(077) Replication of overall response in mNSCLC patients treated with pembrolizumab and chemotherapy using RECIST 1.1 criteria applied to images from the real world

Background: Tumor response is a widely used endpoint for evaluating treatment efficacy, measured often by RECIST 1.1 with images in clinical trials and note abstraction from electronic health records (EHR) in large real world data (RWD) studies. A missed opportunity is to apply RECIST 1.1 to images from routine care and compare it to existing tumor response measures in the same population

Objectives: In this study, we compared responses obtained by applying RECIST 1.1 via central review of images from real world routine care (rwRECIST), to those derived from EHR note-based response (rwR), and to those from a published clinical trial (KN-189: NCT02578680) of patients with metastatic non-small cell lung cancer (mNSCLC) treated with pembrolizumab and chemotherapy

Methods: Within the US nationwide Flatiron Health EHR-derived de-identified database, we selected patients with mNSCLC treated with firstline pembrolizumab and chemotherapy who also had at least one image available prior to treatment, and who met the same KN-189 study criteria including for PDL1, ECOG, ALK and EGFR status. Matching-adjusted indirect comparison (MAIC) was used to further balance any differences in age, gender, smoking, PDL1, brain mets and ECOG status. The dichotomized overall response rate (ORR) was calculated as the proportion of patients with complete or partial response, as measured by: (1) rwR, (2) rwRECIST, and (3) rwRECIST+rwR where missing rwRECIST was imputed with rwR. Patients with missing assessments altogether were considered non-responders for each respective measure. All three were compared to overall ORR from KN-189 and that stratified by PDL1 levels

Results: Of the 255 mNSCLC pts, 13% had no rwR and 27% no rwRECIST. The cohort had similar baseline characteristics to KN-189 after MAIC. Compared to KN-189’s ORR [48% (95%CI: 43-53%)] (n=410), MAIC-adjusted rwRECIST+rwR was similar [48% (95%CI: 41-55%)], while rwR overestimated [60% (95%CI: 53-67%)] and rwRECIST underestimated it [39% (95%CI: 33-46%)]. When stratified by PDL-1 status, rwRECIST+rwR was within 2-4 percentage points of those from KN-189. For patients with both a rwR and rwRECIST (n=184), ORR agreement was 68%, with Cohen’s Kappa of 0.36 (95%CI: 0.23-0.48)

Conclusions: A composite endpoint of RWD images and EHR notes can robustly replicate the treatment response observed in a mNSCLC trial. While rwR is easier to obtain and with less missingness, its agreement with trial RECIST1.1 assessment was moderate, overestimating by approximately 12 percentage points. Although more accurate, the higher occurrence of missingness in rwRECIST, likely due to off-site imaging, leads to an underestimation of the trial ORR, limiting its stand-alone usability