Investigation of Eye-Tracking Scan Path as a Biomarker for Autism Screening Using Machine Learning Algorithms

Abstract

Autism spectrum disorder is a group of disorders marked by difficulties with social skills, repetitive activities, speech, and nonverbal communication. Deficits in paying attention to, and processing, social stimuli are common for children with autism spectrum disorders. It is uncertain whether eye-tracking technologies can assist in establishing an early biomarker of autism based on the children’s atypical visual preference patterns. In this study, we used machine learning methods to test the applicability of eye-tracking data in children to aid in the early screening of autism. We looked into the effectiveness of various machine learning techniques to discover the best model for predicting autism using visualized eye-tracking scan path images. We adopted three traditional machine learning models and a deep neural network classifier to run experimental trials. This study employed a publicly available dataset of 547 graphical eye-tracking scan paths from 328 typically developing and 219 autistic children. We used image augmentation to populate the dataset to prevent the model from overfitting. The deep neural network model outperformed typical machine learning approaches on the populated dataset, with 97% AUC, 93.28% sensitivity, 91.38% specificity, 94.46% NPV, and 90.06% PPV (fivefold cross-validated). The findings strongly suggest that eye-tracking data help clinicians for a quick and reliable autism screening.

Automatic Screening of Diabetic Retinopathy Using Fundus Images and Machine Learning Algorithms

“…A DNN is a feed-forward artificial neural network (ANN) built using numerous artificial neurons, each of which mimics a biological neuron [ 36 ]. As described in Figure 8 a, an artificial neuron has N number of inputs (X i ) to collect the input data and a processing unit with a summing and an activation function to produce an output (Y) [ 37 ], using Equation (15), where Wi*Xi is the weighted inputs, B is the bias, and is an activation function, yields the output Y. The activation function specifies how the weighted sum of the input is turned into an output.…”

“…The activation function chosen has a significant impact on the neural network’s capabilities and performance. There are various activation functions to choose from, but the most popular are sigmoid, tanh, and rectified linear unit (ReLU) [ 37 ]. …”

Intelligent speech technologies for transcription, disease diagnosis, and medical equipment interactive control in smart hospitals: A review

“…With the development of robotics and artificial intelligence (AI) technologies, machines can achieve more efficient and accurate disease diagnosis and assessment in some cases and replace nurses to assist patients in their lives, which alleviate the problem of insufficient medical resources. For example, intelligent image processing methods based on deep learning (DL) have been applied to processing X-ray, CT, ultrasound, and facial images for diagnosing diseases such as COVID-19 detection [ [4] , [5] , [6] ], paralysis assessment [ 7 , 8 ], and autism screening [ 9 ]. In addition, intelligent speech technology (IST) plays a critical role in smart hospitals because language is the most natural mean of communication between doctors and patients and contains much information, such as patients’ identity, age, emotion, and even symptoms of diseases [ 10 ].…”

Employing Classification Techniques on SmartSpeech Biometric Data towards Identification of Neurodevelopmental Disorders

“…Moreover, early detection of developmental disabilities in children is crucial for improving the prognostic procedures for NDs on an individual's development stages [12]. Therefore, there is a need for additional support to diminish the over-or under-diagnosis of NDs in children [11,12,14,15].…”