Deep Learning Models Outperform Traditional SARIMA in U.S. Substance Overdose Mortality Forecasts

Global: Deep Learning Models Outperform Traditional SARIMA in U.S. Substance Overdose Mortality Forecasts

A recent study evaluated methods for estimating excess substance‑overdose deaths in the United States, focusing on the period from 2020 to 2023 when the COVID‑19 pandemic disrupted normal mortality trends. Researchers used national CDC data from 2015‑2019 to train models and projected counterfactual mortality for the pandemic years, aiming to improve public‑health planning by providing more accurate excess‑mortality estimates.

Study Overview

The investigation compared a classic Seasonal Autoregressive Integrated Moving Average (SARIMA) model with three deep‑learning architectures: Long Short‑Term Memory (LSTM), Sequence‑to‑Sequence (Seq2Seq), and Transformer. All models were trained on the same historical dataset and evaluated on their ability to predict mortality under a regime change.

Model Comparison

Performance was measured using Mean Absolute Percentage Error (MAPE) and the coverage of conformal prediction intervals. The LSTM achieved a MAPE of 17.08%, substantially lower than SARIMA’s 23.88%, indicating more precise point forecasts. In terms of uncertainty calibration, the LSTM’s intervals covered 68.8% of observed outcomes, compared with SARIMA’s 47.9% coverage.

Limitations of Attention Models

Both attention‑based models, Seq2Seq and Transformer, demonstrated poorer results. The authors attribute this underperformance to overfitting on historical mean patterns, which limited the models’ ability to capture emergent pandemic‑related trends.

Robust Forecasting Pipeline

The research incorporated a reproducible pipeline that applied conformal prediction techniques and conducted convergence analysis across more than 60 trials per configuration. This systematic approach ensured that reported performance metrics reflected stable model behavior rather than isolated runs.

Practical Deployment

An open‑source framework derived from the study is ready for deployment with fifteen state health departments, offering a scalable tool for local agencies to generate counterfactual mortality estimates in real time.

Implications for Public Health

Findings suggest that carefully validated deep‑learning models can provide more reliable excess‑mortality estimates than traditional statistical methods, especially during periods of structural disruption. The authors emphasize the importance of calibration techniques to maintain confidence in neural forecasts used for high‑stakes public‑health decisions.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.