MACHINE learning models using routinely available clinical and cognitive data demonstrated moderate-to-good performance in classifying Alzheimer’s disease and cognitive impairment, with accuracy reaching 88.1% across three clinically relevant classification tasks.
The findings highlighted the importance of cognitive screening measures in model performance while emphasising the need for further validation before clinical implementation.
Machine Learning Models Assess Alzheimer’s Classification
Researchers conducted a secondary analysis of data from 761 participants classified as cognitively unimpaired, having mild cognitive impairment due to Alzheimer’s disease, or having probable Alzheimer’s dementia.
The study evaluated five machine learning algorithms: random forest, support vector machine, naïve Bayes, decision tree, and k-nearest neighbours.
Models incorporated 10 routinely available demographic, clinical, cognitive, and neuropsychiatric variables, avoiding reliance on neuroimaging or biomarker assessments.
Three classification tasks were assessed: distinguishing between all three cognitive categories, identifying Alzheimer’s dementia versus non-dementia, and differentiating clinically defined cognitive impairment from unimpaired cognition.
Models were developed using an 80% training dataset with repeated five-fold cross-validation and evaluated on a separate 20% internal test dataset.
Cognitive Screening Improves Classification Performance
For classification across all three cognitive categories, standard models achieved accuracies of 63.6–68.2%, with macro-averaged areas under the receiver operating characteristic curve (AUC) of 0.80–0.84.
When distinguishing Alzheimer’s dementia from participants without dementia, models achieved accuracies of 76.8–78.8% and AUC values of 0.83–0.86.
Performance was highest when distinguishing clinically defined cognitive impairment from unimpaired cognition, with accuracies ranging from 78.1–88.1% and AUC values of 0.77–0.92.
Excluding the Mini-Mental State Examination and Montreal Cognitive Assessment reduced AUC values across evaluated models, particularly for classification across all three cognitive categories.
Additionally, algorithm-specific approaches to addressing class imbalance did not consistently improve performance, while one synthetic oversampling strategy significantly worsened multiclass k-nearest neighbours classification.
External Validation Needed Before Clinical Implementation
Although the models demonstrated promising internal classification performance, researchers cautioned that these findings did not establish clinical utility.
The substantial contribution of cognitive screening measures highlighted their importance when developing machine learning approaches for Alzheimer’s disease classification.
Researchers emphasised that the models should be considered adjunctive classification tools within the study cohort rather than validated diagnostic approaches.
Prospective external validation across geographically diverse populations remains necessary to determine whether these methods can support clinical decision-making in routine healthcare settings.
Dakkali MS et al. Machine-learning models for Alzheimer’s disease classification using routinely available clinical, cognitive, and neuropsychiatric variables. BMC Neurology. 2026. https://doi.org/10.1186/s12883-026-05514-z
