Introduction
Multiple myeloma (MM) is a hematological malignancy characterized by clonal plasma cell proliferation [ 1 ]. Although novel agents and hematopoietic stem cell transplantation have improved survival, prognosis remains highly heterogeneous, ranging from durable remission to early progression and ultra-high-risk disease [ 2 ]. Current risk stratification systems, such as the revised International Staging System (R-ISS), incorporate cytogenetics and lactate dehydrogenase (LDH), but still cannot fully explain prognostic differences in real-world patients, especially those at intermediate risk [ 3 ]. Therefore, more comprehensive prediction strategies are needed.
Imaging is essential for the diagnosis, staging, and response assessment of MM. Whole-body low-dose computed tomography (CT), magnetic resonance imaging (MRI), and 18 F-fluorodeoxyglucose positron emission tomography (FDG PET)/CT provide complementary information on osteolytic lesions, bone marrow infiltration, extramedullary disease, and residual lesions [ 4 , 5 ]. Previous studies have also demonstrated the complementary value of MRI and PET/CT in evaluating disease burden and treatment response [ 6 ]. However, conventional imaging assessment mainly relies on visual interpretation and limited qualitative or semi-quantitative parameters, which may not fully capture the prognostic information contained in medical images.
Radiomics and artificial intelligence provide new approaches for extracting prognosis-related imaging phenotypes from high-dimensional medical images. Recent studies have developed MRI- or PET/CT-based radiomics models to predict cytogenetic abnormalities, minimal residual disease, and survival outcomes in MM, suggesting that radiomics may improve traditional prognostic stratification [ 7 ]. However, most studies have focused on a single imaging modality, used relatively small single-center cohorts, or lacked systematic comparison across different survival models [ 8 , 9 ]. Moreover, clinical machine learning models for overall survival or progression-free survival (PFS) prediction have often relied mainly on clinical and laboratory variables, with limited integration of radiomics, psychological status, and quality-of-life indicators [ 10 , 11 ].
To address these limitations, this study enrolled 300 newly diagnosed MM patients and integrated PET, CT, and MRI radiomics features to construct a radiomics-based risk score (Rad-score). The Rad-score was combined with clinical variables to develop fusion feature sets, and seven survival models were compared for predicting 12-month PFS. Model performance was evaluated using time-dependent receiver operating characteristic (ROC) curve analysis, Kaplan-Meier analysis, calibration curves, decision curve analysis (DCA), and Shapley additive explanations (SHAP) analysis. Thus, this study aimed to develop a multimodal, interpretable, and clinically applicable survival prediction strategy for MM.
Materials and Methods
Study Design and Subjects
This retrospective cohort study included 300 patients with MM diagnosed and treated at our hospital from June 2022 to June 2024. All patients underwent pretreatment PET/CT and MRI examinations and had the necessary available baseline clinical, laboratory, imaging, and follow-up data for model development and validation. The study was approved by the Ethics Committee of the Affiliated Hospital Six of Nantong University, Yancheng Third People’s Hospital (no: YTP2022-026, approved on May 11, 2022), and written informed consent was obtained from all patients.
Inclusion and Exclusion Criteria
Patients were included if they were aged ≥18 years, were clinically diagnosed with MM according to International Myeloma Working Group criteria, had undergone pretreatment PET/CT and MRI examinations, and had the essential clinical, imaging, and follow-up data required for model development and validation. Patients were excluded if they had received prior chemotherapy or radiotherapy, had other malignancies, were pregnant or lactating, had poor-quality images, or were participating in any clinical studies.
Treatment Methods
All patients received guideline-based anti-myeloma treatment according to current diagnostic and treatment guidelines. Treatment regimens were selected based on transplant eligibility, disease status, baseline risk stratification, organ function, and physician judgment. To account for potential treatment-related heterogeneity, detailed treatment information was collected and incorporated as candidate clinical variables. These variables included the type of induction regimen, such as bortezomib and dexamethasone (Vd), lenalidomide and dexamethasone (Rd), or bortezomib, lenalidomide, and dexamethasone (RVd); the number of treatment cycles; use of consolidation therapy; receipt of autologous hematopoietic stem cell transplantation; and use of novel agents. Drug types and doses were also reviewed from medical records when available. These treatment-related variables were included in model development to reduce the potential confounding effect of treatment differences on PFS.
