Utilizing Synthetic Individual Patient Level Cohorts from Machine Learning on Clinical Trials to Define Endpoints for Bispecific Trials in Large B-Cell Lymphoma

Background: The treatment landscape for patients with relapsed/refractory (r/r) large B-cell lymphoma (LBCL) is evolving rapidly with new treatments such as chimeric antigen receptor (CAR) T-cell and bispecific antibody therapies. Bispecific antibodies first received accelerated approval in the United States by the FDA based on the results of single-arm trials utilizing historical outcomes data in the chemotherapy era to design these trials. The accelerated approvals of CD20 x CD3 bispecific antibodies established new benchmarks for response rates and progression-free survival (PFS) for patients receiving therapy in the third line or later (3L+), and historical references for response rate and survival outcomes in future trials should be updated. Applying machine learning (ML) algorithms, we simulated virtual individual patient-level from published trials and utilized these synthetic patient-level datasets to interrogate implied statistical relationship among variables that were not directly reported in results sections of source publications.

Methods: Data were extracted from published trials of mosunetuzumab (M, NCT02500407), glofitamab (G, NCT03075696), and epcoritamab (E, NCT03625037) monotherapy trials evaluating bispecific antibody in LBCL along with data from mosunetuzumab + polatuzumab (MP, NCT03671018) as a reference bispecific antibody combination therapy. Virtual patient-level data of patients with r/r LBCL who received various bispecific antibody in 3L+ treatment were generated from tables and figures in the manuscript using ML algorithms created for analyzing publication statistical summaries. Patient characteristics were extracted from table 1, PFS was calculated from K-M survival curve and various subgroup analysis were transcribed from forest plots reported in the source publications papers.

Results: Overall, 498 patients were simulated from four trials (N; M 88, G 154, E 157, MP 98). The median age of all patients was 65 years (range: 20-96). In the original trials, the median lines of treatment in overall patients were 3 (range: 1-13), and 21%, 21%, 33% and 33% of patients received M, G, E and MP, respectively post CAR T-cell therapy. The complete response (CR) rate in virtual cohorts compared to each dose expansion/phase 2 cohort in the real trials was 24% vs. 23.9% (Bartlett, Blood Adv, 2022), 40% vs. 39% (Dickinson, N Eng J Med, 2022), 43% vs. 38.9% (Thieblemont, J Clin Oncol, 2022), and 46% vs. 45.9% (Budde, Nat Med, 2023) for M, G, E and MP, respectively. Median PFS in the virtual cohorts compared to real trials were 3.1 vs. 3.2, 4.8 vs. 4.9, 4.3 vs. 4.4 and 11.4 vs. 11.4 months by M, G, E and MP, respectively, showing concordance between virtual individual patients to patients in trials. Across trials, there were differences in enrollment duration (10, 20, 7, 41 months for M, G, E and MP, respectively) and patient characteristics such as sex, histology, performance status, prior lines of treatment and prior exposure to autologous stem cell transplant and CAR T-cell therapy. Differences in enrollment duration and patient characteristics were associated with differences in PFS and duration of response.

Conclusions: Synthetic generation of patient-level data showed consistent outcomes to published trial reports. This model can potentially perform detailed subgroup analysis and adjust differences in background patient characteristics, enrollment duration and follow up duration among trials, however, need further data from trials and real-world patients to continue refining and validating the model. This approach reuses recent trial data to establish contemporary benchmarks for the design of new trials before real world evidence can be generated with a change in standard of care.