A Multi-agentic AI Framework For Clinical Decision Intelligence

Cardiovascular diagnosis depends on physiological parameters drawn from several modalities at once, yet most clinical AI still evaluates one of them in isolation. As of the end of 2025 the US FDA had approved 1,430 AI-enabled medical devices — only 27 of them Software as a Medical Device focused on cardiovascular disease. This paper documents a purpose-built multi-agent platform for clinical decision intelligence: a blackboard architecture in which a Diagnosis Agent and an Explainability Agent reason together across chest radiographs, echocardiograms, and laboratory data to produce risk stratifications and early warning signals for heart failure. It reports the architecture, the measured performance across three conditions, and — candidly — where the framework still falls short.

What is clinical decision intelligence?

Clinical decision intelligence is the use of coordinated AI agents to combine evidence from multiple diagnostic modalities into a single, explainable risk assessment. Rather than one model scoring one image, specialist agents contribute findings to a shared workspace and reason toward consensus, with their reasoning exposed to the clinician.

Highlights

• Recall above 0.92 on all three conditions – cardiomegaly, pulmonary edema, and pleural effusion — with a sensitivity-first design intended for early detection
• +3 to +6 points of radiologist agreement when the echocardiography agent was added to chest X-ray alone
• 140 paired studies adjudicated by a radiologist with 24 years of experience as the gold standard
• 1,430 FDA-approved AI devices, 27 for CVD — the gap this architecture is built for

What’s inside the paper

• The blackboard architecture — control unit, shared blackboard, and multi-domain expert agents, and why this pattern suits multi-specialty diagnosis better than a monolithic classifier
• Agent design — the Diagnosis Agent’s CXR and Echo sub-agents, the 65/35 weighting applied to their combined output, and the probability thresholds separating positive, uncertain, and negative calls
• Measured results across three conditions — full precision, recall, and agreement tables for CXR-only versus CXR+Echo, showing where multimodal fusion helped and by how much
• Where it underperformed — low precision on edema and effusion traced to class imbalance, and coarse, low-resolution explainability heat maps across all conditions
• A worked false positive — a single case walked through in full, including a false positive edema detection where the agents disagreed with the radiologist
• The precision–sensitivity trade-off — why the framework was tuned for recall, and why ‘alarm fatigue’ is the next problem to solve
• Next steps — a rule-based logic layer for protocol adherence and a RAG agent for automated clinical reasoning and structured reporting

Why it matters for clinical AI teams

Most published clinical AI results report the metric that flatters the model. This paper reports recall above 0.92 alongside precision as low as 0.66, and states plainly that explainability quality remained low across every condition tested. That combination is the useful part. A sensitivity-first system suits preventive screening — in hospitals, trauma centers, and ambulances — where a missed early heart failure signal costs more than a second look. But precision and explainability are exactly what determine whether clinicians trust a system enough to keep using it, and whether it can survive regulatory scrutiny. The paper is a working record of that trade-off, with the architecture, weighting, thresholds, and failure cases documented well enough to replicate or challenge.

Download the Paper

Multi-agent architecture. Multimodal evidence. Results reported in full, including the ones that need work.