77 Background: Pembrolizumab showed promising activity in patients with advanced non–small cell lung cancer (NSCLC) in the KEYNOTE-001 study (NCT01295827) and significantly prolonged overall survival (OS) compared with docetaxel in the randomized KEYNOTE-010 study (NCT01905657). Responses to pembrolizumab appear to be remarkably durable, making long-term survival possible in some patients. Typical parametric survival models do not account for the possibility of long-term survival. An alternative, well-established class of statistical models called long-term survival models can be used to directly estimate the percentage of patients achieving long-term survival (>5 years), called “long-term survival rate.” (Berkson J, Gage RP. J Am Stat Assoc. 1952;47:501-515.)(Tsodikov AD et al. J Am Stat Assoc. 2003;98:1063-1078.) Methods: Data from patients with PD-L1–expressing (tumor proportion score ≥1%), previously treated, advanced NSCLC in KEYNOTE-001 and KEYNOTE-010 were used. KEYNOTE-001 data were used to initially estimate the long-term survival rate of pembrolizumab, while KEYNOTE-010 data were used for subsequent independent validation. Point estimates of long-term survival rates with their 95% CIs based on the model described above are reported. Results: Based on the long-term survival model, the estimated long-term survival rate in pembrolizumab-treated population in KEYNOTE-001 (n = 306) is 25.4% (95% CI, 15.2%-33.3%) and the intention-to-treat population who received pembrolizumab in KEYNOTE-010 (n = 690) is 25.3% (95% CI, 8.9%-36.9%). In contrast, the long-term survival rate of docetaxel arm (n = 343) in KEYNOTE-010 is estimated to be 3.2% (95% CI, 0%-17.4%). Conclusions: In 2 independent data sets, it is estimated that 25% of patients with previously treated PD-L1–expressing NSCLC may achieve long-term benefit from pembrolizumab monotherapy. Long-term survival models can provide an early estimate of long-term benefit from pembrolizumab using data with limited follow-up time. Long-term follow-up from these trials will further validate this finding. MDH and JM are co-first authors.
Long-term survival models with latent activation under a flexible family of distributions