Comparative Evaluation of Machine Learning and Parametric Survival Modelling in Cost-Effectiveness Analysis: A Case Study of First-Line Pembrolizumab Plus Chemotherapy for Non-Small Cell Lung Cancer in India

Objectives: Survival extrapolation is a consequential choice in oncology cost-effectiveness analysis (CEA), and conventional parametric approaches have known limitations under non-proportional hazards, common with immune checkpoint inhibitors. This study tested whether machine learning (ML) survival models offer a useful alternative to parametric extrapolation using KEYNOTE-189 data, and quantified how survival-model choice affects the incremental cost-effectiveness ratio (ICER) of first-line pembrolizumab plus chemotherapy for advanced non-small cell lung cancer (NSCLC) from an Indian healthcare perspective. Methods: Individual patient-level data were reconstructed from published Kaplan–Meier curves using the Guyot algorithm. Five parametric distributions and two ML approaches (random survival forest [RSF] and penalised Cox regression with elastic-net regularisation) were fitted and compared on a held-out test set using Harrell’s C-statistic and the integrated Brier score (IBS). A three-state Markov model used survival-derived transition probabilities, Indian drug-acquisition costs, and published EQ-5D utility values to estimate lifetime costs and quality-adjusted life-years (QALYs), with uncertainty assessed via probabilistic sensitivity analysis (PSA), deterministic sensitivity analysis, and expected value of perfect information (EVPI). Results: The RSF achieved a C-statistic of 0.641 and IBS of 0.1748, comparable to the best-performing parametric models (log-logistic, log-normal) and modestly ahead of the conventionally selected Weibull distribution (C = 0.638). The base-case ICER was ₹1,05,51,767 per QALY with Weibull-derived transitions versus ₹97,96,723 with RSF-derived transitions, a 7.2% reduction. Across 10,000 PSA iterations, the probability of cost-effectiveness at a ₹5,00,000/QALY threshold was near zero under both models; population EVPI was negligible at that threshold but peaked near ₹34,773 crore around ₹1.06 crore/QALY. Drug acquisition cost was the dominant driver of ICER uncertainty, with survival-model choice ranking third among seven parameters examined. Conclusion: Machine learning survival modelling produced a modest but consistent improvement in predictive accuracy over standard parametric distributions, translating into a non-trivial reduction in the estimated ICER. The reimbursement conclusion that pembrolizumab plus chemotherapy is not cost-effective at prevailing Indian prices was robust to survival-model choice, though the price reduction needed to reach cost-effectiveness varied by approach. Survival-model choice merits explicit consideration as a structural uncertainty in oncology health technology assessment, particularly for therapies with non-proportional hazard patterns.