Writing the Signs: An Explainable Machine Learning Approach for Alzheimer's Disease Classification from Handwriting

Alzheimer's disease is a global health challenge, emphasizing the need for early detection to enable timely intervention and improve outcomes. This study analyzes handwriting data from individuals with and without Alzheimer's to identify predictive features across copying, graphic and memory-based tasks. Machine learning models, including Random Forest, Bootstrap Aggregating (Bagging), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Adaptive Boosting (AdaBoost) and Gradient Boosting, were applied to classify patients, with SHapley Additive exPlanations (SHAP) enhancing model interpretability. Time-related features were crucial in copying and graphic tasks, reflecting cognitive processing speed, while pressure-related features were significant in memory tasks, indicating recall confidence. Simpler graphic tasks showed strong discriminatory power, aiding early detection. Performance metrics demonstrated model effectiveness: For memory tasks, Random Forest achieved the highest accuracy ($0.840\pm 0.038$), while Bagged SVC was the lowest ($0.617\pm 0.046$). Copying tasks recorded a peak accuracy of $0.804\pm 0.075$ with Gradient Boost and a low of $0.566\pm 0.032$ for Bagged SVC. Graphic tasks reached $0.799\pm 0.041$ with Gradient Boost and 0.643 ± 0.071 with AdaBoost. For all tasks combined, Random Forest excelled ($0.854\pm 0.033$), while Gradient Boost performed worst ($0.598\pm 0.151$). These results highlight handwriting analysis's potential in Alzheimer's detection.