Supervised machine learning techniques have one of their main objectives to reduce the generalized prediction error, and these methods can be applied to predict health conditions with several purposes. Support vector models (SVM) are a consolidated statistical learning model in the machine learning community because they present properties that are easy to characterize and allow the estimation process with global optimization. However, SVMs present challenges in selecting the appropriate kernel function, as well as the tuning process of choosing its hyperparameters. Random machines is a recent machine learning method that eliminates the need to choose the best kernel function during the tuning process. In this paper, we present a modification of the traditional Random Machines, namely Random Machines with probability outputs, an extension with high predictive performance compared with the traditional Random Machines and other classical machine learning methods. In order to predict the general psychopathological symptoms of the SIPS (Structured Interview for Prodromal Syndromes) on At Risk Mental State (ARMS), the usefulness of Random Machines is demonstrated. For this application, the variables were extracted from facial movement of brief video recordings of 127 patients on medical appointments. Random Machines with probability outputs showed a superior predictive capacity than methods such as Extreme Gradient Boosting, Random Forest, Deep Multilayer Perceptron, SVM and logistic regression.
