[Paper Review] DeepGLEAM: A hybrid mechanistic and deep learning model for COVID-19 forecasting

Dongxia Wu, Liyao Gao+5 more|arXiv (Cornell University)|Feb 12, 2021

Machine Learning in HealthcareComputer Science42 references19 citations

TL;DR

DeepGLEAM is a hybrid model that enhances the mechanistic GLEAM epidemic simulator with a deep learning correction module using a diffusion convolution recurrent neural network (DCRNN), significantly improving short-term COVID-19 mortality forecasts. It outperforms standalone GLEAM, deep learning models, and vector autoregressive models in point prediction and uncertainty quantification, with SG-MCMC showing the best overall performance in RMSE and MAE while providing reliable confidence intervals.

ABSTRACT

We introduce DeepGLEAM, a hybrid model for COVID-19 forecasting. DeepGLEAM combines a mechanistic stochastic simulation model GLEAM with deep learning. It uses deep learning to learn the correction terms from GLEAM, which leads to improved performance. We further integrate various uncertainty quantification methods to generate confidence intervals. We demonstrate DeepGLEAM on real-world COVID-19 mortality forecasting tasks.

Motivation & Objective

• To address the limitations of purely data-driven deep learning models in epidemic forecasting, which lack interpretability and physical consistency.
• To overcome the bias in mechanistic models like GLEAM by learning correction terms from observed surveillance data using deep learning.
• To improve short-term forecasting accuracy for COVID-19 mortality at the state and country level using hybrid modeling.
• To integrate multiple uncertainty quantification methods to generate reliable prediction intervals for public health decision-making.
• To demonstrate that combining mechanistic models with deep learning leads to better performance than either approach alone.

Proposed method

• The model uses the Global Epidemic and Mobility model (GLEAM) as a mechanistic baseline, simulating stochastic, spatially resolved epidemic dynamics across 3,200 subpopulations.
• A diffusion convolution recurrent neural network (DCRNN) is trained to learn correction terms that adjust GLEAM’s predictions to align with real-world observed mortality data.
• The DCRNN models spatiotemporal dependencies using a graph-based adjacency matrix derived from human mobility data, capturing irregular spatial correlations across regions.
• The hybrid model, named DeepGLEAM, produces forecasts by combining GLEAM’s simulation output with the deep learning correction, improving prediction accuracy.
• Multiple uncertainty quantification methods—bootstrap, quantile regression, MIS regression, MC dropout, and SG-MCMC—are evaluated to generate confidence intervals.
• SG-MCMC is selected as the optimal method due to its strong performance in RMSE, MAE, and coverage reliability across different prediction horizons.

Experimental results

Research questions

• RQ1 Can a deep learning model effectively correct systematic biases in a mechanistic epidemic simulator like GLEAM for improved forecasting accuracy?
• RQ2 How do different uncertainty quantification techniques compare in reliability and coverage when applied to a hybrid deep learning and mechanistic model?
• RQ3 Does the integration of deep learning with mechanistic models lead to better short-term mortality forecasts than standalone deep learning or mechanistic models?
• RQ4 Which uncertainty quantification method provides the most robust and well-calibrated prediction intervals across varying forecasting horizons?
• RQ5 Can the hybrid DeepGLEAM model generalize effectively across diverse U.S. states with varying epidemic trajectories?

Key findings

• DeepGLEAM significantly outperforms the standalone GLEAM model in one to three weeks ahead forecasts, with lower RMSE and MAE across all horizons.
• The model achieves the lowest RMSE (58.30) and MAE (29.72) at one-week horizon using SG-MCMC for uncertainty quantification.
• SG-MCMC produces the most accurate mean predictions and best coverage in 95% prediction intervals, with an MIS of 563.77, outperforming bootstrap and MC dropout.
• Quantile regression provides more reliable confidence bounds than SG-MCMC in some state-level cases, such as in Texas and Georgia, due to better coverage despite slightly higher error.
• At the country level, SG-MCMC outperforms quantile regression in both mean prediction accuracy and interval reliability, demonstrating superior overall performance.
• The model maintains strong performance across 41 U.S. states, with consistent improvements in forecast accuracy beyond the training period, as shown in predictions from week 32 onward.