Classifications based on response times for detecting early-stage Alzheimer's disease
This work addresses the problem of early diagnosis for Alzheimer's patients, offering a more accurate and potentially generalizable tool, though it is incremental as it builds on existing task-based datasets.
The paper tackles early-stage Alzheimer's disease detection by using response times from handwriting and drawing tasks, achieving a classification error rate that is two to four times lower than state-of-the-art methods.
Introduction- This paper mainly describes a way to detect with high accuracy patients with early-stage Alzheimer's disease (ES-AD) versus healthy control (HC) subjects, from datasets built with handwriting and drawing task records. Method- The proposed approach uses subject's response times. An optimal subset of tasks is first selected with a "Support Vector Machine" (SVM) associated with a grid search. Mixtures of Gaussian distributions defined in the space of task durations are then used to reproduce and explain the results of the SVM. Finally, a surprisingly simple and efficient ad hoc classification algorithm is deduced from the Gaussian mixtures. Results- The solution presented in this paper makes two or even four times fewer errors than the best results of the state of the art concerning the classification HC/ES-AD from handwriting and drawing tasks. Discussion- The best SVM learning model reaches a high accuracy for this classification but its learning capacity is too large to ensure a low overfitting risk regarding the small size of the dataset. The proposed ad hoc classification algorithm only requires to optimize three real-parameters. It should therefore benefit from a good generalization ability.