A Multimodal Explainable Deep Learning Framework for Autism Severity and Behavioral Profile Prediction Using Clinical and Behavioral Data

Abstract

Autism Spectrum Disorder, or ASD, is a neurodevelopmental condition. People with ASD show differences in how they communicate socially. They may also have distinct behavior patterns. Development can look different from person to person. Getting checked early can help. It can also help teams act sooner. Still, standard screening tools might miss some details. Clinical and behavior signs can add extra clues. Conventional methods may not use all of that combined information. This study proposes a lightweight multimodal explainable deep learning framework for screening-based severity stratification and behavioral profile prediction using clinical and behavioral data. The framework processes clinical attributes through a multilayer perceptron (MLP) and ten behavioral screening responses through a bidirectional long short-term memory (BiLSTM) network. The resulting representations are combined using feature-level fusion and passed to two prediction heads for severity-category and behavioral-profile estimation. SHAPbased analysis is incorporated to examine the contribution of clinical features, while permutation-based analysis is used to assess the importance of individual behavioral screening items. Experiments were conducted on the UCI Adult Autism Screening dataset. After removal of duplicate records and basic data cleaning, 699 samples were retained. The multimodal model scored 99.05% accuracy. It reached 99.22% macro precision. It also got 98.81% macro recall. The macro F1 was 99.00% for the severity task that used screening scores. Its macro ROC-AUC came out to 0.9992. For the behavioral profile task, the results were lower. The micro F1 was 88.98%. The macro F1 was 88.06%. The results demonstrate the feasibility of combining clinical and behavioral information within an explainable deep learning framework. However, because the severity and behavioral-profile targets were derived from the screening responses available in the selected dataset, the findings should be interpreted as a proof-of-concept rather than as independent clinical severity diagnosis. Key Words: Autism Spectrum Disorder, multimodal learning, deep learning, BiLSTM, clinical data, behavioral data, explainable artificial intelligence, SHAP, behavioral profile prediction, severity prediction.