[Paper Review] A Bayesian Approach for Medical Inquiry and Disease Inference in Automated Differential Diagnosis

Hong Guan, Chitta Baral|arXiv (Cornell University)|Oct 15, 2021

Machine Learning in Healthcare21 references4 citations

TL;DR

This paper proposes a Bayesian framework for automated differential diagnosis that integrates Bayesian inference for disease prediction and Bayesian experimental design for active medical inquiry. By leveraging the QMR belief network and extending multi-step look-ahead search, the approach achieves state-of-the-art performance on simulated datasets and competitive results on dialogue datasets, with advantages in interpretability, scalability, and adaptability without retraining.

ABSTRACT

We propose a Bayesian approach for both medical inquiry and disease inference, the two major phases in differential diagnosis. Unlike previous work that simulates data from given probabilities and uses ML algorithms on them, we directly use the Quick Medical Reference (QMR) belief network, and apply Bayesian inference in the inference phase and Bayesian experimental design in the inquiry phase. Moreover, we improve the inquiry phase by extending the Bayesian experimental design framework from one-step search to multi-step search. Our approach has some practical advantages as it is interpretable, free of costly training, and able to adapt to new changes without any additional effort. Our experiments show that our approach achieves new state-of-the-art results on two simulated datasets, SymCAT and HPO, and competitive results on two diagnosis dialogue datasets, Muzhi and Dxy.

Motivation & Objective

• To develop a scalable, interpretable, and adaptive system for automated differential diagnosis that supports both active inquiry and final disease inference.
• To overcome limitations of existing ML and RL-based methods, such as high computational cost, poor adaptability to knowledge base changes, and lack of interpretability.
• To extend Bayesian experimental design from one-step to multi-step look-ahead search for improved inquiry efficiency and accuracy.
• To enable clinicians to manually intervene in the inquiry process while maintaining consistent inference through a fixed graphical model.
• To achieve high diagnostic accuracy with minimal inquiry steps, even under partial observation, by using exact Bayesian probability computation.

Proposed method

• The system uses the Quick Medical Reference (QMR) belief network as a fixed, interpretable graphical model encoding disease-finding dependencies.
• For disease inference, it performs exact Bayesian inference using the QMR network to compute posterior probabilities of diseases given observed findings.
• For medical inquiry, it applies Bayesian experimental design to select the next most informative finding to query, maximizing expected information gain.
• The inquiry phase is enhanced via multi-step look-ahead search, evaluating potential future inquiry paths to select the optimal next question.
• The framework decouples inquiry and inference, allowing clinicians to manually override or modify findings at any step without retraining.
• The approach avoids training by directly using the QMR network and exact probability computation, enabling instant adaptation to new diseases or findings.

Experimental results

Research questions

• RQ1 Can a Bayesian approach with multi-step look-ahead inquiry outperform one-step Bayesian experimental design in reducing diagnostic inquiry steps?
• RQ2 Can a non-training-based system achieve state-of-the-art accuracy on simulated medical diagnosis benchmarks while remaining interpretable and adaptable?
• RQ3 How does the integration of exact Bayesian inference with Bayesian experimental design affect diagnostic accuracy and efficiency in partially observed scenarios?
• RQ4 To what extent can clinicians manually intervene in the inquiry process without degrading system performance?
• RQ5 Does the use of a fixed graphical model (QMR) enable better interpretability and faster adaptation than end-to-end trained models?

Key findings

• The proposed method achieves new state-of-the-art accuracy on two simulated datasets, SymCAT and HPO, with fewer inquiry steps than prior methods.
• On the Muzhi and Dxy dialogue datasets, the approach achieves competitive performance, demonstrating robustness across different data types.
• The multi-step look-ahead search significantly improves diagnostic accuracy and reduces the number of required inquiry steps, approaching the performance of full-observation inference.
• The system maintains high accuracy and efficiency without retraining when new diseases or findings are added to the QMR network.
• The interpretability of results is enhanced by allowing clinicians to inspect the QMR belief network structure to understand diagnostic reasoning.
• The approach is computationally efficient for moderate-sized disease sets and avoids the training overhead and instability of reinforcement learning or deep generative models.