Explainable AI for Early Dementia Diagnosis: A SHAP-Enhanced Gradient Boosting Model on Longitudinal Neuroimaging Data

Early and accurate diagnosis of neurodegenerative disorders, particularly Alzheimer's disease and related dementias, remains a critical healthcare priority as disease-modifying therapies emerge. While deep learning models trained on structural neuroimaging data have demonstrated high diagnostic accuracy, their inherent black-box nature poses substantial barriers to clinical translation. In this study, we propose an explainable machine learning framework using an optimized Gradient Boosting Decision Tree (LightGBM) architecture combined with Tree-based SHapley Additive exPlanations (SHAP) to detect early cognitive impairment and predict progression to dementia using longitudinal magnetic resonance imaging (MRI) derived volumetric metrics. Leveraging longitudinal neuroimaging cohorts, our model extracts annualized rates of atrophy across subcortical structures alongside baseline volumetric features. The optimized ensemble achieved a classification accuracy of 89.4% and an area under the receiver operating characteristic curve (AUC-ROC) of 0.942 in distinguishing stable mild cognitive impairment from progressive neurodegeneration, outperforming conventional support vector machines and random forest baselines. Global SHAP analyses revealed that annualized ventricular expansion rate, bilateral hippocampal volume reduction, and entorhinal cortical thinning constituted the predominant drivers of model decision boundaries. Furthermore, local patient-level explanations successfully delineated individual risk profiles, offering transparent, clinically congruent rationales for each diagnostic projection. This framework bridges high-capacity predictive analytics with verifiable interpretability, demonstrating the viability of explainable data-driven tools for longitudinal clinical decision support.