Can Machine Learning Accurately Predict Payer Behavior?

OBJECTIVES: Machine learning can be applied to predict outcomes. However, within market access, there is limited application due to complexities of payer decision-making. This research aimed to train and test the accuracy of machine learning models based on regression algorithm in forecasting the price of an orphan product in England, France, and Germany.

METHODS: A MCDA model with key attributes reflecting payer drivers for orphan products was developed. Value scores were derived using HTA documents and annual COT (list price) was calculated for all products assessed by each HTA and reimbursed since 2016. Pegcetacoplan, a recently assessed orphan product, was removed for testing. Datasets were analysed using Python libraries and run in JupyterLab. Model training included various algorithm prediction, Spearman correlation analysis, linear and polynomial regression models. Pegcetacoplan price was predicted and compared to actual price.

RESULTS: Most variables were positively correlated to price in France and the UK; UK specific variables (CEA threshold and HST appraisal) provided the greatest correlation. All Germany variables were negatively correlated or had no correlation to price, reflecting evidence uncertainty. Linear regression models had low R² (France=0.17; Germany=0.31; UK=0.41). Trained linear regression models predicted pegcetacoplan price within -24% (Germany), -49% (France) and -113% (UK) of actual price. Polynomial models in France and the UK were overfitting (R²=1.0) and unsuitable for predictions. Power 2 and 3 polynomial models in Germany improved accuracy (R²=0.39 and 0.63), but provided less accurate pegcetacoplan price predictions (-102% and -197%).

CONCLUSIONS: Analyses suggested that inclusion of all reimbursed products can be used to train models with high predictability. However, due to the nuances of payer decision-making, applying these models to real scenarios have varying accuracy. Limitations included small sample, inclusion of high cost ATMPs and confidential discounting. Further research on relevant analogues and in other indications is required.