Research motivation and background
The authors previously presented a proof-of-concept neural network-based metamodel at ISPOR US, designed to map the relationships between the inputs and outputs of a complex model of the prevalence of chronic kidney disease. This model decreased runtime dramatically, facilitating near-instantaneous computation of results. The work presented here extends on this by including SHAP analysis as a form of sensitivity analysis of the model to understand the importance of each input on each output. 

Originality and contribution
Machine learning methods are underutilised within health economics. The authors introduce both neural network-based metamodels and SHAP analysis as an alternative way of understanding the sensitivity of complex health economics models.

Impact and methodology
Neural network-based metamodels can significantly reduce the computational cost of health economic models, which facilitates near-real-time interrogation of results and potentially allows for complex models to be deployed and run in the field. Furthermore, SHAP analysis increases the transparency of models by explicitly indicating the influence of each parameter on the results.

Stakeholder engagement
SHAP analysis provides a universal, flexible, and scalable alternative to traditional sensitivity analyses. Furthermore, by explicitly outlining the influence of each input variable on the outcome, SHAP analysis helps to untangle the key drivers of model outcomes. Lastly, the transparency afforded by SHAP analysis will help to contribute to the uptake of machine learning approaches in health economics and outcomes research.

If you have any questions about the research or would like to discuss your specific challenge and potential solutions, please contact Thomas Padgett, Principal Operational Researcher at HEOR: thomas.padgett@heor.co.uk.