A Retrospective Evaluation of Backfill Estimation Models for Influenza, COVID-19, and RSV Hospital Admissions

Hospital admission data for respiratory illnesses such as influenza, COVID-19, and RSV are critical for real-time public health surveillance and short-term forecasting. The National Healthcare Safety Network (NHSN) which provides a comprehensive dataset for all 52 jurisdictions in United States changed its reporting cadence in November 2024. It began releasing preliminary weekly data on Wednesdays and updated versions on Fridays. Because Wednesday reports are often incomplete, especially for the most recent week, users of this dataset must contend with backfill, that is, subsequent (predominantly) increases in reported values as hospitals complete their submissions. In this study, we evaluate four simple nowcasting models to estimate the magnitude of backfill to the last data point: (1) a NULL model that assumes no backfill; (2) a linear-scaling model that assumes backfill scales linearly with the fraction of hospitals yet to report; (3) a prospective-scaling model used operationally during the season, which estimates backfill as a compromise between the two; and (4) a retrospective-scaling model that uses a fitted regression based on the full season’s data. Using 26 weeks of data from 52 U.S. jurisdictions, we compared model estimates to the final reported values and calculated absolute and relative errors. The linear-scaling model consistently overestimated backfill and underperformed the NULL model in the majority of cases for all three pathogens. In contrast, the prospective-scaling and retrospective-scaling models outperformed the NULL model for influenza and COVID-19 in most locations, but not for RSV. Despite the assumption that backfill should be non-negative, we observed negative backfill in 21 jurisdictions (often early in the season), which reduced model accuracy. Our results suggest that the intuitive linear-scaling approach is unreliable for this dataset, likely due to heterogeneity in reporting-facility catchment areas. While the retrospective-scaling model yielded the most accurate estimates, its reliance on complete-season data limits operational use. The prospective-scaling model, which requires only current data, provided the most reliable real-time estimates for influenza and COVID-19. These findings support the use of prospective-scaling for operational forecasting and highlight the need for probabilistic nowcasting methods.