Veritas-RPM: Provenance-Guided Multi-Agent False Positive Suppression for Remote Patient Monitoring
A new five-layer AI system with six specialist agents achieved a 98% true suppression rate on synthetic patient data.
A team of researchers including Aswini Misro, Vikash Sharma, and Shreyank N Gowda has introduced Veritas-RPM, a novel AI architecture designed to tackle the critical problem of false alarms in remote patient monitoring (RPM). The system employs a sophisticated, provenance-guided multi-agent framework consisting of five distinct processing layers: VeritasAgent for assembling ground-truth data, SentinelLayer for initial anomaly detection, DirectorAgent for routing cases, six domain-specific Specialist Agents for detailed analysis, and a MetaSentinelAgent for final conflict resolution and decision-making.
To rigorously test the system, the team constructed a comprehensive synthetic taxonomy of 98 documented false-positive RPM scenarios. They generated 530 synthetic patient epochs directly from this taxonomy and processed them through the Veritas-RPM pipeline, where ground-truth labels were known for validation. The architecture's performance was measured using key metrics: True Suppression Rate (TSR), False Escalation Rate (FER), and Indeterminate Rate (INDR). The results demonstrated the system's effectiveness, with a reported TSR of 98%, indicating a high capability to correctly identify and suppress false alarms before they reach healthcare providers.
The core innovation lies in the multi-agent design, which moves beyond a single AI model. By decomposing the complex task of medical alert validation into specialized roles—from initial triage to domain-specific analysis and final arbitration—the system mimics a more nuanced, expert-led clinical review process. This structured approach allows Veritas-RPM to trace the 'provenance' or reasoning path of each decision, providing transparency and accountability that is crucial for medical applications. The successful suppression of false positives addresses the pervasive issue of alert fatigue, which can lead to desensitization and missed critical events.
- Five-layer multi-agent architecture with six specialist AI agents for detailed case analysis.
- Tested on a synthetic dataset of 530 patient epochs derived from 98 false-positive scenarios.
- Achieved a 98% True Suppression Rate (TSR), drastically reducing unnecessary clinical alerts.
Why It Matters
Directly reduces clinician alert fatigue, a major patient safety issue, by filtering out up to 98% of false alarms in remote monitoring.