Advancing Cost-Effectiveness Analysis through Generative AI in Health Economics and Outcomes Research (HEOR)

Abstract

This research hypothesizes and empirically employs a model that incorporates generative artificial intelligence (AI) to come up with a cost-effectiveness analysis (CEA) of Health Economics and Outcomes Research (HEOR) by using Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs). The framework produces synthetic patient-level outcomes using nationally representative Medical Expenditure Panel Survey (MEPS) data to address specific limitations of real-world data (sparsity, missingness, no counterfactuals, etc.) and decision analysis (probabilistic). Rigorous preprocessing and model training (Adam, α = 0.001, early stopping, distributional checking using KL/JS divergence) were performed, then synthetic cohorts were mixed with observed data to predict incremental costs, QALYs, and uncertainty. The intervention based on generative-AI was superior to conventional care with reduced mean costs (USD 11,997.74 vs USD 13,150.32) and increased mean QALYs (0.83 vs 0.79), with an ICER of -31,637 USD/QALY. Heterogeneity was indicated by percentile analyses of right-skewed cost and QALY gains with small negative tails. Cost-Effectiveness Acceptability Curves (2,000 bootstraps) showed a high likelihood of cost-effectiveness, at traditional willingness-to-pay (WTP) levels (95% at or below 50,000 USD/QALY; 99% at or below 100,000 USD/QALY). One-way sensitivity analyses ranked WTP and QALY gain as the most significant sources of Incremental Net Monetary Benefit and showed average influence on cost variations; in all studied situations, INMB was positive. Other than the empirical findings, the paper contains a taxonomy of the uses of generative AI in HEOR and an explanation of ethical and methodological issues. The results imply that generative modeling may enhance CEA in terms of robustness, transparency, and policy relevance, but that synthetic data use should be validated in different settings and its use should be subject to regulation.