Application of Reinforcement-Based Learning Feature Selection for Early Detection of Alzheimer's Disease Using Speech

Abstract

The early detection of Alzheimer&s;s disease (AD) enables timely intervention methods. Speech-based biomarkers, when combined with advanced machine learning (ML) models and feature selection, provide scalable and promising diagnostic results. In our study, we present a novel speech-based feature selection framework using deep Q-learning (DQN) to identify the most relevant acoustic features for early detection of AD. Our data included over 6000 features extracted using openSMILE open-source software and processed through a graph convolutional network (GCN)–inspired correlation representation and selected using agent-based reinforcement learning. The selected features are clustered using term frequency-inverse document frequency (TF-IDF) and K-means clustering for interpretability. As part of evaluation, we have used five ML models, namely, support vector machines (SVMs), logistic regression (LR), XGBoost, random forest, and feedforward neural network (FNN), which were used to evaluate the DQN model and visualize the features using Venn diagrams. Considering the amount of data, the results show the enhanced performance of each model and their selected features along with the common features chosen by these models which were beneficial to detect AD. The results of 10-fold cross-validation are more significant than the hold-out method. We believe that this interpretable approach will enable clinicians to better understand the model&s;s decision-making process, while helping to build trust in the model. Moreover, the approach&s;s scalable deployment contributes to equable and socially beneficial healthcare, illustrating the broader applications of health innovation using artificial intelligence (AI).