[Paper Review] Interpretable Sequence Learning for COVID-19 Forecasting

Sercan Ö. Arık, Chunliang Li+13 more|arXiv (Cornell University)|Aug 3, 2020

COVID-19 epidemiological studies42 references59 citations

TL;DR

The paper presents an interpretable, covariate-encoded SEIR-based model that forecasts COVID-19 progression at US state and county levels, outperforming several benchmarks and providing explainable insights.

ABSTRACT

We propose a novel approach that integrates machine learning into compartmental disease modeling to predict the progression of COVID-19. Our model is explainable by design as it explicitly shows how different compartments evolve and it uses interpretable encoders to incorporate covariates and improve performance. Explainability is valuable to ensure that the model's forecasts are credible to epidemiologists and to instill confidence in end-users such as policy makers and healthcare institutions. Our model can be applied at different geographic resolutions, and here we demonstrate it for states and counties in the United States. We show that our model provides more accurate forecasts, in metrics averaged across the entire US, than state-of-the-art alternatives, and that it provides qualitatively meaningful explanatory insights. Lastly, we analyze the performance of our model for different subgroups based on the subgroup distributions within the counties.

Motivation & Objective

• Integrate covariate-encoded learning into an SEIR-style compartmental model to forecast COVID-19 dynamics.
• Extend compartments to include undocumented cases and hospital resources to reflect real-world progression.
• Ensure temporal and spatial generalization via time-varying encoders and cross-location information sharing.
• Achieve superior forecasting accuracy compared to existing models and provide interpretable explanations for epidemiologists and policymakers.

Proposed method

• Extend SEIR with undocumented and hospital-related compartments to better capture disease progression.
• Replace static rates with learnable, covariate-driven encoders that map static and time-varying covariates to transition rates.
• Adopt a generalized additive model style encoder to preserve interpretability and quantify covariate effects.
• Incorporate time-varying covariates and forecast time-series covariates using auxiliary models (e.g., XGBoost) when future covariates are unavailable.
• Regularize and constrain the effective reproduction number to plausible ranges and encourage smooth dynamics via second-order differences.
• Train end-to-end with partially supervised data, using partial teacher forcing and an objective balancing data fit, covariate sharing, and epidemiological priors.
• Include an information-sharing mechanism across locations with a local bias to capture location-specific dynamics and a regularization term to prevent over-reliance on locality.

Experimental results

Research questions

• RQ1 Can a covariate-encoded SEIR framework improve forecast accuracy for COVID-19 at state and county levels compared to existing models?
• RQ2 How do undocumented and hospital-related compartments affect forecasting and interpretability?
• RQ3 Do time-varying covariates and cross-location information sharing enhance generalization under limited data?
• RQ4 What interpretable insights can be extracted about covariate effects on transmission and outcomes?
• RQ5 Can the model provide reliable prediction intervals for policymakers?

Key findings

• The proposed model yields lower forecast errors (MAE) than several benchmarks across state-level 14-day forecasts on multiple dates.
• County-level forecasts show substantially lower MAE than Berkeley CLEP for multiple prediction dates.
• Ablation studies demonstrate benefits from covariate-encoded encoders, extended compartments, partial teacher forcing, and fine-tuning on recent data.
• The model provides interpretable insights, such as the impact of mobility and interventions on transmission rates and the relative roles of documented vs. undocumented infections.
• Prediction intervals with 10th–90th quantiles are demonstrated and appear well-calibrated, widening in non-smooth periods.
• The framework delivers explainable dynamics by explicitly modeling compartment transitions and covariate effects.