Handwriting strokes as biomarkers for Alzheimer's disease prediction: A novel machine learning approach

In recent years, machine learning-based handwriting analysis has emerged as a valuable tool for supporting the early diagnosis of Alzheimer's disease and predicting its progression. Traditional approaches represent handwriting tasks using a single feature vector, where each feature is computed as the mean over elementary handwriting traits or strokes. We propose a novel approach that analyzes each stroke individually, preserving fine-grained movement information that is critical for detecting subtle handwriting changes that may indicate cognitive decline. We evaluated this method on 34 handwriting tasks collected from 174 participants, extracting dynamic and static features from both on-paper and in-air movements. Using a machine learning framework including classification strategies, feature selection techniques, and ensemble methods like ranking-based and stacking approaches, we were able to effectively model stroke-level variations. The ranking-based ensemble achieved the highest accuracy of 80.18% using all features while stacking performed best for in-air movements with 76.67% accuracy. Feature importance analysis through SHAP revealed that certain tasks, particularly sentence writing under dictation, were consistently more predictive. The experimental results demonstrate the effectiveness of our stroke-level analysis approach, which outperformed aggregated statistical methods on 24 out of 34 handwriting tasks, validating the diagnostic value of examining individual movement patterns.