Bayesian Inference for Optimization of Interim Analysis in Clinical Trials By Incorporation of Historical Data: Reanalysis of the HOVON AML 132 Clinical Trial

Background Prospective randomized trials remain pivotal to evaluate the benefit of new therapies. Primary endpoints typically focus on long-term endpoints after study completion, whereas interim analyses are monitoring safety and toxicity. Although the traditional or frequentist approach is widely used to analyze endpoints and anticipate the number of events needed upon trial completion, it is limited by the need for prior assumptions of the expected treatment effect, long-term follow-up, and conservative stopping rules. Bayesian inference might overcome these limitations, for which prior knowledge is combined with new data to compute the posterior distribution, enhancing the strength of the ongoing trial. This study aims to evaluate whether a Bayesian approach using historical data as prior knowledge might support decision-making at interim time points of the HO132 phase 3 clinical trial for patients with acute myeloid leukemia (AML), that was recently reported (Löwenberg et al, 2021).

Methods In the HO132 trial, 927 patients aged 18 to 65 years with newly diagnosed AML were randomized between intensive induction ± lenalidomide. The primary endpoint EFS was not different between experimental and control treatment (HR 0.99, p=0.96). After excluding 64 patients for lenalidomide dose selection, three interim analyses were retrospectively introduced after inclusion of 150, 300 and 600 patients to assess whether the lack of efficacy emerged early during conduct of the trial. The control treatment arm was reinforced using 445 historical control treatment patients from the preceding AML trial (HO102). Patients of both control arms were matched 1:1 using propensity scores based on age and leukemia risk. At each interim analysis, posterior distributions were calculated for endpoints including complete remission (CR) after induction, minimal residual disease (MRD) in CR, early death within 2 months, and EFS. The posterior distributions were summarized to provide point estimates and 95% credible intervals (CI) on the treatment difference and hazard ratio (HR) between both arms. For binary outcomes, the probability of success (range 0-100%) of the control versus the experimental treatment was computed. For EFS, the probability of the HR being below 0.76 was calculated, which was the assumed effect size in the HO132 statistical plan. A probability of <10% was considered as the futility threshold.

Results We compared HO132 experimental treatment with control treatment reinforced with matched historical patients for the previously mentioned endpoints. The proportion of patients obtaining CR was lower with experimental compared with control treatment at interim analysis 1 (Fig 1A, treatment difference: -9.0% [95% CI -19.9 to 1.4]) and the probability for demonstrating a superior CR rate with experimental treatment was 4%. At interim analysis 2 and 3 the treatment difference was -8.1% [95% CI -16.3 to 0.001] and -9.8% [95% CI -18.6 to -4.1] with a probability of success of 3% and 0%, respectively.

Patients in CR and assigned to the experimental treatment arm were less often MRD^neg compared with control at every interim analysis (treatment difference: -10.3% [95% CI -28.7 to 8.0]; -8.1% [95% CI -21.5 to 5.6]; -5.6% [95% CI -14.5 to 3.4], respectively), with a respective probability not meeting the futility stopping rule of 13%, 12% and 11%.

Early deaths were more frequently observed with experimental treatment versus control (treatment difference: 5.5% [95% CI -0.7 to 12.8]), with 4% probability of fewer early deaths with experimental treatment than control at interim analysis 1. The treatment differences for early death at interim analyses 2 and 3 were 3.5% (95% CI -1.7 to 9.2) and 2.0% (95% CI -1.9 to -6.0) with a probability of 10% and 15%, respectively.

EFS was similar between both arms at interim analysis 2 and 3 (HR 1.00, and HR 1.04, respectively, Fig 1B), with a median follow-up time of 10 and 16 months, respectively. At interim analysis 2, this resulted in a probability of 4% reaching the anticipated HR of 0.76, which probability was 0% at interim analysis 3.

Conclusion Reanalyzing the prospective conduct of the HO132 study using Bayesian inference identified a low probability of success of the experimental treatment at three successive putative interim analyses based on important efficacy markers. Consequently, a low probability of efficacy during trial conduct might support early termination of the trial.

Manz:CDR-Life Inc: Consultancy, Current holder of stock options in a privately-held company; University of Zurich: Patents & Royalties: CD117xCD3 TEA. Cloos:Merus: Other: MRD assessments, Research Funding; Astellas: Speakers Bureau; Novartis: Consultancy, Other: MRD assessments, Research Funding; Janssen: Research Funding; Genentech: Research Funding; DC-One: Other: MRD assessments, Research Funding; Navigate: Patents & Royalties: Royalties for MRD analyses; Helsinn: Other: MRD assessments; Takeda: Research Funding. Ossenkoppele:Novartis: Consultancy, Honoraria, Research Funding; Pfizer: Consultancy; BMS/Celgene: Consultancy, Honoraria; Janssen: Consultancy; AGIOS: Consultancy, Honoraria; Amgen: Consultancy; Gilead: Consultancy, Honoraria; Astellas: Consultancy, Honoraria; Roche: Consultancy, Membership on an entity's Board of Directors or advisory committees; Jazz Pharmaceuticals: Consultancy; Merus: Consultancy. Lowenberg:AbbVie: Membership on an entity's Board of Directors or advisory committees; Astellas: Membership on an entity's Board of Directors or advisory committees; Catamaran Bio Inc: Membership on an entity's Board of Directors or advisory committees; Bristol Myers Squibb: Membership on an entity's Board of Directors or advisory committees; Celgene: Membership on an entity's Board of Directors or advisory committees; Clear Creek Bio: Consultancy, Honoraria; F. Hoffmann La Roche: Membership on an entity's Board of Directors or advisory committees.