Substance Use
Advanced Machine Learning for Substance Overdose-Related Mortality Prediction Sukanya Krishna* Sukanya Krishna Marie-Laure Charpignon Maimuna Majumder
In the US, the substance overdose epidemic claimed over 81,000 lives in 2023 alone. More accurate predictive models are needed to guide public health interventions. Traditional statistical methods like Seasonal Autoregressive Integrated Moving Average (SARIMA) have limitations in capturing nonlinear trends, long-term dependencies, and policy interventions. We investigate deep learning (DL) approaches—Long Short-Term Memory (LSTM) networks and Temporal Fusion Transformer (TFT) models—to improve overdose-related mortality predictions.
Using publicly available mortality data from CDC WONDER, this study trains and evaluates LSTM and TFT models against SARIMA, assessing accuracy via mean absolute percentage error (MAPE) and precision via prediction interval (PI) length. Preliminary results show that LSTM models (batch size 5, 9-month lookback, mean squared error loss) achieved comparable MAPE to SARIMA for training data (LSTM: 5.93%; SARIMA: 4.75%) and lower MAPE for validation (LSTM: 2.99%; SARIMA: 4.00%) and testing data (LSTM: 15.24%; SARIMA: 16.23%), suggesting predictive improvements (Figure 1). TFT models will further integrate socioeconomic and behavioral risk factors to enhance explanatory power and capture complex interactions between past trends and socio-behavioral responses. Future work will incorporate uncertainty estimation methods like conformal prediction and Monte Carlo dropout to improve reliability in overdose mortality estimates.
This research aims to demonstrate that DL models may outperform SARIMA in predicting substance overdose mortality and provide insights into future epidemic trends – overall, by substance, and by geography. Our findings have the potential to refine resource allocation and intervention strategies, contributing to a more data-driven response to the overdose crisis. This work represents a critical step toward leveraging machine learning to address complex public health challenges.
