Introduction In the pivotal JULIET trial patients with relapsed or refractory Diffuse Large B cell Lymphoma (DLBC) received a single intravenous infusion of Tisagenlecleucel. In long-term follow-up analysis, an overall response rate (ORR) of 53% and a complete response (CR) of 39% were reported. In 115 evaluable patients, 62% had disease progression or died. At 40.3 months follow up, the median progression-free survival (PFS) was 2.9 months and the median overall survival (OS) was 11.1 months. The median PFS and OS of 35% of the patients with complete response at 3 months, 6 months, or both, were not reached suggesting durable response [1].

Although high metabolic tumor volume (MTV) measured by [18F] FDG PET/CT during CART cell therapy was found to be predictive of early relapse [2], pretreatment available factors - such as high IPI, elevated LDH, low platelets, and MTV - do not fully elucidate which individuals have poor outcome after therapy [1-3]. Accurate prediction of poor outcome in individuals at the treatment-decision stage may lead to more effective patient selection, preventing unnecessary cost and adverse treatment effects.

The aim of this study is to demonstrate feasibility of identifying a subgroup of patients at very high risk of poor outcome (death or disease progression) prior to infusion of Tisagenlecleucel, using Artificial Intelligence (AI) to characterize pre-infusion FDG PET/CT in combination with clinicopathological parameters.

Material and Methods In this secondary analysis of data from the prospective JULIET trial [1,4], 115 FDG PET/CT data sets were included from 115 patients with R-R DLBCL from 27 treatment sites between 2015 and 2018. All patients received a single intravenous infusion of Tisagenlecleucel . The pre-infusion FDG PET/CT images were processed automatically using deep learning. Clinicopathological parameters were added: patient age, IPI, LDH, platelet count, MTV, and LDH.

A novel automated test ("AI signature") was developed to identify a subgroup of patients at very high risk of poor outcome (death or disease progression). In short, an attention-gated convolutional neural network (AG-CNN) was trained to delineate the PET/CT disease foci automatically and consistently across the different treatment centers. MTV was automatically derived, and disease foci were further characterized by their activations in the most densely compressed layer in latent space of the AG-CNN. After additional dimensionality reduction, the AI signature was derived using multivariate Cox regression, random survival forests, Receiver Operating Characteristics (ROC) analysis, and Kaplan Meier modelling. The AI signature was validated using nested 5-fold cross validation: data were partitioned five times into training and testing folds. Models derived from the training folds were tested on the testing folds. Median testing performance and interquartile range (IQR) were reported.

Results The median patient age was 56 years (IQR 46-64). The median follow-up-time was 80 days (IQR 29-554). After 5-fold cross validation, 52.4% of the patients (IQR 39.1-56.5) had a negative AI-signature of whom 100% (IQR 92.9-100) had poor outcome. 47.6% Of the patients (IQR 39.1-56.5) had a positive AI-signature of whom 55.6% (range 53.3-61.5) had poor outcome. Median PFS in long-term follow up was 13.8% (range 11.5-14.6) and 29.8% (range 29.6-33.2) in patients with negative and positive AI-signature, respectively. The median cross-validated area under the ROC curve after multivariate Cox regression was 0.74 (IQR 0.70 - 0.81). The HR was 0.36 (95% CI 0.13 - 0.95; P = . 0.0432) after multivariable correction for age, IPI, LDH, platelet count, MTV, and LDH.

Conclusions Results from the JULIET trial data indicate that an automated test based on AI analysis of pre-infusion FDG PET/CT and clinicopathological parameters is feasible to identify a subgroup of patients at very high risk of poor outcome after Tisagenlecleucel. The test retained significance after multivariable correction for other parameters known to be associated with CART response, IPI, LDH, age, platelet count, MTV, and LDH. Follow-up studies will focus on validating these findings in independent patient cohorts.

References [1] Schuster et al. Lancet Oncol 2021; 22: 1403-15.

[2] Vercellino et al. Blood Adv 2020;4(22): 5607-5615.

[3] Nastoupil et al., J Clin Oncol 2020; 38(27):3119-3128.

[4] Schuster et al., N Engl J Med 2019;380:45-56.

Minnema:Janssen-Cilag: Consultancy, Honoraria, Membership on an entity's Board of Directors or advisory committees; Bristol Myers Squibb: Speakers Bureau; Medscape: Speakers Bureau. Gilhuijs:Novartis Pharma B.V.: Research Funding.

Author notes

*

Asterisk with author names denotes non-ASH members.

Sign in via your Institution