Outcome Measures and Follow-Up
PFS was the primary outcome, defined as the time from diagnosis to the first occurrence of disease progression, clinical relapse, or death from any cause. Patients without these events by the end of the follow-up period were censored at the date of last follow-up. In this study, 12-month PFS was selected as the predefined time point for model evaluation because progression within the first year reflects early disease progression, which may indicate aggressive disease biology, suboptimal treatment response, and poorer prognosis in MM. In addition, all enrolled patients had a potential follow-up window of at least 12 months, supporting stable assessment at this time point.
Disease progression and clinical relapse after complete remission were counted as PFS events according to established criteria [ 12 ]. Disease progression was defined as any of the following changes from the lowest recorded value: an increase of ≥25% in serum M-protein, urine M-protein, the difference between involved and uninvolved serum free light chain, or bone marrow plasma cells, with the required absolute increases; the appearance of new soft tissue plasmacytoma lesions or ≥50% enlargement of existing measurable lesions; or a ≥50% increase in circulating plasma cells when they were the only measurable lesions. Clinical relapse was defined as the appearance of new bone lesions or soft tissue plasmacytomas, clear progression of existing lesions, hypercalcemia, a decrease in hemoglobin of ≥20 g/L unrelated to treatment or non-MM factors, an increase in serum creatinine of ≥176.8 µmol/L after treatment initiation, or hyperviscosity related to serum M-protein.
All progression and relapse events were independently assessed by two experienced hematologists and disagreements were resolved by discussion. Patients were followed through outpatient visits, inpatient re-examinations, and telephone interviews, with regular evaluation of laboratory indicators, imaging findings, and disease status. The follow-up cutoff date was June 30, 2025, ensuring a potential follow-up window of at least 12 months for all enrolled patients. During follow-up, PFS events occurred for 123 patients in the overall cohort, including 86 patients in the training cohort and 37 patients in the test cohort.
Dataset Division and Prevention of Information Leakage
The whole cohort was randomly divided into a training cohort and an independent test cohort at a ratio of 7:3 before any data-driven analysis. The training cohort included 210 patients, of whom 86 had experienced PFS events, whereas the independent test cohort included 90 patients, of whom 37 had experienced PFS events. The event proportion was comparable between the two cohorts, supporting the representativeness of the split. All preprocessing and model-development procedures were performed in a strictly training-set-only manner. Specifically, missing-value imputation models, normalization parameters, variance filtering, correlation filtering, radiomics feature selection, Rad-score construction, and hyperparameter tuning were fitted exclusively in the training cohort. The derived parameters, retained feature set, and final model coefficients or model structures were then directly applied to the independent test cohort. No test-set information was used for feature selection, parameter optimization, model fitting, or threshold determination.
To reduce overfitting, several safeguards were used. First, dimensionality reduction was performed before model training; second, radiomics features were screened by reproducibility, variance, correlation, and least absolute shrinkage and selection operator (LASSO)-Cox selection; third, hyperparameters were optimized by cross-validation within the training cohort only; and fourth, the final model was evaluated once in an independent test cohort.
Overall Study Workflow
The overall workflow of patient enrollment, image acquisition, radiomics feature extraction, feature selection, model construction, validation, and interpretation is summarized in Figure 1 .
Collection and Preprocessing of Clinical Variables
Baseline clinical variables included demographic data, Eastern Cooperative Oncology Group (ECOG) score, psychological and quality-of-life indicators, laboratory results, high-risk cytogenetic abnormalities, and treatment-related factors. Psychological indicators included anxiety and depression scores. Anxiety and depression were assessed using the Hospital Anxiety and Depression Scale-Anxiety subscale (HADS-A) and Hospital Anxiety and Depression Scale-Depression subscale (HADS-D), respectively. Quality of life was evaluated using the European Organisation for Research and Treatment of Cancer Quality of Life Questionnaire-Core 30 (EORTC QLQ-C30). All psychological and quality-of-life assessments were completed at baseline before treatment, and scores were calculated according to the standard scoring instructions for each questionnaire. Candidate variables were assessed for completeness and multicollinearity. Variables with >20% missing data were excluded, while variables with ≤20% missing data were imputed using multiple chained equations based on the training set. The imputation model derived from the training cohort was subsequently applied to the test cohort, and no information from the test cohort was used during imputation model fitting. A total of 23 clinically relevant and stable variables were retained for modeling.
