LEARNING models using routinely available clinical, cognitive, and neuropsychiatric information could support the classification of Alzheimer’s disease (AD) and cognitive impairment, according to new research involving 761 participants
Researchers found that several machine-learning algorithms demonstrated moderate-to-good performance in distinguishing Alzheimer’s dementia, mild cognitive impairment (MCI), and unimpaired cognition. However, the models relied heavily on established cognitive screening assessments, highlighting the importance of conventional clinical measures in AI-assisted classification.
Exploring Accessible Approaches to Alzheimer’s Classification
Alzheimer’s disease is the most common cause of dementia, making accurate identification of cognitive impairment an important priority in clinical practice.
Although machine-learning approaches have demonstrated potential for Alzheimer’s classification, many existing models depend on neuroimaging, specialised biomarkers, or extensive neuropsychological testing, which may not be readily accessible across healthcare settings.
To investigate whether more widely available information could support cognitive-status classification, researchers conducted a secondary analysis of data from 761 participants.
Participants were classified into three clinically defined groups: cognitively unimpaired individuals, those with MCI due to Alzheimer’s disease, and those with probable Alzheimer’s dementia.
Researchers evaluated five machine-learning algorithms: random forest, support vector machine, naïve Bayes, decision tree, and k-nearest neighbours.
The models incorporated 10 routinely available demographic, clinical, cognitive, and neuropsychiatric variables.
Data were divided into an 80% training set and a 20% held-out internal test set, with repeated five-fold cross-validation used during model development.
AI Models Demonstrate Moderate-to-Good Accuracy
Researchers assessed model performance across three clinically relevant classification tasks.
The first involved distinguishing between cognitively unimpaired participants, those with MCI, and those with Alzheimer’s dementia.
For this three-category classification task, standard machine-learning models achieved accuracies ranging from 63.6% to 68.2%.
Macro-averaged areas under the receiver operating characteristic curve (AUCs) ranged from 0.80 to 0.84, indicating moderate-to-good discrimination despite the challenges of distinguishing three cognitive-status categories.
The second task focused on identifying Alzheimer’s dementia compared with participants without dementia.
For this analysis, standard models achieved AUC values of 0.83–0.86, with classification accuracies ranging from 76.8% to 78.8%.
The third task examined whether models could distinguish individuals with clinically defined cognitive impairment from cognitively unimpaired participants.
This analysis produced the strongest overall accuracy results, ranging from 78.1% to 88.1%, with AUC values between 0.77 and 0.92.
These findings suggest that routinely collected information may help distinguish broader categories of cognitive impairment, although separating MCI from Alzheimer’s dementia and unimpaired cognition remains more challenging.
Cognitive Screening Measures Drive Performance
A key finding was the substantial contribution of established cognitive screening tools to machine-learning performance.
Researchers investigated the effects of removing the Mini-Mental State Examination (MMSE) and Montreal Cognitive Assessment (MoCA) from the models.
Excluding these assessments reduced AUC values across the evaluated algorithms, with the greatest reductions observed in the three-category classification task.
This suggests that cognitive screening measures provided important discriminatory information that could not be fully replaced by the remaining demographic, clinical, and neuropsychiatric variables.
The researchers also explored whether techniques designed to address imbalances between diagnostic groups could improve performance.
However, algorithm-specific imbalance-adjustment strategies did not consistently produce better results.
Notably, the Synthetic Minority Over-sampling Technique for Nominal and Continuous variables (SMOTENC) significantly reduced the performance of the multiclass k-nearest neighbours model.
These findings demonstrate that more complex data-processing approaches do not necessarily improve classification accuracy.
Implications for AI-Assisted Dementia Assessment
The findings highlight the potential for machine-learning models to support cognitive-status classification using information already collected during routine clinical assessments.
Such approaches could eventually contribute to clinical decision-support systems, particularly where access to advanced diagnostic investigations is limited.
However, the researchers cautioned that the models demonstrated classification performance within an existing cohort rather than established clinical utility.
Importantly, the study evaluated clinically defined cognitive-status categories and did not establish that the models could independently confirm Alzheimer’s pathology.
The reliance on MMSE and MoCA also raises questions about how much additional value machine learning provides beyond conventional cognitive screening.
Furthermore, performance was assessed using an internal held-out test set, meaning the findings have not yet been independently validated in other populations or healthcare settings.
The researchers concluded that machine-learning models using routinely available variables demonstrated moderate-to-good discriminatory performance, but emphasised that prospective external validation across geographically diverse populations is necessary before clinical implementation.
