Artificial intelligence and machine learning in relapsed/refractory multiple myeloma: A systematic review.

Abstract

e19500 Background: Relapsed/refractory multiple myeloma (RRMM) remains incurable with heterogeneous outcomes. Traditional prognostic systems (ISS/R-ISS) offer limited discriminatory power and no treatment guidance in heterogeneous RRMM. The heterogeneity of RRMM makes it suitable for AI/ML applications; however, these approaches have not been systematically reviewed. Methods: We systematically searched PubMed, Embase, and Web of Science (till December 2025) for studies applying AI/ML to RRMM populations. Included studies required: (1) AI/ML algorithm application, (2) RRMM patients included or validated, and (3) clinical outcome prediction. Studies were categorized by application domain. Performance metrics (C-index, AUC), validation methods, and clinical applicability were extracted. Results: We identified 20 studies applying AI/ML to >20,000 patients across six RRMM-relevant domains. Survival prediction (n=7): Transformer, Random Forest, and ensemble models achieved C-indices of 0.62–0.82; the EMN-HARMONY model (N=14,345) showed 6-variable score outperforming R-ISS, while POD24 was strongest mortality predictor. CAR T-cell therapy (n=3): The MyCARe logistic regression model stratified 5-month relapse risk (7% vs 27% vs 53%; p<0.001) using 4 clinical variables; single-cell multiomics identified CD4+ cytotoxic T cells as predictors of both response and toxicity. Treatment response (n=4): IsomiR-based classifiers achieved high accuracy (AUC 0.98) for daratumumab response; multiomics network models (mmSYGNAL) outperformed ISS and cytogenetics across disease trajectory. Toxicity prediction (n=3): ML nomograms predicted carfilzomib-associated heart failure (C-index 0.78) and daratumumab-related pneumonia (AUC >0.80) using ECOG, hemoglobin, and comorbidities. Minimal residual disease (MRD) (n=1): AI-based pattern classification of longitudinal MRD significantly stratified PFS (p<10⁻⁷). Imaging radiomics (n=2): 3D-CNN on whole-body MRI (AUC 0.804) and DeepSurv on PET/CT (C-index 0.657) identified prognostic imaging features. External validation: 11/20 studies (55%). Conclusions: AI/ML demonstrates substantial potential for RRMM risk stratification, CAR T optimization, treatment response prediction, and toxicity prevention. Three models (MyCARe, EMN-HARMONY, carfilzomib cardiotoxicity nomogram) use routine clinical variables and are immediately implementable. Standardized external validation and prospective trials are needed for widespread clinical application. Study Method Finding Metric Mosquera 2025 Ensemble ML 6-variables vs R-ISS C-index:0.66 Gagelmann 2024 Logistic MyCARe CAR T relapse score 7% vs 53% relapse Gómez-Martín 2025 IsomiR Daratumumab response AUC:0.98 Qiao 2025 XGBoost carfilzomib cardiotoxicity C:0.78 Martinez-Lopez 2024 AI class MRD dynamics p10⁻⁷ Morita 2024 3D-CNN whole-body-MRI prognosis AUC:0.80