S12. A MACHINE LEARNING FRAMEWORK FOR ROBUST AND RELIABLE PREDICTION OF SHORT- AND LONG-TERM CLINICAL RESPONSE IN INITIALLY ANTIPSYCHOTIC-NAïVE SCHIZOPHRENIA PATIENTS BASED ON MULTIMODAL NEUROPSYCHIATRIC DATA

Abstract

Background: Approximately one-third of patients with schizophrenia fail to respond to antipsychotic treatment, but markers of clinical response are missing. Machine learning approaches have shown promise in prediction of clinical outcome, but the reproducibility of machine learning analyses in computational psychiatry is a growing concern. This presentation will discuss a workflow aiming at reducing bias and overfitting in a multimodal neuropsychiatric dataset of antipsychotic-naive, first-episode schizophrenia patients. As a novel feature in psychiatric research, simulated data were included in the design process. Moreover, analyses were conducted in two independent machine learning (ML) approaches, one based on a single algorithm and the other incorporating an ensemble of algorithms. The aims were to (1) classify patients from controls, (2) predict short- and long-term treatment response, and (3) validate the methodological framework. Method(s): Data included 138 antipsychotic-naive, first-episode schizophrenia patients (44 females /94 males), who had undergone assessments of psychopathology, cognition, electrophysiology, structural magnetic resonance imaging (MRI), before and after their first antipsychotic treatment period. Perinatal data and longterm outcome measures were obtained from Danish registers. Baseline diagnostic classification algorithms also included data from 151 matched controls (52 females/99 males). Short-term treatment response in patients was defined as change in PANSS score after initial treatment. Long-term response was a binary outcome (good vs. poor) based on data from Danish registers. Result(s): The two ML approaches both significantly classified patients from healthy controls. The single algorithm approach yielded a balanced accuracy (BACC) of 64.2% (confidence interval (CI): [51.7, 76.7]), and the ensemble approach yielded a BACC of 63.8% (CI: [50.8, 76.7]). Post hoc analyses showed that the classification primarily was driven by the cognitive data. Neither approach predicted short- nor long-term treatment response. Validation of the framework showed that choice of algorithm and parameter settings in the real data was successfully guided by results from the simulated data. Conclusion(s): This rigorous modeling framework involving simulated data and two parallel ML approaches significantly discriminated patients from controls. However, the extensive neuropsychiatric data from antipsychotic-naive patients were not predictive of treatment response. Validation of the framework showed that the ranking of the algorithms and parameter settings in the simulated was maintained in the real data.