Jørgensen, Rasmus Rask Kragh

ABSTRACT One quarter of patients with diffuse large B‐cell lymphoma (DLBCL) who achieve a complete response following front‐line therapy will eventually experience disease relapse. Existing prognostic models use baseline factors to predict outcomes from the start of initial therapy and are not intended for use following completion of treatment. We developed a conditional event‐free survival (cEFS) tool that predicts a patient's risk of lymphoma recurrence after completing front‐line therapy taking into consideration baseline risk characteristics plus the time the patient has been in remission. Model development used pooled and harmonized data from two observational cohorts and four randomized trials of patients with newly diagnosed DLBCL treated with immunochemotherapy and either achieving a complete response or not experiencing an event within 12 weeks of end of therapy. Cox proportional‐hazards regression modeling with backwards stepwise selection was utilized for model development. External validation was performed using a large independent data registry. In total, data from 2757 patients were used for model development. The risk of experiencing an event at a time point after having survived a period of time could be computed parsimoniously using four basic predictors: age, ECOG performance status grade, Ann Arbor stage, and lactate dehydrogenase level. C‐index in the model development cohort was 0.65 (95% CI, 0.63–0.68), compared to 0.64 (95% CI, 0.62–0.66) in the external validation cohort. We believe that this dynamic risk prediction tool will not only aid clinicians in providing data‐driven recommendations for surveillance, but also assist patients themselves with key decisions, such as family planning.

PURPOSE Patients diagnosed with advanced-stage Hodgkin lymphoma (aHL) have historically been risk-stratified using the International Prognostic Score (IPS). This study investigated if a machine learning (ML) approach could outperform existing models when it comes to predicting overall survival (OS) and progression-free survival (PFS). PATIENTS AND METHODS This study used patient data from the Danish National Lymphoma Register for model development (development cohort). The ML model was developed using stacking, which combines several predictive survival models (Cox proportional hazard, flexible parametric model, IPS, principal component, penalized regression) into a single model, and was compared with two versions of IPS (IPS-3 and IPS-7) and the newly developed aHL international prognostic index (A-HIPI). Internal model validation was performed using nested cross-validation, and external validation was performed using patient data from the Swedish Lymphoma Register and Cancer Registry of Norway (validation cohort). RESULTS In total, 707 and 760 patients with aHL were included in the development and validation cohorts, respectively. Examining model performance for OS in the development cohort, the concordance index (C-index) for the ML model, IPS-7, IPS-3, and A-HIPI was found to be 0.789, 0.608, 0.650, and 0.768, respectively. The corresponding estimates in the validation cohort were 0.749, 0.700, 0.663, and 0.741. For PFS, the ML model achieved the highest C-index in both cohorts (0.665 in the development cohort and 0.691 in the validation cohort). The time-varying AUCs for both the ML model and the A-HIPI were consistently higher in both cohorts compared with the IPS models within the first 5 years after diagnosis. CONCLUSION The new prognostic model for aHL on the basis of ML techniques demonstrated a substantial improvement compared with the IPS models, but yielded a limited improvement in predictive performance compared with the A-HIPI.

Purpose: Improved predictive tools are desired in classic Hodgkin lymphoma (cHL) to improve prognostication and to better risk stratify patients (pts) for clinical trials. In younger pts who tolerate intensive therapies, focus has been on limiting late treatment toxicity through risk-adapted, PET-guided treatment strategy. In older pts, acute treatment toxicities are limiting factors for optimal disease control. Due to these fundamental differences, predictive tools may perform differently in younger and older pts. In particular, tools developed in younger pts may be suboptimal in older populations and vice versa. The primary objective of this study was to assess the performance of the A-HIPI (development was limited to 18-65 years of age) against the international prognostic score (IPS) for modelling overall survival (OS) and progression-free survival (PFS) in younger and older pts with cHL. Methods: This study utilized data from the population-based, nationwide Danish National Lymphoma registry. Pts with advanced-stage cHL (Ann Arbor stage III-IV or IIB with extra nodal disease) were included and categorized by age in two groups: ages ≤65 years vs. >65 years. In the primary analysis, all pts regardless of therapy were included. A sensitivity analysis restricting to pts starting on BEACOPP or ABVD was performed (Only ABVD in the >65 cohort). Concordance index (C-index) and calibration were used to assess model performance and generalizability. Calibration was visualized using calibration plots and quantified by calibration in the large (CITL) and calibration slope (CS) at 5-years post diagnosis. Results: A total of 893 advanced-stage HL pts fulfilled the inclusion criteria (634 in the ≤65 cohort and 259 in the >65 cohort). The 5-year OS and PFS in the >65-year cohort were 44% (Confidence intervals (CI); 37-51%) and 35% (CI; 30-42%), respectively. Fifty-six pts in the >65 cohort were classified as low-intermediate-risk (IPS; 1-2), 128 intermediate-risk (IPS; 3-4), and 75 high-risk (IPS; 5-7). For the ≤65 cohort, the 5-year OS and PFS was 88% (CI; 85-90%) and 76% (CI; 73-80%), respectively. Twenty-three pts were classified as low-risk (IPS; 0), 244 low-intermediate-risk, 286 intermediate-risk, and 81 high-risk. The A-HIPI had a C-index of 0.74 and 0.63 for OS and PFS, respectively in the ≤65 years cohort vs. 0.65 and 0.64 for OS and PFS, respectively, in the >65 cohort. The IPS had a C-index in the ≤65 cohort of 0.66 and 0.60 for OS and PFS, respectively, vs. 0.58 and 0.59 for OS and PFS, respectively, in the >65 cohort (i.e., the IPS performed better in the ≤65 cohort in terms of OS and PFS but was inferior to the A-HIPI). When restricting to pts treated with ABVD or BEACOPP, the A-HIPI C-index in the >65 cohort (116 pts) were reduced to 0.55 for OS and 0.54 for PFS. For the ≤65 cohort (595 pts), the A-HIPI C-index was stable at 0.75 and 0.62 for OS and PFS, respectively. In terms of A-HIPI calibration in the >65 cohort, CITL and CS were -0.78 and 0.39 for OS vs. -0.98 and 0.80 for PFS. In the ≤65 cohort, CITL and CS were -0.06 and 0.99 for OS vs. 0.21 and 1.18 for PFS. Visual inspection of the calibration plot (Figure 1) for the >65 cohort showed that the A-HIPI underestimated the observed proportion for both OS and PFS (i.e., the observed events were higher than the predicted). Calibration plot for the ≤65 cohort (Figure 2) indicated that the A-HIPI accurately modelled the observed proportion for both OS and PFS (i.e., the predicted probability was consistent with the observed proportion). Conclusion: The newly developed A-HIPI was validated in a large population-based cohort of Danish pts with advanced-stage cHL. Performance of the A-HIPI was high in the ≤65 cohort and superior to the IPS for OS prediction. Calibration of A-HIPI was excellent in pts ages ≤65 years. However, A-HIPI performed significantly worse among pts >65 years and attained poor performance in pts receiving ABVD or BEACOPP, which is not surprising since patients older than 65 were not included in the development. This study emphasizes the importance of further efforts to develop or update models that apply specifically in older individuals. The inclusion of additional variables reflecting comorbidities and pts fitness into a prognostic model may be necessary to effectively deliver accurate prognoses and risk assessments for the older pt.