Deep learning assisted scoring of point of care lung ultrasound for acute decompensated heart failure in the emergency department - since the onset of the covid-19 pandemic, the practice of "boarding"…

Deep learning assisted scoring of point of care lung ultrasound for acute decompensated heart failure in the emergency department - since the onset of the covid-19 pandemic, the practice of "boarding" patients admitted to the hospital in the emergency department (ED) has reached unprecedented levels. For critically ILL patients including those with acute decompensated heart failure (adhf), ED boarding worsens outcomes as patients spend hours in the ED waiting to be transferred to the appropriate inpatient ward for specialized care. Given the unabated increase in ED boarding, length of ED stay, and subsequent time to specialist evaluation and management, developing new technologies to enable rapid reassessment of adhf patients during these protracted ED stays is critical for improved care and patient outcomes. In a typical workflow in the emergency department, physicians perform bedside lung ultrasound once, at time of initial patient presentation, and use the presence or absence of 'b- lines' in the images as a biomarker for pulmonary congestion. Often assessed by ED physicians in a binary manner, the presence of b-lines is used in conjunction with a clinical exam and blood tests to rule in acute adhf. While detecting b-lines can be as easy as looking at two lung zones to make a clinical decision of adhf, counting b-lines requires both skill and training in b-line identification, and in aggregating b-line counts over 8+ lung zones for accuracy. For a busy ED physician this is prohibitive given constraints on time, training, and cognitive load. To ease this problem, ED physicians need tools that can automatically count and aggregate the b-lines to quantify the severity of the congestion. Without this automation, IT is entirely possible that either suboptimal or even no treatment will be initiated for adhf patients in the ED leading to increased hospital length of stay, further perpetuating the ED boarding. The creation of tools for automatic quantification has the potential to enable workflows with reassessment to meet the changing patient care needs. Our long-term goals are to develop computational tools that mitigate the operator-dependence endemic to ultrasound image acquisition and interpretation. The objective of this trailblazer R21 application is to develop and validate computational methods for quantifying pulmonary congestion from bedside lung ultrasound in the ED, which will be achieved by (1) developing and evaluating explainable tools for automated quantification of pulmonary congestion using retrospective lung ultrasound data and (2) validating the performance of the trained models in a workflow demonstrated by a prospective observational study in which patients presenting to the ED with adhf will be assessed with lung ultrasound both pre-and post-therapeutic intervention, and findings typically used to measure pulmonary congestion on inpatient services will be recorded for both time points.