Data and code for "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"

Derived results and reproduction code for a two-cohort zero-shot benchmark of time-series foundation models (TiRex-2, Chronos-Bolt, TimesFM-2.5, Moirai-1.1-R) against task-trained baselines (Temporal Fusion Transformer, PatchTST) for forecasting intraoperative mean arterial pressure and predicting impending hypotension (MAP < 65 mmHg) over 1–15 min, developed on VitalDB and externally validated on MOVER. The bundle contains per-window forecasts, aggregate metrics, embeddings, precomputed tables, and a notebook (reproduce_paper.ipynb) that regenerates every manuscript and supplementary figure and prints every table's numbers with no model retraining and no foundation-model inference. The raw VitalDB and MOVER datasets are not redistributed (subject to their data-use terms); only derived artifacts are included. A companion GitHub repository maintains the living codebase.