OBJECTIVES: In Germany all new innovative medicines are subject to an early benefit assessment by the German Federal Joint Committee (G-BA) with subsequent price negotiation and optional arbitration.

This study aims to explore the various data-based modeling techniques to predict post-AMNOG (Arzneimittelmarkt-Neuordnungsgeset) annual cost of treatment (COT) for oncology products.

METHODS: The modeling approach utilizes historical pricing data and G-BA ratings available within Eversana’s NAVLIN database.

The oncology dataset includes 321 data-points spanning 40 products and 71 indications launched from 2018-2023, that have undergone the AMNOG process and have an annual COT.

Initially, a One-Way Analysis of Variance (ANOVA) was conducted on G-BA ratings with respect to annual COT. Next, machine learning predictive algorithms – including four different types of regression methods, k-nearest neighbors (KNN) and decision trees – were tested on standardized data.

Models were cross-validated and trained on historical data from 2018-2020, and further tested on product launches starting 2021. Accuracy and reliability of models was assessed using adjusted R-squared (R-sq), Mean Absolute Error (MAE), Root Mean Squared Error (RMSE) and Mean Absolute Percentage Error (MAPE).

RESULTS: ANOVA test established a relationship between G-BA and post-AMNOG COT (F-value 2.2107 and p-value 0.0419). On training dataset, Decision Tree model provided the best adjusted R-sq (0.901) and lowest MAPE (10.6%) scores. The gamma regression model had the lowest comparative performance based on all metrics.

On test-data (2021-2022), the Huber regression model performed the best with adjusted R-sq of 0.987 and MAPE of 8.9%. The gamma model had the lowest performance, while KNN and decision tree models also performed poorly.

CONCLUSIONS: The Huber Regression model displayed best performance in predicting post-AMNOG price for products launched in 2021 -22.

This flexible yet rigorous framework can be modified to include more independent variables, understand their effect on launch prices and evaluate algorithms for predictive modeling of COT.