HEARTS: Benchmarking LLM Reasoning on Health Time Series

Summary

HEARTS is a comprehensive benchmark for evaluating large language model reasoning over health time series — including ECG, PPG, accelerometer data, and sleep signals. It provides standardised tasks, datasets, and metrics for zero-shot, few-shot, and fine-tuned LLM evaluation on physiological signal interpretation. The benchmark enables systematic comparison of LLM reasoning capabilities on wearable health data, establishing whether language models can serve as health signal interpreters without task-specific training.