C35-05 Evaluating Real World Impact: A Computer Vision CT Algorithm Improves Emphysema Detection and Averts Acute Care Utilization

Abstract Rationale Chronic obstructive pulmonary disease (COPD) and emphysema affect approximately 15.7 million Americans and impose significant clinical burden. More than half of adults with impaired lung function may be undiagnosed. Using a real-world dataset, this analysis modeled how a computer vision algorithm analyzing computed tomography (CT) scans could potentially improve emphysema detection and longitudinal patient outcomes. Methods The ClearRead CT | Lung Tissue Analysis Prototype algorithm (Riverain Technologies) was applied retrospectively to patients with chest CT scans, no prior documented diagnosis of emphysema, and minimum six-months of pre- and post-CT follow-up data from linked Electronic Health Records (EHR) within Dandelion Health (a real-world longitudinal clinical data platform). Algorithm predictions, validated by an external panel of thoracic radiologists, identified patients who were not clinically diagnosed (based on the presence of ICD10 code, J43%, or radiology report mention) after their scan despite having emphysema present. To understand the impact of missed diagnoses, real world respiratory-related healthcare resource utilization (HRU) was used to compare trajectories of patients with diagnosed versus undiagnosed emphysema following their chest CT scan. After balancing cohorts using inverse probability treatment weighting, two clinical pathways were modeled: an ‘Algorithm Pathway’ using the algorithm’s 61.9% sensitivity from the validation study and a ‘Conventional Pathway’ using literature-derived 40.6% sensitivity. Each pathway’s outcomes were calculated as weighted averages of observed patient HRU based on their respective sensitivity rates (i.e., ‘Algorithm Pathway’: 61.9% diagnosed, 38.1% undiagnosed). Subgroup analyses evaluated differences by sex, age, race/ethnicity and COVID-19 pandemic timing. Results The 21.3% absolute improvement in emphysema diagnosis within the ‘Algorithm Pathway’ was associated with 51.1 averted acute care events per 10,000 patients annually as compared to the ‘Conventional Pathway’: 23.5 for acute respiratory failure and 27.6 for COPD exacerbations. Inpatient admissions for acute respiratory failure and COPD exacerbations were reduced by an average of 17.3% per year, although mean length of stay was similar for both pathways (3.8 vs 3.3 days). Emergency department visits for these same conditions decreased by 32.9% (p 0.05) in the ‘Algorithm Pathway’. Undiagnosed emphysema was particularly associated with increased acute care utilization among patients who were non-white, male, under 65, and scanned pre-COVID-19. Conclusions Incidental emphysema detection facilitated by artificial intelligence (AI)-enhanced CT interpretation was associated with reductions in resource-intensive respiratory care, particularly among underrepresented populations. This study establishes a framework for evaluating the real-world impact of AI by retrospectively applying algorithms to imaging data linked to longitudinal clinical outcomes. This abstract is funded by: Dandelion Health, Riverain Technologies