PRM55 - A NEW ANALYSIS METHODOLOGY TO FIND ATYPICAL COST PROFILES IN PATIENTS LIVING WITH HIV FROM A NATIONAL CLAIM INSURANCE DATA (SNIIRAM)

Abstract

This study is a large-scale extension of a published study on the economic burden of HIV in France that was solely focused on hospital costs. Through the application of a new analytical methodology, the objective is to find the profiles of patients living with HIV (PLHIV) who overexpress the cost of care, using all reimbursement data of the French health insurance (SNIIRAM database). We performed a retrospective cohort study. The SNIIRAM contains individualized data on all reimbursed health expenses for the entire French population. PLHIV were included between 01/01/2013 and 31/12/2013 by the presence of a care related to HIV. Each patient was followed one year after his inclusion. The annual cost per patient was assessed for the economic burden. The search for patient profiles overexpressing these costs is based on an innovative methodology with a machine learning algorithm. A Boosted Decision Tree regression model was customized to match the size and complexity of health data and to find meaningful profiles (a profile is a combination of dozens of variables). A medical reading of the data identified tens of variables of interest to describe the 98 000 PLHIV included in the study: sociodemographic information, hospital admissions, presence of opportunistic infections, comorbidities, frequency of viral load testing, consultations, and the cost of care. The dedicated machine learning algorithm identified more than ten relevant profiles. An interactive data-visualization tool allowed the restitution of these complex results in a clear way. This study shows that in-depth uses of large healthcare databases require innovative analysis methods. We combined a standard epidemiological study methodology (cohort definition and monitoring, medical review, economic burden) with an innovative mathematical model (Machine Learning) to obtain accurate results regarding patient profiles. These profiles are usable to objectify action plans to improve care policies.