Streamlining patients' opioid prescription dosage: an explanatory bayesian model

Nearly half a million people died between 1999 and 2019 from overdosing on both prescribed and illicit opioids. Thus, much research has been devoted to determining the factors affecting the dosages of opioid prescriptions. In this study, we build a probabilistic data-driven framework that develops Tree Augmented Naive Bayes (TAN) models to predict patients' opioid prescription dosage categories and investigate the conditional interrelations among these predictors. As this framework is rooted in the CDC's prescription guidelines, it can be applied in clinical settings by focusing primarily on pre-discharge pain assessments. Following data acquisition and cleaning, we utilize Elastic Net (EN) and Genetic Algorithm (GA) to identify the most important predictors. Next, Synthetic Minority Oversampling Technique (SMOTE), and Random Under Sampling (RUS) are employed to overcome the data imbalance problem present in the dataset. A patient's gender, income level, smoking status, BMI, age, and length of stay at the hospital are identified as the most significant predictors for opioid prescription dosage. In addition, we construct a Bayesian Belief Network (BBN) model, which reveals that the effect of smoking status and gender in predicting opioid prescription dosage depends on the patient's income level. Finally, a web-based decision support tool that can help surgeons better assess and prescribe appropriate opioid dosages for patients is built.