52: PROSPECTIVE REAL-TIME VALIDATION OF A LUNG ULTRASOUND DEEP LEARNING MODEL IN THE INTENSIVE CARE UNIT

Abstract

Introduction: Lung ultrasound (LUS) shows diagnostic superiority to chest radiography in the ICU setting however its widespread deployment is challenged by limited user base and its dependence on clinical expertise. Artificial intelligence models have been developed to overcome these barriers through LUS automation; however, bedside pragmatic performance has never been demonstrated. In this study, we aimed to evaluate the accuracy of a bedside, real-time deployment of a deep learning model capable of distinguishing between normal (A line pattern) and abnormal (B line pattern) lung parenchyma on lung ultrasound in critically ill patients. Methods: Prospective, observational study evaluating the performance of a previously-trained LUS deep learning (DL) model on 100 admitted patients to an academic intensive care unit (ICU). Enrolled patients received an anterior zone LUS exam (4 views) with frame-by-frame DL model predictions using a portable device. Clip-level model predictions were analyzed and compared to blinded expert review for A vs B line pattern. Four clip-level prediction cutoffs were applied to maximize model sensitivity and specificity at bedside. Results: A total of 100 unique ICU patients (400 clips) were enrolled from two tertiary care sites. The majority of patients were receiving mechanical ventilation (56/100), 12 patients were in the weaning process with ongoing trach-mask trials, 4 patients were on High-Flow Nasal Cannula, 4 were on non-invasive ventilation (BiPAP), and 25 patients were on nasal prongs. The Mindray M9 and Sonosite X-Porte ultrasound machines were used to scan 50 patients each. When compared to gold-standard expert annotation, the real-time inference yielded an accuracy of 95%, sensitivity of 93%, and specificity of 96% for identification of the B line pattern. Varying the prediction cutoffs reveal that diagnostic performance may be readily customized to suit the clinical environment. Conclusions: Promoting the feasibility and direct application of bedside AI in the ICU, we have shown that a binary DL ultrasound model demonstrates excellent bedside accuracy in distinguishing between normal and abnormal lungs when deployed in real-time in the ICU.