Diagnostic Accuracy of Automated Pneumothorax Detection Via Novice-Acquired Ultrasound After Chest Tube Removal: Comparison With Expert Interpretation and Chest Radiography.
Authors
Affiliations (4)
Affiliations (4)
- Schulich School of Medicine, Western University, London, ON, Canada.
- Deep Breathe Inc, London, ON, Canada.
- Centro de Investigaciones Clínicas, Fundación Valle del Lili, Cali, Colombia.
- Division of Critical Care Medicine, Western University, London, ON, Canada.
Abstract
Pneumothorax (PTX) is a frequent complication after chest tube removal, and timely detection is important to inform monitoring and potential intervention. Chest radiograph (CXR) remains the standard modality after chest-tube-removal PTX detection, despite its limited sensitivity and frequent delays in acquisition. Lung ultrasound (LUS) has superior accuracy and portability but is highly operator dependent, limiting its usability. The objective of this study is to evaluate whether artificial intelligence-assisted LUS (AI-LUS) enables novice users to accurately detect PTX after chest tube removal, compared with expert interpretation and CXR. Does AI-LUS improve the ability of novice users to detect findings associated with PTX after chest tube removal compared with expert interpretation and CXR? We conducted a prospective diagnostic accuracy study in adult patients undergoing chest tube removal at a tertiary academic hospital. LUS clips were acquired by novice operators using a handheld ultrasound device. A previously trained artificial intelligence model was then calibrated and used to detect the absence or presence of lung sliding. The reference standard was expert consensus LUS interpretation, with CXR serving as a secondary reference standard. Sensitivity and specificity were calculated at 2 time points: immediately after removal and after routine CXR. A total of 76 patients were enrolled, yielding 848 LUS clips across 2 time points. Data from the first 12 patients were used to calibrate the model, with the remaining 64 forming the validation cohort. Compared with expert LUS interpretation, AI-LUS demonstrated a sensitivity of 0.775, a specificity of 0.831, and a negative predictive value of 0.96 for identifying absent lung sliding. When compared with CXR, AI-LUS achieved a sensitivity of 1.0 immediately after removal and 0.923 after CXR for PTX detection. Our results show that novice-performed AI-LUS demonstrated moderate diagnostic accuracy for detecting absent lung sliding after chest tube removal. Its very high sensitivity and excellent negative predictive value for identifying cases with absent lung sliding associated with PTX relative to CXR highlights a potential role for AI-LUS as a rapid triage tool that may reduce reliance on routine CXR, while acknowledging that PTX inference requires clinical correlation and additional ultrasound findings.