Health Reasoning over Time Series

A unified benchmark for evaluating hierarchical reasoning capabilities of LLMs over general health time series.

HEARTS benchmark overview HEARTS benchmark overview — animated
What's new

Latest from HEARTS

Leaderboard

Current Leader

Overall score
About HEARTS

The first diverse benchmark for health time-series reasoning.

With the broadest coverage of sequence length, sampling frequency, and time span to date, HEARTS tests whether models truly reason over physiological signals — not just match surface patterns.

16Real-world
datasets
12Health
domains
20Signal
modalities
110Reasoning
tasks

Distribution across 12 health domains

    Broadest coverage to date

    Sequence length
    601M+
    Sampling frequency
    daily48 kHz
    Time span
    secondsyears

    20 signal modalities

    EEGECGPPGSpO₂ Blood PressureHeart RateEMGIMU TemperatureEye TrackingEOGPERG CGMEDAThoracicRespiration Air FlowAudioAnnotationAggregated

    Organized into four levels of reasoning

    01

    Perception

    Identify physiological markers that ground downstream analysis.

    2 task types
    • Statistical Calculation
    • Feature Extraction
    02

    Inference

    From precise temporal grounding to holistic subject characterization.

    3 task types
    • Event Localization
    • Physiological Classification
    • Subject Profiling
    03

    Generation

    Point-by-point synthesis of complete physiological sequences.

    3 task types
    • Future Forecasting
    • Signal Imputation
    • Cross-modal Translation
    04

    Deduction

    Arrow-of-time reasoning over longitudinal, multi-session signals.

    2 task types
    • Temporal Ordering
    • Trajectory Analysis
    HEARTS-LIGHTENING A cost-efficient evaluation subset — our standard track for future evaluations.

    To balance evaluation comprehensiveness with computational efficiency, we introduce HEARTS-LIGHTENING, a cost-efficient subset comprising 5,417 test samples. This streamlined configuration preserves the complete taxonomy of all 110 tasks while reducing the number of test cases per task, thereby substantially lowering computational overhead without compromising the statistical stability of our evaluations.

    Re-evaluating the models under this lightweight setting confirms that the overall performance tiers remain stable, with only minor ranking shifts observed among models with closely matched scores. Given its efficiency and strong alignment with the full benchmark, we adopt the HEARTS-LIGHTENING configuration as a standard track for future evaluations.

    What we learned

    Genuine reasoning, or surface patterns and priors?

    5 key findings

    Click any card to see the supporting evidence, redrawn from the results.

    Contribute

    Put your model on the board

    Submission is handled through the official Google form. Contributors should submit either model weights or a model API endpoint. The HEARTS team runs the benchmark tests internally and publishes validated leaderboard updates.

    1

    Submission

    Submit model weights or an API endpoint through the official Google form.

    2

    Team evaluation

    The HEARTS team runs the benchmark internally and verifies submission validity.

    3

    Leaderboard update

    Accepted runs are published with standardized, comparable reporting.

    Before you submit — checklist & details

    Checklist

    • Use the official Google submission form.
    • Submit either model weights or a model API endpoint (one is required).
    • Provide clear access instructions so the team can run evaluation.
    • Include model name, version, and expected inference configuration.
    • Include contact details for follow-up and troubleshooting.

    What to include

    • Submission type: model weights or API endpoint.
    • Model family, checkpoint / version, and provider.
    • Access details (download path or endpoint / auth instructions).
    • Inference requirements (context window, tool usage, limits).
    • Any constraints or caveats to know before running tests.

    Contact

    Questions, feedback, or interested in contributing? Reach the HEARTS team — we're happy to help with setup, access, or proposing new tasks and datasets.

    Cite

    Citation

    @article{hearts2026,
        title={HEARTS: Benchmarking LLM Reasoning on Health Time Series},
        author={Sirui Li and Shuhan Xiao and Mihir Joshi and Ahmed Metwally and Daniel McDuff and Wei Wang and Yuzhe Yang},
        journal={arXiv preprint arXiv:2603.06638},
        year={2026}
    }