Fig 1.
(A) Volume of the ITS network. Volume was determined by dividing the total number of ITS-reported influenza tests in a specified region per epi week by the total number of CDC-reported influenza tests in the same region per epi week and averaging this metric over the length of the analysis. (B) ITS Model 1 adjusted R2 values. Correlation between the estimated proportion of tests positive in the CDC surveillance system (WHO/NREVSS and ILINet) based on ITS Model 1 and the actual proportion of tests positive in the CDC surveillance system. (C) ITS Model 2 adjusted R2 values. Correlation between the estimated ILI-related proportion of physician visits in the CDC surveillance system (WHO/NREVSS and ILINet) based on ITS Model 2 and the actual ILI-related proportion of physician visits in the CDC surveillance system.
These three continental U.S. maps show the volume of the ITS network, the adjusted R2 values for ITS Model 1, and the adjusted R2 values for ITS Model 2. While the volume of the ITS network varies per region, the adjusted R2 values for ITS Model 1 and ITS Model 2 are all high and above 0.80. Hawaii is in Region 9, and Alaska is in Region 10. The ITS network includes fewer influenza tests than the CDC surveillance system, and varies based on HHS region (a). The high correlation between the ITS Models’ estimates of the CDC metrics and the actual CDC metrics suggest that the ITS Models fit the true data well (b and c). Map obtained from [22].
Fig 2.
The estimated influenza tests that are positive developed from the ITS Model 1 track well the actual influenza tests that are positive as reported by the CDC.
ITS Model 1 estimates the CDC proportion of influenza tests that are positive (ILIppt(ta)) by using the proportion of influenza tests that are positive as recorded by ITS (Vppt(ta)), the CDC proportion of influenza tests that are positive with a 1-week lag (ILIppt(t(a−1))), and the absolute value of the difference between the proportion of tests that are positive as recorded by ITS with a 1-week lag and the proportion of influenza tests that are positive as reported by the CDC with a 1-week lag (|Vppt(t(a−1)) − ILIppt(t(a−1))|). Each graph shows two peaks with each peak relating to one flu season which occurs in the winter. The proportion of influenza tests that are positive is along the y-axis. The CDC proportion of influenza tests that are positive is in black, the ITS model estimates are in red, and the 95% prediction intervals are outlined by dark red dotted lines. The epidemiological week (epi week) is along the x-axis and spans from epi week 36 in 2015 to epi week 19 in 2017, except ITS data collection (and thus analysis) began later for Region 2 (epi week 13 in 2016 to epi week 19 in 2017) and Region 10 (epi week 2 in 2016 to epi week 19 in 2017). See Figure C in S1 Text for a visualization of raw data and estimates.
Fig 3.
The estimated weighted ILI-related proportion of physician visits developed from the ITS Model 2 track well the actual weighted ILI-related proportion of physician visits as reported by the CDC.
ITS Model 2 estimates the CDC weighted ILI-related proportion of physician visits (ILIprop(ta)) by using a metric developed from ITS data (the total number of influenza test results divided by the total number of test machines, Vtotal(ta)), the weighted ILI-related proportion of physician visits as reported by the CDC with a 1-week lag (ILIprop(t(a−1))), and the ITS metric (the total number of influenza test results divided by the total number of test machines) with a 1-week lag (Vtotal(t(a−1))). Each graph shows two peaks with each peak relating to one flu season. The weighted ILI-related proportion of physician visits is along the y-axis. The CDC weighted ILI-related proportion of physician visits are in black, the ITS model estimates are in red, and the 95% prediction intervals are outlined by dark red dotted lines. The epidemiological week (epi week) is along the x-axis and spans from epi week 36 in 2015 to epi week 19 in 2017, except ITS data collection (and thus analysis) began later for Region 2 (epi week 13 in 2016 to epi week 19 in 2017) and Region 10 (epi week 2 in 2016 to epi week 19 in 2017). See Figure D in S1 Text for a visualization of raw data and estimates.
Fig 4.
Comparing ITS data based forecast to other real-time data sources.
Each panel compares the log-score difference of ITS forecasts to forecasts based on another data source. Red color: > 0.1 average log-score improvement associated with ITS; Orange: (0, 0.1]; Grey: [−0.1, 0); Black: < −0.1 worse. Values of zero did not occur and white color stands for Reg 2 and 10 in 2016 when no forecasts where calculated. First row: we see clear improvement of ITS compared with ILI as in the Fig 5. Second row: ITS compared with nowcasting source Wikipedia. Third row: ITS compared with ILINearby. Third row: Comparing an ensemble of ITS + ILINearby with Nearby. These comparisons demonstrate that ITS is a valuable addition to presently available nowcasting data sources.
Fig 5.
ITS data can be used to improve the forecasting accuracy of current influenza trends.
Total seasonal log-score gain for predictions using real-time estimates using the ITS data compared to CDC data with a 1 week lag for different forecasting horizons for the 2016-2017 and 2017-2018 season for one region. The model shows the most improvement (i.e., gain in log-score) compared to the CDC data with a 1 week lag when predicting 1 week into the future, but even in predictions with larger horizons, predictions using ITS data were better than predictions using CDC data with a 1 week lag. Boxplots summarizes results from the 11 different geographical areas. Figure B in S1 Text shows details for all areas.