Abstract
This data-driven computational psychiatry research proposes a novel machine learning approach to developing predictive models for the onset of first-episode psychosis, based on artificial neural networks. The performance capabilities of the predictive models are enhanced and evaluated by a methodology consisting of novel model optimisation and testing, which integrates a phase of model tuning, a phase of model post-processing with ROC optimisation based on maximum accuracy, Youden and top-left methods, and a model evaluation with the k-fold cross-testing methodology. We further extended our framework by investigating the cannabis use attributes' predictive power, and demonstrating statistically that their presence in the dataset enhances the prediction performance of the neural network models. Finally, the model stability is explored via simulations with 1000 repetitions of the model building and evaluation experiments. The results show that our best Neural Network model's average accuracy of predicting first-episode psychosis, which is evaluated with Monte Carlo, is above 80%.