A Formal Multi-Agent Framework for Trustworthy Clinical Decision Support: Architecture, Verification, and Empirical Validation in Pulmonology

Keywords: multi-agent system, clinical decision support, applicative frame, combinatory logic, reference center, pulmonology, RAG, verification

TL;DR: A multi-agent system with a combinatory logic foundation enables provably correct clinical decision support, validated in pulmonology with state-of-the-art results on limited data

Abstract: Modern hybrid artificial intelligence systems in medicine combine neural networks, symbolic components, and retrieval-augmented generation (RAG). However, the absence of a unified formal foundation complicates their verification and explainability. This paper proposes a formal multi-agent model based on applicative frames and multidimensional combinatory logic. Each agent is represented as a tuple (G, A1, . . . , AM, φ, E), the inference process is encoded as a combinator term using the basis {B, I, K}, and decision-making is identified with term reduction to normal form. Theorems of correctness and relative completeness of the reduction procedure are proved. The model is implemented in the Pulmo.Sechenov.AI system for a pulmonology reference center. Retrospective validation on 300 clinical cases (COPD, asthma, pneumonia) demonstrates that on anamnestic data alone, the system achieves sensitivity of 98% for COPD, 96% for asthma, and 78% for pneumonia with 100% specificity, confirming the practical applicability of the formal approach.

Submission Number: 70