Multimodal Autism Detection Using Genetic Algorithm-Optimized Neural Networks

Rabiaa Tbibe, Khaoula Bouazzi, Sadok Bouamama, Patrick SIARRY

Keywords: Autism Detection, Multimodal Deep Learning, Genetic Algorithm Optimization, Neural Networks, Medical Diagnosis

Abstract: Resulting from rapid advancements in artificial intelligence, especially machine learning and deep learning, the healthcare sector has undergone a significant transformation. The value of these technologies lies in their excellent ability to analyze large amounts of medical data, leading to better diagnoses and early detection of complex conditions like autism spectrum disorder. This study develops a multimodal deep learning framework that combines behavioral questionnaires, neurophysiological signals, and facial features, using data from thousands of participants across multiple extensive datasets. The baseline neural network initially achieved moderate accuracy. After optimization with a genetic algorithm, performance improved greatly, reaching excellent accuracy, a very high area under the curve (AUC)-ROC score, a strong F1 score, and a notable performance boost. The genetic algorithm identified optimal hyperparameters, including appropriate neuron counts, effective dropout rates, and suitable learning rates, resulting in very high sensitivity for clinical use. The optimized framework surpasses existing methods in computational efficiency and has the potential to be applied in clinical settings for early ASD detection.

Primary Area: optimization

Submission Number: 18746