Research map / Time series / Foundation models for time series

OpenMHC: Accelerating the Science of Wearable Foundation Models

NeurIPS 2026 · Time series › Foundation models for time series · approach: Time-series language reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Releases OpenMHC, a large open dataset of 11,894 participants' Apple HealthKit wearable data (67M hours) with a predictive, imputation, and forecasting benchmark, plus open reimplementations of WBM and LSM-2 wearable foundation models and baselines.

Problem

Large wearable health datasets are private or access-gated, and leading wearable foundation models are rarely open-weight or reproducible, which blocks open research on behavioral wearable data.

Approach

Curate a broadly accessible longitudinal HealthKit dataset from the My Heart Counts study and build a three-track benchmark (predictive, imputation, forecasting) with skill-score and average-rank aggregation, then reimplement WBM and LSM-2 and compare them against classical, neural, and time-series foundation model baselines.

Key results

• LSM-2 achieves the best average rank (R=2.20) and skill score (S=+15.1) on the predictive track, ahead of XGBoost (S=+11.6).
• LSM-2-Sparse (7-day) attains the best imputation skill score (+64.7) versus LOCF reference, with daily LSM-2 at +61.4.
• Fine-tuned Chronos-2 is best on forecasting (R=3.56, S=+37.6), with non-fine-tuned Chronos-2 at S=+36.4 and DLinear at S=+35.9.

Surprising findings

• A well-tuned XGBoost with 495 hand-crafted features is the second-best predictive model, trailing only LSM-2.
• Constant mode-based baselines score highest on some activity channels because those channels are mostly zero.

Limitations

• Outcomes are self-reported and noisy, so tasks are retrospective characterization rather than clinical prediction.
• Cohort is skewed toward white, male, US, late-thirties participants.
• WBM and LSM-2 are best-effort reimplementations that may not match the original proprietary models' performance.

Experimental setup

Datasets: OpenMHC

Benchmarks: OpenMHC predictive tasks (32 self-reported outcomes) OpenMHC daily and long-context imputation OpenMHC 24-hour forecasting

Metrics: AUPRC Spearman rho Pearson r MAE ROC AUC AUROC skill score average rank bootstrap 95% CI

Baselines: XGBoost MultiRocket Linear GRU-D WBM Chronos-2 Toto LOCF BRITS DLinear FEDformer TimesNet

Builds on: LSM-2 WBM Chronos-2 Toto XGBoost DLinear BRITS MultiRocket GRU-D

Closest papers by meaning

• Beyond Sensor Data: Foundation Models of Behavioral Data from Wearables Improve Health Predictions (ICML 2025 · similarity 0.90)
• Scaling Wearable Foundation Models (ICLR 2025 · similarity 0.89)
• Large-scale Training of Foundation Models for Wearable Biosignals (ICLR 2024 · similarity 0.88)
• Inertia-1: An Open Exploration of Wearable Motion Foundation Models (NeurIPS 2026 · similarity 0.86)
• WildPPG: A Real-World PPG Dataset of Long Continuous Recordings (NeurIPS 2024 · similarity 0.83)
• CLIMB: Data Foundations for Large Scale Multimodal Clinical Foundation Models (ICML 2025 · similarity 0.83)