Многомерное профилирование и модели риска, основанные на клиническом развертывании АИ, в МГУС и миеломе.

Introduction

Monoclonal gammopathy of undetermined significance (MGUS), smoldering multiple myeloma (SMM), and multiple myeloma (MM) represent a spectrum of plasma cell disorders that range from benign to malignant [ 1 ]. While MGUS and SMM are considered precursor states to MM, the progression from MGUS to active myeloma remains poorly understood and difficult to predict. As a result, there is a critical need for improved methods of risk stratification to better identify patients at high risk of progression and guide clinical management. Traditional risk models have relied on clinical, laboratory, and cytogenetic markers, but these models often fall short in providing personalized predictions, especially in light of the complex molecular underpinnings of the disease [ 2 ].

Recent advancements in multi-omics technologies, such as genomics, transcriptomics, proteomics, and metabolomics, have revealed the intricate molecular landscape of these disorders, offering unprecedented insights into disease progression. Furthermore, the integration of artificial intelligence (AI) and machine learning (ML) techniques has opened new avenues for developing more accurate and dynamic risk models [ 3 ]. AI-based approaches can leverage the vast amounts of multi-omics data to identify previously unrecognized patterns and predict patient outcomes with higher precision. The combination of AI with multi-omics data promises to enhance our ability to stratify risk at a molecular level, enabling more personalized treatment strategies and improving patient outcomes [ 4 ].

Despite the potential of these technological advancements, several challenges remain in their clinical implementation. Issues such as data integration across diverse platforms, the need for standardized protocols, and concerns around data privacy and security are significant barriers that must be addressed. Additionally, ethical considerations, such as ensuring equitable access to AI-based tools and mitigating algorithmic bias, are critical for the widespread adoption of these technologies. Regulatory frameworks will also need to evolve to accommodate continuously learning AI systems, ensuring their safety and efficacy in clinical practice [ 5 ]. Existing MGUS and SMM risk models such as the Mayo Clinic score for MGUS, which uses M-protein concentration, isotype and serum free light chain ratio, and the Spanish PETHEMA models, which rely on flow-cytometric aberrant plasma cells and immunoparesis or DNA ploidy, have been highly influential in defining clinical risk groups [ 6 – 8 ]. However, they capture only a small number of baseline clinical or immunophenotypic variables, assume relatively homogeneous biology within each risk category and do not systematically incorporate cytogenetic or genomic high-risk lesions such as 1q gain or TP53 abnormalities, which are now recognized as key determinants of outcome [ 9 – 11 ]. These constraints contribute to substantial residual heterogeneity in progression risk within each group and to miscalibration of predicted risk in contemporary cohorts treated with modern therapies [ 12 , 13 ].

At the same time, multi-omics and AI-based models introduce new challenges, including limited interpretability of complex “black-box” architectures, risk of bias and dataset shift, and inadequate external validation or transparent reporting, all of which can undermine reproducibility and clinical trust [ 14 ]. Addressing these issues is essential if AI-integrated risk tools are to move beyond proof-of-concept and achieve safe, widespread deployment.

In this review, we first summarize the current MGUS and SMM risk stratification systems and highlight where they fall short in capturing underlying biological heterogeneity. We then synthesize recent multi-omics and AI-based approaches along the MGUS–SMM–MM continuum, focusing on how genomic, transcriptomic and microenvironmental features can be integrated with clinical variables. Rather than proposing yet another single “signature”, our aim is to distill the design principles of AI-driven, clinically deployable risk models, including choice of training cohort, feature integration, calibration and validation strategies, and workflow integration, that can ultimately generate robust, interpretable and actionable risk scores for use in routine practice.

Current understanding of precursor stages and risk progression

Monoclonal gammopathy of undetermined significance (MGUS), smoldering multiple myeloma (SMM), and active multiple myeloma (MM) represent a continuum of plasma cell disorders, with distinct diagnostic criteria that have evolved over time [ 1 ]. MGUS is characterized by the presence of a monoclonal protein (M-protein) in the serum, without evidence of organ damage or malignancy, typically defined as a serum monoclonal protein level of < 3 g/dL, less than 10% clonal plasma cells in the bone marrow, and no clinical signs of multiple myeloma. Although MGUS is asymptomatic, it is considered a precursor to multiple myeloma, as approximately 1% of individuals with MGUS progress to MM annually [ 15 ].

