Towards cost-effective cognitive impairment diagnosis systems by emulating doctors' reasoning with deep reinforcement learning.

backgroundCognitive impairment diagnosis in aging adults is of great significance but faces challenges like high costs and complex clinical reasoning. Traditional clinical procedures often require comprehensive neuropsychological assessments involving numerous cognitive tests, interviews, or biomarker analyses. While these approaches maximize diagnostic accuracy, they are inherently time-consuming, expensive, and place a heavy burden on both patients and clinicians.

methodsThis study used data from the CHARLS-HCAP program, which included 1,478 participants classified as dementia. We propose a new framework for cost-effective cognitive impairment diagnosis system using deep reinforcement learning. The diagnostic workflow is formalized as a sequential decision-making process in which the agent dynamically determines which features (i.e. cognitive tests and questionnaires) to acquire at each step. Combined with a supervised classifier serves as a virtual ‘clinician’, the currently available information is used to provide probabilistic feedback on the patient's state. The Rainbow DQN-based agent is trained end-to-end to learn an optimal query strategy that mimics a physician's reasoning: it adaptively selects subsequent cognitive assessments and ultimately renders a categorical diagnosis of dementia severity, aiming to maximize expected diagnostic accuracy while minimizing cumulative costs, representing patient burden and resource utilization.

resultsIn our study, the reinforcement learning–based diagnostic system achieved an area under the receiver operating characteristic curve (AUROC) of 0.877 (micro-average) and 0.823 (macro-average) during autonomous inquiry. It attained state-of-the-art diagnostic accuracy while using only an average of 14.76 questionnaire items per patient—reducing, the total estimated assessment time by 48.3% compared to full-feature supervised learning baselines. This reduction corresponds to a meaningful decrease in assessment time cost and cognitive burden for patients. Ablation experiments also validated the effectiveness of our components, highlighting the critical role of reward exploration. Finally, through system inquiry quality analysis and policy distillation visualization, the parameters of the neural network were approximated and converted into diagnostic pathways, thereby enhancing clinicians’ and patients’ trust in the model.

conclusionsWe developed a reinforcement learning–based diagnostic system for cognitive status within a population cohort. This model effectively optimizes the long-standing economic and promotional issues of community dementia screening, demonstrating ​strong accuracy, economic efficiency, and methodological advances. Furthermore, by analyzing the diagnostic logic, we enhanced clinicians’ and patients’ trust in the model.