Performance evaluation of transfer learning models for ASD prediction using non-clinical analysis

One of the significant problems the world is currently dealing with is "autism spectrum disorder"(ASD). As a "behavioral disorder,"ASD causes individuals to have trouble interacting and communicating with others. Young children often exhibit ASD symptoms, particularly in the first two years, and will persist until appropriate therapy is given. ASD can be diagnosed both clinically and non-clinically. Clinical analysis is performed by examining variations in brain imaging of the children. Whereas, Non-clinical analysis can be performed using gaze analysis of the eye, where various attributes like eye position, eye movement, and attention are considered for predicting autistic traits. This paper represents the performance evaluation of Transfer Learning (TL) methods used for the non-clinical analysis. All used TL methods are recommended and famous for feature extraction in facial attributes. We used a dataset provided by Zenodo, which is open-source and contains images of children's faces looking at an area of interest in an image. The investigation uses well-known transfer learning models such as VGG16, VGG19, InceptionV3, AlexNet, and ResNet50 and compares their performance to the CNN model. All models are well-tuned for hyperparameters and then used to predict autistic traits rather than data classification. InceptionV3 outperformed all other models with prediction accuracy of 87.99% and 84.33% on the validation and test datasets, which is significantly higher than other proposed models. Also, this is the only work that has attempted to use a transfer learning model on the dataset proved by Zenodo. This is a novel addition to the paper.