Presentation (CTI)

OBJECTIVES: List prices may diverge from net prices after rebates, discounts, clawbacks, confidential agreements, and concessions. This systematic review classified empirical, statistical, and machine-learning methods used to estimate net prices, gross-to-net differences, and list-to-net price trends, and identified evidence gaps. By assessing whether methods can provide reproducible, validated estimates of net-price trends, this review supports more reliable assessment of affordability, budget impact, and whether discounts translate into value for patients and health systems.
METHODS: Searches covered MEDLINE, Embase, EconLit, Scopus/Web of Science, ISPOR abstracts, and grey literature. Eligible records were empirical or methodological analyses of prescription medicines describing an approach for estimating net prices or list-to-net changes. Extraction captured geography, therapeutic area, price definition, data source, estimation approach, rebate treatment, volume weighting, transparency, reproducibility, and cross-market applicability. For machine-learning approaches, fields captured target variable, features, model class, training data, validation, interpretability, missing-data handling, and uncertainty assessment. Methods were grouped by primary strategy and machine-learning use.
RESULTS: Five categories were identified: (i) observed net-price approaches using net sales, volume, claims, or transaction data; (ii) assumption-based gross-to-net and scenario models; (iii) public proxy approaches using tender, reimbursement, ex-manufacturer, or wholesale acquisition cost prices; (iv) financial statement or market-level inference; and (v) machine-learning-enabled estimation. Machine learning was most relevant where net prices were unavailable, supporting imputation of confidential discounts, prediction of net-price ranges, product-mix assessment, triangulation across sources, and uncertainty assessment. Key limitations included inconsistent price definitions, limited rebate data, weak validation, product-mix effects, limited volume weighting, and poor transferability.
CONCLUSIONS: The central challenge is not only confidential data access, but the lack of validated methods for translating partial, proxy, or market-level evidence into reproducible estimates. Machine-learning-enabled approaches address this gap by supporting imputation, triangulation, validation, and uncertainty assessment where data are unavailable. A clearer taxonomy can distinguish true inflation or erosion from method or data artefacts.