Analyzing and Forecasting Pediatric Fever Clinic Visits in High Frequency Using Ensemble Time-Series Methods After the COVID-19 Pandemic in Hangzhou, China: Retrospective Study

Abstract

Background: The COVID-19 pandemic has significantly altered the global health and medical landscape. In response to the outbreak, Chinese hospitals have established 24-hour fever clinics to serve patients with COVID-19. The emergence of these clinics and the impact of successive epidemics have led to a surge in visits, placing pressure on hospital resource allocation and scheduling. Therefore, accurate prediction of outpatient visits is essential for informed decision-making in hospital management.

Objective: Hourly visits to fever clinics can be characterized as a long-sequence time series in high frequency, which also exhibits distinct patterns due to the particularity of pediatric treatment behavior in an epidemic context. This study aimed to build models to forecast fever clinic visit with outstanding prediction accuracy and robust generalization in forecast horizons. In addition, this study hopes to provide a research paradigm for time-series forecasting problems, which involves an exploratory analysis revealing data patterns before model development.

Methods: An exploratory analysis, including graphical analysis, autocorrelation analysis, and seasonal-trend decomposition, was conducted to reveal the seasonality and structural patterns of the retrospective fever clinic visit data. The data were found to exhibit multiseasonality and nonlinearity. On the basis of these results, an ensemble of time-series analysis methods, including individual models and their combinations, was validated on the data set. Root mean square error and mean absolute error were used as accuracy metrics, with the cross-validation of rolling forecasting origin conducted across different forecast horizons.

Results: Hybrid models generally outperformed individual models across most forecast horizons. A novel model combination, the hybrid neural network autoregressive (NNAR)-seasonal and trend decomposition using Loess forecasting (STLF), was identified as the optimal model for our forecasting task, with the best performance in all accuracy metrics (root mean square error=20.1, mean absolute error=14.3) for the 15-days-ahead forecasts and an overall advantage for forecast horizons that were 1 to 30 days ahead.

Conclusions: Although forecast accuracy tends to decline with an increasing forecast horizon, the hybrid NNAR-STLF model is applicable for short-, medium-, and long-term forecasts owing to its ability to fit multiseasonality (captured by the STLF component) and nonlinearity (captured by the NNAR component). The model identified in this study is also applicable to hospitals in other regions with similar epidemic outpatient configurations or forecasting tasks whose data conform to long-sequence time series in high frequency exhibiting multiseasonal and nonlinear patterns. However, as external variables and disruptive events were not accounted for, the model performance declined slightly following changes in the COVID-19 containment policy in China. Future work may seek to improve accuracy by incorporating external variables that characterize moving events or other factors as well as by adding data from different organizations to enhance algorithm generalization.

Keywords: COVID-19; hospital management; long sequence; outpatient visits; pediatric fever clinic; time-series forecasting; visits in high frequency.

Figure 1

Fever clinic deployment instructions in the designated hospital.

Figure 2

The ratio of fever clinic visits to the total number of outpatient visits over the past 4 years.

Figure 3

Time-series plot of hourly fever clinic visits from January 23, 2020, to May 23, 2023.

Figure 4

Daily seasonal patterns in the hourly fever clinic visit time series.

Figure 5

Weekly seasonal patterns in the hourly fever clinic visit time series.

Figure 6

Yearly seasonal patterns in the hourly fever clinic visit time series.

Figure 7

Autocorrelation function (ACF) and partial autocorrelation function (PACF) of the hourly fever clinic visit time series.

Figure 8

Autocorrelation function (ACF) and partial autocorrelation function (PACF) of the seasonal differential hourly fever clinic visit time series.

Figure 9

Seasonal-trend decomposition using Loess on the hourly fever clinic visit time series.

Figure 10

Comparison of 15-day forecasts generated using single models with actual observations. NNAR: neural network autoregressive; SARIMA: seasonal autoregressive integrated moving average; SNaive: seasonal naive; STLF: seasonal and trend decomposition using Loess forecasting.

Figure 11

Comparison of 15-day forecasts generated using hybrid models with actual observations. NNAR: neural network autoregressive; SARIMA: seasonal autoregressive integrated moving average; STLF: seasonal and trend decomposition using Loess forecasting.

Figure 12

Illustration of the rolling forecasting origin cross-validation methodology.

Figure 13

Root mean squared error (RMSE) values calculated for different forecast horizons ranging from 1 to 30 days ahead. NNAR: neural network autoregressive; SARIMA: seasonal autoregressive integrated moving average; SNaive: seasonal naive; STLF: seasonal and trend decomposition using Loess forecasting.

Figure 14

Mean absolute error (MAE) values calculated for different forecast horizons ranging from 1 to 30 days ahead. NNAR: neural network autoregressive; SARIMA: seasonal autoregressive integrated moving average; SNaive: seasonal naive; STLF: seasonal and trend decomposition using Loess forecasting.

Figure 15

Monthly aggregated root mean squared error (RMSE) and mean absolute error (MAE) box plot.