Multimodal Image Data Acquisition and Lesion Delineation
All patients underwent pretreatment 18 F-FDG PET/CT and MRI examinations. PET/CT images were obtained using a Biograph mCT scanner (Siemens, Munich, Germany). According to the routine institutional protocol, PET imaging was performed approximately 60 min after intravenous injection of 18 F-FDG. MRI was performed using a 3.0 T Magnetom Skyra scanner (Siemens). The MRI protocol entailed conventional anatomical and functional sequences, including T1-weighted imaging, T2-weighted imaging, short tau inversion recovery, and diffusion-weighted imaging. Digital Imaging and Communications in Medicine (DICOM) images were imported to ITK-SNAP software for manual delineation of volumes of interest. For patients with multiple lesions, target lesions were selected according to lesion visibility, measurable size, image quality, and consistency across PET/CT and MRI. Lesions with unclear boundaries, severe artifacts, or insufficient size for reliable segmentation were excluded. When multiple eligible lesions were present, the representative lesion with the largest measurable size or highest metabolic activity was preferentially selected. Lesions were independently segmented by two experienced radiologists blinded to clinical data, and discrepancies were reviewed by a senior radiologist. Segmentation was performed with reference to both anatomical abnormalities and PET metabolic uptake to ensure consistency of lesion localization across modalities.
Radiomics Feature Extraction and Stability Screening
Before feature extraction, PET, CT, and MRI images were preprocessed using a standardized workflow to reduce inter-modality variability. All images were resampled to a uniform voxel spacing before feature extraction. PET images were converted to standardized uptake values, CT images were analyzed in Hounsfield units, and MRI signal intensities were normalized to reduce scanner- and sequence-related variability. Gray-level discretization was performed using consistent settings across all patients. After image preprocessing, radiomics features were extracted from PET, CT, and MRI images using the open-source Pyradiomics package. Extracted features included shape, histogram, and texture features. Feature stability was assessed using the intraclass correlation coefficient (ICC), and features with ICC of >0.75 were retained. To avoid data leakage, subsequent low-variance filtering and correlation-based feature reduction were performed only in the training cohort. The same retained feature set was then applied to the test cohort. After this process, 146 stable radiomics features were retained for subsequent analysis.
The radiomics feature-selection strategy followed a predefined sequential procedure. First, features with insufficient segmentation reproducibility were removed according to the ICC threshold. Second, low-variance features were excluded in the training cohort. Third, highly correlated features were reduced using correlation analysis to decrease redundancy. Fourth, the remaining radiomics features were entered into LASSO-Cox regression for supervised dimensionality reduction. Importantly, all data-driven feature filtering steps after the initial reproducibility assessment were fitted only in the training cohort, and the identical retained feature set was subsequently applied to the test cohort.
Establishment of the Rad-Score
The 146 retained features were subjected to LASSO-Cox regression for further dimensionality reduction. LASSO-Cox feature selection was performed exclusively in the training cohort. The optimal penalty parameter λ_min was selected through 10-fold cross-validation within the training cohort, and features with non-zero coefficients were used to construct the Rad-score as follows :
Here, β i represents the non-zero regression coefficient in the LASSO-Cox model, Feature i is the corresponding radiomics feature, and k represents the number of features finally retained. The Rad-score formula established in the training cohort was then directly applied to the test cohort for model evaluation. This procedure ensured that radiomics feature selection and Rad-score construction were fully nested within the training cohort, thereby reducing the risk of information leakage. As a unified multimodal imaging phenotype, the Rad-score comprehensively reflected the morphological, metabolic, and textural information in PET, CT and MRI images.
Construction of Fusion Feature Set
To reduce dimensionality and improve model stability, the Rad-score was combined with the 23 clinical variables to form a fusion feature set containing 24 features. All variables were standardized using Z-score normalization before model training. The mean and standard deviation used for Z-score normalization were calculated from the training cohort and then applied unchanged to the test cohort.
