A Novel Clinical Trial in Rare Paediatric B-Cell Non-Hodgkin Lymphoma: Evaluating the Glo-Bnhl Adaptive Bayesian Design Using Data from the Sparkle Trial

Background

Outcomes for relapsed or refractory (r/r) paediatric B-cell non-Hodgkin Lymphoma (B-NHL) are extremely poor with long term cure rates below 30%. A collaborative pragmatic innovative approach to trial design was required for a new trial to efficiently evaluate promising novel agents showing benefit in adults.

Study Overview

Glo-BNHL (NCT05991388) is an international, platform clinical trial consisting of three treatment arms each investigating a different class of novel agent. Currently these classes of agents are bispecific antibodies, antibody-drug conjugates (ADC), and CAR T-cell products.

Each treatment arm has a primary outcome measure of either complete or objective response, allowing quick evaluation. Analysis utilises Bayesian beta-binomial conjugate models with each treatment arm analysed independently.

Treatment arms follow a two-stage design with pre-defined clinically meaningful target response rates based on the treatment and population being investigated. Once 15 patients are recruited to the initial stage, a posterior probability of at least 0.8 that the true response rate exceeds the target is required to continue to an expansion stage. At least 15 further patients will be recruited before analysing data from both stages. A successful outcome will require a posterior probability of at least 0.95 that the true response rate exceeds the target.

Regular interim analyses will take place after every 3 or 5 patients in the initial and expansion stages respectively, allowing early stopping for futility. The predicted probability of success (PPoS) is the probability of a positive result at the next decision point based on the current observed data. PPoS will be used at interim analyses to indicate that a treatment arm should close early if there is insufficient probability of efficacy.

Design Evaluation

To evaluate performance of the trial design, data from the SPARKLE trial (Blood Adv, 2023) was sampled and analysed as the population for Treatment Arm II (TAII) of Glo-BNHL. TAII was chosen since the ADC arm is most similar to the therapy investigated in SPARKLE. TAII seeks to find evidence of a complete response (CR) rate exceeding 20%. In the SPARKLE trial, CR rate was 18% overall, and in the Burkitt and non-Burkitt subgroups was 23% and 14% respectively.

A program was written which sampled and analysed response data according to the TAII design, calculating PPoS and posterior probabilities as required and applying decision criteria. Trials could end in a ‘Go’ decision if they showed sufficient evidence of efficacy at the final analysis, or in a ‘No Go’ decision if at the final analysis there was insufficient evidence of efficacy or if the trial stopped early for futility. 1000 simulations were conducted for each of the three populations: Burkitt, non-Burkitt, all patients.

In the overall group 98% of simulated trials ended in a No Go with 89% failing to reach the expansion stage. Similarly in the non-Burkitt group 99.5% ended in a No Go with 95% failing to reach the expansion stage. In the Burkitt group 92% of trials ended in a No Go but a higher proportion (23%) reached the expansion stage. This indicates that the design correctly identifies and closes treatment arms with a CR rate below 20% and that a CR rate of 23% is not sufficient for a Go decision.

To evaluate the design in scenarios with underlying CR rates above the target, the trial design program was run on randomly generated data from a binomial distribution. With underlying CR rates of 40% and 50%, trials ended in a Go decision 62% and 82% of the time respectively. Whilst if the CR rate was 60% almost all trials resulted in a Go. However, in trials with a true CR rate of 30%, only 56% proceeded to the expansion stage and 40% ended in a Go.

Conclusion

Using the SPARKLE study data as well as simulated data, the Glo-BNHL trial design has been shown to reliably identify underperforming and overperforming treatment arms, allowing early decisions to be made regarding efficacy and futility. However, due to the restricted sample size, where the true response rate is <20% greater than the target, it is less reliable in identifying a positive result.

In a setting such as rare paediatric r/r B-NHL where it is imperative that the minimal number of children are exposed to ineffective treatments and rapid evaluation of agents is required, an adaptive Bayesian design with a response outcome and frequent interim analyses allows meaningful assessment with few patients.

Qin:Johnson & Johnson Innovative Medicine: Current Employment, Current equity holder in publicly-traded company. Deshpande:Johnson & Johnson Innovative Medicine: Current Employment, Current equity holder in private company. Burke:AstraZeneca: Consultancy; Novartis: Consultancy.