123 Deep-Learning Generated B-Line Score Mirrors Clinical Progression of Disease for Patients Admitted With Shortness of Breath or Hypoxia

Abstract

Thoracic ultrasound can evaluate the lungs for fluid in the alveolar and interstitial spaces of the lung, based on the presence of artifacts known as B-lines. Degree of B-lines may provide an independent measure of severity of pulmonary edema in congestive heart failure. Artificial intelligence can be used to provide objective data, improve interpretations by novice users, and allow processing of large amount of data for research. The aim of this study is to determine the association between a B-line score using a deep learning algorithm with the degree of pulmonary congestion based on clinical assessment and its variation with response to treatment. This single-center, prospective, observational study enrolled adult patients presenting with dyspnea or hypoxia and suspected to have congestive heart failure at an urban academic ED from July 2018 to May 2019. Subjects found to have any B-lines on initial lung ultrasound were followed through their hospital course and included in this analysis. Admitted patients underwent daily ultrasonography. Clips of 8 lung zones (right and left anterior/lateral and superior/inferior) were obtained using the Philips LumifyTM ultrasound system with sector probe. Demographics, vital signs, lab values, and imaging results were also collected. A previously published deep learning algorithm was refined to rate the severity of B-lines on a scale of 0-4 in 0.5 increments. Average severity for all 8 zones at each time-point for each patient was calculated. A clinical congestion score (0-18 scale), based on presence of dyspnea, orthopnea, fatigue, JVD, rales and pedal edema was calculated. Mixed effects modeling was used to investigate the association between B-line score as independent variable and clinical congestion score as dependent variable, accounting for variation on the person level and longitudinal data. Missing data was excluded from analysis. Covariates tested included patient and exam level data (sex, age, presence of selected comorbidities, baseline sodium and hemoglobin, creatinine, vital signs, oxygen delivery amount and delivery method, diuretic dose). There were 242 unique subjects with 722 complete sets of 8-zone clips included in the analysis. For the mixed effects model, chosen covariates were: sex, baseline sodium, baseline hemoglobin, respiratory rate, CHF, chronic obstructive pulmonary disease, fraction of inspired oxygen, oxygen delivery method (eg, nasal cannula, non-rebreather, ventilator), total loop diuretic dose on day of exam, and the 8-zone DL obtained B-line score. There was no collinearity. The 8-zone B-line score was significantly associated with clinical congestions core, with coefficient of 0.73 (p<0.001). A deep learning generated B-line score mirrors clinical course for subjects admitted patients with shortness of breath or hypoxia, independent of other patient and clinical factors. B-line quantification using AI can provide an objective and independent measure of pulmonary congestion in patients with congestive heart failure and can be used to help monitor clinical progress.