We performed a search of electronic medical records for patients with advanced or recurrent NSCLC treated at 47 affiliated institutions of West Japan Oncology Group (WJOG). Eligible patients had histologically or cytologically confirmed advanced NSCLC who received cytotoxic chemotherapy as a first-line treatment. Patients with recurrent disease who had received curative surgery or chemoradiotherapy were included. The following regimens were not counted as a line of therapy: preoperative or postoperative adjuvant chemotherapy, chemotherapy associated with curative radiotherapy, epidermal growth factor receptor (EGFR) tyrosine kinase inhibitors (TKIs) for EGFR mutation–positive patients and anaplastic lymphoma kinase (ALK) TKIs for ALK rearrangement–positive patients. Patients who were treated with nivolumab or pembrolizumab in the second-line setting and subsequently with S-1, with pemetrexed, or with docetaxel with or without ramucirumab as the third-line treatment between December 1, 2015 and July 31, 2017 were included in the CAP cohort (see online supplementary figure S1). The clinical outcomes for the CAP cohort were compared with those for a control cohort of patients treated with second-line cytotoxic chemotherapy including either S-1, pemetrexed or docetaxel with or without ramucirumab—without preceding ICI therapy—between April 1, 2014 and July 31, 2017. The patients in the control cohort were included from April 1, 2014 in order to collect data on such chemotherapy because nivolumab and pembrolizumab became practically available in Japan from December 2015 and December 2016, respectively, and were then widely used as a second-line treatment. Patient eligibility was confirmed by the WJOG data center.

The primary end point of the study was ORR. The secondary end points were progression-free survival (PFS) from the first day of treatment with S-1, pemetrexed or docetaxel with or without ramucirumab until disease progression or death due to any cause, OS from the first day of such treatment until death due to any cause, and safety. ORR was assessed by the investigators according to Response Evaluation Criteria in Solid Tumors (V.1.1) and was calculated only for patients with measurable lesions. Safety evaluations included assessment of treatment-related select adverse events (AEs), which were defined as AEs with a potential immunologic basis.18 AEs were graded according to the National Cancer Institute Common Terminology Criteria for Adverse Events (V.4.0).

Comparisons between the two cohorts were performed with Fisher’s exact test for categorical variables and with the Wilcoxon test for continuous variables. Given that we assumed that imbalances in patient characteristics between the two cohorts might exist, we applied propensity score analysis with the inverse probability weighting (IPW) method to minimize the bias due to measured confounders.19 The propensity score for each patient was calculated as a probability from a logistic regression model that included all covariates deemed likely to have affected treatment decisions and response—including age, sex, smoking status, Eastern Cooperative Oncology Group performance status, histology, EGFR or ALK genetic alterations, brain metastasis, history of curative radiotherapy and type of chemotherapy (docetaxel with or without ramucirumab, S-1 or pemetrexed). Stabilized weights were calculated for each patient on the basis of the estimated propensity score.19 When weighted regression analyses were performed, a robust sandwich variance was used to account for the weighted nature of the sample.19

Survival curves for the two cohorts were created with IPW-adjusted Kaplan-Meier plots, and IPW-adjusted HR was calculated with IPW-weighted Cox’s proportional hazard models.

We conducted subgroup analysis based on each chemotherapy regimen for the two cohorts. For subgroup analysis of efficacy, the propensity score was calculated by excluding the covariates of histology and history of curative radiotherapy for patients with EGFR or ALK genetic alterations, and by excluding performance status and EGFR or ALK genetic alterations for those with history of curative radiotherapy. We adopted this approach because the logistic regression model used to yield propensity scores was unstable when these covariates were included in the explanatory variables.

All p values <0.05 were considered statistically significant. Clinical data were managed by the WJOG data center. Statistical analysis was performed by an outside contract research organization (EPS, Tokyo, Japan) with SAS (V.9.4) software.

A total of 1626 patients at 47 institutions was assessed for study eligibility (figure 1). Of these patients, we excluded 187 (12%) individuals who did not meet the inclusion criteria. The remaining 1439 patients were included in the unadjusted analysis. The CAP cohort consisted of 243 patients, with 105 (43%) having received docetaxel, 77 (32%) docetaxel plus ramucirumab, 49 (20%) S-1 and 12 (5%) pemetrexed, whereas the control cohort consisted of 1196 patients, with 778 (65%) having received docetaxel, 94 (8%) docetaxel plus ramucirumab, 174 (15%) S-1 and 150 (13%) pemetrexed. The median follow-up time was 8.1 months (95% CI 7.5 to 9.4 months) in the CAP cohort and 9.3 months (95% CI 8.7 to 9.9 months) in the control cohort. Unadjusted patient characteristics and comparisons between the two cohorts are shown in table 1 and online supplementary table S1. The two cohorts differed significantly with respect to age, histology, EGFR or ALK genetic alterations, PD-L1 tumor proportion score (TPS), brain metastasis and type of chemotherapy. To correct for potential imbalances, we performed propensity score analysis. The distribution of propensity score in each cohort is shown in online supplementary figure S2. We evaluated covariate balance with the use of standardized difference.20 A standardized difference <0.1 would indicate good balance. After IPW adjustment, the covariates were well balanced between the CAP and control cohorts (see online supplementary table S2). Treatment data for first-line chemotherapy, PD-1 inhibitor treatment after first-line chemotherapy and poststudy systemic therapy are provided in online supplementary tables S3 and S4.

