VitalAgent
tool-augmented ECG/PPG monitoring agent (Univ. of Melbourne, Ting Dang group)
ECGPPGText
Key facts
- Architecture
- LLM agent (DeepSeek-V4 Flash via OpenAI-compatible API in the paper) + 29 deterministic signal/lookup/knowledge tools + longitudinal physiological memory; ECGFounder used inside the ecg_diagnosis tool
- Pretraining Data
- None — no model trained; evaluated on VitalBench (Icentia11k, AF-PPG-ECG, PPG-DaLiA, WESAD)
- Venue
- arXiv 2026-05-28; NeurIPS 2026 GenAI4Health Workshop
Summary
Reactive QA: Tier A 0.860 / Tier B 0.557 accuracy (leakage-free). Per-window AF from tool features is weak (balanced acc ~0.67). VitalBench is non-commercial.
Reported benchmarks
AFib / arrhythmia
Long-term ECG and wrist PPG with annotated AF episodes (AFPPGECG) · Bal. accuracy ↑
#1 of 1
0.672
Per-window 10 s, deterministic RR-variability tools, no training; sens 78.3 / spec 56.1; 44 pts; only row on this axis
AFib / arrhythmia
icentia11k · Bal. accuracy ↑
#1 of 1
0.678
Same pipeline on ECG; sens 71.9 / spec 63.6; 25 pts; only row on this axis