An interpretable deep neural network approach for autism spectrum disorder detection and diagnosis

Abstract

Autism Spectrum Disorder (ASD) is a neurological condition that affects communication and has become a significant global health concern. The number of children diagnosed with ASD is increasing rapidly, particularly in middle-income and low-income countries, due to various social and economic factors. Early detection of ASD is crucial for timely intervention, especially in regions where there is a shortage of trained healthcare professionals. In such settings, an automated detection system based on machine learning (ML) and deep learning (DL) could provide an effective solution. While several studies have focused on ML-based methods for ASD detection, deep learning approaches remain under-explored. To address this gap, we propose a deep learning-based system for ASD diagnosis. The process begins with preprocessing the dataset to remove unnecessary information and artifacts, as well as handling missing values. Next, we apply the Synthetic Minority Over-sampling Technique (SMOTE) on the training set to balance the class distribution. We then utilize a Deep Neural Network (DNN) with a fully connected layer for feature extraction and classification, achieving an impressive accuracy of 99.00 %. Finally, to enhance the model's interpretability and facilitate clinical adoption, we apply SHAP (SHapley Additive exPlanations) to identify the key features that influence the predictions. This study not only improves diagnostic accuracy but also provides clinicians with a transparent and reliable tool for ASD diagnosis.