P164 Using artificial intelligence to interrogate multi-national imaging datasets to determine the mechanism of COVID-19 pneumothorax. (11th November 2022)

• Record Type: Journal Article
• Authors: Selby, IA; Driggs, D; Majcher, V; Roberts, M; Escudero Sanchez, L; Rudd, JHF; Sala, E; Bibiane-Schönlieb, C; Marciniak, SJ; Babar, J
• Is Part Of: Thorax. Volume 77(2022)Supplement 1
• Page Start: A169
• Page End: A170
• Publication Date: 2022-11-11
• Subjects: Chest -- Diseases -- Periodicals; Thorax; Chest -- Diseases; Periodicals; Periodicals; 617.54
• DOI: 10.1136/thorax-2022-BTSabstracts.298
• Languages: English
• ISSNs: 0040-6376

Abstract

Abstract : Introduction: Pneumothorax is a rare but important complication of COVID-19. 1 Although barotrauma may account for some cases, many affected patients have not received positive-pressure ventilatory (PPV) support 1 . The pathophysiology of COVID-pneumothorax is challenging to investigate because imaging data exist in diverse silos and only 0.97% of patients admitted for COVID-19 experience this complication. 1 To provide mechanistic insight, we used artificial intelligence at scale to identify cases for detailed analysis from 4 large imaging datasets across 26 centres in 7 countries. Methods: A convolutional neural network was trained to detect pneumothorax on chest x-rays (CXRs) using the open-source CheXpert dataset, which includes 17, 313 pneumothoraces. Testing was performed on labelled subsamples of the COVID-19 datasets. After running the model on all COVID-positive CXRs, predicted pneumothoraces were reviewed and the incidence of COVID-pneumothorax was estimated. Available CTs for patients with pneumothorax were assessed by radiologists. Radiology reports were used to curate additional CTs for two datasets. Results and Discussion: Quantitative results are summarised in figure 1 . Adjusting for model sensitivity, the estimated incidence of COVID-pneumothorax was 0.97%, consistent with previous research. 1 45 pneumothorax patients with CTs were identified; however, 13 unrelated to COVID-19, and 9 iatrogenic cases (except barotrauma) were excluded. Almost all remaining patients displayed diffuse, moderate-to-severe pneumonitis. Most pneumothoraces in patients on PPV were likely related to an interplay of barotrauma and COVID-19, with an acute lung injury pattern on CT. A high proportion demonstrated emphysema and three patients developed cystic abnormalities. One case followed a cavitating pulmonary infarction secondary to pulmonary embolism. Patients who had not received PPV, or had but were stepped down, developed pneumothoraces later in the disease. CT showed patterns consistent with the absorption stage of COVID-19, where consolidation is reduced but ground glass opacification persists with development of irregular bronchial dilatation. Such pneumothoraces perhaps represent increased parenchymal resistance. Conclusion: There are multiple mechanisms of COVID-pneumothorax. Barotrauma in patients with acute lung injury is most common, whilst pneumothorax in the absence of PPV most commonly occurs in the sub-acute, absorption stage of the disease. Reference: Marciniak SJ, et al . COVID-19 Pneumothorax in the United Kingdom. ERJ 2021. Please refer to page A215 for declarations of interest related to this abstract.