data), which is an optimistic upper bound compared to prospective deployment where only the planned infusion is known. Additionally, the models were evaluated on invasive arterial lines; portability to non-invasive monitoring or other care settings (ICU, ward) remains unproven. The study concludes that while discrimination is label-free, calibration requires a small, local labeled subset for recalibration.

In summary, the paper argues that for institutions lacking the infrastructure for large-scale bespoke model training, zero-shot foundation models offer a viable, "off-the-shelf" solution for hemodynamic monitoring that matches the performance of specialized supervised models at critical time horizons.