← News

ResearchSources: arXiv · 19 Nov 2026

ProbMedTOD pairs small language models with Bayesian networks for patient history taking

Researchers describe a task-oriented dialogue system that synthesizes a clinical Bayesian network from a few clinical notes and uses it to guide multi-turn diagnostic questioning with small language models.

In plain words

The described system, ProbMedTOD, builds a probabilistic model of diagnoses from a small set of clinical notes, then lets a small language model conduct patient interviews. The model tracks how likely each diagnosis is as questions are asked, and chooses questions that are most informative. The authors report it beat larger-model and retrieval-based baselines on two datasets.

What was announced

The arXiv abstract frames patient history taking as an iterative process in which each question updates beliefs over candidate diagnoses. It argues that privacy, cost, and latency constraints make small language models a preferable backbone, but that these models lack explicit mechanisms for maintaining and updating diagnostic uncertainty over time. ProbMedTOD addresses this by automatically synthesizing a clinical Bayesian network in three stages: ontology building, structure prediction, and parameter estimation, avoiding the expert-curated structure, dependencies, and parameters that make such networks hard to construct.

In operation, a BayesNet Agent maintains real-time posterior distributions over diagnoses while a Policy Agent selects diagnostically informative questions during multi-turn dialogue. On the DDXPlus and MIMIC-IV Notes datasets, the authors report that ProbMedTOD outperformed LLM-only and retrieval-based baselines, with the strongest gains among smaller models. The abstract presents these as the authors' own experimental results, and the excerpt does not include the full paper's methods or detailed figures.

What it does not show

  • Evidence is limited to the paper's abstract; the full paper's methods, detailed results, and peer-review status are not available in the supplied passages.
  • Performance claims are the authors' own reports on DDXPlus and MIMIC-IV Notes and have not been independently confirmed.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.