Presentation (CTI)

OBJECTIVES: Artificial intelligence (AI) has been used for disease detection, but evidence of its impact is limited. We present a methodology for evaluating the potential benefit of clinical AI using real-world data and health economic modeling, applied to an imaging algorithm detecting incidental emphysema.
METHODS: The ClearRead CT | LTA prototype algorithm (Riverain Technologies) was applied retrospectively to chest CT scans of patients from Dandelion Health (a longitudinal, multimodal clinical data platform) without a prior emphysema diagnosis. Patients were identified as true-positive (“Early Diagnosis”) (n = 212), diagnosed with emphysema after CT, and false-negative (“Delayed Diagnosis”) (n = 90), not diagnosed after scan but detected by the algorithm with confirmation from a panel of radiologists. Healthcare utilization and estimated costs were calculated for each cohort after balancing with inverse probability treatment weighting. A hybrid decision-tree Markov model with four states (stable, moderate, severe emphysema, and death) projected five-year outcomes from the payer perspective. Model inputs incorporated sensitivity to represent “Early Diagnosis” and “Delayed Diagnosis” pathways. Deterministic and probabilistic sensitivity analyses were conducted by sex, age, race/ethnicity, and pre/post-COVID-19.
RESULTS: The algorithm demonstrated 61.9% sensitivity versus literature-reported 40.6% for conventional radiology. Base case incremental cost-effectiveness ratio was $122,714/QALY. At a $150,000/QALY threshold, the base case had net monetary benefit (NMB) of $624 per patient with 53% probability of cost-effectiveness. Value was higher in non-white patients (NMB: $5,763, 71% probability), males ($4,388, 68%), patients less than 65 years old ($3,665, 64%), and pre-COVID-19 patients ($3,290, 66%). Severe-state costs were the strongest driver of value.
CONCLUSIONS: An imaging algorithm identifying incidental emphysema was found to be cost-effective in multiple scenarios. Health economic modeling combined with multimodal data can assess the potential benefit of clinical AI and identify where deploying AI is cost-effective. As use of clinical AI expands, determining the impact of AI will become increasingly important.