(A-208) Development and Validation of a Modern Prescription-Based Risk Score Predicting Healthcare Spending and Utilization

Background: Prescription-drug based risk scores are an important alternative to diagnosis-based risk scoring systems when diagnostic information is unavailable; however, their value depends upon a given score’s ability to capture the current therapeutic landscape.

Objectives: To use machine learning to develop and validate a prescription-drug based risk scoring system to predict healthcare spending and utilization.

Methods: We performed a retrospective cohort study using Merative MarketScan administrative claims data from January 2021 through December 2022. We included individuals aged 18 years or older on January 1, 2022 who were continuously enrolled during the 24 months of interest. We used prescription claims from 2021 and classified drugs into 273 categories using RedBook drug classifications. We measured five outcomes in 2022: (1) total healthcare expenditures; (2) hospitalization expenditures; (3) outpatient expenditures; (4) total number of hospitalizations; (5) and total number of outpatient visits. We used descriptive statistics to characterize the cohort, drug use and outcomes. We will fit Random Forest, Gradient Boosting, penalized regression and potentially other machine learning models to predict each outcome, using 60% of the cohort to develop the model and the remaining 40% of the cohort to validate it.

Results: After applying the exclusion criteria to the over 25 million enrollees in the database from 2021 to 2022, 7,441,574 adult enrollees were included in the cohort. The average (standard deviation [SD]) age in 2021 was 43 (SD 13) years, and most of the enrollees were female (52.2%). A total of 71.7% of the enrollees had at least one prescription in 2021. Over half of the enrollees (57.4%) were exposed to two or more drug classes and 29.5% of patients were exposed to five or more drug classes. Among the enrollees, 271,459 (3.6%) and 6,560,561 (88.2%) received inpatient and outpatient services, respectively. The mean (SD) total expenditures was $6,963 ($29,083), mean number of inpatient services were 1.71 (15.64), and the mean number of outpatient visits were 35.78 (49.95).

Conclusions: In this preliminary analysis, prescription medication use was common. In future analysis, we will apply machine learning to develop and validate the risk score. This newly developed risk score will be useful to characterize pharmacologic therapy among contemporary study cohorts as well as to adjust for confounding in pharmacoepidemiologic investigations.