Causal reinforcement learning for personalized adaptive interventions in mild cognitive impairment

Mild cognitive impairment (MCI) prevention demands personalized, adaptive strategies, yet current clinical guidelines rely on static, experience-based approaches. We developed a causal reinforcement learning (CRL) framework integrating individual treatment effect (ITE) estimation via T-learner with Conservative Q-Learning (CQL) to optimize adaptive MCI interventions in continuous action spaces. Data were derived from a Sequential Multiple Assignment Randomized Trial (SMART) involving 61 participants (mean age 71.0 ± 6.0 years; 67.2% female) across three modalities: Virtual Reality Taichi, Offline Taichi, and Computerized Cognitive Training. Baseline cognitive scores, comorbidities, and sociodemographic factors defined the state space; ITEs served as reward signals. In an exploratory continuous-action-space analysis, CRL-derived strategies were estimated to improve cognitive outcomes by an average of 1.03 Memory Guard score points over dynamic treatment regimens across four machine learning algorithms (all P < 0.001), while requiring substantially lower intervention intensity (15–90 vs 120–240 min/week). SHAP analysis highlighted diabetes status, age, and baseline cognition as key treatment allocation drivers. Exploratory heterogeneous analyses revealed diminished benefits in hypertensive, older, male, and higher educated participants, whereas higher baseline cognitive reserve predicted greater gains. This framework may offer a data-driven approach for personalized MCI prevention, enabling adaptive interventions that account for individual treatment heterogeneity. The underlying trial was registered with the Chinese Clinical Trial Registry (registration number ChiCTR2100042748; registered on 27 January 2021).