Development of a Machine Learning Model for Autism Diagnosis 

         Autism spectrum disorder (ASD) is a developmental disorder characterized by social and communication impairments, that affects 1 in 59 US children. Early diagnosis and intervention are key to improving learning for autistic individuals. However, currently, there are no medical tests available to diagnose ASD. The only methods for diagnosing ASD are standardized tests which require prolonged diagnostic time, increased medical costs, and are often based on subjective observations. This research improved upon existing diagnostic models of ASD, using the predictive power of phenotypic data available from ABIDE-II, an open-sourced, anonymous dataset.  

       This research developed a machine learning model to diagnose ASD using a patient's phenotypic data, specifically IQ and SRS (social response) scores. These results were compared to using brain imaging metrics to diagnose autism. The data was used to test supervised machine learning models, including k-nearest neighbors, decision tree, and logistic regression, and compared their classification performance on accuracy, sensitivity and specificity. Each was coded using Python. 

       SRS proved to be the most effective of all the diagnostic data, in comparison to IQ, gender, and brain imaging metrics. In fact, removing gender from the list of attributes that was used to diagnose autism, saw no change in the accuracy of the diagnostic model. This is contrary to previous research indicating that boys are more likely to be diagnosed with autism than girls as gender didn’t improve the accuracy of the model. Unfortunately, not enough research has been done to see whether there is a correlation between brain imaging metrics and autism, so next steps are to use an unsupervised model to do so. 

      This work is significant because it leads the way for future app development for patients and caregivers to use as a self-service tool for diagnosis, thereby enabling early detection of autism and reducing medical costs associated with diagnosis significantly.