Autism detection using Machine Learning through auditory features

Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder affecting more than 75 million people in the world. Early detection is crucial for reducing lifetime care costs and improving the quality of life of the individual with Autism. Current diagnostic methods rely heavily on behavioral observations and thus lack the precision required for early detection. The study explores the use of speech combined with machine learning for early Autism detection. The study utilizes ASDBank and CHILDES, clinical databases within TalkBank, to obtain audio recordings of children with and without ASD. The children were instructed to complete the same task to ensure consistency and fairness across the trials. Auditory features- Spectral Centroid, Spectral Rolloff, Zero Crossing Rate, and Root Mean Square-are first extracted from the audio files. Then we compared 3 models-Support Vector Machine (SVM), Random Forest (RF) classifier, and a Neural Network (NN)- to determine the most appropriate model for autism detection using the 4 auditory features. The SVM model achieved the highest accuracy. The second part of the paper focuses on an ablation study to determine which of the features is the most impactful for this detection. The results indicated that Zero Crossing Rate was the most impactful feature.