Hybrid Deep-Learning Model for Predicting Meltdown Probability

Abstract

Autistic children (Autism Spectrum Disorder or ASD) are susceptible to meltdowns as complex behavioral events that occur through the interplay of emotional, sensory and contextual stressors. Anticipation of these incidents is a major concern to the caregivers and clinicians. In this paper, the author presents Hybrid Deep-Learning Model (HDLM) that combines Long Short-Term Memory (LSTM) and Extreme Gradient Boosting (XGBoost) networks to predict the likelihood of occurrence of a meltdown based on the multimodal IoT sensor data. The hybrid architecture is a composite of temporal sequence learning and features level gradient optimization, which allows the accurate and interpretable forecasting. Real-world behavioral datasets experimental evaluation had 92% accuracy, 0.94 ROC-AUC and reduced latency by 35% over traditional single-model baselines. The findings affirm that HDLM framework has the ability to offer early, explainable risks alerts in adaptive interventions in the care of Autism.