SMM is a step closer to full-blown MM and is defined by higher levels of monoclonal protein (≥ 3 g/dL) and a higher percentage of clonal plasma cells in the bone marrow (≥ 10%), but without the characteristic end-organ damage seen in MM (disseminated bone lesions, anemia, hypercalcemia, or renal impairment) [ 16 ]. Importantly, SMM represents a risk state where progression to active MM is highly probable if left unchecked, with progression rates between 10 and 20% annually in certain high-risk subgroups [ 8 ].

Active MM is a malignant disease characterized by significant clonal proliferation of plasma cells in the bone marrow, with or without extramedullary disease, and associated with end-organ damage [ 17 ]. Diagnostic criteria for active MM include the presence of ≥ 10% clonal plasma cells in the bone marrow, M-protein in the serum or urine, and evidence of end-organ damage (e.g., renal failure, bone lesions, or anemia) as defined by the International Myeloma Working Group (IMWG). The presence of high-risk cytogenetic abnormalities and genetic mutations, as well as clinical symptoms, distinguishes active MM from its precursor stages [ 18 ].

The epidemiology of MGUS, SMM, and MM reveals distinct population-level patterns, with MGUS being more common in older adults and men, particularly those of African ancestry [ 19 ]. The prevalence of MGUS increases with age, affecting approximately 3% of individuals aged 50 years and older, and it is more frequent in African Americans compared to Caucasians. While MGUS itself is generally asymptomatic, it carries a 1% per year risk of progressing to multiple myeloma, with genetic ancestry playing a pivotal role in influencing progression rates [ 20 ]. Specifically, individuals of African descent have a higher risk of developing MM from MGUS, with studies indicating that the incidence of progression to MM is significantly higher in African American populations compared to Caucasians [ 21 ].

The natural history of MGUS and SMM reflects the complex interplay between genetic factors and environmental influences. While some individuals with MGUS remain stable over decades, others progress to SMM or MM, underscoring the importance of understanding individual risk factors [ 22 ]. Studies have shown that genetic risk factors, such as the presence of certain single nucleotide polymorphisms (SNPs) and mutations in genes such as TP53 and KRAS, can influence the likelihood of progression. Furthermore, environmental factors such as occupational exposures, autoimmune diseases, and infections may also contribute to the progression of these precursor conditions [ 23 ].

Over the years, several risk models have been developed to predict the progression of MGUS and SMM to active MM. These models are primarily based on laboratory and cytogenetic parameters, offering valuable prognostic information, though they remain limited by their reliance on relatively few variables [ 24 ]. The IMWG model, for example, identifies the presence of specific cytogenetic abnormalities, such as del(17p), t(4;14), and t(14;16), as significant risk factors for progression from MGUS to MM. However, this model focuses primarily on cytogenetic alterations, neglecting other potentially important molecular and immune factors [ 25 ].

The Mayo Clinic model incorporates serum M-protein levels and the percentage of plasma cells in the bone marrow, identifying individuals with high M-protein levels and an elevated bone marrow plasma cell burden as being at greater risk of progression [ 26 ]. The Spanish PETHEMA model uses similar parameters but also takes into account the presence of clonal plasma cells in extramedullary sites and the role of high-risk cytogenetic abnormalities. While these models provide important insights, they are limited in their ability to accurately predict progression on an individual level due to their reliance on a small number of parameters [ 27 ].

Comparative data illustrate both the strengths and limitations of these clinical scores. For MGUS, the Mayo model stratifies patients into four categories with 20-year progression risks ranging from about 5% in the lowest-risk group to nearly 60% in the highest-risk group, underscoring its ability to separate clearly low- and high-risk patients while still leaving substantial residual risk within categories [ 28 ]. In SMM, the original Mayo 2008 model and the Spanish PETHEMA immunophenotypic model both identify high-risk subgroups with roughly 50–70% risk of progression within 5 years, but they rely on different variables and show only moderate concordance when applied to the same cohorts [ 29 ]. The more recent IMWG 2/20/20 model achieves similar enrichment, with high-risk patients having around a 50% probability of progression within 2 years, and benefits from using simple, widely available clinical measurements [ 30 , 31 ]. However, independent validations indicate that these scores can be miscalibrated in contemporary, heterogeneously treated cohorts, often overestimating near-term progression risk, and they remain largely agnostic to underlying genomic drivers and microenvironmental features [ 32 ]. Overall, current models are easy to implement and useful for defining broad risk categories, but provide only moderate discrimination and limited biological granularity at the individual-patient level.

