Predicting SMC Appraisal Outcomes Using Machine Learning: A Model to Support Early Market Access Strategy
Author(s)
Samuel Aballea, MSc, PhD1, Zeineb Hammami, ENG2, Przemyslaw Lipka, .1, Bechir Mlika, ENG2, Mondher Toumi, MD, Msc, PhD3.
1InovIntell, Kraków, Poland, 2InovIntell, Tunis, Tunisia, 3Laboratoire de Santé Publique, Aix-Marseille University, Marseille, France.
1InovIntell, Kraków, Poland, 2InovIntell, Tunis, Tunisia, 3Laboratoire de Santé Publique, Aix-Marseille University, Marseille, France.
OBJECTIVES: For health technology developers (HTDs), the ability to anticipate HTA outcomes during early development stages can improve decision-making. While standard approaches for predicting HTA outcomes rely on qualitative methods, machine learning (ML) offers an opportunity to build scalable, data-driven predictive tools. This study aimed to develop and validate a ML model to predict the outcomes of appraisals by the Scottish Medicines Consortium (SMC), based on structured information about the product and its target indication.
METHODS: We used data from the NaviHTA database, which contains structured information extracted from European HTA reports since 2018. A total of 364 SMC appraisals comprising 385 decisions were included, with 14% resulting in product rejection. The binary outcome was classified as "recommended" or "rejected", with separate predictions generated for distinct subpopulations when relevant. The model used Python with tree-based gradient boosting (CatBoost) and 70/30 train-test cross-validation. Sixty-six predictive features were initially considered, including standardized treatment effect measures. Models were developed with and without incremental cost-effectiveness ratio (ICER) as predictive feature. Feature importance was assessed using Shapley values.
RESULTS: The final model without ICER included 34 predictive features. It achieved 87% accuracy and detected 53% of rejected cases. When ICER was included, the recall rate for rejections increased to 71%. Of the 11 misclassified cases, 8 were linked to incomplete information in SMC reports, and could be resolved by providing complete input sets to the model. Key features driving predictions included target population size, comparator relevance, ICER, type of hypothesis tested, and study design.
CONCLUSIONS: This ML model demonstrates robust predictive capability for SMC decisions using information typically available in target product profiles, combined with ICER estimates. The model predictions and associated influence score can support HTDs in refining clinical and market access strategies, ultimately improving alignment with payer expectations and patient benefit.
METHODS: We used data from the NaviHTA database, which contains structured information extracted from European HTA reports since 2018. A total of 364 SMC appraisals comprising 385 decisions were included, with 14% resulting in product rejection. The binary outcome was classified as "recommended" or "rejected", with separate predictions generated for distinct subpopulations when relevant. The model used Python with tree-based gradient boosting (CatBoost) and 70/30 train-test cross-validation. Sixty-six predictive features were initially considered, including standardized treatment effect measures. Models were developed with and without incremental cost-effectiveness ratio (ICER) as predictive feature. Feature importance was assessed using Shapley values.
RESULTS: The final model without ICER included 34 predictive features. It achieved 87% accuracy and detected 53% of rejected cases. When ICER was included, the recall rate for rejections increased to 71%. Of the 11 misclassified cases, 8 were linked to incomplete information in SMC reports, and could be resolved by providing complete input sets to the model. Key features driving predictions included target population size, comparator relevance, ICER, type of hypothesis tested, and study design.
CONCLUSIONS: This ML model demonstrates robust predictive capability for SMC decisions using information typically available in target product profiles, combined with ICER estimates. The model predictions and associated influence score can support HTDs in refining clinical and market access strategies, ultimately improving alignment with payer expectations and patient benefit.
Conference/Value in Health Info
2025-11, ISPOR Europe 2025, Glasgow, Scotland
Value in Health, Volume 28, Issue S2
Code
HTA272
Topic
Health Technology Assessment, Methodological & Statistical Research
Topic Subcategory
Decision & Deliberative Processes
Disease
No Additional Disease & Conditions/Specialized Treatment Areas