Artificial intelligence to improve efficiency of administration of gross motor function assessment in children with cerebral palsy

Abstract

Aim: To create a reduced version of the 66-item Gross Motor Function Measure (rGMFM-66) using innovative artificial intelligence methods to improve efficiency of administration of the GMFM-66. Method: This study was undertaken using information from an existing data set of children with cerebral palsy participating in a rehabilitation programme. Different self-learning approaches (random forest, support vector machine [SVM], and artificial neural network) were evaluated to estimate the GMFM-66 score with the fewest possible test items. Test agreements were evaluated (among other statistics) by intraclass correlation coefficients (ICCs). Results: Overall, 1217 GMFM-66 assessments (509 females, mean age 8y 10mo [SD 3y 9mo]) at a single time and 187 GMFM-66 assessments and reassessments (80 females, mean age 8y 5mo [SD 3y 10mo]) after 1 year were evaluated. The model with SVM predicted the GMFM-66 scores most accurately. The ICCs of the rGMFM-66 and the full GMFM-66 were 0.997 (95% confidence interval [CI] 0.996–0.997) at a single time and 0.993 (95% CI 0.993–0.995) for the evaluation of the change over time. Interpretation: The study shows that the efficiency of the full GMFM-66 assessment can be increased by using machine learning (self-learning algorithms). The presented rGMFM-66 score showed an excellent agreement with the full GMFM-66 score when applied to a single assessment and when evaluating the change over time.