Risk Prediction of Opioid Dependency Using Machine Learning Based on Electronic Health Record

Opioid abuse and dependence have emerged as a pressing global concern, posing significant challenges to public health and society. Early identification and prediction of opioid dependency represent crucial steps in mitigating its abuse impact on individuals and communities. The application of machine learning techniques to analyze medical data has opened new avenues for achieving this goal. While this field and the prediction is still in its infancy, our research explores the potential of several machine learning algorithms including LightGBM for this risk prediction. To tackle the inherent class imbalance in the MIMICIII dataset, we implemented the Synthetic Minority Oversampling Technique (SMOTE). We developed predictive models using four distinct algorithms: decision trees, random forests, support vector machines, and LightGBM. These models were meticulously evaluated to assess their performance. Ultimately, our findings revealed that the LightGBM model outperformed the other algorithms, demonstrating superior accuracy and achieving a higher Area Under the Curve value. This outcome underscores the potential of LightGBM as a valuable algorithm in the early prediction of the risk of opioid dependence, thereby offering substantial benefits to both patients and society at large.