Assessing the consistency of overall survival extrapolations in Swedish health technology assessment reports of oncology treatments using extended follow-up data.
Long-term overall survival (OS) extrapolations underpin cost-effectiveness appraisals, projecting survival beyond trial follow-up. We aimed to compare extrapolations presented in manufacturer submissions and appraisals published by the Swedish Dental and Pharmaceutical Benefits Agency (TLV) with extrapolations using OS data with longer follow-up available after the original appraisal.
We retrieved publicly available appraisals for oncology compounds from 2009 to 2019. For each appraisal, Kaplan-Meier curves of OS and base case extrapolation assumptions were collected. Longer-term OS data were identified through literature review and supplemented by manufacturers, then digitized to create pseudo-individual patient data. The extrapolation approach used in each appraisal was reapplied on the follow-up data cut to produce long-term OS predictions. Predictions from the original and follow-up data cuts were compared by area under the curves (AUC) and Akaike information criterion (AIC)-based ranking.
Forty reports were included. The average duration of the original data cut was 28.1 months. Using the follow-up data cut, with an average follow-up of 54.2 months, the average AUC of extrapolated OS estimates increased by 8.8% (95% CI 5.2% to 12.6%) under TLV assumptions and by 10.9% (95% CI 6.9% to 15.1%) under manufacturer assumptions. Thirteen (33%) TLV cases and 14 (35%) manufacturer cases showed AUC increases greater than 10%. In the AIC model-rank analysis, 18 models ranked worse when fitted to the longer-term data than to the original data, 16 ranked the same or better, and six were excluded.
Long-term survival estimates generally increased when original assumptions were applied to follow-up OS data, with material increases (>10%) in roughly one-third of cases, indicating original extrapolations were frequently conservative for the active arm. The findings support revisiting survival model fit, treatment-waning assumptions, and alternative scenarios when longer-term data become available. Validation across both arms is warranted to assess incremental outcomes and cost-effectiveness.
We retrieved publicly available appraisals for oncology compounds from 2009 to 2019. For each appraisal, Kaplan-Meier curves of OS and base case extrapolation assumptions were collected. Longer-term OS data were identified through literature review and supplemented by manufacturers, then digitized to create pseudo-individual patient data. The extrapolation approach used in each appraisal was reapplied on the follow-up data cut to produce long-term OS predictions. Predictions from the original and follow-up data cuts were compared by area under the curves (AUC) and Akaike information criterion (AIC)-based ranking.
Forty reports were included. The average duration of the original data cut was 28.1 months. Using the follow-up data cut, with an average follow-up of 54.2 months, the average AUC of extrapolated OS estimates increased by 8.8% (95% CI 5.2% to 12.6%) under TLV assumptions and by 10.9% (95% CI 6.9% to 15.1%) under manufacturer assumptions. Thirteen (33%) TLV cases and 14 (35%) manufacturer cases showed AUC increases greater than 10%. In the AIC model-rank analysis, 18 models ranked worse when fitted to the longer-term data than to the original data, 16 ranked the same or better, and six were excluded.
Long-term survival estimates generally increased when original assumptions were applied to follow-up OS data, with material increases (>10%) in roughly one-third of cases, indicating original extrapolations were frequently conservative for the active arm. The findings support revisiting survival model fit, treatment-waning assumptions, and alternative scenarios when longer-term data become available. Validation across both arms is warranted to assess incremental outcomes and cost-effectiveness.
Authors
Szilcz Szilcz, Mogard Mogard, Ekman Ekman, Lindgren Lindgren, Nilsson Nilsson, Johannesen Johannesen, Medin Medin
View on Pubmed