OpenMHC

Evaluation API and reference implementations for the MyHeartCounts Datasets & Benchmarks (MHC D&B) leaderboard. Wearable sensor data from a real-world cardiovascular cohort, with three benchmark tracks: outcome prediction, imputation, and forecasting.

• Leaderboard: https://myheartcounts.stanford.edu/benchmark
• Submit a result: open a submission issue
• Paper: OpenMHC: Accelerating the Science of Wearable Foundation Models (NeurIPS 2026)

Install

git clone https://github.com/AshleyLab/myheartcounts-dataset.git
cd myheartcounts-dataset
pip install -e .

Python ≥ 3.10. Installs the openmhc package and its evaluation dependencies.

Quickstart

Download the tiny version of the dataset (small subset for reviewers and quickstart):

import openmhc
openmhc.download_dataset(version="tiny")

Then evaluate a model. Models implement one of two duck-typed protocols — no inheritance required.

Track 1 — outcome prediction (Encoder)

import numpy as np
import openmhc

class MeanPoolEncoder:
    def encode(self, weekly_tensors: np.ndarray) -> np.ndarray:
        # weekly_tensors: (B, 168, 38). Return (B, D) embeddings.
        return weekly_tensors[:, :, :19].mean(axis=1).astype(np.float32)

results = openmhc.evaluate_downstream(MeanPoolEncoder())
print(results.summary())
print("global score (mean AUROC over binary tasks):", results.global_score)

Track 2 — imputation (Imputer)

import numpy as np
import openmhc

class MeanImputer:
    def fit(self, data, masks):
        self.means = np.nanmean(data, axis=(0, 2))

    def impute(self, data, observed_mask, target_mask):
        out = data.copy()
        for ch in range(19):
            target = target_mask[:, ch, :] == 1
            out[:, ch, :][target] = self.means[ch]
        return out.astype(np.float32)

results = openmhc.evaluate_imputation(MeanImputer())
print(results.summary())

A more complete walkthrough is in notebooks/quickstart.ipynb.

Submit to the leaderboard

body = results.to_submission_yaml(
    method_name="My Method",
    submitter_team="Stanford CS",
    paper_url="https://arxiv.org/abs/...",
    code_url="https://github.com/...",
)
print(body)

to_submission_yaml returns a paste-ready body matching the textareas in the submission issue template. For Track 2 imputation, skill scores and per-category subgroup scores are computed locally against the frozen LOCF baseline; for Track 1, those fields are filled in by the maintainers from raw_metrics during ingestion. The HuggingFace Space ingests merged submissions and the public leaderboard rebuilds automatically.

Submissions must follow the standard evaluation protocol — same split file, masking config, and label-validity criterion as the paper. The submission template enforces required fields.

Repo layout

Path	What's there
src/openmhc/	Public API (evaluate_downstream, evaluate_imputation, download_dataset, …)
src/downstream_evaluation/	Track 1 internals (linear probes, time-window selection, metrics)
src/imputation_evaluation/	Track 2 internals (masking scenarios, per-channel metrics)
src/labels/	Label registry + type lookup
data/labels/	Schema-only registry files (label types, ordinal vocab, validity config)
notebooks/quickstart.ipynb	End-to-end example
.github/ISSUE_TEMPLATE/submission.yml	Leaderboard submission form

The actual participant data lives separately — see DATASET.md for download instructions and split conventions.

Citation

@inproceedings{openmhc2026,
  title     = {OpenMHC: Accelerating the Science of Wearable Foundation Models},
  author    = {MyHeartCounts Team},
  booktitle = {NeurIPS Datasets and Benchmarks},
  year      = {2026}
}

License

Code: MIT. Dataset: governed by a separate Data Use Agreement — see DATASET.md.