Autism Spectrum Disorder detection in children by calculating the distances between facial landmarks

Currently, the Machine Learning models that are being used in autism detection have high computational requirements, which in turn makes the detection time consuming. This paper proposes a model that will detect whether a person has ASD (Autism Spectrum Disorder) or is TD (Typically Developing) based on the distances between specific facial landmarks. Haar Cascade algorithm was used for the detection of eyes and MediaPipe framework was used for detection of facial landmarks on the face. There are four hundred and sixty-eight facial landmarks detected by MediaPipe out of which the detection of right landmarks was done. Six relevant distances were calculated. The Euclidean distance was calculated between these detected landmarks for autistic faces as well as Typically Developing faces. The extracted distances were then passed through various classifiers such as Random Forest, KNN, XGBoost, SVM, Logistic Regression and Decision Tree, after which a machine learning model was trained. The maximum accuracy was achieved by making use of XGBoost classifier. The final model obtained an accuracy of 83.5% which was given by XGBoost classifier.