Machine-learning models for Alzheimer’s disease classification using routinely available clinical, cognitive, and neuropsychiatric variables

Alzheimer’s disease (AD) is the most common cause of dementia. Although machine-learning models have shown promise for AD classification, many approaches rely on neuroimaging, biomarker assessments, or extensive neuropsychological evaluations that are not routinely available in clinical practice. This study evaluated multiple machine-learning algorithms using routinely available demographic, clinical, cognitive, and neuropsychiatric variables for classification of clinically defined cognitive-status categories. A secondary analysis was conducted using data from 761 participants classified as cognitively unimpaired (CU), mild cognitive impairment (MCI) due to AD, or probable AD dementia. Ten routinely available demographic, clinical, cognitive, and neuropsychiatric variables were used to develop five machine-learning models: random forest, support vector machine, naïve Bayes, decision tree, and k-nearest neighbors. Three clinically relevant classification tasks were evaluated: multiclass classification of CU, MCI, and AD dementia; discrimination of AD dementia from non-dementia participants; and discrimination of participants with clinically defined cognitive impairment from CU participants. Models were developed using a stratified 80% training set with repeated five-fold cross-validation and evaluated once on a 20% held-out internal test set. On the held-out internal test set, accuracy among the standard models for multiclass classification ranged from 63.6% to 68.2%, with macro-averaged area under the receiver operating characteristic curve (macro-AUC) values of 0.80–0.84. For discrimination of AD dementia from non-dementia participants, standard models achieved AUC values of 0.83–0.86 and accuracies of 76.8–78.8%. Standard models distinguishing participants with clinically defined cognitive impairment from CU participants achieved AUC values of 0.77–0.92 and accuracies of 78.1–88.1%. Algorithm-specific imbalance-adjustment strategies did not consistently improve classification performance, and the Synthetic Minority Over-sampling Technique for Nominal and Continuous (SMOTENC) significantly reduced multiclass k-nearest neighbors’ performance. Exclusion of the Mini-Mental State Examination (MMSE) and Montreal Cognitive Assessment (MoCA) reduced AUC across the evaluated models, with the largest reductions observed in multiclass classification. The models demonstrated moderate-to-good internal discriminative performance across three clinically relevant cognitive-status classification tasks. Model performance depended substantially on cognitive screening measures, particularly MMSE and MoCA. These findings should be interpreted as adjunctive classification performance within this cohort rather than evidence of established clinical utility. External validation in prospective, geographically diverse cohorts is required before clinical implementation.