Leveraging Machine Learning for Predicting Health Technology Assessment Outcomes
Speaker(s)
Lin R1, Eaves K1, Mills M2, Kanavos P3
1Hive Health Optimum Ltd., Pimlico, LON, UK, 2Hive Health Optimum Ltd., LONDON, LON, UK, 3London School of Economics and Political Science, London, LON, UK
Presentation Documents
OBJECTIVES: Heterogeneity in HTA outcomes across settings contributes to disparities in patient access to innovative medicines. This study develops a predictive model for HTA outcomes of innovative medicines across 5 countries, using advanced machine learning techniques.
METHODS: Data on drug characteristics, clinical evidence, economic evidence, disease characteristics, regulatory pathways, and firm characteristics were extracted for HTA decisions, spanning 2009 - 2024 from an internal HTA database (HTA-HIVE). Scope was limited to HTA agencies conducting cost-effectiveness analysis. Both supervised and unsupervised models were investigated to determine their effectiveness in predicting HTA outcomes. The data was preprocessed and split into training and testing sets, with additional steps to address class imbalances. Hyperparameter tuning and cross-validation were employed to ensure model generalisation and prevent overfitting. Multivariate logistic regression, random forest, and gradient boosting analyses were conducted to identify the most influential factors in predicting HTA outcomes. Feature importance was evaluated, and unsupervised learning models, including PCA and k-means clustering, were applied to uncover patterns and group similarities.
RESULTS: In a preliminary analysis on 550 HTA outcomes, 15% received a positive recommendation, 71% received a restricted recommendation, and 14% a negative recommendation. Key predictors of HTA outcome included therapeutic area, resubmissions, study design, ICER and HTA agency. Amongst the preliminary predictive models, logistic regression models had the highest performance (with F1 scores ranging from 85-90%), followed by random forest models (65-75%). Unsupervised learning identified clusters with high positive outcome rates, particularly those involving agencies like CADTH, with therapeutic areas predominantly in oncology.
CONCLUSIONS: Improving the predictability of HTA outcomes can lead to more efficient and effective assessments, ensuring timely access to new treatments for patients. This study highlights the potential of machine learning in HTA data analysis, offering a framework for more accurate and timely evaluations, ultimately benefiting patient care and health outcomes.
Code
HTA64
Topic
Health Technology Assessment, Methodological & Statistical Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics, Decision & Deliberative Processes, Systems & Structure
Disease
No Additional Disease & Conditions/Specialized Treatment Areas