Estimating the Efficacy of Common Treatments in Children and Young Adults Diagnosed with Cerebral Palsy Using Three Machine Learning Algorithms

Background: Orthopedic and neurological deformity are often treated in children and young adults with cerebral palsy (CP). Due to challenges arising from combinatorics, research funding priorities, and medical practicalities, the efficacy of these treatments is not well studied. Objectives: Our goal was to estimate the efficacy of 13 common orthopedic and neurological treatments at four different levels of outcome in children and young adults diagnosed with CP. The outcome levels considered were anatomy and physiology, gait parameter, overall gait pattern, and function. Methods We used three well-establish causal inference approaches (direct matching, virtual twins, and Bayesian causal forests) and a large clinical gait analysis database to estimate the average treatment effect on the treated (ATT). We then examined the efficacy across treatments, methods, and outcome levels. Results The median ATT of 13 common treatments in children and young adults with CP, measured as Cohen's D, bordered on medium at the anatomy and physiology level (median [IQR] = 0.42 [.05, .60]) and became smaller as we moved along the causal chain through gait parameter (0.21 [.01, .33]), overall gait pattern (0.09 [.03, .19]), and function (-0.01 [-.06, .13]). Conclusions Current treatments have medium effects on anatomy and physiology, but modest to minimal efficacy on gait and function. Further work is needed to understand the source of heterogeneous treatment effects, which are large in this patient population. Replication or refutation of these findings by other centers will be valuable to establish the generalizability of these results and for benchmarking of best practices.