Nowcasting Models for Public Health Decision-Making Show Promise with Dynamic Parameter Selection

Strategic Insights -

In public health, data availability significantly limits decision-making, particularly during epidemics. Traditional methods may not accurately reflect the trajectory of an outbreak. Nowcasting models can estimate eventual case counts by accounting for reporting delays but have limitations when applied to multiple surveillance systems. Researchers sought to address these problems by developing a flexible nowcasting framework. They used a Bayesian nowcasting tool to dynamically estimate delay probabilities and selected parameters such as maximum delay and training window based on recent data.

Key Takeaways:

  • The use of recent data to estimate dynamic delay and training window parameters resulted in nowcasts with less error compared to those made with static parameters for long historic periods.
  • Nowcasts likely to fail can be predicted a priori using metrics such as the relative width of prediction intervals and permutation entropy of the epidemic trend.
  • Dynamic parameter selection improves nowcast performance, while more complex models may not necessarily lead to better results.
  • Collaboration with surveillance colleagues is necessary to implement data-driven choices that enhance the utility of predictions for decision-making.
  • Implementing a system to suppress nowcasts likely to fail can improve the reliability of public health decision-making.
  • Flexible nowcasting models can be applied to various epidemic diseases, including COVID-19.

Statistics:

  • 321 nowcasts were generated and evaluated for COVID-19 cases in six U.S. states and dengue cases in Puerto Rico.
  • 90%, 95%, and 99% quantile distribution of the most recently reported data were used to set maximum delay values.
  • The maximum delay values were increased by 1.5 or 2.0 times the most recent delay value in some scenarios.
  • Logarithmic scoring and coverage metrics were used to assess prediction error and precision for the most recent three weeks of predictions.
  • Permutation entropy of the epidemic trend was used to predict nowcasts likely to fail.

Sources:

  • medrxiv.org/content/10.1101/2024.11.09.24315999v2