Predictive extrapolation : a framework to reduce uncertainty around long-term treatment effects of innovative cancer therapies

Immuno-oncology (IO) is a novel cancer treatment that activates the immune system to target tumor cells. IO has demonstrated a superior efficacy in clinical trials, including patients with metastatic non-small cell lung cancer (mNSCLC). However, assessing the cost-effectiveness of IO is challenging due to uncertainty around IO real-world impacts and its long-term treatment benefits beyond the trial period. Standard parametric distribution (SPD) models, commonly used for survival extrapolation, may have limitations in survival projections based on solely the immature trial data. Furthermore, although the treatment landscape for mNSCLC has evolved, the costs associated with novel treatments remain insufficiently understood. This thesis aims to address these challenges and explore methodological solutions applicable to health economic evaluations of novel treatments.

Paper I investigated treatment patterns following the introduction of IO for mNSCLC and its impact, using the Flatiron Health oncology database comprising over 10,000 patients. The study compared real-world overall survival (rwOS) between patients receiving IO and those treated with chemotherapy (CT), using propensity score matched cohorts. The HR for rwOS was 0.887 (95% CI: 0.850- 0.978) in the first-line therapy and 0.775 (95% CI: 0.737-0.875) in the second-line therapy after adjusting for crossover effects. IO was associated with statistically significant survival benefits sustained over several years, though uncertainty increased with longer follow-up. These findings support the clinical effectiveness of IO in real-world practice while highlighting challenges in long-term outcome evaluation.

Paper II evaluated survival extrapolation methods in 11,224 patients with HER2- positive breast cancer using the National Breast Cancer Register in Sweden. SPD models were compared to a novel modelling approach using excess hazard (EH) models with or without a cure assumption. An EH model is a method to implement the relative survival framework by incorporating background mortality in modelling to prevent unrealistic survival projections. Survival models were developed by stage based on truncated datasets to reflect immature trial data, and compared to Kaplan-Meier data observed up to 15 years. Extrapolations were then extended over a 50-year horizon to assess between-model variance in restricted mean survival time (RMST). EH cure models closely aligned with observed mid-term survival in stage I - III, while EH no cure models performed better in stage IV. In long-term survival projections, SPD models produced implausible estimates in early-stage cancer, exceeding survival in the general population. EH models substantially decreased between-model variance for RMST, for instance, by 93.7% using EH no cure models and 83.5% using EH cure models, compared to SPD models in stage II. These results suggest that the greater internal consistency of EH models may reduce structural uncertainty in long-term survival extrapolation based on immature data.

Paper III examined treatment cost trends in 17,107 patients initially diagnosed with mNSCLC in Sweden from 2011 to 2020, using data from the National Registers. Patients were grouped by diagnosis period and treatment type including EGFR- targeted, ALK-targeted, IO, and CT drugs alone. The use of IO increased sharply after 2016, overtaking CT drugs alone by 2019, while EGFR- and ALK-targeted therapies showed gradual uptake. First-year mean costs rose from €26,640 in 2011-2013, €37,199 in 2014 - 2016 to €61,053 in 2017-2020, mainly due to increased costs of IO drug and inpatient care. IO patients had the highest first- year mean cost (€105,287), with costs declining in subsequent years. In contrast, costs for other groups remained relatively stable beyond the first year.

Paper IV assessed cost-effectiveness using a partitioned survival model comparing pembrolizumab with CT in mNSCLC, based on digitally reconstructed data derived from the KEYNOTE-024 study. SPD and EH models were fitted to both the 11-month interim and the 5-year update data. With the interim data, exponential and log-normal models performed equally well in terms of the AIC and the BIC, but only log-normal projections aligned with the 5-year data. However, the SPD log-normal models projected long survival tails, deemed overly optimistic. In contrast, the EH log-normal models provided plausible extrapolations with tapering tails. EH models yielded consistent ICERs across data maturities, as SEK 412,611 per QALY using the interim data and SEK 458,946 per QALY using the update data, and reduced uncertainty compared to SPD models in probabilistic sensitivity analyses.

The research highlights the challenges and advantages of using EH models compared to SPD models, notably in achieving more consistent long-term survival extrapolation and reducing uncertainty in ICER estimation. The findings also emphasize the importance of understanding long-term real-world treatment benefits and substantial costs of novel treatments.

List of scientific papers

I. Long-term real-world survival of immunotherapy compared to chemotherapy for metastatic non-small cell lung cancer: a propensity score-matched analysis. Kim K, Sweeting M, Jönsson L, Wilking N. Thorac Cancer. 2025 Jan;16(1):e15535. https://doi.org/10.1111/1759-7714.15535

II. General population mortality adjustment in survival extrapolation of cancer trials: exploring plausibility and implications for cost- effectiveness analyses in HER2-positive breast cancer in Sweden. Kim K, Sweeting M, Wilking N, Jönsson L. Med Decis Making. 2024 Oct;44(7):843-853. https://doi.org/10.1177/0272989x241275969

III. Cost of treatment of metastatic non-small lung cancer in Sweden, 2011-2023. Kim K, Sweeting M, Wilking N, Jönsson L. [Manuscript]

IV. A re-evaluation of cost-effectiveness of immunotherapy in advanced non-small cell lung cancer, using excess hazard models, in a case study of the KEYNOTE-024 trial. Kim K, Sweeting M, Wilking N, Jönsson L. [Manuscript]