An Interpretable Random Forest Model for Predicting Anosognosia in Preclinical Alzheimer's Disease: Findings from the A4 Study

Anosognosia, reduced awareness of cognitive difficulties, is common across the Alzheimer's disease (AD) continuum yet remains difficult to detect before objective impairment appears; We developed an explainable machine learning model to predict a Cognitive Function Index (CFI) discrepancy–based operational indicator of anosognosia ΔCFI-defined reduced awareness) in amyloid-positive, cognitively unimpaired older adults from the A4 study. Anosognosia was defined as a negative discrepancy between participant- and study-partner-reported CFI scores (ΔCFI<0); the comparison group (ΔCFI>0) reflects hypernosognosia, over-reporting of difficulty relative to the study partner, rather than intact awareness. Participants with ΔCFI=0 or incomplete data were excluded, yielding a final sample of 1,058. A Random Forest classifier was trained on nine features with SHapley Additive exPlanations (SHAP) values computed to quantify predictor contributions at the individual level. The model achieved 65.88% accuracy, 53.45% sensitivity, 70.59% specificity. Hippocampal volume (HV) was the highest predictor by permutation importance, followed by age and study-partner cohabitation status; FCSRT and PACC contributed meaningfully, while MMSE, GDS, and education showed near-zero importance. SHAP analyses confirmed that lower HV, older age, and lower cognitive scores jointly drove predictions of anosognosia, while misclassifications arose from competing demographic and neuroimaging signals. These findings demonstrate that an accessible combination of clinical and neuroimaging variables can identify meaningful predictors of anosognosia in preclinical AD. HV was the leading predictor, consistent with the established role of medial temporal lobe integrity in self-monitoring. The asymmetry among sensitivity, specificity, and positive predictive value underscores the need to incorporate additional biomarkers, particularly tau pathology, functional connectivity, and longitudinal data.