Data Mining for Insights into Accidental Drug Mortality Trends in the U.S.

This research project addresses the use of data mining techniques for studying accidental opioid-related deaths in the United States using the "Accidental_Drug_Related_Deaths" dataset. More specifically, we seek to identify the distinctions in patterns and associated factors between accidental overdose incidents and intentional overdose incidents (suicide, undetermined intent, homicide). During the time period being analyzed, we determined that there were more than 90% of overdose deaths that were classified as accidental in 2022. The main goal of this project is to utilize classification models to predict whether or not opioids are likely to be involved in an incident of overdose based on the identifying features of the demographic and situational features. This prediction will allow for more timely and targeted medical interventions with the individuals experiencing an overdose incident. The analysis considers class imbalance, using SMOTE-Tomek Links, models each of the classification in a single analysis, including, Random Forest, XGBoost, Logistic Regression, SVM, and MLP Classifier, while hyper-parameter tuning the models toward maximum recall. In addition to the classification models, we applied several algorithms for anomaly detection (OneClassSVM, Isolation Forest, Elliptic Envelope) to describe the possibility of rare drug combinations that could be potentially dangerous. Overall, we demonstrated that when handled carefully, there is a range of preprocessing, feature selection, and machine learning approaches for anomaly detection that made it possible to better identify risk factors for opioid-related overdose, which could be used to support a public health-based area of intervention, data-driven programming for states that are facing the opioid crisis.