To determine whether historical behavior data can predict the occurrence of high-risk behavioral or seizure events in individuals with profound Autism Spectrum Disorder (ASD), thereby facilitating early intervention and improved support. To our knowledge, this is the first work to integrate the prediction of seizures with behavioral data, highlighting the interplay between adverse behaviors and seizure risk.
Approach: We analyzed nine years of behavior and seizure data from 353 individuals with profound ASD. Using a deep learning-based algorithm, we predicted the following day's occurrence of seizure and three high-risk behavioral events (aggression, self-injurious behavior (SIB), and elopement). We employed permutationbased statistical tests to assess the significance of our predictive performance.
Main results: Our model achieved accuracies 70.5% for seizures, 78.3% for aggression, 80.2% for SIB, and 85.7% for elopement. All results were significant for more than 85% of the population. These findings suggest that high-risk behaviors can serve as early indicators, not only of subsequent challenging behaviors but also of upcoming seizure events.
Significance: By demonstrating, for the first time, that behavioral patterns can predict seizures as well as adverse behaviors, this approach expands the clinical utility of predictive modeling in ASD. Early warning systems derived from these predictions can guide timely interventions, enhance inclusion in educational and community settings, and improve quality of life by helping anticipate and mitigate severe behavioral and medical events.
Keywords: adverse behaviors; artificial intelligence; autism spectrum disorder; digital health records; seizure.