Community-acquired pneumonia identification from electronic health records in the absence of a gold standard: A Bayesian latent class analysis

Abstract

Community-acquired pneumonia (CAP) is common and a significant cause of mortality. However, CAP surveillance commonly relies on diagnostic codes from electronic health records (EHRs), with imperfect accuracy. We used Bayesian latent class models with multiple imputation to assess the accuracy of CAP diagnostic codes in the absence of a gold standard and to explore the contribution of various EHR data sources in improving CAP identification. Using 491,681 hospital admissions in Oxfordshire, UK, from 2016 to 2023, we investigated four EHR-based algorithms for CAP detection based on 1) primary diagnostic codes, 2) clinician-documented indications for antibiotic prescriptions, 3) radiology free-text reports, and 4) vital signs and blood tests. The estimated prevalence of CAP as the reason for emergency hospital admission was 13.6% (95% credible interval 13.3-14.0%). Primary diagnostic codes had low sensitivity but a high specificity (best fitting model, 0.275 and 0.997 respectively), as did vital signs with blood tests (0.348 and 0.963). Antibiotic indication text had a higher sensitivity (0.590) but a lower specificity (0.982), with radiology reports intermediate (0.485 and 0.960). Defining CAP as present when detected by any algorithm produced sensitivity and specificity of 0.873 and 0.905 respectively. Results remained consistent using alternative priors and in sensitivity analyses. Relying solely on diagnostic codes for CAP surveillance leads to substantial under-detection; combining EHR data across multiple algorithms enhances identification accuracy. Bayesian latent class analysis-based approaches could improve CAP surveillance and epidemiological estimates by integrating multiple EHR sources, even without a gold standard for CAP diagnosis.

Fig 1. Study inclusion and exclusion flowchart.

Fig 2. Predicted sensitivity, specificity, PPV, and NPV of the four algorithms (a, b) and algorithms combinations (c, d) under Model-3 in the primary analysis.

Algorithm combinations used an ‘OR’ rule for three or four positive outcomes. PPV: positive predictive value; NPV: negative predictive value. Codes: CAP primary diagnostic codes; Indication: CAP antibiotic indication; Radiology: chest X-ray report or CT scan report; Tests: shortness of breath and elevated C-reactive protein levels.

Fig 3. Observed and predicted frequency of each algorithm combination of community-acquired pneumonia (CAP) in Model-3 in the primary analysis.

The columns indicate specific combinations, with filled circles denote an algorithm being positive in each combination. Codes: CAP primary diagnostic codes; Indication: CAP antibiotic indication; Radiology: chest X-ray report or CT scan report; Tests: shortness of breath and elevated C-reactive protein levels.

Fig 4. Proportion of positive results in each algorithm and latent prevalences by year.

Codes: CAP primary diagnostic codes; Indication: CAP antibiotic indication; Radiology: chest X-ray report or CT scan report; Tests: shortness of breath and elevated C-reactive protein levels; Latent: latent prevalence estimated from the model. Shaded area shows the 95% credible intervals for the latent prevalence estimated from the model.