TY - JOUR T1 - Analysis of infection biomarkers within a Bayesian framework reveals their role in pneumococcal pneumonia diagnosis in HIV patients JF - bioRxiv DO - 10.1101/070144 SP - 070144 AU - Austin G. Meyer Y1 - 2016/01/01 UR - http://biorxiv.org/content/early/2016/08/18/070144.abstract N2 - Background HIV patients are more likely to contract bacterial pneumonia and more likely to die from the infection. Unfortunately, there are few tests to quickly diagnosis the etiology of these dangerous infections. Several biomarkers may be useful for diagnosing the most common pneumonia-causing organism, S. pneumoniae, but studies utilizing the standard statistical approach provide little concrete guidance for the HIV-infected population.Methodology and Findings Using a Bayesian approach, I analyze data from a cohort of 280 HIV patients with x-ray confirmed community acquired pneumonia. First, I use a variety of techniques to establish predictor significance and to identify their optimal cutoffs. Next, in lieu of cutoffs, I find the continuous and combined likelihood ratios for every value of each biomarker, and I compute the associated posttest probabilities. As expected, I find the three biomarkers with good clinical yield and a statistically significant association with S. pneumoniae are C-reactive protein (CRP), procalcitonin (PCT), and lytA gene PCR (lytA). Based on Bayesian clinical yield, optimal cutoffs are largely equivocal. The optimal dichotomous cutoff for CRP is essentially any value between 10 mg/dL and 30 mg/dL (△pPosttest ≈ 0.49). The optimal cutoff for PCT is any value between 2 ng/mL and 40 ng/mL (△pposttest ≈ 0.35). The optimal cutoff for lytA is any value less than 6 log10 copies/mL (△pposttest ≈ 0.45). Further, I find that continuous likelihood ratios provide much more accurate posttest probabilities than dichotomous cutoffs. For example, starting with the empirical pretest probability, a lytA approaching 0 copies/mL lowers the probability of S. pneumoniae infection to less than 15%, while a result of 10 copies/mL raises the probability to greater than 65%. However, a lytA value just above or below the suggested cutoff of 8000 copies/mL or my new optimal cutoff of 30,000 copies/mL leaves the posttest probability of infection essentially unchanged from the pretest probability.Conclusion CRP, PCT, and lytA all provide significant value in diagnosing the etiology of pneumonia in HIV patients. The optimal dichotomous cutoffs for lytA, CRP, and PCT need to be adjusted for pneumococcal diagnosis in this population. However, continuous and combined likelihood ratios avoid discarding valuable quantitative information, and a combined likelihood ratio can be easily computed without the need for prior logistic regression. Importantly, there is significant overlap between these biomarkers such that only one of the three biomarkers at a time should be used to update clinical probabilities. Thus, it is ill-advised to combine the likelihood ratios of different biomarkers to produce a posttest probability. Finally, I provide a simple web application to quantitatively calculate the posttest probability of S. pneumoniae infection in HIV patients with x-ray confirmed pneumonia: http://meyerapps.org/pneumococcal_etiology_hiv. ER -