0325 Contactless Sleep Monitoring Technology to Predict Adverse Behaviors in Autism

Abstract Introduction More than 40% of people with Autism experience disrupted sleep, which is meaningfully higher than the general population. Poor sleep quality is linked to heightened adverse behaviors during wake. However, the ability to predict these behaviors based on overnight sleep patterns has been challenging, as it requires humans in the loop. Also, conventional monitoring approaches using wearable sensors may prove intolerable for those who experience sensory processing difficulties. We developed contactless systems and assessed whether we could successfully predict challenging daytime adverse behavior and whether those behaviors affect each other. Methods Over a span of 24 months, we deployed affordable infrared cameras and microphones in the bedrooms of 14 individuals with autism, aged 15 to 22. To ensure privacy, the system converted the video footage into aggregated motion statistics at the edge and Mel-frequency cepstral coefficients for audio. Then we uploaded only the extracted features. We analyzed more than 2,000 nights of data using the developed system to identify temporal patterns. Using machine learning models, we correlated overnight features with behavior labels that occurred the following day. The events were logged by the behaviorist-trained staff members. Such adverse events include aggression, self-injury, and disruption. Results Our model identified adverse behaviors in the morning with statistical significance(p< 0.05) for 7 of 14. Morning predictions' accuracy outperformed afternoon predictions (71% versus 54% balanced accuracy), showing sleep influence diminishes as the day progresses. The model differentiated between behavior types, accurately predicting aggression in 6 participants and self-injury in 3 participants. Next, we incorporated 7-days data, which led to a marginal benefit over single-night analysis. This shows the immediately preceding night carries primary predictive weight. Cohen's d effect size of 0.45 suggests moderate clinical utility for the responsive subgroup. Conclusion The off-body sensing pipeline offers a practical, non-invasive approach for sleep monitoring in Autism and has the potential for predicting daytime adverse behavior. Such an approach could enable care teams to preemptively modify environmental factors and therapeutic interventions based on objective overnight data and lead to enhancing behavioral outcomes. Support (if any) Yashar Kiarashi was supported by the Center for Discovery and Thrasher Research Fund Early Career Award.