View Abstract

A5514 - A Prediction Model for Pneumonia Risk Incorporating Clinical and Proteomic Features
Author Block: M. Lee1, Y. Zuo2, J. P. Mizgerd3, K. A. Steiling4, A. J. Walkey5; 1Pulmonary Center, R304, Boston University School of Medicine, Boston, MA, United States, 2Center for Clinical Translational Epidemiology and Comparative Effectiveness Research, Boston University School of Medicine, Boston, MA, United States, 3Boston Univ School of Med, Boston, MA, United States, 4Boston University School of Medicine, Boston, MA, United States, 5Pulmonary, Boston University School of Medicine, Boston, MA, United States.
RATIONALE: Improved understanding of the risk factors for pneumonia can better align human disease with animal models, increase understanding of pneumonia pathophysiology, and may allow for innovations in efforts to prevent pneumonia. We sought to develop a model for estimating 10-year risk of developing pneumonia using clinical and serum protein measurements. METHODS: Participants (age≥65 years) in the Framingham Heart Study (Offspring Cohort) were linked to their Centers for Medicare Services (CMS) claims data. Subjects without link to CMS data, with prior pneumonia or who were not alive at the exam 7 protein immunoassay panel measurement (conducted from 1998 to 2001) were excluded. We abstracted demographics, clinical comorbidities, measures of functional status, and levels of 88 serum protein immunoassays from exam 7 data. Subjects were followed for up to 10 years for an incident pneumonia diagnosis identified via Medicare claims. We used Fine-Grey competing risk models to account for competing risk of death, multiple imputation to account for missing data (10%), and the least absolute shrinkage and selection operator machine approach to select model covariates. Finally, we tested whether adding protein biomarkers to clinical data could improve pneumonia prediction.
RESULTS: We identified 1370 study participants with immunoassay data and linkage to CMS claims data. During 10 years of follow up, 501 (36%) had a pneumonia diagnosis. We found baseline age (subdistribution hazard ratio [SHR], 1.06; 95% CI, 1.04-1.08), atrial fibrillation (SHR, 1.43; 95% CI, 1.06-1.93), chronic pulmonary disease (SHR, 1.87; 95% CI, 1.33-2.61), diabetes (SHR, 1.36; 95% CI, 1.05-1.75), heart failure (SHR, 1.74; 95% CI, 1.10-2.74), hospitalization within 1 year (SHR, 1.34; 95% CI, 1.09-1.65) and current smoking status (SHR, 1.79; 95% CI, 1.31-2.45) were predictive of pneumonia development. Five of the 88 baseline serum protein measurements, including C-reactive protein (CRP) (SHR, 1.17; 95% CI, 1.06-1.28), were predictive of the pneumonia development. The addition of baseline CRP to a clinical model improved the predictive model (Akaike information criterion [AIC] value of 4950 from 4960 and C-statistic of 0.64 from 0.62). CONCLUSION: Predictive models can identify novel risk factors for pneumonia, including serum protein measures, and identify high-risk patients for prevention efforts. CRP, a routinely available biomarker of chronic inflammation, improves pneumonia risk prediction over clinical factors alone.