A Novel Prognostic Model That Is Superior to FLIPI-1 and FLIPI-2 and Integrates Clinical and Treatment Factors to Predict Progression-Free Survival and Early Treatment Failure in Patients with Follicular Lymphoma in the GALLIUM Trial

Introduction: Progression of disease within 24 months of initial therapy (POD24) is associated with poor survival in patients with follicular lymphoma (FL). Existing prognostic models, such as FLIPI-1 and FLIPI-2, show poor sensitivity for POD24, and are derived from cohorts lacking bendamustine-treated patients. More accurate predictive models based on current standard therapies are needed to identify patients with high-risk disease. The Phase III GALLIUM trial (NCT01332968) compared the safety and efficacy of standard chemotherapy regimens plus rituximab (R) or obinutuzumab (G) in patients with previously untreated FL. Using GALLIUM data, we developed a novel risk stratification model to predict both PFS and POD24 in FL patients after first-line immunochemotherapy.

Methods: Enrolled patients were aged ≥18 years with previously untreated FL (grades 1-3a), Stage III/IV disease (or Stage II with bulk), and ECOG PS ≤2, and required treatment by GELF criteria. Patients were randomized to receive either G- or R-based immunochemotherapy, followed by maintenance with the same antibody in responders. The chemotherapy arm (CHOP, CVP, or bendamustine) was selected by each study center. POD24 was defined as progressive disease or death due to disease within 24 months of randomization (noPOD24 = no progression or lymphoma-related death in that period). The most strongly prognostic variables, based on PFS hazard ratios, were estimated using penalized multivariate Cox regression methodology via an Elastic Net model. Selected variables were given equal weights, and a clinical score was formed by summating the number of risk factors for each patient. Low- and high-risk categories were determined using a cut-off that provided the best balance between true- and false-positives for PFS. PFS correlation and sensitivity to predict POD24 were assessed. The data used are from an updated GALLIUM efficacy analysis (data cut-off: April 2018; median follow-up: 57 months).

Results: 1202 FL patients were enrolled. Based on data availability and biological plausibility (i.e. could reasonably be linked with high-risk disease), 25 potential clinical and treatment-related prognostic variables were entered into the Elastic Net model (Table). A model containing 11 factors was retained by the methodology and chosen as the best model (Table). Patients were categorized as 'low risk' if they scored between 0 and 3 (n=521/1000 patients with complete data) and as 'high risk' if they scored between 4 and 11 (n=479/1000 patients).

At 2 years, the PFS rate was 84.5% in the whole FL population. Using our model, 2-year PFS for high-risk patients was 77% compared with 79.9% for FLIPI-1 and FLIPI-2. In low-risk patients, 2-year PFS was 92% compared with 87.9% for FLIPI-1 and 87.6% for FLIPI-2 (low-intermediate-risk patients). Our model increased the inter-group difference in 2-year PFS rate from 8% (FLIPI-1) and 7.7% (FLIPI-2) to 15%. At 3 years, the inter-group difference increased from 6.9% (FLIPI-1) and 9% (FLIPI-2) to 17% (Figure). Sensitivity for a high-risk score to predict POD24 was 73% using our model compared with 55% for FLIPI-1 and 52% for FLIPI-2 (based on 127 POD24 and 873 noPOD24 patients with complete data). Excluding patients who received CVP, which is now rarely used, resulted in an inter-group difference in PFS of 15% at 2 years and 16.8% at 3 years. A sensitivity analysis showed that inclusion of the 9 clinical factors only (i.e. removal of CVP and R treatment as variables) formed a more basic scoring system (low-risk patients, 1-3; high-risk patients, 4-9); the inter-group difference in PFS was 16.5% at 2 years and 17.6% at 3 years. However, sensitivity for POD24 decreased to 56%.

Conclusion: Our clinical prognostic model was more accurate at discriminating patients likely to have poor PFS than either FLIPI-1 or FLIPI-2, and its prognostic value was sustained over time. Our model also identified the FL population at risk of POD24 with greater sensitivity. Variables such as age and bone marrow involvement were not retained by our model, and thus may not have a major impact in the current era of therapy. Factors such as sum of the products of lesion diameters were included, as this captures tumor burden more accurately than presence of bulk disease. Future studies will aim to improve the accuracy of the model by considering gene expression-based prognostic markers and DNA sequencing to form a combined clinico-genomic model.

View largeDownload slide

View largeDownload slide

Close modal