Efficient Alzheimer’s Diagnosis Through Sequential Decision-Making with Reinforcement Learning

french Alzheimer’s disease (AD) is the leading cause of dementia worldwide, with diagnosis often requiring a combination of cognitive assessments, neuroimaging, and biomarker analysis. These procedures, while effective, are resource-intensive, invasive, and time-consuming. This paper investigates reinforcement learning (RL) as a means of optimizing the diagnostic process, aiming to reduce cost and patient burden without compromising accuracy. We formulate AD diagnosis as a sequential decision-making problem and evaluate three approaches: a Deep Q-Network (DQN), a hybrid Proximal Policy Optimization with XGBoost classifier (PPO–XGB), and a hybrid Long Short-Term Memory network with Random Forest classifier (LSTM–RF). Using a subset of the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset, the DQN achieved the best balance between accuracy (89.2%) and efficiency, favoring high-yield cognitive assessments over costly modalities. PPO–XGB demonstrated competitive accuracy (86.93%) with fewer tests (5.06 on average), while LSTM–RF performed strongly in temporal pattern recognition but with lower overall accuracy (80.9%). Results highlight RL’s potential to serve as a cost-aware, adaptive controller for diagnostic test selection, offering a scalable framework for resource-efficient clinical decision-making in Alzheimer’s disease.