Despite the utility of existing risk models, several limitations hinder their ability to provide precise predictions at the individual level. One of the most significant challenges is the lack of longitudinal data, as many models rely on cross-sectional data that do not capture the temporal evolution of the disease. For example, the Mayo Clinic and IMWG models primarily consider baseline measurements of serum M-protein and bone marrow plasma cell percentages, but they do not incorporate dynamic changes over time, which may be crucial for accurately predicting progression [ 32 ].

Another limitation is the poor adaptability of these models to incorporate new molecular markers, such as genetic mutations, epigenetic modifications, and immune cell dynamics. Advances in multi-omics technologies have revealed that disease progression in MGUS and SMM is driven by a complex network of genomic, epigenomic, and immune-related changes [ 33 , 34 ], yet traditional models have been slow to integrate these factors into their predictive frameworks. As a result, current models often fail to account for the heterogeneity of disease progression, particularly among individuals with high-risk molecular features [ 35 ].

Furthermore, the precision of these models at the individual patient level remains suboptimal. Many risk models lack sufficient granularity to stratify patients based on the molecular complexity of their disease, leading to the over-simplification of progression risks. This underscores the need for more personalized and dynamic approaches to risk stratification that incorporate a broader range of omics data and better reflect the complexity of disease evolution [ 36 , 37 ]. In summary, while current risk models have provided important insights, they need refinement to integrate emerging molecular markers and better predict progression at an individualized level (Table 1 ).

Molecular layers in MGUS to myeloma transformation

Progression from a pre-malignant MGUS clone to full-blown myeloma is an evolutionary process driven by accumulating molecular abnormalities. It is now appreciated that virtually all genomic events defining overt MM can already be found in rare cells at the MGUS stage. However, progression is marked by quantitative expansion of subclones and qualitative shifts in the genomic landscape and surrounding ecosystem [ 38 ].

Genomic landscape

The founding genetic lesions in MGUS typically include either hyperdiploidy (multiple trisomies of odd-numbered chromosomes) or a translocation involving the immunoglobulin heavy chain locus (e.g. t(11;14), t(4;14)), along with other early hits like chromosome 13 deletion [ 39 ]. These primary drivers are common to MGUS and MM, implying that the initial transformation of a normal plasma cell to a neoplastic clone does not yet confer malignancy [ 40 ]. Whole-genome sequencing studies have shown that as MGUS progresses, the tumor accumulates secondary mutations (e.g. activating mutations in KRAS/NRAS, BRAF, and other MAPK pathway genes), copy number abnormalities (like 1q21 amplifications), and structural variations such as chromothripsis or chromoplexy [ 23 ]. The overall mutation load increases and correlates with more aggressive behavior. For example, serial sequencing of paired MGUS and MM samples reveals that most non-synonymous mutations present in MM were already detectable at low frequency in antecedent MGUS, but additional novel mutations, often linked to APOBEC-induced mutagenesis, emerge during the transition, fueling clonal expansion [ 41 ]. High-risk genomic features such as TP53 deletions (17p−) or biallelic inactivation, complex structural events, and Myc translocations are rarely seen in stable MGUS but are enriched in progressing cases and frank myeloma [ 42 ]. Clonal evolution in SMM often follows a branching pattern: multiple subclones compete, and a more malignant subclone eventually dominates and triggers symptomatic disease. This dynamic interplay underlines the importance of assessing clonal architecture and genomics, beyond static clinical markers, in order to gauge progression risk [ 43 ].

