One Step at a Time: Exploring Multiple Ways to Improve a Machine Learning Model to Predict a Future Autism Diagnosis

Early autism screening is an important part of pediatric primary care that can help direct children into services. Currently, most autism screening is conducted via the M-Chat, a provider administered questionnaire. While useful, the M-Chat is prone to inconsistent and misuse. Epidemiologic work has shown that autistic children have early life clinical experiences that can be indicative of a future autism diagnosis. Such encounter data can be used to promote an automated, health system based, screening tool. Nonetheless, predicting a future autism diagnosis is challenging due to the rarity and complexity of the outcome. In this talk, I discuss a variety of ways we are working to improve a machine learning based prediction model for autism. We touch on topics of data representation, borrowing from related conditions & data sources, and addressing algorithmic bias.