This study was a retrospective pooled analysis of individual participant data from the per-protocol populations of the clinical trials OAK16 17 (NCT02008227, July 7, 2016 data cut-off), POPLAR18 (NCT01903993, May 8, 2015 data cut-off), BIRCH4 (NCT02031458, May 28, 2015 data cut-off), and FIR5 (NCT01846416, Jan 7, 2015 data cut-off). Secondary analysis of anonymized clinical trial data was deemed negligible risk research by the Southern Adelaide Local Health Network, Office for Research and Ethics and was exempt from review. Data were accessed according to Roche’s policy and process for clinical study data sharing.19

OAK and POPLAR were randomized trials of atezolizumab 1200 mg intravenous every 3 weeks versus docetaxel 75 mg/m² intravenous every 3 weeks for patients with advanced NSCLC who have failed platinum-containing therapy.16 18 Primary analyses within this study were pooled analyses of OAK and POPLAR which due to randomized design allow the comparison of assessed associations between treatments (atezolizumab vs docetaxel); unless otherwise stated results refer to OAK and POPLAR.

BIRCH and FIR were single-arm studies of atezolizumab 1200 mg intravenous every 3 weeks in PD-L1-positive patients with advanced NSCLC .4 5 Pooled analyses of BIRCH and FIR were used to demonstrate consistency of identified associations within an external cohort of patients treated with atezolizumab.

The coprimary outcomes assessed were OS and PFS. Secondary outcomes assessed were time to deterioration (TDD) in health-related quality-of-life (HRQoL) and physical function (PF).

Best overall response (BOR) and PFS was investigator assessed for POPLAR and OAK and defined by RECIST V.1.1.16 18 An independent review facility assessed BOR and PFS via RECIST V.1.1 for BIRCH,4 and they were investigator assessed per modified RECIST for FIR.5

For all trials, patient-reported outcomes (HRQoL and PF) were assessed via the European Organisation for Research and Treatment of Cancer (EORTC) QLQ-C30 and QLQ-LC13. Study participants completed the QLQ-C30 and QLQ-LC13 at baseline and day 1 of each cycle. PF was measured using a 4-point scale, ranging from ‘not at all’ to ‘very much’. The global health status scale was used to measure HRQoL, and comprised of two questions on a 7-point scale reflecting the patients’ overall health and QoL.17 Responses were scored according to the EORTC scoring manual third edition (standardizing raw responses to a 0–100 point range).20 A 10-point change in a score is perceived by patients as clinically meaningful.21 Time to confirmed symptomatic deterioration was therefore defined as the time from baseline to the first time the score showed a ≥10-point increase, which was confirmed over at least two consecutive cycles or accompanied by death within 3 weeks of the last assessment.

Tumor assessment and the calculation of SLD of target-lesions (per RECIST) was carried out at 6-week intervals within OAK,16 POPLAR,18 BIRCH,4 and FIR.5 Early changes in tumor size (∆TS) were defined by the percentage change in pretreatment (screening) SLD at the 6-week visit.

All analyses were based on the per-protocol populations (ie, assessment of participants based on actual treatment received).

Univariable and multivariable associations between ∆TS and OS, PFS, TDD in HRQoL and PF were modeled using Cox proportional hazards regression and reported as HRs with 95% CIs and concordance statistics (c). All multivariable analyses were adjusted for pretreatment Eastern Cooperative Oncology Group performance status (ECOG PS), age, sex, race, smoking status, histology, count of prior treatments, PD-L1 expression, serum lactate dehydrogenase (LDH) levels and the presence of liver, lung or brain lesions—adjusted analyses were conducted to assess the independence of associations from other factors. All statistical regression analyses were stratified by study. Potential non-linear associations between ∆TS and OS/PFS were explored using restricted cubic splines, with model fit assessed via the Akaike information criterion (AIC). Analyses were conducted with a focus on facilitating clinical use and interpretability. ∆TS was categorized based on model fit (assessed via the AIC), observed non-linear effects, and interpretable reference cut-points. A sensitivity recursive partitioning analysis was conducted to identify within atezolizumab-treated patients who experienced a decrease in pretreatment SLD at the 6-week visit, the ∆TS cut-off associated with robust prediction of patients who will achieve improved OS and PFS.

The distribution of pretreatment characteristics and BOR for patients with advanced NSCLC with and without ETS (defined by a 10% or greater decrease in the pretreatment SLD at the 6-week visit) following the initiation of atezolizumab or docetaxel were explored using the Fisher test for categorical data and Wilcoxon test for continuous data. Explored pretreatment characteristics included age, sex, race, smoking status, histology, ECOG PS, count of prior treatments, time since diagnosis, stage, count of tumor sites, presence of liver/lung/brain lesions, LDH levels, PD-L1 expression, EGFR/KRAS mutation status, and EMLA-ALK rearrangement.

Exploratory analysis of the prognostic effect of ETS on OS/PFS by pretreatment characteristic subgroups was modeled by Cox proportional hazards regression and presented in a forest plot (online supplementary figures S3 and S4). Heterogeneity of treatment effect on OS/PFS according to patients with and without ETS was assessed using a treatment-by-biomarker interaction term in a Cox proportional regression model. A sensitivity analysis of the association between ETS status at 12 weeks with OS and PFS was conducted.