Epigenomic and transcriptomic changes

In addition to DNA sequence alterations, epigenetic reprogramming accompanies the MGUS-to-MM transition. Global DNA methylation patterns show progressive hypomethylation from normal plasma cells to MGUS to MM, particularly at intergenic regions, reflecting a more dedifferentiated state. Regions controlling tumor suppressor genes may become aberrantly methylated and silenced, while oncogenic pathways can be activated via chromatin changes [ 44 ]. Similarly, changes in histone modification landscapes (e.g. gains of active marks at proliferation genes) occur as the clone evolves. These epigenetic shifts lead to distinct gene expression profiles when comparing MGUS vs. MM plasma cells [ 45 , 46 ]. Multiple studies have attempted to identify mRNA signatures predictive of progression: for instance, genes involved in MYC targets, E2F signaling (cell cycle progression), NF-κB signaling, and DNA repair have been found upregulated in SMM or MM relative to MGUS [ 47 – 49 ]. Conversely, certain differentiation markers and apoptotic regulators may be downregulated. A recent multi-omics analysis integrated transcriptomic and proteomic data, identifying that MGUS→MM progression was associated with enhanced translation machinery and stress response pathways (e.g. EIF2 signaling and ribosomal quality control were upregulated), whereas cell adhesion and surface immune proteins were lost in MM [ 50 ]. Notably, a subset of newly diagnosed myelomas exhibited an “MGUS-like” proteomic profile and had better outcomes, reinforcing that molecular phenotype, not just clinical stage, matters [ 51 ]. Overall, transcriptome and proteome alterations indicate that progressing clones shift toward a state of higher protein synthesis (to support rapid growth) while shedding surface antigens (potentially to evade immune detection). These changes can potentially serve as biomarkers; indeed, one study found that an in vitro “protein translation fitness” score separated MGUS from MM and correlated with progression risk [ 52 ].

Microenvironment and immunological factors

The bone marrow microenvironment plays an integral role in constraining or facilitating myeloma development. In precursor stages, the immune system often keeps the clonal plasma cells in check – for example, T cells from MGUS patients show a more robust cytotoxic profile compared to those in MM [ 53 ]. However, even at the MGUS stage, subtle microenvironmental changes can foreshadow progression. Single-cell RNA sequencing and mass cytometry studies have revealed early alterations such as an increase in exhausted T cells and regulatory T cells and a decrease in memory T cells and dendritic cell populations in MGUS/SMM compared to healthy individuals [ 54 ]. In one study, transcriptional profiling of immune cells showed that pro-inflammatory signaling and innate immune activation were already dampened in high-risk SMM, suggesting an evolving immune escape [ 55 ]. Moreover, MGUS patients with reduced polyclonal immunoglobulin levels (immunoparesis), a sign of broader B-cell suppression, have a higher likelihood of progression [ 56 ]. As the disease progresses, bone marrow stromal cells also co-evolve: mesenchymal stromal cells in myeloma acquire an inflammatory, tumor-promoting phenotype (for instance, secreting high levels of IL-6, a growth factor for myeloma cells) [ 57 ]. The net effect is a permissive niche wherein malignant plasma cells face less immune surveillance and receive more growth and survival signals. This evolving microenvironment is both a consequence of the tumor’s presence and a driver of its further progression. Importantly, certain microenvironmental biomarkers have predictive value, for example, an increased ratio of immune suppressive to cytotoxic cells in SMM has been correlated with shorter time to progression [ 58 , 59 ].

Metabolic and other alterations

Metabolic reprogramming in the precursor states is less studied, but emerging evidence suggests differences in energy metabolism and bone interactions. Proteomic analyses indicate that MGUS cells rely more on oxidative metabolism, whereas MM cells upregulate glycolytic and mitochondrial stress pathways [ 60 ]. Additionally, interactions with the bone microenvironment change, progressing lesions more often show osteoclast activation and bone destruction, likely via factors like RANKL and DKK1 secreted as the clone expands. While clinical biomarkers of bone turnover (e.g. bone lesions on imaging) are part of defining active myeloma, these are late changes; future metabolomic or secretomic markers might detect more subtle shifts earlier [ 61 ].

In summary, progression from MGUS to myeloma is a multi-step, multi-omic process: the clonal plasma cells acquire additional driver lesions, their gene expression and protein output adapt to support unchecked growth, and the surrounding immune-stromal environment gradually shifts from restraining to enabling tumor expansion [ 62 ]. Each of these layers, genome, epigenome, transcriptome, proteome, metabolome, and microenvironment, contributes pieces to the puzzle of risk. This complexity helps explain why any single clinical marker is insufficient for strong prediction. It also provides the rationale that an integrated multi-omics approach, capturing many of these alterations simultaneously, could dramatically improve the stratification of which MGUS/SMM patients are truly on the trajectory toward myeloma (Table 2 ).

