Zero-shot time-series foundation models for intraoperative hypotension prediction

Code and derived results for the manuscript:

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. Max Haberbusch, Medical University of Vienna.

The frozen code + data snapshot for this work is archived on Zenodo: https://doi.org/10.5281/zenodo.21340658

We benchmark four zero-shot time-series foundation models (TiRex-2, Chronos-Bolt, TimesFM-2.5, Moirai-1.1-R) against two task-trained baselines (Temporal Fusion Transformer, PatchTST) for forecasting mean arterial pressure (MAP) and predicting impending hypotension (MAP < 65 mmHg) over 1–15 min. The foundation models are applied with no task-specific training and no labels. Development is on VitalDB (2,708 cases); external validation is on the independent MOVER cohort (1,827 cases). The two cohorts are always reported stratified, never pooled.

---

Reproduce every figure and table — one notebook

Everything a reviewer needs to verify the paper is in reproduce_paper.ipynb. It regenerates every main and supplementary figure and prints every table's numbers from the released result files. No model is retrained and no foundation-model inference is run — the notebook reads the per-window forecasts and aggregate metrics we provide and rebuilds the figures/tables from them.

# 1. create an environment (Python 3.11+)
pip install -r requirements.txt

# 2. get the released result files (see "Where the data lives" below), then:
jupyter notebook reproduce_paper.ipynb        # Kernel -> Restart & Run All

The notebook's final cell prints a figure/table → result-file provenance map for auditability.

Where the data (result files) lives

The result files split into two tiers:

• Curated small results (~0.4 MB) are committed to this repository: every manuscript table (results/tables/) plus the aggregate CSVs/JSONs the notebook reads. A bare clone therefore already contains every table's numbers.
• Bulk results (~1.4 GB: per-window forecasts, per-model embeddings, model checkpoints) are not in git. They are archived on Zenodo: https://doi.org/10.5281/zenodo.21340658 Download that bundle and unpack it into the repository root so its results/ files sit alongside the committed ones, next to reproduce_paper.ipynb.

The full figure/table regeneration in the notebook needs the Zenodo bulk files; browsing the table values does not. This GitHub repository is the living codebase; the Zenodo record is the citable, frozen snapshot of code + results at publication.

The raw source datasets are not redistributed here (data-use terms):

• VitalDB — openly available at https://vitaldb.net and via PhysioNet.
• MOVER — public-access, from the UCI Machine Learning Repository (https://archive.ics.uci.edu/dataset/877/mover).

Reproducing the figures/tables does not require the raw datasets — only the released results/. Re-running the models from raw data (optional) is documented in notes/REPRODUCE.md.

Repository layout

reproduce_paper.ipynb     one-shot figure/table regeneration (start here)
requirements.txt          runtime dependencies for the notebook
scripts/                  figure/table generators + shared forecasting-pipeline code
  paper_figures.py          Fig 1–5, several tables
  transfer_figure.py        Fig 6 + transfer table
  decision_curves_figure.py FigS decision curves
  external_table.py         external-validation table
  stats_tests.py            paired significance tests table
  compute_footprint.py      per-model compute footprint (GPU; measurement only)
  compute_footprint_table.py aggregates footprint JSON -> table (stdlib, no GPU)
  explainability/           representation-analysis figures (UMAP, RSA/CKA)
  baselines/                TFT/PatchTST models, training, zero-shot adapters
datasets/vitaldb/         VitalDB loader, cohort builder, configs, DATA_NOTES.md
datasets/mover/           MOVER loader + configs
configs/eval.yaml         shared evaluation protocol (horizons, threshold, bootstrap)
slurm/                    cluster job scripts for the heavy runs (training, inference, embeddings)
notes/REPRODUCE.md        end-to-end recipe to regenerate results from raw data
notes/CLUSTER.md          how the heavy runs execute on the SLURM/A100 cluster
MOVER_SCHEMA_REPORT.md    MOVER schema and how it maps onto the VitalDB pipeline

Run pipeline scripts with PYTHONPATH=scripts:datasets/vitaldb:datasets/mover so the shared pipeline finds the dataset loaders.

Cluster (optional — re-running the heavy steps)

Model training, zero-shot inference, embedding extraction and the compute-footprint measurement run on an A100 GPU via SLURM + a Pyxis/enroot container. See notes/CLUSTER.md and the scripts in slurm/.

License

Code is released under the license in LICENSE.txt. Released result files are distributed under CC BY 4.0. The raw VitalDB and MOVER datasets remain under their respective terms.

Citation

If you use this code or the archived results, please cite the Zenodo record:

Haberbusch, M. (2026). 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" (v1.0.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.21340658

@misc{HaberbuschZenodo2026,
  author    = {Haberbusch, Max},
  title     = {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''},
  year      = {2026},
  publisher = {Zenodo},
  version   = {1.0.0},
  doi       = {10.5281/zenodo.21340658},
  url       = {https://doi.org/10.5281/zenodo.21340658}
}