With the IPW method, the ORR for the CAP and control cohorts was found to be 18.9% and 11.0%, respectively, with the ORR for the CAP cohort being significantly higher than that for the control cohort (ORR ratio 1.71 with 95% CI 1.19 to 2.46, p=0.004) (table 2). IPW-adjusted subgroup analysis according to chemotherapy regimen revealed the ORR for the CAP and control cohorts to be 17.6% and 11.4%, respectively, for patients treated with docetaxel, and 20.9% and 18.3%, respectively, for those treated with docetaxel plus ramucirumab. Patients treated with docetaxel or docetaxel plus ramucirumab in the CAP cohort thus had a numerically higher ORR compared with those in the control cohort.

With regard to the results of IPW-adjusted survival analysis, the median PFS for the CAP and control cohorts was 2.8 and 2.7 months, respectively. PFS thus did not differ significantly between the CAP and control cohorts (IPW-adjusted HR 0.95 with 95% CI 0.80 to 1.12, p=0.55) (figure 2). IPW-adjusted Kaplan-Meier curves showed that the median OS for the CAP and control cohorts was 9.2 and 10.4 months, respectively. There was thus no difference in OS between the CAP and control cohorts (IPW-adjusted HR 1.05 with 95% CI 0.86 to 1.28, p=0.63) (figure 3).

In the unadjusted analysis, ORR was 17.9% for the CAP cohort and 10.8% for the control cohort (ORR ratio 1.66 with 95% CI 1.21 to 2.28, p=0.003) (see online supplementary table S5). Unadjusted Kaplan-Meier curves showed that median PFS was 3.2 and 2.7 months (unadjusted HR 0.93 with 95% CI 0.81 to 1.08, p=0.36) (see online supplementary figure S3) and that median OS was 9.5 and 10.4 months (unadjusted HR 1.08 with 95% CI of 0.91 to –1.28, p=0.40) (see online supplementary figure S4) for the CAP and control cohorts, respectively.

We performed sensitivity analysis with the use of alternative approaches to evaluate robustness with our findings regarding estimated treatment effect. Multivariable analyses were conducted with a log-linear regression model for response and Cox’s proportional hazard model for survival, and with the same covariates as used in the IPW method being included as the explanatory variables. These approaches yielded similar results with those for IPW adjustment (see online supplementary table S6).

We further evaluated efficacy in the CAP cohort with regard to several factors that potentially could have influenced the response to chemotherapy after ICI treatment (table 3). Data on PD-L1 expression in tumor cells were available for 39.5% (96 of 243) of patients in the CAP cohort. Of these 96 patients, 23 (24.0%) individuals had a TPS for PD-L1 ≥50%, 41 (42.7%) had a TPS of 1%–49% and 32 (33.3%) had a TPS <1%. The ORR and PFS did not differ between patients with a PD-L1 TPS ≥50% and those with a TPS <1%. Patients with a PD-L1 TPS of 1%–49% had a lower ORR than did those with a TPS <1%, whereas PFS did not differ between these two subgroups. There was no significant difference in ORR or PFS between subgroups of the CAP cohort classified according to duration of PD-1 inhibitor treatment, type of response to PD-1 inhibitor treatment, the interval between the last dose of PD-1 inhibitor and the start of subsequent chemotherapy, EGFR or ALK alteration status or history of curative radiotherapy.

Finally, we examined whether the CAP cohort experienced increased toxicity, given that small series of patients with melanoma or NSCLC were previously found to experience severe systemic toxicities during treatment with TKIs subsequent to that with ICIs.21–23 Treatment-related select AEs for chemotherapy, PD-1 inhibitors and each chemotherapy regimen separately are listed in table 4 and online supplementary tables S7–11. AEs of any grade for chemotherapy that showed a significantly higher incidence in the CAP cohort than in the control cohort included stomatitis (14.4% vs 8.8%, p=0.009) and hypothyroidism (1.2% vs 0%, p=0.005). Subgroup analysis according to chemotherapy regimen revealed no significant differences in the incidence of any-grade stomatitis or hypothyroidism between the CAP and control cohorts, whereas the incidence of an increase in total bilirubin level of any grade in patients treated with docetaxel or of hyperglycemia and increases in aspartate aminotransferase (AST) and alanine aminotransferase (ALT) levels of any grade in those treated with S-1 was significantly higher in the CAP cohort than in the control cohort. All treatment-related deaths were due to pneumonitis: one patient (0.4%) in the CAP cohort and seven patients (0.6%) in the control cohort.