P-118: Machine learning assisted risk stratification of newly diagnosed multiple myeloma with respect to the VRd treatment

Introduction: Bortezomib plus lenalidomide plus dexamethasone (VRd) regimen has been administered as the first-line treatment for newly diagnosed multiple myeloma (NDMM). While the regimen generally offers a high response rate and improved survival, as much as 15% of the patients suffered from early death and/or disease progression in our retrospective analysis. We thus developed the machine learning (ML) models predicting survival of the NDMM patients following the VRd treatment so as to assist personalized treatment selection.

Methods: We used the Multiple Myeloma Research Foundation's dataset (IA17) and identified 254 transplant ineligible NDMM treated by the VRd regimen as the first-line therapy. ML models were developed using the XGBoost (eXtreme Gradient Boosting) method. From the total of 148 demographic and clinical variables, a minimal set was chosen via sequential forward feature selection. The overall survival (OS) model was trained to compute the time-course of changes in the probability of being alive using accelerated failure time. The MMRF cohort was divided into the training and validation set during 5-fold cross-validation via the StratifiedKFold that minimizes the differences in the composition of classes such as international staging system (ISS) across the folds. The predictive performance of the trained ML models was assessed with respect to the ROC–AUC recorded during validation. We used the computed probability of survival (t = 36 months) during validation for the VRd-specific risk stratification.

Results: Clinical characteristics of the VRd treatment group within the MMRF dataset were as follows. The median age was 68 years, with 63% of the patients being male. IgG isotype was the most dominant (38%), and the majority belonged to the ISS stage I (38%) and II (40%). The median progression-free survival (PFS) and OS of the group were 49 and 94 months, respectively. The ML models achieved the ROC-AUC of up to 0.886 when predicting OS by the three years (36 months) following the first administration of VRd regimen. The median OS of the high and low risk groups were significantly different (29 months vs not reached, P < 0.0001), recording the hazard ratio of 7.14 (95% CI, 4.48-11.38). The following covariates were identified as being the most important when predicting the risk of early death following the VRd treatment: UAMS 70 gene index, CD56 expression, serum lambda level, white blood cell counts, FISH results with respect to 8p22 and CYLD.

Conclusions: In conclusion, we used the ML method to stratify the transplant ineligible NDMM into the high and low risk subgroups with respect to administration of the VRd regimen as the first-line treatment. Together with the previously developed ML models predicting response and survival following the VMP or Rd treatment, the proposed ML model can assist personalized treatment selection for the transplant ineligible NDMM.