Biosensor prediction of aggression in youth with autism using kernel-based methods

Aggression to others by youth with autism is a significant problem since their difficulties self-reporting distress can lead to behaviors that appear to occur without warning. To address this issue, we recently demonstrated that biosensor data combined with linear classification algorithms (i.e., ridge-regularized logistic regression) can be used to predict aggression up to 1 minute before it occurs using 3 minutes of data from the past with an average area under the curve (AUC) of 0.71-0.84 depending on whether population versus individual models are used. In the present study, we both extend and enhance these prior results through the use of principal component analysis and a nonlinear kernel-based classifier (Support Vector Machines). Our results illustrate that these newly applied methods yield significant improvements, predicting aggression up to 3 minutes before it occurs with an average AUC of 0.98 in both population and individual models. Furthermore, we extend our prior work by evaluating aggression prediction performance across varying observed aggression intensities and find that moderate and high intensity aggression episodes are detectable with 2 to 5% higher average AUC than low-intensity aggression episodes.