DA1 Artificial Neural Network Meta-Models in Cost-Effectiveness analysis of intensive Blood-Glucose Control: A Case Study Applied to the UK prospective Diabetes Study (UKPDS) Individual Patient outcome Simulation Model

Abstract

Within cost effectiveness analysis, joint uncertainty in costs and effects is commonly dealt with using probabilistic sensitivity analysis (PSA). Although economic models using patient level data can simulate more complex disease processes than cohort-based models, the computational time required to eliminate 1st-order uncertainty often makes extensive PSA impossible. To overcome this, a non-parametric artificial neural network (ANN) simulation meta-modelling method is presented using a case study that evaluates the cost-effectiveness of intensive blood-glucose monitoring in patients with type 2 diabetes. A complex individual patient simulation model (UKPDS Outcome Model version 1.3) was used with quality adjusted life years (QALY) and cost of complications as model outputs. To reduce 1st-order uncertainty, 1000 patients were simulated for each input combination selected. ANN simulation meta-models using a sample of 200 individual runs were developed and cross-validated to approximate the original simulation as these do not require any specific input-output functional relationship and can handle any number of input parameters. Performance was compared with a Gaussian Process (GP) meta-model, and a valid and better predictive meta-model was then used for PSA. From ANN meta-models, the mean absolute percentage error (defined as positive difference between the predicted and true output divided by the range in true output) was 3.8 % for costs and 1.4% for QALYs compared with 5.1% and 2.1% in GP meta-models. The distribution of errors was approximately symmetrical around zero meaning that mean costs and QALYs for an intervention are unlikely to be affected by the small inaccuracies associated with ANN approximations. ANN produces better predictive capability than GP meta-models in estimating costs and QALYs from the UKPDS outcome model. A PSA carried out using the ANN meta-model demonstrated the potential for ANN in analysing complex health economic models.