Emulating clinician-assigned diagnostic patterns of Alzheimer's disease in a tertiary neurology hospital using interpretable Bayesian machine learning.

Abstract

Background: Accurate differentiation of Alzheimer's disease (AD) from other causes of cognitive impairment remains a major challenge, particularly in low- and middle-income countries where access to advanced biomarkers such as positron emission tomography or cerebrospinal fluid analysis is limited. Interpretable machine learning models using routinely available clinical and neuroimaging features may support care by standardizing and replicating expert clinical decision-making.

Objective: To evaluate an interpretable Bayesian machine learning framework for analyzing and emulating expert clinician-assigned diagnostic patterns for AD versus non-AD cases.

Methods: We conducted a retrospective cross-sectional analysis of 687 adults evaluated for cognitive impairment at a tertiary neurology hospital in Bangladesh. Demographic, clinical, cognitive, and MRI-based visual rating scales including medial temporal atrophy and Fazekas white-matter hyperintensity scores were used to model the clinical AD diagnoses. Logistic regression was compared with two Bayesian machine learning approaches. Model performance was assessed using discrimination metrics, balanced accuracy, and decision curve analysis. Model interpretability was examined using Shapley Additive Explanations.

Results: Among the evaluated models, Bayesian additive regression trees demonstrated the highest discriminative performance (AUC = 0.903) in reconstructing the clinical diagnostic process. Medial temporal atrophy, Fazekas score, age, Mini-Mental State Examination score, depression, and slow gait emerged as the most influential features driving the clinical AD diagnosis. Bayesian models showed improved balance between sensitivity and specificity compared with logistic regression.

Conclusions: Interpretable Bayesian machine learning models can successfully capture and emulate expert clinical diagnostic patterns. Such models offer a valuable quantitative clinical audit tool to make implicit diagnostic weightings transparent and standardize dementia evaluation.