Statistics, Not Scale: Modular Medical Dialogue with Bayesian Belief Engine

2026-04-23 · Source: cs.LG updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Data Science & Analytics · Depth: Expert, extended

Summary

The BMBE (Bayesian Medical Belief Engine) framework introduces a modular diagnostic dialogue system that strictly separates natural language communication from probabilistic reasoning in medical AI. Unlike autonomous large language models (LLMs) that conflate these capabilities, BMBE uses an LLM solely as a "sensor" to parse patient utterances into structured evidence and verbalize questions. All diagnostic inference, including Bayesian belief updates, question selection via expected information gain, and stopping rules, resides in a deterministic, auditable Bayesian engine. This architecture ensures privacy by design, as patient data never enters the LLM, and allows the statistical backend to be replaced per target population without retraining. BMBE demonstrates superior performance, cost-effectiveness, and robustness compared to frontier standalone LLMs across empirical and LLM-generated knowledge bases, offering calibrated selective diagnosis with a continuously adjustable accuracy–coverage tradeoff.

Key takeaway

For AI Architects designing medical diagnostic systems, you should prioritize architectural separation of language and reasoning. Implement a modular design where LLMs handle only communication, while a deterministic Bayesian engine performs all probabilistic inference. This approach will deliver higher diagnostic accuracy, better cost-effectiveness, and enhanced privacy and auditability compared to monolithic LLM solutions, allowing you to tune the system's accuracy-coverage tradeoff to match specific clinical risk tolerances without retraining.

Key insights

Strictly separating language processing from probabilistic reasoning in medical AI yields superior, auditable, and private diagnostic systems.

Principles

• Diagnostic reasoning is fundamentally probabilistic inference.
• LLMs excel at language, not calibrated probabilistic reasoning.
• Architectural separation enhances privacy and adaptability.

Method

BMBE decomposes medical dialogue into an LLM-based language interface for parsing and verbalization, and a Bayesian reasoning engine for all diagnostic inference, communicating via structured evidence triples.

In practice

• Use an inexpensive LLM as a sensor for structured data extraction.
• Implement a Bayesian engine for auditable, calibrated diagnostic inference.
• Adjust the confidence threshold τ for desired accuracy-coverage tradeoff.

Topics

• Bayesian Medical Belief Engine
• Diagnostic Dialogue Systems
• Large Language Models
• Probabilistic Reasoning
• Medical AI Architecture

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Scientist, Research Scientist, AI Architect

Related on AIssential

• Active Evidence-Seeking and Diagnostic Reasoning in Large Language Models for Clinical Decision Support — LLMs exhibit a substantial performance gap in active, multi-turn clinical diagnosis compared to static, full-context evaluations.
• SymptomWise: A Deterministic Reasoning Layer for Reliable and Efficient AI Systems — SymptomWise improves AI diagnostic reliability by separating language understanding from deterministic, knowledge-driven reasoning.
• TRIAGE: Dialectical Reasoning for Explainable Risk Prediction on Irregularly Sampled Medical Time Series with LLMs — TRIAGE uses dialectical LLM reasoning to provide calibrated, continuous risk scores and interpretable rationales for medical time series.
• MedRLM: Recursive Multimodal Health Intelligence for Long-Context Clinical Reasoning, Sensor-Guided Screening, Evidence-Grounded Decision Support, and Community-to-Tertiary Referral Optimization — MedRLM recursively processes heterogeneous patient data as an external clinical environment for auditable, workflow-aware decision support.
• Language Models as Interfaces, Not Oracles: A Hybrid LLM-ML System for Pediatric Appendicitis — Hybrid AI systems improve clinical decision support by using LLMs for feature extraction and validated ML models for reliable, auditable prediction.
• Compositional Reasoning Depth Predicts Clinical AI Failure: Empirical Evidence Consistent with Transformer Compositionality Limits in Electronic Health Record Question Answering — LLM accuracy in EHR QA systematically declines with increased compositional reasoning depth, quantified by "hop count."

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.