openmhc-wbm-dp

OpenMHC Outcome Prediction — WBM

Track 1 (outcome prediction) reference checkpoint for the MyHeartCounts / OpenMHC wearable-health benchmark.

This checkpoint is the WBM encoder — a bi-directional Mamba2 contrastive self-supervised model that maps a week of wearable sensor data (168 hourly steps, 19 channels) to a 256-d representation. The reported WBM model pairs this encoder (per-user pooled → PCA-50 → linear probe) with a Linear fallback for users without a weekly embedding.

This is an OpenMHC reimplementation of Apple's WBM (Wearable Behavior Model), introduced in Beyond Sensor Data: Foundation Models of Behavioral Data from Wearables Improve Health Predictions (Apple, 2025; see references below).

Pretrained with a contrastive objective on the MHC training split.

• Checkpoint format: PyTorch Lightning checkpoint (model.ckpt) + normalization_stats.json (canonical hourly z-score constants; channels 0–6 normalized, 7–18 identity).
• Outcome-prediction tasks: 33 health & behavior labels (classification, ordinal, regression).

Model & implementation

• Paper: Beyond Sensor Data: Foundation Models of Behavioral Data from Wearables Improve Health Predictions (Apple, 2025)
• Apple Machine Learning Research summary

Requirements

Running the encoder needs the CUDA-only Mamba2 kernels (mamba-ssm) and a GPU.

Usage

import openmhc
from openmhc.encoders import WBM

# pip install "openmhc[hf]"  (+ mamba-ssm on a CUDA machine)
enc = WBM.from_release("hf://MyHeartCounts/[email protected]")
results = openmhc.evaluate_prediction(enc, version="full")

See openmhc_manifest.json for provenance (source W&B artifact, training details) and architecture metadata.

Citation

If you use this checkpoint, please cite the OpenMHC benchmark.