Machine learning risk stratification strategy for multiple myeloma: Insights from the EMN-HARMONY Alliance platform

Abstract

Traditional risk stratification in multiple myeloma (MM) relies on clinical and cytogenetic parameters but has limited predictive accuracy. Machine learning (ML) offers a novel approach by leveraging large datasets and complex variable interactions. This study aimed to develop and validate novel ML-driven prognostic scores for newly diagnosed MM (NDMM), with the goal of improving upon existing ones. To this end, we analyzed data from the EMN-HARMONY MM cohort, comprising 14,345 patients, including 10,843 NDMM patients enrolled across 16 clinical trials. Three ML models were developed: (1) a comprehensive model incorporating 20 variables, (2) a reduced model including six key variables (age, hemoglobin, β2-microglobulin, albumin, 1q gain, and 17p deletion), and (3) a cytogenetics-free model. All models were internally validated using out-of-bag cross-validation and externally validated with data from the Myeloma XI trial. Model performance was evaluated using the concordance index (C-index) and time-dependent area under the receiver operating characteristic curve (ROC-AUC). The comprehensive model achieved C-index values of 0.666 (training) and 0.667 (test) for overall survival (OS) and 0.620/0.627 for progression-free survival (PFS). The reduced model maintained accuracy (OS: 0.658/0.657; PFS: 0.608/0.614). The cytogenetics-free model showed C-index values of 0.636/0.643 for OS and 0.600/0.610 for PFS. Incorporating treatment type and best response to first-line treatment further improved performance. The new prognostic models improved over the International Staging System (ISS), Revised International Staging System (R-ISS), and Second Revision of the International Staging System (R2-ISS) and were reproducible in real-world and relapsed/refractory MM, including daratumumab-treated patients. This ML-based risk stratification strategy provides individualized risk predictions, surpassing traditional group-based methods and demonstrating broad applicability across patient subgroups. An online calculator is available at https://taxonomy.harmony-platform.eu/riskcalculator/.

Figure 1

Graphical workflow of the study, outlining the methodology used for data preprocessing, model development, validation, and performance evaluation. ISS, International Staging System; OS, overall survival; PFS, progression‐free survival; R‐ISS, Revised International Staging System; R2‐ISS, Second Revision of the International Staging System.

Figure 2

Kaplan–Meier curves for the cytogenetics‐based model, with patients stratified into quartiles (Q1–Q4, determined by the 25th, 50th, and 75th percentiles) of continuous risk scores calculated separately for overall survival (OS) and progression‐free survival (PFS) prediction. (A) OS in the training cohort, (B) OS in the test cohort, (C) PFS in the training cohort, and (D) PFS in the test cohort.

Figure 3

Kaplan–Meier curves for the cytogenetics‐free model, with patients stratified into quartiles (Q1–Q4, determined by the 25th, 50th, and 75th percentiles) of continuous risk scores calculated separately for overall survival (OS) and progression‐free survival (PFS) prediction. (A) OS in the training cohort, (B) OS in the test cohort, (C) PFS in the training cohort, and (D) PFS in the test cohort.

Figure 4

Time‐dependent area under the receiver‐operating‐characteristic curves (AUCs) evaluating the accuracy of the baseline cytogenetics‐based and cytogenetics‐free models for overall survival (OS) and progression‐free survival (PFS) prediction in transplant‐eligible and transplant‐ineligible patients.

Figure 5

Variable‐importance analysis based on the permutation metric implemented in the vimp function of the randomForestSRC R package. Bars represent the drop in out‐of‐bag prediction accuracy observed after permuting each predictor, thereby quantifying its relative contribution to model performance. Results are shown for overall survival and progression‐free survival predictions generated by the cytogenetics‐based and cytogenetics‐free Random Survival Forest models.

Figure 6

Example output of the risk score calculator for a patient with hemoglobin (Hb) 10 g/dL, beta‐2 microglobulin (B2‐mg) 3.0 mg/dL, albumin 2.9 mg/dL, presence of 17p deletion, and absence of 1q gain. (A, B) The predicted risk distributions for immunomodulatory imide (IMiD)‐based therapy (A) and proteasome inhibitor (PI)–IMiD combination therapy (B). For each regimen, the patient's Random Forest–derived risk score is converted into a percentile rank relative to the corresponding distribution in the training cohort. Percentiles are displayed overall and stratified by transplant eligibility status in the training set (transplant candidate vs. noncandidate), providing an intuitive frame of reference for clinicians. Lower percentiles indicate lower relative risk (e.g., for progression‐free survival), whereas higher percentiles indicate higher relative risk. Because the metric is percentile‐based, it represents relative—not absolute—risk.