Presentation (CTI)

OBJECTIVES: Patient-level and discrete-event simulation (DES) models capture heterogeneity and event-history dependence that cohort Markov models cannot, but their runtime makes full probabilistic sensitivity analysis (PSA) and value-of-information (VOI) computation - particularly EVPPI/EVSI - frequently infeasible. We developed and validated a machine-learning emulator (surrogate model) that makes these analyses practical.
METHODS: Using a space-filling (Latin hypercube) design over the model's input distributions, we ran the DES at 1,000 parameter sets and trained emulators (Gaussian process regression and gradient-boosted trees) to predict incremental costs, QALYs and net benefit. Predictive accuracy was assessed by out-of-sample cross-validation and probabilistic calibration/coverage. The validated emulator then generated the full PSA, cost-effectiveness acceptability curves, EVPI and EVPPI, with EVPPI cross-checked against established non-parametric regression methods (GAM; GP/INLA). Case study: a patient-level DES of treatment sequencing in metastatic colorectal cancer.
RESULTS: Emulators predicted incremental net benefit with out-of-sample R²=0.98 and 95% prediction-interval coverage of 94%. Full PSA requiring ~38 hours of direct simulation (10,000 runs) was reproduced by the emulator in under 10 minutes; emulator-based EVPPI agreed with GAM/GP reference estimates within 3%, while nested Monte Carlo EVPPI on the native model was computationally infeasible (estimated >2 months). Predictive accuracy degraded only in the extreme tails of time-to-event parameters.
CONCLUSIONS: A validated ML emulator reproduces patient-level simulation outputs with high fidelity and cuts the cost of rigorous uncertainty and value-of-information analysis by orders of magnitude, bringing EVPPI/EVSI within reach for complex individual-level models. This supports better-targeted research prioritisation and generalises across disease areas and model types.