Machine Learning Classifies Alzheimer’s Stages Using Routine Clinical and Cognitive Data

Machine-learning models trained strictly on routinely available demographic, clinical, cognitive, and neuropsychiatric variables demonstrated moderate-to-good internal discriminative performance when categorizing patients across clinically defined cognitive-status groups, according to a secondary analysis published in BMC Neurology. Researchers evaluated multiple algorithmic approaches designed to categorize cognitive stages without depending on advanced neuroimaging, invasive biomarker assessments, or extensive neuropsychological testing batteries. By focusing entirely on accessible patient information, the investigation explored whether standard computational models could differentiate between cognitively unimpaired individuals, patients experiencing mild cognitive impairment due to Alzheimer’s disease, and those diagnosed with probable Alzheimer’s disease dementia, achieving macro-averaged area under the curve values up to 0.84 on internal testing.

Although computational algorithms and machine-learning models have demonstrated considerable promise for Alzheimer’s disease classification in academic literature, many existing technical approaches rely heavily on neuroimaging modalities, specialized biomarker assessments, or lengthy neuropsychological evaluations. Consequently, the research team sought to evaluate the discriminative capability of multiple machine-learning algorithms using only routinely available demographic, clinical, cognitive, and neuropsychiatric variables that clinicians can readily collect in standard clinical workflows.

The investigation was designed as a secondary analysis utilizing a clinical dataset of 761 participants who were clinically categorized into three distinct diagnostic cohorts: cognitively unimpaired (CU), mild cognitive impairment (MCI) attributable to Alzheimer’s disease, and probable Alzheimer’s disease dementia. To construct predictive frameworks, the investigators identified ten routinely available variables spanning demographic characteristics, clinical evaluations, cognitive assessments, and neuropsychiatric measurements. These routine data inputs were then utilized to develop and benchmark five well-established machine-learning models, specifically random forest, support vector machine, naïve Bayes, decision tree, and k-nearest neighbors algorithms, testing their capacity to recognize disease patterns from routine measures alone.

To comprehensively examine algorithmic capability, the authors structured three distinct, clinically relevant classification tasks reflecting common diagnostic questions in neurology. The first challenge required multiclass classification across all three patient categories, simultaneously separating cognitively unimpaired participants, individuals with mild cognitive impairment due to Alzheimer’s disease, and patients with probable Alzheimer’s disease dementia. The second task evaluated binary discrimination, isolating participants with probable Alzheimer’s disease dementia from non-dementia participants. The third task assessed the models’ ability to distinguish any participant presenting with clinically defined cognitive impairment from those categorized as cognitively unimpaired. For model development and validation, the researchers partitioned the cohort into a stratified 80 percent training set utilizing repeated five-fold cross-validation, and subsequently evaluated final performance once on a held-out 20 percent internal test set.

Evaluation on the held-out internal test set revealed moderate-to-good discriminative accuracy across all three diagnostic tasks for standard models. In the multiclass classification challenge, accuracy among standard models ranged from 63.6 percent to 68.2 percent, accompanied by macro-averaged area under the receiver operating characteristic curve (macro-AUC) values between 0.80 and 0.84. In the binary task discriminating probable Alzheimer’s disease dementia from non-dementia participants, standard algorithms achieved AUC values ranging from 0.83 to 0.86 and overall accuracies between 76.8 percent and 78.8 percent. For distinguishing individuals with clinically defined cognitive impairment from cognitively unimpaired participants, standard models attained AUC values spanning 0.77 to 0.92 alongside classification accuracies ranging from 78.1 percent to 88.1 percent.

The authors also investigated technical modifications to address data distribution, discovering that algorithm-specific imbalance-adjustment strategies did not consistently enhance classification performance across the models. In particular, applying the Synthetic Minority Over-sampling Technique for Nominal and Continuous (SMOTENC) significantly impaired the multiclass classification performance of the k-nearest neighbors algorithm. Furthermore, feature ablation experiments demonstrated that model performance depended substantially on routine cognitive screening tools. When investigators excluded the Mini-Mental State Examination (MMSE) and the Montreal Cognitive Assessment (MoCA) from the feature set, AUC values dropped across all evaluated machine-learning architectures, with the most pronounced performance reductions observed during multiclass classification.

In their conclusions, the investigators stressed that these findings demonstrate adjunctive classification performance within this specific retrospective cohort rather than evidence of established clinical utility. Because the predictive models rely substantially on cognitive screening measures such as the MMSE and MoCA, their utility reflects internal classification properties rather than standalone diagnostic capability. The researchers cautioned that while routine variables offer practical advantages over neuroimaging and biomarker assays, external validation in prospective, geographically diverse patient cohorts is strictly required before any of these machine-learning models can be considered for real-world clinical implementation.

Subject of Research: Machine learning for Alzheimer's disease classification using routine clinical variables

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

Article References: Dakkali, M. S., Azarbad, R., Sotoudeheian, M., Dakkali, N., Gheisipour, D., Hashemzehi, Z., & Arish, M. (2026). Machine-learning models for Alzheimer’s disease classification using routinely available clinical, cognitive, and neuropsychiatric variables. BMC Neurology. https://doi.org/10.1186/s12883-026-05514-z

Image Credits: AI Generated

DOI: 10.1186/s12883-026-05514-z

Keywords: Alzheimer's disease, Mild cognitive impairment, Cognitively unimpaired, Machine learning, Random forest, Support vector machine, Naïve Bayes, Decision tree, K-nearest neighbors, Cognitive screening, Mini-Mental State Examination, Montreal Cognitive Assessment, Multiclass classification, SMOTENC, Receiver operating characteristic, Cross-validation, Clinical neurology, Neuropsychiatric assessment