Establishment and Assessment of Prognostic Models
Prognostic models were developed using three feature sets: clinical features, radiomics features, and fusion features. Seven survival models were constructed for each feature set, including the Cox, LASSO-Cox, gradient boosting (GB)-Cox, CoxBoost, gradient boosting model (GBM), random survival forest, and support vector censored regression (SVCR) models. Hyperparameters were optimized using fivefold cross-validation in the training set, and model performance was evaluated in the independent test set.
For machine-learning survival models, hyperparameter optimization was performed using a grid search combined with fivefold cross-validation within the training cohort. The candidate hyperparameter space was prespecified before model evaluation. For tree-based models, the tuned parameters included the number of trees, maximum tree depth, learning rate, minimum samples per terminal node, and/or subsampling parameters, as appropriate for each algorithm. For the CoxBoost model, the number of boosting steps and penalty-related parameters were tuned. For SVCR, the kernel-related and regularization parameters were optimized. The hyperparameter combination with the best cross-validated performance in the training cohort was selected, after which the final model was refitted using the full training cohort and evaluated in the independent test cohort only.
Model Performance Assessment and Interpretability Analysis
Model performance was assessed using time-dependent ROC curves and 12-month area under the curve (AUC), sensitivity, specificity, accuracy, precision, and F1 score values. To provide a more comprehensive evaluation of model performance, the concordance index (C-index) and Brier score were also calculated. The 95% confidence intervals (CIs) for performance metrics were estimated using bootstrap resampling. The model with the best descriptive performance was further evaluated by Kaplan-Meier survival analysis, calibration curve, and DCA. SHAP analysis was used to assess feature importance and improve model interpretability.
Statistical Analysis
Statistical analyses were performed using IBM SPSS Statistics 26.0 (IBM Corp., Armonk, NY, USA) and Python 3.9 (Python Software Foundation, Wilmington, DE, USA). Continuous variables were expressed as mean ± standard deviation or median with interquartile range (IQR), depending on the distribution, and compared using t-tests or Mann-Whitney U tests. Categorical variables were expressed as frequencies and percentages and compared using chi-square or Fisher exact tests. A two-sided value of p<0.05 was considered statistically significant.
Results
Baseline Characteristics of Patients
A total of 300 patients with MM were included. The mean age was 64.9±8.1 years, 55.7% were male, and the mean body mass index value was 23.5±3.5 kg/m 2 . The median ECOG score was 1 (IQR: 0-2). High-risk cytogenetic abnormalities were present in 26.3% of patients. Key disease-related indicators, including ISS stage, R-ISS stage, β2-microglobulin, LDH, hemoglobin, blood calcium, and creatinine, were also recorded and are summarized in Table 1 . Vd, Rd, and RVd regimens accounted for 47.0%, 28.0%, and 25.0% of treatments, respectively. Autologous hematopoietic stem cell transplantation was performed for 64.7% of patients and 47.7% received novel agents. No significant differences were observed between the training and test cohorts in baseline clinical characteristics, disease-related indicators, or treatment-related variables, indicating that the two cohorts were generally comparable. Overall, the cohort showed heterogeneity in clinical and treatment characteristics, supporting subsequent prognostic model development.
Time-Dependent AUC Analysis for Different Feature Sets
Seven survival models were developed for each of the clinical, radiomics, and fusion feature sets. Time-dependent ROC analysis showed that the clinical and fusion feature sets had similar predictive performance, whereas the radiomics feature set showed lower AUC values across most time points ( 2A , 2B , 2C ). Although the fusion feature set showed slightly higher AUC values at some time points, the differences were limited. Therefore, further evaluation at a fixed clinically relevant time point was performed.
Comparison of Model Performance at 12 Months
At the 12-month time point, a clinically meaningful time point corresponding to the highest relapse density [ 13 ], 21 models based on clinical, radiomics, and fusion feature sets were compared in the independent test cohort. The test cohort included 90 patients, among whom 37 experienced PFS events within 12 months. Model performance was evaluated using AUC, sensitivity, specificity, accuracy, precision, F1 score, the Harrell C-index, and the 12-month Brier score.
