SCH: Collaborative and Grounded Al Agents for Accountable Bacteremia Decision - Bloodstream infections (bacteremia) are a frequent precursor to sepsis, a leading cause of in-hospital mortality in the United States and worldwide. Their management depends on a multi-step workflow: risk stratification, contamination discrimination, organism and resistance identification, blood-culture allocation, and antibiotic selection, handled by different clinical roles. Current processes are slow, costly, prone to contamination, and inconsistently supported by computational tools. Existing artificial intelligence (AI) systems address isolated steps, lack transparency, and are rarely validated across sites, limiting clinical adoption. The broad, long-term objective of this project is to develop and validate an end-to-end, physician-centered AI system that improves the accuracy, equity, and safety of bacteremia diagnosis and treatment, thereby reducing sepsis mortality and strengthening antimicrobial stewardship. The project pursues three specific aims. Aim 1 will develop a role-aligned multi-agent architecture whose specialized agents mirror clinical roles “triage, risk stratification, contamination discrimination, organism and antibiogram interpretation, culture allocation, and therapy recommendation” coordinated by a critic and self-refinement agent that detects contradictions, propagates uncertainty, and produces calibrated recommendations. Aim 2 will establish knowledge-grounded, physician-incorporated reasoning through a hybrid graph- and vector-based retrieval framework with patient-state anchoring and temporal weighting that adapts to evolving resistance, with interfaces that let clinicians inject decision rules and enforce safety constraints. Aim 3 will build a framework for multi-site verification, fairness testing, domain-shift adaptation, and privacy-preserving deployment that maintains accuracy across diverse populations while complying with HIPAA. Methods integrate large language model–based agents, reinforcement learning from clinician feedback, retrieval-augmented generation, domain-shift recalibration, and privacy-preserving computation. The system will be evaluated on public and de-identified, IRB-approved clinical data using discrimination, calibration, fairness, consistency, and usability measures, including silent-mode prospective testing, to deliver reliable, explainable, and verifiable decision support for bloodstream infections.
