Acute diarrheal syndromic surveillance: effects of weather and holidays

Abstract

Objective: In an effort to identify and characterize the environmental factors that affect the number of patients with acute diarrheal (AD) syndrome, we developed and tested two regional surveillance models including holiday and weather information in addition to visitor records, at emergency medical facilities in the Seoul metropolitan area of Korea.

Methods: With 1,328,686 emergency department visitor records from the National Emergency Department Information system (NEDIS) and the holiday and weather information, two seasonal ARIMA models were constructed: (1) The simple model (only with total patient number), (2) the environmental factor-added model. The stationary R-squared was utilized as an in-sample model goodness-of-fit statistic for the constructed models, and the cumulative mean of the Mean Absolute Percentage Error (MAPE) was used to measure post-sample forecast accuracy over the next 1 month.

Results: The (1,0,1)(0,1,1)7 ARIMA model resulted in an adequate model fit for the daily number of AD patient visits over 12 months for both cases. Among various features, the total number of patient visits was selected as a commonly influential independent variable. Additionally, for the environmental factor-added model, holidays and daily precipitation were selected as features that statistically significantly affected model fitting. Stationary R-squared values were changed in a range of 0.651-0.828 (simple), and 0.805-0.844 (environmental factor-added) with p<0.05. In terms of prediction, the MAPE values changed within 0.090-0.120 and 0.089-0.114, respectively.

Conclusion: The environmental factor-added model yielded better MAPE values. Holiday and weather information appear to be crucial for the construction of an accurate syndromic surveillance model for AD, in addition to the visitor and assessment records.

Keywords: Surveillance; diarrhea; emergency service hospital; environment; forecasting.

Fig. 1

Overview of model construction and evaluation process. NEDIS = National Emergency Department Information System; ED = Emergency Department; KMA = Korea Meteorological Administration; MAPE = Mean Absolute Percentage Error; AD = Acute Diarrhea.

Fig. 2

A schematized method for the gradual period extensions. We gradually increased the data for model training with 1-month intervals (starting at May 1, 2008, which became the 13^th month from the initial data) to the month just prior to the target prediction period, and constructed new best-fitting prediction models for each of the following months.

Fig. 3

Sequence graph of actual visiting AD patient (A), and after applying the first-order seasonal differencing (B). The number of visiting patients had a tendency to increase with passing time, and thus we applied the logarithm transformation. Additionally, first-order seasonal differencing was also applied to remove the 7-day periodicity: a pattern graph that moved around the 0 point, as shown in graph (B). AD = Acute Diarrhea.

Fig. 4

Ratio of AD patients by day of the week. The ratio of AD patients to total ED patients evidenced a tendency to increase on weekends, particularly on Sundays. AD = Acute Diarrhea; T = Korean Thanksgiving period (lunar); NL = New Year’s Day (lunar).

Fig. 5

In-model goodness-of-fit values (stationary R-squared) for the constructed models. The stationary R-squared values changed between 0.651 and 0.828 (for the simple model) and between 0.805 and 0.844 (for the environmental factor-added model) with p<0.05.

Fig. 6

Cumulative mean graphs for MAPEs of actual and predicted visits during the prediction period (May 1, 2008 to April 30, 2009). The result and estimation of predictions of the numbers of visiting patients from May, 2008 to April, 2009 were shown with cumulative mean of MAPE. The environmental factor-added model predicted the number of visiting patients more accurately than the simple models. MAPE = Mean Absolute Percentage Error; AD = Acute Diarrhea.