Among all candidate models, the fusion feature set-based GBM showed the best overall descriptive performance, with the highest F1 score of 0.825 and a high AUC of 0.892. In the independent test cohort, this model achieved sensitivity of 0.892, specificity of 0.811, accuracy of 0.844, precision of 0.767, and F1 score of 0.825. The 12-month AUC of the fusion GBM was 0.892 (95% CI: 0.818-0.966) ( Table 2 ). Survival-specific evaluation further showed a Harrell C-index value of 0.804 (95% CI: 0.745-0.860) and 12-month Brier score of 0.143 (95% CI: 0.107-0.180), indicating good discriminative ability and acceptable prediction error.
Although the clinical Cox model showed a slightly higher AUC, the fusion GBM had the highest F1 score and a favorable balance between sensitivity and specificity. Therefore, the fusion feature set-based GBM was selected as the candidate model with the best overall descriptive performance for predicting 12-month progression and was used for subsequent risk stratification, calibration, DCA, and SHAP interpretation.
Risk Stratification Ability of the Fusion Feature Set-Based GBM
Patients in the independent test set were divided into low-, intermediate-, and high-risk groups according to tertiles of the risk scores generated by the fusion feature set-based GBM. Kaplan-Meier analysis showed clear separation of PFS curves among the three groups ( Figure 3 ). The high-risk group had the poorest PFS, whereas the low-risk group had the most favorable outcome. The difference was statistically significant (log-rank p<0.001), indicating good risk stratification ability.
Assessment of Calibration Ability of the Fusion Feature Set-Based GBM
The 12-month calibration curve showed good agreement between predicted and observed PFS probabilities in the test set ( Figure 4 ). The curve was close to the ideal reference line across most risk intervals, suggesting satisfactory calibration and reliable probability estimation.
Clinical Net Benefit of the Fusion Feature Set-Based GBM Evaluated by DCA
DCA showed that the fusion feature set-based GBM provided greater net benefit than the “treat-all” and “treat-none” strategies across a threshold probability range of approximately 0.10-0.70 ( Figure 5 ). It also showed higher descriptive net benefit than models based on clinical or radiomics features alone across most threshold ranges, suggesting its potential value for individualized risk assessment.
Feature Importance Analysis of the Fusion Feature Set-Based GBM
SHAP analysis showed that the Rad-score was among the most important predictors, indicating that radiomics provided additional prognostic information ( Figures 6A and 6B ). Important clinical predictors included β2-microglobulin, age, LDH, blood calcium, platelet count, and hemoglobin, which are recognized prognostic factors in MM [ 14 , 15 ]. High-risk cytogenetic abnormalities, renal function, performance status, and emotional and quality-of-life indicators also contributed to model prediction. These findings suggest that the fusion model integrated imaging and clinical information for multidimensional prognostic assessment.
Discussion
The present study explored the integration of PET, CT, and MRI radiomics features with 23 clinical variables and compared seven machine learning-based survival models. The fusion feature set-based GBM achieved the highest F1 score and showed balanced overall performance among the evaluated models, although some clinical models had comparable or slightly higher AUC values. SHAP analysis further showed that the Rad-score ranked among the important predictors, suggesting that multimodal radiomics may provide complementary prognostic information beyond conventional clinical factors.
Previous studies have reported the potential value of radiomics in MM prognosis. Guinea-Pérez et al. [ 16 ] constructed a PET/CT-based survival model and found that PET/CT fusion was associated with better predictive performance than single-modality models, but MRI and clinical variables were not incorporated. Another study developed an MRI-based nomogram for the prediction of overall survival, but it remained limited to a single imaging modality [ 17 ].
In addition, most machine learning studies in MM have mainly relied on clinical and laboratory variables, with limited integration of multidimensional information such as radiomics, psychological status, and quality of life [ 18 ].
In contrast to those studies, the present study attempted to provide a more comprehensive integration of multimodal imaging and clinical information, together with systematic feature construction, model comparison, and performance evaluation. However, the incremental value of this multimodal strategy should be interpreted cautiously because no formal statistical comparison with previously reported models was performed.
