mobayes-clinical-decision

name: mobayes-clinical-decision description: Modular Bayesian Framework (MoBayes) for separating reasoning from language in clinical decision support. Addresses the architectural limitation of LLMs conflating next-token prediction with probabilistic medical reasoning. Activation: clinical reasoning, medical LLM, Bayesian clinical, decision support, MoBayes, 临床决策.

MoBayes: Modular Bayesian Clinical Decision Support

Overview

MoBayes (arXiv:2604.20022) proposes a modular Bayesian framework that separates language generation from clinical reasoning in conversational healthcare systems. The core insight: LLMs conflate token prediction with probabilistic decision making, creating an architectural gap in clinical settings where accuracy matters more than fluency.

Core Principles

The Separation Problem

LLMs trained for conversational clinical support mix two distinct functions:

Language generation - producing fluent, contextually appropriate responses
Probabilistic reasoning - making accurate clinical inferences from symptoms and test data

This conflation means: (a) reasoning quality depends on language model training data, (b) fluent-sounding responses may mask poor reasoning, (c) no dedicated uncertainty quantification mechanism exists.

MoBayes Architecture

User Query → [Language Module] → Clinical State → [Bayesian Reasoning Module] → [Language Module] → Response

Language Module: LLM for interface — understanding queries, formatting outputs
Bayesian Reasoning Module: Dedicated probabilistic model for clinical inference
- Encodes medical knowledge as conditional probability distributions
- Produces calibrated uncertainty estimates
- Independent of language model training distribution

Implementation Pattern

class MoBayesClinical:
    def __init__(self, language_model, bayesian_network):
        self.llm = language_model
        self.bayes = bayesian_network
    
    def diagnose(self, patient_query):
        # Step 1: Extract clinical features via LLM
        features = self.llm.extract_clinical_features(patient_query)
        
        # Step 2: Reason via Bayesian network (separate from language)
        diagnosis, confidence, differential = self.bayes.infer(features)
        
        # Step 3: Generate response via LLM with reasoning results
        return self.llm.generate_response(diagnosis, confidence, differential)

Key Advantages

Calibrated uncertainty: Bayesian module provides proper probability estimates, not token-level logits
Knowledge updatability: Medical knowledge can be updated without retraining the LLM
Auditability: Reasoning path is explicit and inspectable (vs. opaque LLM generation)
Safety guarantees: Hard constraints can be encoded in the Bayesian module
Domain specialization: Medical reasoning is decoupled from general language capabilities

When to Use

Clinical decision support systems using LLMs
Medical triage chatbots requiring calibrated confidence
Healthcare AI systems needing auditability
Any domain where reasoning accuracy must be separated from language fluency

Pitfalls

LLM feature extraction must be validated against ground truth clinical annotations
Bayesian network requires careful elicitation of conditional probabilities
The interface between modules must handle uncertainty propagation correctly
Not suitable for open-ended medical conversation without reasoning constraints