A Practical Application of Data Mining Methods to Build Predictive Models for Autism Spectrum Disorder Based on Biosensor Data From Janssen Autism Knowledge Engine (JAKE®)

Abstract

The Janssen Autism Knowledge Engine (JAKE®) collects a large number of features from five biosensors across a range of tasks. The application of data mining methods to these data may be a useful approach to enable objective discrimination between autism spectrum disorder (ASD) and typically developing (TD) participants. Following a prospective observational study using JAKE, ASD participants classified as “moderate” or “severe” based on total scores on the Social Responsiveness Scale, and TD participants were used to build models, using repeated cross-validation, to identify biosensor features contributing to diagnosis. Four different models (partial least squares, random forest, elastic net, and C5.0) were chosen to build diagnostic classifiers using the training set, and the fitted models were evaluated on the test set. Model performance on the training set, based on receiver operating characteristics (ROC), was moderate (area under ROC curve = 0.61–0.72), and model performance on the test set based on kappa statistic was between 0.40 and 0.46 across the four models. Data mining methods applied to biosensor data can lead to models that discriminate ASD from TD. This method may prove useful in creating new diagnostic tests for ASD.