Model interpretability is important for the clinical translation of radiomics and machine learning. Existing PET/CT radiomics models for MM are often limited by black-box characteristics and unstable validation performance [ 9 , 19 ]. In contrast, the present study used SHAP analysis to quantify the contribution of individual features and illustrate how the Rad-score and key clinical indicators may contribute to survival risk estimation.
This approach is consistent with the increasing emphasis on the use of explainable artificial intelligence in hematological malignancies and may help to improve the interpretability of model-based risk prediction [ 20 ].
The model also has potential clinical implications. MM shows marked prognostic heterogeneity, and early identification of high-risk patients is crucial for individualized treatment [ 21 ]. Existing systems such as the ISS and R-ISS mainly rely on baseline laboratory and cytogenetic variables, including β2-microglobulin, albumin, LDH, and high-risk cytogenetic abnormalities. Although these systems are widely used, they may not fully reflect tumor spatial heterogeneity, metabolic activity, focal lesion characteristics, or whole-body disease burden captured by multimodal imaging. PET/CT- or MRI-based radiomics models have been explored for MM prognosis, but most previous studies focused on single imaging modalities or did not specifically address early progression risk within a clinically relevant time window [ 19 , 22 ]. Therefore, the proposed fusion model should be considered a potential complement to, rather than a replacement for, the ISS or R-ISS. By focusing on 12-month PFS prediction, the fusion GBM may help to identify patients at increased risk of early progression and support more individualized follow-up, including closer laboratory monitoring, earlier imaging reassessment, and careful evaluation of treatment response. For selected high-risk patients, the model may also provide additional information when considering risk-adapted treatment strategies, such as intensified induction therapy, earlier transplantation, or novel antibody-based agents, together with established staging systems and clinical judgment. Nevertheless, because this study lacked external validation and did not prospectively evaluate whether the model changes decision-making compared to the ISS or R-ISS, its clinical applicability remains preliminary. Future studies should examine whether adding this model to existing risk systems can improve risk classification, treatment decisions, and patient outcomes.
Study Limitations
The present study has some limitations to be noted. First, this was a single-center retrospective study with a relatively small sample size and no independent external validation cohort, which may limit the generalizability of the proposed model. This is particularly relevant for radiomics-based machine learning models, as radiomics features may be affected by differences in imaging protocols, reconstruction settings, preprocessing procedures, and lesion segmentation across institutions. Although standardized preprocessing and manual delineation were performed by experienced radiologists, potential protocol-related and segmentation-related variability could not be completely avoided. Second, treatment regimens were heterogeneous and residual confounding may remain despite the inclusion of key treatment-related variables in the model. Third, although ISS and R-ISS information was included in the baseline assessment, direct predictive performance comparison or incremental-value analysis against the ISS, R-ISS, or Second Revision of the ISS (R2-ISS) was not performed. Therefore, the proposed model should be regarded as a potential complement to, rather than a replacement for, established prognostic systems. In addition, the R2-ISS score could not be fully evaluated because complete data required for its calculation were unavailable in this retrospective cohort. Future multicenter prospective studies with external validation are needed to directly compare this model with established prognostic systems and further assess its robustness, generalizability, and clinical applicability. Integration of multi-omics data, including genomics, may also improve model performance.
Conclusion
In the present study, various machine learning-based survival models were constructed based on radiomics and clinical features, among which the fusion feature set-based GBM showed numerically favorable performance in the prediction of 12-month PFS, with relatively good discriminative ability, calibration performance, and potential clinical net benefits. The model’s transparency and interpretability were further supported by SHAP analysis. The integration of multimodal imaging data and clinical information may contribute to early risk prediction in MM and provide a potential reference for individualized treatment planning after further validation.
Ethics
Ethics Committee Approval: The study was approved by the Ethics Committee of the Affiliated Hospital Six of Nantong University, Yancheng Third People’s Hospital (no: YTP2022-026, approved on May 11, 2022).
Informed Consent: Written informed consent was obtained from all patients.
Authorship Contributions: Surgical and Medical Practices: S.S., W.G.; Concept: S.S., Z.H.; Design: S.S., Z.H.; Data Collection or Processing: Y.T., W.G., F.S.; Analysis or Interpretation: S.S., Z.H., Y.T., L.W.; Literature Search: L.W.; Writing: S.S., F.S.
