Nowcasting officially reported highly pathogenic avian influenza outbreak counts under heterogeneous reporting delays

Abstract

Official animal disease series are often incomplete due to delays in reporting. This retrospective observational study evaluated whether increased model complexity enhances probabilistic nowcasts of eventual officially reported outbreak-equivalent counts. We reconstructed historical information states from a fixed extraction from the World Organisation for Animal Health (WOAH) World Animal Health Information System (WAHIS), which included 46,860 highly pathogenic avian influenza (HPAI) and 72,690 African swine fever (ASF) outbreaks. Twelve prespecified methods were assessed over 78 locked cutoffs for 2024–2025 using the Weighted Interval Score (WIS), calibration diagnostics, block bootstrap, simulation, and a separately frozen USDA APHIS transfer evaluation. The primary HPAI analysis included 35,880 observations from 115 country–host units. At week closure, only 7.6% of eventual HPAI and 6.7% of ASF outbreak equivalents were visible, with completeness varying by host, region, year, season, and workload. The prespecified primary hypothesis was not supported: the delay-aware temporal network scored 6.9% worse than the hierarchical Bayesian comparator, and the relational graph changed WIS by only 0.4%. Gradient-boosted quantile regression improved WIS by 18.5% (absolute difference −0.0397, 95% interval −0.0553 to −0.0255), but the improvement reached 26.0% among high-volume reporters and reversed to a 6.3% deficit among units with minimal reporting history. Conditioning on the partial count was 24.7% worse than a matched no-delay network at cohort age 0 and performed better thereafter. The slower-reporting ASF system reversed the pooled contrast, whereas the APHIS module improved WIS by 10.1% after system-specific retraining. Adding an explicit mass at zero enhanced the transparent comparator by 18.8% and resulted in no detectable difference from gradient boosting; the neural model remained 17.4% worse under the same recalibration. In exploratory analyses, gradient boosting cleared the prespecified 5% gate under all six cohort definitions and all three maturity horizons. Performance deteriorated markedly for high-volume countries and regions excluded from training. These estimates pertain to eventual officially reported outbreak equivalents under the surveillance conditions studied, not undetected infections.