Predicting Early Relapse for Patients with Multiple Myeloma through Machine Learning

AK+NG eq. cont. Introduction Management of multiple myeloma (MM) improved dramatically in the last decade. However, prognosis is still poor for high-risk MM patients experiencing early relapse (ER) within 12 months after primary diagnosis. While some patients achieve long lasting remission with first-line triple combinations including immunomodulatory drugs (IMiDs) and proteasome inhibitors (PI) followed by high dose therapy plus autologous stem cell transplantation (ASCT), ER has been recognized as an independent risk factor of resistance to rescue treatments and shorter overall survival (OS). Individual treatment response duration is influenced by several disease-specific and patient-related features. Despite the availability of several MM prognostic scores, the prediction and prevention of ER in MM is still challenging. Our goal was to construct a holistic MM ER prediction model based on machine learning (ML) algorithms. Methods In this study, we applied ML algorithms to sociodemographic, clinical, and cytogenetic data associated with MM ER collected by the Multiple Myeloma Research Foundation (MMRF) CoMMpass consortium. Our model included the following covariates at baseline: gender, age, BMI, ECOG, creatinine, hemoglobin, platelets, neutrophils, calcium, beta-2-microglobuline, albumin, first-line therapy classification (doublet versus triplet, IMiD- versus PI-based versus combination), ASCT within the first year of primary diagnosis, high-risk cytogenetics as well as best response during first year of therapy. Our core patient characteristics in relation to ER are summarized in Table 1. Only complete case documentations were included in the data set (N = 563). To reduce dimensionality and overfitting, continuous data was converted to ordinal data bins according to standard clinical reference ranges (e.g. American Board of Internal Medicine (ABIM) reference ranges for laboratory tests). To ensure that only relevant patient characteristics are used for ER prediction, the Boruta algorithm (version 7.0.0) was used for feature selection using R 4.0.2. Feature independence was assessed by the calculation of correlation between the selected features using the Spearman correlation coefficient for nominal features and the point-biserial correlation coefficient for the comparison of nominal and dichotomous features. Because only few patients experienced ER, the Synthetic Minority Oversampling Technique (SMOTE) was used to correct for imbalance in the target var