When and where: Spatiotemporal machine learning forecasts of respiratory infection-related primary care visits.

OBJECTIVES

Viral respiratory infections (VRI) spread through both temporal and spatial processes, yet most short-term forecasts rely solely on temporal trends. We assessed whether incorporating regional signals improves forecasts of primary care visits for respiratory syndromes.

METHODS

Using population-based administrative health data from British Columbia, Canada, we extracted daily visits of VRI-based syndromes across local health areas from February 2022 to April 2024. Temporal and geospatial lagged features capturing spatiotemporal dependence were incorporated into machine learning models, including boosted trees, random forests, neural networks, and regularized linear models. Models were trained using blocked time-series cross-validation, and 28-day-ahead forecasts were evaluated using the mean absolute scaled error relative to a naïve temporal baseline. Additional sensitivity analyses were conducted including type of geospatial features.

RESULTS

The addition of geospatial features improved the forecast accuracy for respiratory symptoms, primarily for COVID-19 for both adult and pediatric populations. Boosted trees often achieved the highest forecast accuracy relative to other models and the naïve forecaster. The degree of model performance varied by population density.

CONCLUSIONS

Incorporating regional signals can enhance short-term forecasts of respiratory-related primary care visits, though benefits are context-specific and depend on the resolution of the signal. Accounting for spatial structure in epidemic forecasting may support more effective public health planning and surveillance.