Prediction of Drug-Related Risks Using Clinical Context Information in Longitudinal Claims Data.

OBJECTIVES: To develop and internally validate prediction models for medication-related risks arising from overuse, misuse, and underuse that utilize clinical context information and are suitable for routine risk assessment in claims data (i.e., medication-based models predicting the risk for hospital admission apparent in routine claims data or MEDI-RADAR).

METHODS: Based on nationwide claims from health-insured persons in Germany between 2010 and 2012, we drew a random sample of people aged ≥65 years (N = 22,500 randomly allocated to training set, N = 7500 to validation set). Individual duration of drug supply was estimated from prescription patterns to yield time-varying drug exposure windows. Together with concurrent medical conditions (ICD-10 diagnoses), exposure to the STOPP/START (screening tool of older persons' potentially inappropriate prescriptions/screening tool to alert doctors to the right treatment) criteria was derived. These were tested as time-dependent covariates together with time-constant covariates (patient demographics, baseline comorbidities) in regularized Cox regression models.

RESULTS: STOPP/START variables were iteratively refined and selected by regularization to include 2 up to 11 START variables and 8 up to 31 STOPP variables in parsimonious and liberal selections in the prediction modeling. The models discriminated well between patients with and without all-cause hospitalizations, potentially drug-induced hospitalizations, and mortality (parsimonious model c-indices with 95% confidence intervals: 0.63 [0.62-0.64], 0.67 [0.65-0.68], and 0.78 [0.76-0.80]).

CONCLUSIONS: The STOPP/START criteria proved to efficiently predict medication-related risk in models possessing good performance. Timely detection of such risks by routine monitoring in claims data can support tailored interventions targeting these modifiable risk factors. Their impact on older peoples' medication safety and effectiveness can now be explored in future implementation studies.