Conflict of Interest: No conflict of interest was declared by the authors.
Financial Disclosure: This study was financially supported by the 2024 Yancheng Science and Technology Bureau Project (no. YCBE202457) and 2024 Research Project of the Yancheng Municipal Health Commission (no. YK2024162).
References
1 Fotiou D Katodritou E From biology to clinical practice: the bone marrow microenvironment in multiple myeloma. J Clin Med 2025 14 2 327 10.3390/jcm14020327 39860333 PMC11765558
2 Davies FE Pawlyn C Usmani SZ San-Miguel JF Einsele H Boyle EM Corre J Auclair D Cho HJ Lonial S Sonneveld P Stewart AK Bergsagel PL Kaiser MF Weisel K Keats JJ Mikhael JR Morgan KE Ghobrial IM Orlowski RZ Landgren CO Gay F Caers J Chng WJ Chari A Walker BA Kumar SK Costa LJ Anderson KC Morgan GJ Perspectives on the risk-stratified treatment of multiple myeloma. Blood Cancer Discov 2022 3 4 273 284 10.1158/2643-3230.BCD-21-0205 35653112 PMC9894570
3 Rajkumar SV Multiple myeloma: 2022 update on diagnosis, risk stratification, and management. Am J Hematol 2022 97 8 1086 1107 10.1002/ajh.26590 35560063 PMC9387011
4 Spinnato P Filonzi G Conficoni A Facchini G Ponti F Sambri A De Paolis M Cavo M Salizzoni E Nanni C Skeletal survey in multiple myeloma: role of imaging. Curr Med Imaging 2021 17 8 956 965 10.2174/1573405617666210126155129 33573573
5 Wu F Bernard S Fayad LM Ilaslan H Messiou C Moulopoulos LA Mulligan ME Updates and ongoing challenges in imaging of multiple myeloma: AJR expert panel narrative review. AJR Am J Roentgenol 2021 217 775 785 33978464 10.2214/AJR.21.25878
6 Michalska-Foryszewska A Rogowska A Kwiatkowska-Miernik A Sklinda K Mruk B Hus I Walecki J Role of imaging in multiple myeloma: a potential opportunity for quantitative imaging and radiomics? Cancers 2024 16 23 4099 10.3390/cancers16234099 39682285 PMC11640347
7 Chen H Han J Huang H He Q Ren X Yu F Chang C Ding X Luo Q A feasibility study of [18F] FDG PET/CT radiomics in predicting high-risk cytogenetic abnormalities in multiple myeloma. EJNMMI Res 2025 15 1 131 10.1186/s13550-025-01321-8 41091252 PMC12528638
8 Sachpekidis C Goldschmidt H Edenbrandt L Dimitrakopoulou-Strauss A Radiomics and artificial intelligence landscape for [ 18 F] FDG PET/CT in multiple myeloma. Semin Nucl Med 2025 55 3 387 395 10.1053/j.semnuclmed.2024.11.005 39674756
9 Klontzas ME Triantafyllou M Leventis D Koltsakis E Kalarakis G Tzortzakakis A Karantanas AH Radiomics analysis for multiple myeloma: a systematic review with radiomics quality scoring. Diagnostics (Basel) 2023 13 12 2021 10.3390/diagnostics13122021 37370916 PMC10296889
10 Romero M Mosquera Orgueira A Mejía Saldarriaga M How artificial intelligence revolutionizes the world of multiple myeloma. Front Hematol 2024 3 1331109
11 Ren L Xu B Xu J Li J Jiang J Ren Y Liu P A machine learning model to predict survival and therapeutic responses in multiple myeloma. Int J Mol Sci 2023 24 6683 37047654 10.3390/ijms24076683 PMC10095137
12 Rajkumar SV Dimopoulos MA Palumbo A Blade J Merlini G Mateos MV Kumar S Hillengass J Kastritis E Richardson P Landgren O Paiva B Dispenzieri A Weiss B LeLeu X Zweegman S Lonial S Rosinol L Zamagni E Jagannath S Sezer O Kristinsson SY Caers J Usmani SZ Lahuerta JJ Johnsen HE Beksac M Cavo M Goldschmidt H Terpos E Kyle RA Anderson KC Durie BG San Miguel JF International Myeloma Working Group updated criteria for the diagnosis of multiple myeloma. Lancet Oncol 2014 15 12 e538 e548 10.1016/s1470-2045(14)70442-5 25439696
