ECG Reasoner
Explainable wearable cardiac monitoring that pairs deterministic ECG signal processing with multi-agent clinical reasoning under the OpenClaw safety harness.
Motivation
Large language models are starting to reason about clinical data, but in cardiology a confident-sounding wrong answer is not a harmless one. Most medical AI systems remain hard to interpret, weakly grounded in actual physiology, and difficult to trust in high-risk settings. ECG Reasoner takes the opposite stance: every diagnosis must be traceable to a specific physiological measurement, every uncertain case must fail conservatively, and the whole thing must run on hardware a person already owns — a watch, a laptop, and a safety layer in between.
How it works
ECG waveforms are captured from a consumer wearable such as an Apple Watch and passed through a deterministic DSP layer built on the classical Pan–Tompkins QRS detection algorithm — bandpass filtering, derivative enhancement, squaring, moving-window integration, and adaptive thresholding — for transparent, reproducible R-peak localization without any black-box inference. From the detected beats the system extracts an interpretable feature bundle: heart rate, heart rate variability (HRV/SDNN), RR-interval variability, ectopic-beat ratio, and R-peak amplitude.
A rule-based arrhythmia classifier first normalizes these features into safety-oriented risk categories (normal, watch, concern, critical) before any LLM is involved. The structured features then go to an AG2 multi-agent consultation, where a SignalAnalyst agent proposes ranked diagnostic hypotheses and a Cardiologist agent commits to a single diagnosis — with every conclusion required to cite explicit feature values. Wrapping the entire reasoning layer is OpenClaw, a safety harness enforcing output guardrails, evidence verification, consistency checks, conservative fallback policies, and audit logging. The final output is an evidence-grounded diagnosis delivered in dual modes: a technical clinician view and a plain-language patient view.
Funding & recognition
🏆 2nd Place Award — CVPR 2026 Workshop, GigaBrain Challenge (PhysClaw Track), Denver, CO, June 2026.
Awarded to Team OpenClaw CardioHarness: Zhipeng Jiang, Xinyu Zhang, Hanzhang Fang, Jing Huang, Ying Zhou.
Official challenge page: GigaBrain Challenge 2026