Leveraging observational and prospective data to develop an opioid exposure detection model

Abstract

Experimental studies are widely considered as the gold standard for discovering new evidence. However, advances in computational science provide an opportunity to take advantage of large clinical datasets in cases where randomized experiments are not practical. In this study, we used a large clinical database to train a model capable of detecting exposure to opioid medication (AUROC=0.76). We designed and implemented a clinical study to measure the performance of the model against the unseen data from the study. Our results show that the model based on hospital patient data exhibited promising performance against the retrospective clinical study data (AUROC=0.68).