13 Landgren O Prior TJ Masterson T Heuck C Bueno OF Dash AB Einsele H Goldschmidt H Knop S Li C Mellqvist UH McFadden I Oprea C Ross JA Talpes M Hydren JR Ahlstrom JM Kazandjian D Weinhold N Zhang R Stetler-Stevenson M Marti G Devlin SM EVIDENCE meta-analysis: evaluating minimal residual disease as an intermediate clinical end point for multiple myeloma. Blood 2024 144 4 359 367 10.1182/blood.2024024371 38768337 PMC11418064
14 Zhang T Lin Z Zheng Z Wang Q Zhou S Zhang B Zheng D Chen Z Zheng S Zhang Y Lin X Dong R Chen J Qian H Hu X Zhuang Y Zhang Q Jin Z Jiang S Ma Y Prognostic significance of β2-microglobulin decline index in multiple myeloma. Front Oncol 2024 14 1322680 38562177 10.3389/fonc.2024.1322680 PMC10982376
15 Du C Li L Fan H Mao X Liu J Xu Y Sui W Deng S Li C Zhao J Yi S Hao M Zou D Zhao Y Qiu L An G The age-dependent changes in risk weights of the prognostic factors for multiple myeloma. Hematology 2023 28 1 2258686 10.1080/16078454.2023.2258686 37724573
16 Guinea-Pérez J Ceresi A Fernández García A Arroyo Galende B Belmonte-Hernández A Peluso S Alvarez F Radiomics feature analysis for survival prediction in multiple myeloma: an automated PET/CT approach. Comput Methods Programs Biomed 2025 271 109019 10.1016/j.cmpb.2025.109019 40865375
17 Li Y Liu Y Yin P Hao C Sun C Chen L Wang S Hong N MRI-based bone marrow radiomics nomogram for prediction of overall survival in patients with multiple myeloma. Front Oncol 2021 11 709813 34926240 10.3389/fonc.2021.709813 PMC8671997
18 Sharma R Desai K Dadhania N Croce P Thalji M Hanif A Mirza W ElManzalawi Y Systematic review of currently available predictive machine learning models for multiple myeloma. Blood 2024 144 Suppl 1 7518 10.1182/blood-2024-211834
19 Ni B Huang G Huang H Wang T Han X Shen L Chen Y Hou J Machine learning model based on optimized radiomics feature from 18F-FDG-PET/CT and clinical characteristics predicts prognosis of multiple myeloma: a preliminary study. J Clin Med 2023 12 2280 36983281 10.3390/jcm12062280 PMC10059677
20 Nazha A Elemento O Ahuja S Lam B Miles M Shouval R McWeeney S Sirhan S Srisuwananukorn A Haferlach T Artificial intelligence in hematology. Blood 2025 146 19 2283 2292 10.1182/blood.2025029876 40845137
21 Schavgoulidze A Perrot A Leleu X Cazaubiel T Chretien ML Feugier P Belhadj K Manier S Roussel M Brechignac S Orsini-Piocelle F Mohty M Schiano de Collela JM Macro M Adiko D Dib M Fontan J Luttiau-Motard C Bouscary D Pascal L Roland V Lifermann F Bakala J Montes L Kennel C Rey P Richez V Keddar F Frenzel L Calmettes C Chaleteix C Plantier I Chalayer E Schmitt A Roul C Demarquette H Cerutti C Pavageau L Derrier L Avet-Loiseau H Corre J High-risk genomic consensus validation for patients with newly diagnosed multiple myeloma using next-generation sequencing. Blood 2026 147 3 266 275 10.1182/blood.2025029999 40991836 PMC13292120
22 Yu T Chen M Huang B Liu J Li X Gu J Li Y Li J Clinical feature and predictive model for transplanted patients with functional high-risk multiple myeloma. Arch Med Res 2026 57 103291 40912194 10.1016/j.arcmed.2025.103291