Performance Comparison of Deep Learning Methods with Traditional Machine Learning Methods in Early Childhood Autism Spectrum Disorder Detection

Abstract

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition characterized by social communication deficits and restricted, repetitive behaviors. In order to improve long term developmental outcomes by prompt therapeutic intervention, early detection means ideally before the age of three is essential. Conventional diagnostic approaches rely on clinical observation and standardized behavioral assessments, which are time intensive and often result in delayed identification, particularly in limited resource settings. This study presents a rigorous, systematic benchmarking of six machine learning models and six deep learning models for early ASD screening in children aged 18 to 60 months. This study evaluate both traditional machine learning (TML) classifiers  including Random Forest, Support Vector Machines, XGBoost, Naive Bayes, k-Nearest Neighbors, and Logistic Regression  and contemporary deep learning (DL) architectures including CNNs, LSTM, Bi-LSTM with Attention, ResNet-50, EfficientNet-B3, and BERT-based Transformer models. Experiments are conducted on five publicly available and clinically validated datasets ABIDE I & II, SFARI Gene, ADOS-2, Q-CHAT, and NDAR (NIMH). Multimodal data including fMRI imaging, behavioral observation scores, eye tracking sequences, and genomic markers are preprocessed using standardized pipelines. Models are evaluated under 10-fold stratified cross validation using accuracy, precision, recall, F1-score, and AUC-ROC as performance metrics. ROC curves are generated to visualize sensitivity-specificity trade-offs across all classifiers. Deep learning models consistently outperformed traditional counterparts across all metrics. The Bi-LSTM with Attention and BERT-based Transformer achieved the highest AUC-ROC of 0.97. ROC curve analysis confirmed visually and statistically that DL model curves lie uniformly closer to the upper left ideal corner, with DeLong's test confirming significant pairwise AUC differences (p < 0.001) between DL and TML groups. DL models demonstrate superior diagnostic discrimination for ASD screening and hold significant promise for clinical decision support. Future work should address explainability and federated learning for privacy preserving deployment.