Distributional Impacts of AI-Enhanced Telerehabilitation on Functional Recovery: A Recentered Influence Function Quantile Regression Decomposition Analysis

Background: Conventional evaluations of digital health interventions typically assess mean treatment effects, potentially masking heterogeneous impacts across the functional recovery distribution. Patients at the lower and upper tails of recovery trajectories may respond differently to AI-enhanced telerehabilitation, yet standard regression approaches cannot capture these distributional nuances. Objective: This study applied Recentered Influence Function (RIF) quantile regression with Oaxaca-Blinder decomposition to examine how AI-enhanced telerehabilitation differentially affects functional recovery outcomes across the entire distribution, and to decompose observed disparities into explained (composition) and unexplained (structure) components. Methods: We analyzed data from 486 post-stroke patients across three rehabilitation centres in Singapore (January 2023 to December 2025). Patients received either AI-enhanced telerehabilitation (n=241) incorporating natural language processing-based progress monitoring and adaptive exercise prescription, or standard care (n=245). RIF-quantile regressions were estimated at the 10th, 25th, 50th, 75th, and 90th quantiles of the Functional Independence Measure (FIM) score distribution. Oaxaca-Blinder decomposition at each quantile partitioned group differences into composition effects (attributable to differences in observable characteristics) and structure effects (attributable to differential returns to those characteristics). Results: The AI-enhanced telerehabilitation group demonstrated significantly greater FIM improvements across all quantiles, with the largest effects at the 10th quantile ({beta} = 12.74, 95% CI: 8.92 to 16.56, p < 0.001) and 25th quantile ({beta} = 9.83, 95% CI: 6.71 to 12.95, p < 0.001), diminishing at the 90th quantile ({beta} = 3.21, 95% CI: 0.88 to 5.54, p = 0.007). RIF decomposition revealed that at the 10th quantile, 68.3% of the treatment-control gap was attributable to structure effects, indicating that AI-enhanced telerehabilitation fundamentally altered recovery mechanisms for lower-performing patients rather than merely leveraging differences in patient characteristics. Conclusions: AI-enhanced telerehabilitation produces its most pronounced benefits among patients at the lower end of the functional recovery distribution, suggesting a potential mechanism for reducing outcome inequality in stroke rehabilitation. RIF-quantile regression decomposition offers a methodologically rigorous framework for understanding distributional treatment effects that are invisible to conventional mean-focused analyses.