mSignals: Multi-Agent ICU Patient State Alignment
ICU patients generate data across five independent clinical systems simultaneously — and no existing system aligns them. mSignals applies multi-agent data alignment architecture to acute care monitoring: assigning independent agents to vital signs, laboratory results, pharmacy, clinical notes, and patient history, enforcing freshness and confidence contracts on each, and synthesizing an aligned patient state where inter-agent conflicts are surfaced as first-class clinical signals. When a vital signs agent reports hemodynamic stability and a laboratory agent reports rising lactate — that divergence is the sepsis early warning. mSignals makes it visible. No existing score-fusion system can.
94–99% are false positives
US healthcare system
Sepsis detection at 6h horizon
Vitals · Labs · Pharmacy · Notes · History
The mSignals validation study runs compound detection rules retrospectively against MIMIC-IV (40,000+ de-identified ICU stays, Beth Israel Deaconess Medical Center) — the standard dataset for ICU AI research. The study compares conflict-aware synthesis against single-parameter threshold alarms, the Epic Deterioration Index, and SOFA score trajectory. The MIMIC-IV validation artifact serves simultaneously as thesis proof, enterprise sales evidence, and the clinical evidence package for FDA 510(k) submission.
Each agent fetches its clinical domain independently. The Governor layer applies freshness TTL contracts and staleness penalties before any agent output reaches synthesis. The ICU Coordinator runs five compound detection rules — sepsis, AKI, respiratory failure, drug-physiology conflict, and fluid overload — and produces an Aligned Patient State where inter-agent disagreements are named, attributed, and resolved with a directive. Every action is logged to an immutable audit trail satisfying HIPAA §164.312(b) and the 21st Century Cures Act CDS attribution requirement.
NLA · HIA
Independent fetches
Governor contracts
LRA lactate rising
↕ Divergence = signal
not noise
Agent attribution
Human Gate tier
Audit log entry
