A precision health approach to medication management in neurodevelopmental conditions: a model development and validation study using four international cohorts

Background Psychotropic medications are commonly used for children with neurodevelopmental conditions, but their effectiveness varies, making prescribing challenging and potentially exposing individuals to multiple medication trials. This study developed artificial intelligence (AI) models to predict prescribing patterns of stimulants, anti-depressants, and anti-psychotics. Methods Models were built to predict cross-sectional medication usage from the Child Behaviour Checklist. Data from the Province of Ontario Neurodevelopmental (POND) network ( N =598) trained and tested the models, while data from the Healthy Brain Network (HBN N =1,764) and Adolescent Brain Cognitive Development (ABCD N =2,396) studies were used for external validation. For the electronic medical record (EMR) cohort, data from the Psychopharmacology Program (PPP N =312) at Holland Bloorview Kids Rehabilitation Hospital were used to predict longitudinal success. Stacked ensemble models were built separately for each medication class, and area under the receiver operating characteristic curve (AU-ROC) evaluated performance. Results The research cohorts demonstrated feasibility, with internal testing (POND) achieving an AU-ROC (median [IQR]) of 0.75 [0.73,0.80] for stimulants, 0.83 [0.78,0.87] for anti-depressants, and 0.79 [0.72,0.86] for anti-psychotics. Performance in external testing sets (HBN and ABCD) confirmed generalizability. In the EMR cohort (PPP), AU-ROCs were 0.84 [0.81,0.88] for stimulants, 0.82 [0.77,0.87] for anti-depressants, and 0.87 [0.83,0.91] for anti-psychotics. Limitations The study’s findings should be interpreted with caution due to several factors, including the use of medication choice as a proxy for treatment response and the omission of key sociodemographic variables, which may introduce bias. The model was trained solely on North American data, limiting global generalizability, and predictions were restricted to broad medication classes rather than within-class choices. Conclusions Findings demonstrate that AI can accurately learn expert prescribing patterns and predict medication prescribing decisions, supporting the potential of data-driven tools to guide personalized medication management for neurodevelopmental conditions and reducing the trial-and-error burden in clinical practice.