Background: Budget impact models (BIMs) forecast the financial implications of adopting new technologies and the potential need for budget reallocation, thus playing a crucial role in reimbursement decisions. Despite the importance of accurate forecasts, studies indicate large discrepancies between estimates and reality. We are developing an artificial intelligence-based clinical decision tool to identify patients with non-small cell lung cancer who are most likely to benefit from immunotherapy.
Objective: To evaluate the budgetary implications and describe a systematic literature review of published lung cancer BIMs.
Methods: We searched PubMed and EMBASE for studies published between 2010 and 2023 that include BIMs that describe lung cancer interventions. Forward and backward reference searches were performed for all qualifying studies. We extracted author and publication year, country, interventions, disease stages, time horizon, analytical perspective, modeling methods used, types of costs included, sensitivity analyses conducted, and data sources used. We then evaluated adherence to the Professional Society for Health Economics and Pharmacoeconomics Research best-practice guidelines.
Results: A total of 25 BIMs were identified, spanning 14 different countries. Model structure could not be ascertained definitively for nearly half of the models. The cost calculator approach was most common among the others. Time horizons ranged from 1 to 5 years, in line with recommendations. Most models compared drugs, 4 compared nondrug interventions, and 7 compared diagnostic technologies. Assumptions about market uptake were poorly documented and poorly motivated. Inclusion of cancer-related costs was rare. Adherence to best practices was variable and did not appear to improve over time.
Conclusions: The number of published BIMs for lung cancer exceeded expectations. There were modest trends toward publication frequency and model quality over time. Our analysis revealed variability across the models, as well as their adherence to best practices, indicating substantial room for improvement. Although none of the models were individually suitable for the purpose of evaluating an artificial intelligence-based treatment selection tool, some models provided valuable insights.