Clinical Reasoning over Tabular Data and Text with Bayesian Networks

TL;DR

This paper tackles enabling clinical reasoning over both structured tabular data and unstructured clinical text by extending Bayesian networks with neural text representations. It introduces two architectures: BN-gen-text, a generative approach modeling the text as P(T|D0,D1,S0,S1,S2) with Gaussian mixtures over BioLORD embeddings, and BN-discr-text, a discriminative approach with neural text classifiers for each parent configuration, allowing joint inference of D0 and D1 given B,S,T. In a synthetic primary-care pneumonia use-case, both approaches improve posterior diagnostic probabilities over a text-free BN, with BN-discr-text approaching the BN++ upper bound when text and symptoms are observed; ablation confirms the critical role of text-linked edges. The results demonstrate the value of preserving raw text for clinical decision support and illustrate how neuro-symbolic BN integration can enhance interpretability and robustness to missing data in medical reasoning.

Abstract

Bayesian networks are well-suited for clinical reasoning on tabular data, but are less compatible with natural language data, for which neural networks provide a successful framework. This paper compares and discusses strategies to augment Bayesian networks with neural text representations, both in a generative and discriminative manner. This is illustrated with simulation results for a primary care use case (diagnosis of pneumonia) and discussed in a broader clinical context.

Figures (3)

• Figure 1: Key steps in generating the artificial dataset, where each sample consists of both tabular variables and corresponding clinical text descriptions. With help of an expert, we define a Bayesian network (BN) simulating the pneumonia use case (step 1). We sample the tabular variables (background, diagnoses and symptoms) from the distribution defined by this BN (step 2), prompting a large language model (GPT3.5 instructGPT) to generate realistic but fictitious consultation notes based on the sampled symptoms (step 3). We repeat steps 1 to 3 to generate 4000 training samples and 1000 test samples. Finally, to induce property (ii) of realistic medical data (see Section \ref{['sec:data_gen']}), we remove two symptoms, $fever$ and $pain$, from the tabular portion of the data, ensuring they are never encoded and only observed through the text (step 4a). From now on, when we talk about symptoms, we take this to mean the symptoms $dysp$, $cough$, $nasal$, unless explicitly stated otherwise. For the training set only, we partially mask out the remaining symptoms (step 4b) and the text (step 4c) in a complementary subset of the training samples. Each sample now represents a fictional patient encounter, consisting of one background feature ($season$), two diagnoses ($pneu$ and $inf$), three symptoms ($dysp$, $cough$ and $nasal$, partially unobserved) and a textual description ($text$, partially unobserved). The text contains additional context on the three encoded symptoms, as well as describing two additional unencoded symptoms $fever$ and $pain$.
• Figure 2: Schematic depiction of all models. The top row presents our baselines BN, BN++ and FF-discr-text. The bottom row shows BN-gen-text and BN-discr-text, two variants of a Bayesian network augmented with text representations.
• Figure A1: Conditional probability tables (CPTs) for all parent-child relations in the ground truth Bayesian network, as defined by an expert general practitioner.