Why do psychosis patients discontinue their antipsychotics? Using AI to unravel the roles of efficacy and side effects.

Although most people with First Episode Psychosis show some improvement after their first antipsychotic treatment, 73-77% discontinue their first antipsychotic [1]. This discontinuation is associated with more frequent and longer psychiatric admissions [2]. Two key reasons that patients discontinue their medication are lack of efficacy or intolerable side effects [3]. Although several factors have been found to be associated with time to treatment discontinuation [4], it is not yet possible to determine which patients are at the greatest risk of discontinuing their treatment. It would be beneficial to predict the likelihood of a patient discontinuing their antipsychotic medication prior to treatment onset to guide clinical decision-making and find the most appropriate treatment. Artificial intelligence methodologies have the potential to solve this problem by finding multivariate patterns in data collected pre-treatment that can predict treatment discontinuation at the individual level. The aim of this research is to develop AI-based prediction tools for treatment discontinuation. We will first predict all-cause discontinuation and then see whether this prediction can be improved by separating the two groups into discontinuation due to tolerability and discontinuation due to efficacy. The sample consists of 364 individuals with a diagnosis of a psychotic disorder for up to 7 years from the European Long-acting Antipsychotics in Schizophrenia Trial (EULAST) [5]. Participants in this trial were randomised to either long-acting injectable (LAI) paliperidone, LAI aripiprazole, or the respective oral formulations of these antipsychotics and were followed up for 18 months. We trained support vector machine learning models within a 10-by-10 repeated nested cross-validation on the whole sample using psychopathology, sociodemographic, education and employment, and substance use data. The chosen outcomes were all-cause discontinuation, symptomatic remission at one month after baseline measured by the Positive and Negative Syndrome Scale PANSS) as a proxy for discontinuation for efficacy reasons, and movement side effects at once month after baseline measured by the Abnormal Involuntary Movement Scale (AIMS) as a proxy for discontinuation for tolerability reasons. Model performances were given in Balanced Accuracy (BAC). We covaried for inpatient status as baseline because patients in hospitals would receive very different treatment to outpatient care. The rate of all-cause discontinuation was 54.3% in our sample: this outcome was poorly predicted (BAC=52.2%). Symptomatic remission was successfully predicted with a much higher performance with a BAC of 67%. Movement side effects were predicted poorly by the clinical data with a BAC of 55.4%. Here we present the first evidence that all-cause discontinuation is a poor label for prediction modelling, and that this may represent too broad of an outcome for accurate prediction. This is supported by the fact that symptomatic remission can be well predicted in the same sample. Further analyses will investigate whether blood biomarkers will improve prediction, whether metabolic side effects can be predicted, and how treatment modifies model performance.