Addressing the Challenges of Applying Machine Learning for Predicting Mental Disorders and Their Prognosis Using Two Case Studies

Abstract

One of the principal applications of machine learning in psychiatry is to build automated tools that can help clinicians predict the diagnosis and prognosis of mental disorders using available data from patients’ profiles. Here, in two different studies, we investigate ways to use machine learn- ing to produce models that can predict mental disorders and their prognosis, using different neuroimaging modalities, genotype data, and clinical information.The first study addresses the challenge of producing a classifier that a human clinician can interpret and potentially use in clinical practice. In this study, we were seeking a simple and accurate classifier that can correctly distinguish Alzheimer’s disease (AD) patients from healthy controls (HC). We wanted to learn this classifier from the data in the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database, using just a fairly small set of input features, including grey matter volumes of 33 regions of interest derived from brain structural MRI, as well as the APOE genotype. Running our overall learner, involving standard feature selection processes and three simple base-learners on these features, produced a 7-feature elastic net model that achieved an accuracy of 89.28% on the test set. Next, we ran the same overall learner using two more-complex base-learners over the same initial dataset. The accuracy of the best model here (SVM-RBF over 23 features) was 90.47%, which was not statistically different from the performance of our much simpler linear model, over just 7 features. We, therefore, introduce this simple 7-feature model as our accurate and simple classification model.Our second study explored the utility of machine learning methods in predicting the response of a group of schizophrenia patients (n=51 to 90, depending on the response criterion) to a specific treatment, given their functional magnetic resonance imaging data, structural magnetic resonance imaging data, diffusion tensor imaging data, and clinical information. In this study, we explored various clinical measures for defining treatment response, various feature types for imaging and non-imaging data, various machine learning tasks and learning algorithms, but (probably due to the small sample size), we were not able to obtain any significant results.