Diagnostic probability change across follow-up: a Bayesian EHR cohort study in psychiatry

Diagnosis is often interpreted through a Bayesian lens as a point-in-time update from prior to posterior probability after new clinical evidence. In routine care, however, diagnostic evidence may accumulate across repeated encounters. This study applied a Bayesian EHR-based framework to estimate diagnostic probability change from the initial to final observed visit across major psychiatric diagnostic categories.

We analyzed electronic health records from 17,057 psychiatric outpatients at Changhua Christian Hospital from 2016 to 2024. A multivariate Bayesian generalized linear mixed model was used to estimate the final-versus-initial change in diagnostic recording probability across 10 major ICD-10 categories, controlling for demographics, visit frequency, and within-patient correlations.

The final-versus-initial odds-scale effect varied substantially across diagnostic categories. The largest effects were observed for F7x (Intellectual developmental disorder; Odds Ratio [OR]=7.4, 95 % Credible Interval [CrI] 5.2-10.8) and F9x (Behavioural and emotional disorders with onset in childhood/adolescence; OR=5.7, 95 % CrI 4.5-7.3), whereas F4x showed the smallest effect (OR=1.2, 95 % CrI 1.1-1.4). When translated to the probability scale under specified baseline probabilities, these odds-scale differences corresponded to larger probability changes for diagnoses requiring longitudinal developmental, functional, or collateral information.

This framework provides a probability-scale approach for characterizing diagnostic updating across follow-up in routine EHR data. It may help identify diagnostic categories in which longitudinal evidence accumulation is especially important and inform the design of structured follow-up and assessment workflows.