PMH63 USE OF ARTIFICIAL INTELLIGENCE METHODS TO BETTER DEFINE TREATMENT LINES IN SCHIZOPHRENIA

Abstract

Large longitudinal administrative databases are an important source of information on treatment pathways, but clinical information is often incomplete or clouded by incidental events (e.g., loss of follow-up, lack of compliance, misreporting). Therefore the analysis of treatment pathways requires assumptions. The objective of this study is to assess the ability of Artificial Intelligence (AI) methods to classify prescriptions into treatment lines and help analysts defining rules for the analysis of treatment pathways, based on a sample of data from patients with schizophrenia. The study was conducted using the Japan Medical Data Center (JMDC) claims database. Patients aged ≥15 years with ≥2 claims associated with schizophrenia diagnosis and a first antipsychotic prescription recorded between January 2009 and March 2013 were included. Prescriptions were first classified into treatment lines based on assumptions including number of days allowed between consecutive refills and minimum number of overlap days to consider combinations (“standard approach”). Secondly, the dataset was split into training (80%) and test (20%) datasets. AI algorithms were developed on the training dataset to predict treatment lines defined previously on the test dataset. When available, the 20 previous and following prescriptions were used as inputs. AI implemented algorithms were Neural Networks, Decision trees, Random Forest and Extreme Gradient Boosting. 3142 patients met inclusion criteria and were eligible to be included in this study. Measures of algorithm accuracy were 33% for Neural Networks, 68% for Decision trees, 71% for Random Forest and 77% for Extreme Gradient Boosting. On 100 prescriptions with discrepancy between standard approach and Extreme Gradient Boosting method, 48% were deemed misclassified with the standard approach. The Extreme Gradient Boosting algorithm was the best performing method and helped to detect outlier patients. These results support the use of AI methods to check and refine assumptions for defining treatment pathways.