A cost-effectiveness analysis of AI-assisted OCT classification for age-related macular degeneration

Background: The accuracy of an artificial intelligence diagnostic solution, regardless of how high it may be, holds limited practical value without economic viability. A highly accurate system that proves financially unsustainable offers little benefit to healthcare institutions operating under resource constraints. To demonstrate this principle, age-related macular degeneration (AMD) was selected as an example condition for analysis. AMD represents the leading cause of irreversible blindness in adults over 50 years old in developed countries, and optical coherence tomography (OCT) has become the gold standard imaging modality for its diagnosis and monitoring. However, manual OCT interpretation by retinal specialists faces challenges including inter-observer variability, time constraints, and delayed detection of early-stage disease. While machine learning approaches have demonstrated promising diagnostic performance for AMD detection, comprehensive economic evaluation remains essential to justify implementation in real-world clinical environments. Our goal is to evaluate whether machine learning-assisted OCT classification offers a cost-effective alternative to traditional specialist interpretation for AMD detection.

Methods: A cost-effectiveness analysis was conducted comparing machine learning-assisted AMD classification to traditional manual OCT interpretation using deliberately conservative parameter estimates. Performance metrics were degraded from validated results (99.50% sensitivity, 99.80% specificity) to account for real-world deployment variability. Annual costs were modeled across institutional volumes ranging from 5,000 to 100,000 OCT scans, incorporating conservative implementation costs ($1.2 million) and annual maintenance ($175,000). Monte Carlo simulation with 10,000 iterations assessed model robustness across clinically plausible parameter ranges, while tornado analysis identified key cost drivers.

Results: For a representative institution processing 10,000 annual scans with 67% disease prevalence, machine learning assistance generated $7.78 million in annual savings (29.7% cost reduction). False negative errors decreased from 6% baseline to 0.50%, with false negative cost reduction accounting for 97.8% of total error cost savings. The system achieved a 1.85-month payback period and 549% first-year return on investment. Monte Carlo simulation demonstrated 100% probability of positive return across all iterations (95% confidence interval: $6.18 to $8.94 million). Savings scaled linearly across all evaluated volumes without minimum volume thresholds.

Conclusions: Machine learning-assisted OCT classification for AMD generates substantial, robust cost savings through improved diagnostic accuracy that reduces costly false negative errors. The rapid payback period and strong return on investment establish compelling economic justification for implementation across healthcare settings of varying sizes, even under conservative assumptions that account for real-world performance degradation. These findings illustrate the broader principle that rigorous economic analysis is essential for validating any AI diagnostic system prior to clinical deployment.