A data-driven, algorithmic approach to recommending hours of ABA for individuals with ASD

Determining the precise number of therapy hours a patient needs is a critical clinical decision. Too few hours can reduce overall progress and likely keeps the individual in treatment longer than necessary. Too many hours can cause the individual to spend unnecessary time and money they could have spent on other activities that increase their happiness and well-being. Too many hours also reduce the hours the provider has available to see other clients further exacerbating access issues prominent in mental health today. Despite its importance, little research exists to show how specific patient profiles and intake assessments can lead to replicable and precise therapeutic recommendations. In this study, we show how patient clustering algorithms can be combined with predictive modeling to create a data-driven, algorithmic system that generates dose-response curves relating hours per week of therapy to patient progress while considering the patient’s unique profile. Specifically, we used 48 variables spanning hours and characteristics of therapy, treatment goal characteristics, and patient characteristics to predict goals mastered for 39,475 individuals with ASD receiving applied behavior analysis (ABA) services from 833 service providers. Unsupervised machine learning identified 18 distinct patient clusters. Across clusters, top performing regression models predicted patient progress for all patients with r2=.97 and MAE=0.003 and with r2 for individual clusters ranging between .95-.99 (~.20-.24 points higher than past research) and MAE ranging between <0.001-0.25. Once designed, the resulting patient specific dose-response curves can be used to identify the optimal hours of week that maximizes progress while reducing unnecessary time in treatment. Though designed specifically for predicting ABA hours for individuals with ASD, the current method offers a generalizable approach clinicians can use to develop their own data-driven, algorithmic approach to determine the hours of therapy that optimize patient progress.