Predicting Prescribing of Brand-Name vs. Generic Drugs: Machine Learning Analysis with Large-Scale Data (Preprint)

Autor: Jamison Meindl, Rishi Desai, Jessica Franklin, Joyce Lii, Aaron Kesselheim
Rok vydání: 2020
DOI: 10.2196/preprints.19605
Popis: BACKGROUND The use of brand-name prescription drugs instead of cheaper generic alternatives results in greater rates of cost-related medical nonadherence and poor patient outcomes, as well as higher costs to the health care system. However, because of negative perceptions of generic drugs, some patients switch back to a brand-name drug after switching to a newly-available generic. OBJECTIVE In this study, a machine learning model was developed to predict the likelihood of a patient who receives a generic drug switching back to the brand-name version. METHODS The model created was a stacking ensemble model using neural networks, k-nearest neighbors, and AdaBoost. The model was trained, using both validation and testing stages, on insurance claims data—which contain limited demographic data as well as information about a patient’s diagnoses, procedures, and medications—to predict the probability of a switchback. RESULTS The models performed poorly at predicting the outcome using insurance claims data, with the maximum area under the receiver operating characteristic curve of 0.578. The machine learning models also failed to outperform the basic logistic regression. CONCLUSIONS Switchbacks are primarily motivated by characteristics not found in large insurance claims databases, such as physician preferences and patient experiences. To improve prediction results, the model developed in this study could be used with data drawn from historical patient records or further demographic data.
Databáze: OpenAIRE