Task-Optimized Machine Learning for High-Accuracy Alzheimer’s Diagnosis from Handwriting Data

Training complex models on Alzheimer’s Disease (AD) datasets is challenging due to the costly process of extracting features from a wide range of patient tasks. Developing high-performance AD detection models that rely on a small number of tasks can help reduce dataset acquisition costs and improve the interpretability of the AD detection model. To address this, we propose a two-stage forward-backward feature selection approach to identify the most relevant tasks and features for predicting AD with high accuracy. We evaluate a range of machine learning methods, including Extreme Gradient Boosting (XGBoost), Random Forest, K-Nearest Neighbors, Support Vector Machine, Multi-Layer Perceptron, and Logistic Regression, to determine the best classification model for feature selection and downstream prediction tasks. Given the limited sample size, we assess model performance using Leave-One-Out-Cross-Validation (LOOCV) to ensure robust results. Our method was compared with multiple state-of-the-art approaches for feature selection. The results of our analysis indicate that combining our proposed methods for feature selection with the XGBoost classifier, using only four tasks, produces a model that is both more interpretable and high-performing compared to other approaches. This suggests focusing on these four tasks, rather than collecting extensive task data from patients, can yield a reliable predictor for diagnosis of AD with an accuracy of 91.37%, 93.94% recall, 89.77% precision, and 91.32% F1 score - surpassing other classification methods. This research represents a significant advancement in the efficiency and reliability of AD diagnosis, improving patient prognosis and offering potential benefits to healthcare systems.