Predicting the hotspots of age-adjusted mortality rates of lower respiratory infection across the continental United States: Integration of GIS, spatial statistics and machine learning algorithms. (October 2020)

• Record Type: Journal Article
• Authors: Mollalo, Abolfazl; Vahedi, Behrooz; Bhattarai, Shreejana; Hopkins, Laura C.; Banik, Swagata; Vahedi, Behzad
• Is Part Of: International journal of medical informatics. Volume 142(2020)
• Publication Date: 2020-10
• Subjects: Accuracy assessment -- Decision trees -- GIS -- Hotspots -- Lower respiratory infections -- US; Medical informatics -- Periodicals; Information science -- Periodicals; Computers -- Periodicals; Medical technology -- Periodicals; Medical Informatics -- Periodicals; Technology, Medical -- Periodicals; Computers; Information science; Medical informatics; Medical technology; Electronic journals; Periodicals; Electronic journals; 610.285
• DOI: 10.1016/j.ijmedinf.2020.104248
• Languages: English
• ISSNs: 1386-5056

Abstract

Highlights: Lower respiratory infections (LRI) are the cause of a significant number of hospitalizations in the US. No previous nationwide study examined geographic variations of LRI mortality rates and their association with underlying factors. There was a shift in the location of LRI hotspots from west coast to southeast over time. Decision tree classifiers could predict LRI mortality hotspots with high accuracies. Higher spring temperature and increased precipitation during winter were among the most substantial predictors of presence or absence of LRI hotspots. Abstract: Objective: Although lower respiratory infections (LRI) are among the leading causes of mortality in the US, their association with underlying factors and geographic variation have not been adequately examined. Methods: In this study, explanatory variables (n = 46) including climatic, topographic, socio-economic, and demographic factors were compiled at the county level across the continentalUS.Machine learning algorithms - logistic regression (LR), random forest (RF), gradient boosting decision trees (GBDT), k-nearest neighbors (KNN), and support vector machine (SVM) - were employed to predict the presence/absence of hotspots (P < 0.05) for elevated age-adjusted LRI mortality rates in a geographic information system framework. Results: Overall, there was a historical shift in hotspots away from the western US into the southeastern parts of the country and they were highly localized in a few counties. The two decision tree methods (RF and GBDT) outperformed the other algorithms (accuracies: 0.92; F1-scores: 0.85 and 0.84; area under the precision-recall curve: 0.84 and 0.83, respectively). Moreover, the results of the RF and GBDT indicated that higher spring minimum temperature, increased winter precipitation, and higher annual median household income were among the most substantial factors in predicting the hotspots. Conclusions: This study helps raise awareness of public health decision-makers to develop and target LRI prevention programs.