Predicting the start week of respiratory syncytial virus outbreaks using real time weather variables

Abstract

Background: Respiratory Syncytial Virus (RSV), a major cause of bronchiolitis, has a large impact on the census of pediatric hospitals during outbreak seasons. Reliable prediction of the week these outbreaks will start, based on readily available data, could help pediatric hospitals better prepare for large outbreaks.

Methods: Naïve Bayes (NB) classifier models were constructed using weather data from 1985-2008 considering only variables that are available in real time and that could be used to forecast the week in which an RSV outbreak will occur in Salt Lake County, Utah. Outbreak start dates were determined by a panel of experts using 32,509 records with ICD-9 coded RSV and bronchiolitis diagnoses from Intermountain Healthcare hospitals and clinics for the RSV seasons from 1985 to 2008.

Results: NB models predicted RSV outbreaks up to 3 weeks in advance with an estimated sensitivity of up to 67% and estimated specificities as high as 94% to 100%. Temperature and wind speed were the best overall predictors, but other weather variables also showed relevance depending on how far in advance the predictions were made. The weather conditions predictive of an RSV outbreak in our study were similar to those that lead to temperature inversions in the Salt Lake Valley.

Conclusions: We demonstrate that Naïve Bayes (NB) classifier models based on weather data available in real time have the potential to be used as effective predictive models. These models may be able to predict the week that an RSV outbreak will occur with clinical relevance. Their clinical usefulness will be field tested during the next five years.

Figure 1

Copy of one page of the survey given to the expert panel to determine the start dates of the RSV outbreaks. A similar page was provided for each of the years considered in this study.

Figure 2

Schema showing how the 23 years of data available for this study were segregated into high and low peak seasons and training and test data sets. An RSV season was defined as September 20th to July 15th of the following year.

Figure 3

Visual representation of how the input data for the NB classifier are structured.

Figure 4

The true positive fraction (y-axis) is graphed against false positive fraction (x-axis) for NB classifier models for high peak seasons and for every possible variable combination (n = 255) of the weather variables used in this study. The size of each circle represents the number of NB classifier models that obtained the given accuracy. Each model predicts an RSV outbreak the Sunday preceding the outbreak (Week 0), one week in advance (Week 1), two weeks in advance (Week 2), and three weeks in advance (Week 3).

Figure 5

The true positive fraction (y-axis) is graphed against false positive fraction (x-axis) for NB classifier models for low peak seasons and for every possible variable combination (n = 255) of the weather variables used in this study. The size of each circle represents the number of NB classifier models that obtained the given accuracy. Each model predicts an RSV outbreak the Sunday preceding the outbreak (Week 0), one week in advance (Week 1), two weeks in advance (Week 2), and three weeks in advance (Week 3).