This is the code repository for the paper:
Radiological predictors of PCP in HIV-positive adults in South Africa: a retrospective cohort study (RadPredict)
Nicola K Wills, Jared Tavares, Qonita Said-Hartley, Sean Wasserman
Chest X-ray (CXR) is a cost-effective tool for PCP recognition. Accurate definition of discriminatory CXR features that correlate with laboratory-confirmed PCP in HIV-positive adults are needed to refine current diagnostic approaches.
We conducted a retrospective analysis of medical records, with independent blinded review of admission CXRs by a specialist radiologist, of laboratory-confirmed PCP cases and non-PCP controls, matched by CD4 count and admission-year, and identified through systematic screen of samples sent for Pneumocystis jirovecii testing from HIV-positive adults admitted with respiratory complaints at Cape Town Metro hospitals (2012 – 2020). We explored CXR features associated with definite and severe PCP (defined by marked hypoxia, ICU referral/admission, and/or in-hospital death). Using a backward stepwise approach, we explored the performance of binary logistic regression models, incorporating select clinical and CXR predictors, for PCP diagnosis and severe PCP.
104 adults were included (52 PCP cases matched to 52 non-PCP controls), with significantly higher median respiratory rate (34 versus 28 breaths/minute, p = 0.003) and hypoxia (44% versus 28%, p = 0.0003) in the PCP group. Diffuse versus patchy ground glass opacification was associated with a significantly increased odds of PCP (OR 5.7, 95% confidence interval (CI) 1.6 – 28.9, p = 0.01) and severe PCP (OR 4.5, 95%CI 1.6 – 14.4, p = 0.008). Any consolidation was associated with severe PCP (OR 3.3, 95%CI 3.3 (1.2 – 11.0), p = 0.03). Pleural effusion (OR 0.1, 95%CI 0.0 – 0.4, p = 0.01), reticular/reticulonodular changes and lymphadenopathy were seen infrequently in PCP. Two regression models, incorporating either hypoxia or elevated respiratory rate (≥ 30 breaths/minute) with diffuse ground glass changes, absence of pleural effusion or absence of reticular/reticulonodular changes (respiratory rate model only) were found to perform well in predicting laboratory-confirmed PCP (area under the receiver operating characteristic curve (AUC) of 0.828 (hypoxia model) and 0.857 (respiratory rate model)), with robust internal validation. There were no candidate variables sufficiently predictive to build a severe PCP prognostic model.
CXR evaluation, together with bedside clinical cues, offers good accuracy (AUC > 0.8) in discriminating definite PCP from other HIV-associated respiratory diseases.
A metadata clone of the RedCap project can be found in the RadPredict_2024-02-07_1337.REDCap XML file.