Learning Physician's Treatment for Alzheimer's Disease based on Electronic Health Records and Reinforcement Learning

Kritib Bhattarai, Sivaraman Rajaganapathy, Trisha Das, Yejin Kim, Yongbin Chen, Qiying Dai, Xiaoyang Li, Xiaoqian Jiang, Nansu Zong

Abstract: Alzheimer's Disease (AD) is a progressive neurological disorder that necessitates physicians with sophisticated skills and knowledge to effectively care for AD patients. In this study, we adopted reinforcement learning (RL) to learn a physician's treatment plan for AD by utilizing Electronic Health Records (EHR). By defining states, actions, and rewards, we modeled the data of 1,736 patients from the Alzheimer's Disease Neuroimaging Initiative (ADNI) into an RL problem. We evaluated the RL-based learning model across four patient cohorts: the entire dataset, AD-only data, AD-hypertension data, and AD-hypertension-depression data. The RL learning models demonstrated promising outcomes in generating an optimal physician policy, which represents the treatment plan, in comparison to the clinician policy obtained from transitional probability. For instance, the q-learning-based policy achieved a score of -2.48, whereas the clinician policy scored -3.57. This research highlights the potential of RL-based treatment learning to enhance the management of Alzheimer's Disease.