Post-pandemic forecasting of pediatric acute respiratory infections with deep learning: a multi-pathogen,...

Area of Science:

• Epidemiology
• Infectious Diseases
• Data Science

Background:

• Acute respiratory infections (ARIs) are a major global cause of childhood illness and hospitalization.
• The post-COVID-19 era presents new challenges for forecasting ARIs due to shifting pathogen patterns.
• Accurate forecasting is crucial for effective clinical and public health preparedness.

Purpose of the Study:

• To compare the predictive performance of traditional statistical and deep learning models for pediatric ARIs.
• To enhance the accuracy of ARI forecasting for improved public health and clinical decision-making.
• To evaluate forecasting models across various pathogens, time horizons, and error metrics.

Main Methods:

• Utilized 29,260 pediatric hospitalization records (2021-2024) with pathogen screening data.
• Excluded Chlamydia pneumoniae due to low case numbers, analyzing 10 respiratory pathogens.
• Evaluated 13 time-series models (including SARIMA and 11 deep learning models) for short-, medium-, and long-term forecasting.

Main Results:

• Deep learning models showed lower error rates than SARIMA in 72.2% of comparisons.
• DLinear excelled in short-term forecasting (e.g., human bocavirus, human coronavirus); TSMixer led in long-term forecasting.
• No single model consistently outperformed others for medium-term forecasting; performance varied by pathogen and horizon.

Conclusions:

• Deep learning models offer substantial improvements for pediatric ARI pathogen surveillance in the post-pandemic era.
• Tailored model selection based on pathogen and forecasting horizon is essential for timely interventions and resource allocation.
• Enhanced forecasting supports clinical decisions, public health strategies, and vaccination planning.