Implementing a Decision Support Mechanism for Assessing the Probability of Future Alzheimer's Disease Based on the Medical and Demographic Information of Patients

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder that gradually impairs memory and cognitive function. Early detection of AD is vital, as it allows for timely clinical intervention and potentially slows disease progression. In recent years, machine learning (ML) has emerged as a powerful tool in medical diagnostics, offering support in identifying early indicators of AD. This study investigates the performance of several ML models in classifying early-stage AD using clinical data. Seven ML algorithms were evaluated: Categorical Boosting (CatBoost) and eXtreme Gradient Boosting (XGBoost), which are both types of Gradient Boosting Machine (GBM), Logistic Regression (LR), Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Random Forest (RF), and Multi-Layer Perceptron (MLP) Neural Network (ΝΝ). These models were compared based on their accuracy, precision, recall, F1-score, specificity, Cohen’s kappa, and Receiver Operating Characteristic (ROC) Curve-Area Under the Curve (AUC), key metrics for assessing classification performance. Among all models tested, CatBoost, which is a GBM-based classifier, achieved the highest overall performance, with an accuracy of 96.74%, precision of 96.86%, recall of 94.48%, F1-score of 95.65%, specificity of 98.13%, Cohen’s kappa of 93.05%, and ROC-AUC of 96.30%. RF also performed strongly, with 95.81% accuracy and 94.34% F1-score. XGBoost, another GBM classifier, showed comparable effectiveness with 95.58% accuracy and 94.08% F1-score. In contrast, traditional classifiers like LR and SVM delivered lower accuracy (84.42% and 83.72%, respectively) and F1-scores near 77%. MLP outperformed these traditional methods with 85.58% accuracy and 78.91% F1-score. KNN had the weakest performance, particularly in recall (49.08%) and F1-score (62.26%), despite relatively high precision (85.11%) and specificity (94.76%). Overall, the findings affirm that the models CatBoost and RF provide the most reliable and generalizable solutions for the classification task in this study, exhibiting high discriminative capability and stability across all evaluated metrics.