MyHeartCounts/openmhc-mixlinear-fc

OpenMHC Forecasting — MixLinear

Track 3 (forecasting) reference checkpoint for the MyHeartCounts / OpenMHC wearable-health benchmark (NeurIPS 2026).

This checkpoint is a MixLinear model. MixLinear is an extremely lightweight forecaster combining segment-based linear modeling in the time domain with adaptive low-rank filtering in the frequency domain.

Trained from scratch on the MHC training split using the PyPOTS implementation (requires pypots>=1.2).

• Checkpoint format: PyPOTS checkpoint (model.pypots) + standard_scaler_stats.json + training_config.json
• Forecasting task: 24-hour-ahead, 19 sensor channels, hourly resolution.

Model & implementation

• PyPOTS toolkit (implementation)
• PyPOTS documentation
• Paper: MixLinear: Extreme Low Resource Multivariate Time Series Forecasting with 0.1K Parameters (Ma et al., 2024)

Usage

import openmhc
from openmhc.forecasters import MixLinearForecaster

# pip install "openmhc[pypots]"
fc = MixLinearForecaster.from_release("hf://MyHeartCounts/openmhc-mixlinear-fc@v1.1")
results = openmhc.evaluate_forecasting(fc, version="full")

The same bundle also loads in the evaluation harness via model.release_dir=hf://MyHeartCounts/openmhc-mixlinear-fc@v1.1. See openmhc_manifest.json for provenance (training run, base model, fine-tuning details) and architecture metadata.

Citation

If you use this checkpoint, please cite the OpenMHC benchmark and the original MixLinear work (linked above).