OBJECTIVES: Knowledge and understanding of why a model reaches a certain outcome is defined as model explainability; this considers the relative weight of each variable in reaching the final outcome. The explainability of a cost-effectiveness model is often limited to the use of one-way sensitivity analysis (OWSA). However, the OWSA is also limited to the estimation of the variation of the model output based on a certain percentual variation of a single input parameter. By its nature, OWSA does not account for either correlations between input parameters or non-linear relationships between input parameters and model outputs. An optimal alternative to assessing explainability is the SHAP (SHapley Additive exPlanations) approach derived from coalitional game theory. This method is increasingly employed to explain the output of machine learning models. SHAP values represent the individual contribution of each model parameter to the model prediction, similar to the coefficients of each variable in a linear regression.

METHODS: We present the use of SHAP in explaining cost-effectiveness model outcomes. For this purpose, the results of the probabilistic sensitivity analysis (PSA) of the Markov model were used to calculate SHAP for the estimated ICER. SHAP results are visualised through the beeswarm plot to describe the distribution of SHAP for each input variable and force plots to describe the contribution of each input variable to the calculated ICER for a full set of inputs.

RESULTS: Marked differences were observed in the ranking of the variables importance, and relative conclusions drawn, between OWSA and the proposed SHAP approach. Different sets of baseline inputs/model structure led to different outcomes.

CONCLUSIONS: The SHAP approach could provide a new and more appropriate way than OWSA to explain the role of each variable in the calculation of cost-effectiveness model outcomes.