What question did this study set out to answer?

The aim is to compare various forecasting models for disease burden using GBD 2023 data across high-income countries.

April 14, 2026Open Access

Foundation models, deep learning, and statistical methods for disease burden forecasting: a 10-model comparison using GBD 2023 data

Key Points

The aim is to compare various forecasting models for disease burden using GBD 2023 data across high-income countries.
Evaluated 9 forecasting models from statistical, machine learning, and deep learning paradigms.
Constructed a naive persistence baseline for comparison.
Tested all possible inverse-MAE-weighted ensemble combinations to enhance predictive accuracy.
Models were retrained on historical data (1990-2023) for projections through 2040.
PatchTST had the lowest individual model MAE at 36.70 DALYs/100,000.
A 5-model ensemble achieved the best overall MAE of 33.01.
Chronos model provided the best-calibrated prediction intervals with 62.9% coverage.
Projected increases in anxiety and dementia rates by 2040, while decreases in IHD and stroke were noted.

Abstract

Background: Forecasting disease burden informs health system planning, yet most projections use single statistical methods. Modern deep learning architectures and pretrained foundation models have not been systematically compared for this task. Methods: We obtained GBD 2023 disability-adjusted life year (DALY) rates for 18 causes across five high-income countries (1990–2023; 90 time series). Nine forecasting models spanning four paradigms — statistical (auto-ARIMA, ETS, Prophet), machine learning (XGBoost), deep learning (temporal fusion transformer, N-BEATS, N-HiTS, PatchTST), and a pretrained foundation model (Chronos, zero-shot) — were evaluated against a naive persistence baseline. All 502 possible inverse-MAE-weighted ensemble combinations were tested. Models were retrained on 1990–2023 to project Australian burden through 2040 with sex and age stratification. Results: PatchTST achieved the lowest individual-model MAE (36.70 DALYs/100,000; skill score 0.144), followed by Chronos zero-shot (38.81; 0.095) and ETS (39.64; 0.075). A 5-model ensemble achieved the best overall MAE (33.01; skill 0.230). Chronos — requiring no training — produced the best-calibrated prediction intervals (62.9% empirical coverage for nominal 80%), far exceeding TFT (19.6%) and Prophet (23.1%). For Australia, anxiety (+111.5%) and dementia (+25.0%) showed the largest projected increases to 2040, while IHD (−41.9%) and stroke (−30.9%) showed the largest decreases. Female liver cancer DALYs were projected to increase 43.2% versus a 6.4% male decrease. Conclusions: PatchTST and pretrained foundation models outperformed all conventional methods for GBD forecasting, while multi-model ensembles provided the best overall accuracy. The zero-shot performance of Chronos demonstrates that pretrained time-series models transfer effectively to epidemiological domains without task-specific training. Code and data: https://doi.org/10.5281/zenodo.19528301

问 AI

Bookmark

View Full Paper