Kaplan-Meier analysis was used for plotting and estimating probabilities of OS, PFS, TDD, HRQoL and PF. All p values <0.05 were considered statistically significant. Statistical analyses were performed using the R statistical environment (V.3.4, survival and rms packages).

On univariable and multivariable analysis, ∆TS (175 to 30 vs 30 to 10 vs 10 to −10 vs −10 to −30 vs −30 to −100) was significantly associated with OS (p<0.001), PFS (p<0.001), TDD in HRQoL (p<0.01) and PF (p<0.001) in patients treated with atezolizumab from OAK and POPLAR (table 1). Online supplementary figure S1 presents Kaplan-Meier estimates of OS, PFS, TDD in HRQoL and PF for atezolizumab-treated patients by ∆TS from OAK and POPLAR. Increases in ∆TS were associated with worse outcomes in patients treated with atezolizumab (compared with ∆TS of 10 to −10), with outcomes worst with increases greater than 30%. Conversely, decreases in ∆TS were associated with improved outcomes, with best outcomes achieved with decreases greater than 30%. As a biomarker to identify patients likely to achieve improved survival and patient-reported outcomes from atezolizumab, a decrease cut-off of 10% or greater in the pretreatment SLD at the 6-week visit appears appropriate—consistent with Wang et al.9

Online supplementary figure S2 presents the continuous relationship (log relative hazard ±95% CI) between ∆TS with OS and PFS for patients treated with atezolizumab and docetaxel from OAK/ POPLAR. The decrease cut-off of 10% or greater is observed as a key inflection point for patients who will achieve varying OS and PFS outcomes from atezolizumab therapy. Recursive partitioning analysis identified that for atezolizumab-treated patients from OAK/ POPLAR who experienced tumor shrinkage, the ∆TS cut-off associated with robust prediction of improved OS and PFS was decreases greater than 10.5% and 10.3%, respectively—further supporting a 10% or greater cut-off.

For atezolizumab-treated patients from BIRCH and FIR, exploratory analysis indicated an association between ∆TS and OS (p<0.001), PFS (p<0.001), TDD in HRQoL (p<0.001) and PF (p<0.001, online supplementary table S1). Consistent with the identified associations in the OAK and POPLAR cohort.

A significant univariable association between ETS (defined by a 10% or greater decrease in the pretreatment SLD at the 6-week visit) and improved OS (HR 0.33, p<0.001, c=0.58) and PFS (HR 0.31, p<0.001, c=0.60) was observed in patients treated with atezolizumab from OAK and POPLAR (table 2). The association between ETS and improved OS (HR 0.32, p<0.001, ∆c=0.02) and PFS (HR 0.31, p<0.001, ∆c=0.05) was maintained on adjusted analysis (table 2)—demonstrating an independence of ETS from other known prognostic factors. Table 2 presents 24-month probabilities of OS, PFS, TDD in HRQoL and PF for atezolizumab-treated patients with and without ETS. Figure 2 presents Kaplan-Meier estimates of OS, PFS, TDD in HRQoL and PF for atezolizumab-treated patients with and without ETS.

For atezolizumab-treated patients from BIRCH and FIR, ETS was associated with improved OS (p<0.001) and PFS (p<0.001), aligning with the identified associations in the OAK and POPLAR cohort (online supplementary table S2 and figure S5). ETS status at the 12-week visit was similarly associated with improved OS (p<0.001) and PFS (p<0.001) within patients treated with atezolizumab from OAK/ POPLAR and BIRCH/ FIR (online supplementary tables S3 and S4).

Pretreatment characteristics of patients with advanced NSCLC with and without ETS following the initiation of atezolizumab are summarized in online supplementary tables S5 and S6. Within the OAK and POPLAR cohort of patients with advanced NSCLC treated with atezolizumab, the presence of ETS was associated with being a current smoker, lower tumor sites count, no lung tumors and PD-L1 expression (p<0.05). Within BIRCH and FIR, the presence of ETS was associated with a lower ECOG PS, less prior treatments, no liver tumors and PD-L1 expression (p<0.05). Online supplementary tables S7 and S8 summarize the BOR of patients with advanced NSCLC according to ETS status following the initiation of atezolizumab within OAK/POPLAR and BIRCH/FIR, respectively.

The presence of ETS was significantly associated with improved OS (univariable HR 0.61, p<0.001; multivariable HR 0.65, p<0.001) and PFS (univariable HR 0.61, p<0.001; multivariable HR 0.64, p<0.001) in patients treated with docetaxel (online supplementary table S9 and figure S6). However, the size of the effects within the docetaxel-treated patients was significantly smaller than that in the atezolizumab-treated patients (P(interaction)=0.002 for OS and P(interaction)<0.001 for PFS)—indicative that the magnitude of OS and PFS benefit from atezolizumab (compared with docetaxel) was greater in patients with ETS.

For patients with ETS, the 24-month OS probability was 55% (95% CI; 44% to 69%) for atezolizumab-treated patients compared with 32% (23%–44%) for docetaxel, while 24-month PFS probability was 31% (23%–41%) vs 4% (2%–10%), respectively. For patients without ETS, the 24-month OS probability was 23% (95% CI; 19% to 29%) for atezolizumab-treated patients compared with 20% (16%–26%) for docetaxel, while 24-month PFS probability was 3% (2%–6%) vs 1% (0%–5%), respectively. Figure 3 presents Kaplan-Meier estimates of the effect of atezolizumab (compared with docetaxel) on OS and PFS in patients with and without ETS.