Machine learning for early diagnosis of autism spectrum disorder

Context: Autism Spectrum Disorder (ASD) is a developmental disorder that affects communication, social interaction and behaviour. By building a machine learning model that predicts the probability of ASD through certain behaviours, demographic information and clinical history. We will be able to contribute to moving forward with getting a diagnosis for individuals with ASD even earlier. The study and resulting neural network that exists were created to be a universally available, scalable approach that can help with early diagnosis in both clinical and non-clinical situations. Objective: The main objective of this project is to build a comprehensive AI-based model for the early detection of ASD. Our approach is designed to augment early intervention efforts using a cloud-based web interface and machine learning techniques that deliver insights in an easy-to-use manner. Methods: The dataset used for the study was the "Autism Dataset for Toddlers". High-dimensional assessment of ASD traits was done using several machine learning techniques, like K-Nearest Neighbors (KNN), Decision Trees (DT), Random Forests (RF), Support Vector Machines (SVM), XGBoost and LightGBM. We created and evaluated the performance of the models using accuracy, precision, recall, F1-score and ROC AUC after feature selection-based techniques such as ANOVA-SVM. Results: XGBoost was the best classifier as it had 99.6% accuracy and ROC AUC was even better than the Decision Tree, and Random Forest even though they achieved an accuracy of 98.8%. With a close 98.10% with Support Vector Machine followed up, with K-Nearest Neighbors at 96.68%. Because the system runs on a cloud-based interface, this processing occurs in real time and enables early ASD screening. Altogether, our XGBoost model holds great potential for early autism screening as it provides a viable option for both clinicians and families.