Taken together, not all molecular events along the MGUS–SMM–MM continuum have the same predictive weight. Among cytogenetic abnormalities, gain/amplification of 1q, del(17p)/TP53 inactivation, and IgH translocations such as t(4;14) and t(14;16) consistently associate with higher risk of progression and inferior survival and are increasingly incorporated into clinical high-risk definitions [ 63 ]. In contrast, many recurrent mutations in genes such as KRAS, NRAS, BRAF, DIS3 and FAM46C are very common in established MM but show more variable and often context-dependent associations with progression, limiting their immediate utility as stand-alone biomarkers in precursor conditions [ 64 ]. Emerging genomic features—including high genome-wide mutational burden, APOBEC-driven mutational signatures, complex structural events (e.g. chromothripsis), and composite “MM-like” or “genomic MM” scores—appear to capture a truly malignant biology and identify MGUS/SMM cases at particularly high risk of near-term transformation, but they remain largely research-grade and require further standardization and prospective validation before routine clinical use [ 65 , 66 ]. Overall, the current evidence suggests that a small set of well-established cytogenetic lesions already have actionable prognostic value, whereas most other molecular events are best integrated into multi-parameter genomic or multi-omics scores rather than interpreted in isolation.

Multi-omics integration and AI modeling strategies

Harnessing the wealth of information across different “omics” requires computational strategies to integrate diverse data types into a cohesive predictive model. Unlike traditional statistical methods that struggle with high-dimensional data, modern ML and deep learning algorithms are well-suited to identify complex, multivariate patterns indicative of progression risk [ 67 ]. In recent years, researchers have developed various integration techniques to merge multi-omics inputs and feed them into predictive models, each with advantages and drawbacks.

Integration strategies

A fundamental design question is how and when to combine data from multiple omics layers during model training. In an early integration (also known as “concatenation-based”) approach, all features from genomics, transcriptomics, proteomics, etc., are concatenated into one large feature vector per sample from the start , and a single model (e.g. a classifier or regression model) is trained on this combined feature set [ 68 ]. Early integration is straightforward but can lead to extremely high-dimensional inputs, exacerbating the “curse of dimensionality” and introducing noise or multicollinearity (since different omics may measure correlated biological phenomena) [ 69 ]. At the other extreme, late integration (a “model-based” or ensemble approach) involves first training separate models on each data modality (for instance, a genomics model, an epigenomics model, etc.) and then combining their outputs (through voting, averaging, or a second-level model) to produce a final prediction [ 70 ]. Late integration can leverage modality-specific modeling strengths and is modular (one can plug in or remove a data type), but it might miss cross-talk patterns between omics and often requires large sample sizes to train multiple models [ 71 ].

Increasingly popular are intermediate (mixed) integration strategies, which seek a middle ground. One common approach is to first apply dimensionality reduction or embedding techniques to each omic separately, for example, using principal component analysis (PCA), autoencoders, or other feature extraction on each dataset, and then merge those lower-dimensional representations for the final predictive model [ 72 ]. This can dramatically reduce noise and feature count while retaining the most informative signals from each layer. A variant of this is hierarchical integration, where one model learns representations from one omic, then incorporates the next omic in a stepwise fashion (often used when one data type is considered primary). There are also network-based methods that create a multilayer graph of interactions (genes, proteins, etc. as nodes with multi-omic annotations) and use algorithms like graph neural networks to propagate information through the graph [ 73 , 74 ]. This graph-based integration can capture complex relationships (e.g. regulatory networks) that flat feature matrices cannot. Ultimately, no single integration strategy is universally best, the optimal choice depends on the data properties and the specific prediction task. For MGUS/SMM, where sample sizes are relatively modest and data types numerous, methods that effectively reduce dimensionality and fuse information without overfitting are crucial [ 67 ].

Dimensionality Reduction and Feature Selection: Multi-omics datasets often encompass tens of thousands of features (e.g. gene expression values, mutation flags, methylation sites, etc.) for each sample. Dimensionality reduction is therefore a critical step to distill signals and avoid overfitting. Unsupervised techniques like PCA and t-distributed stochastic neighbor embedding (t-SNE) have been useful for visualization and to confirm that MGUS vs. MM samples, for instance, occupy different regions of molecular space [ 75 ]. More sophisticated methods include variational autoencoders (VAEs) and other neural network embeddings: these “learn” a compressed latent representation of multi-omic data by trying to reconstruct the original data from a bottleneck layer [ 76 ]. The OmiEmbed framework is a notable example, using a multi-omics VAE to embed high-dimensional profiles into a lower-dimensional latent vector that still contains the relevant information for downstream tasks. In OmiEmbed, separate encoders for each omic type transform the data, then concatenation and further encoding yield a unified latent space, which subsequently informs prediction tasks [ 75 ]. By training this embedding with multi-task objectives (e.g. simultaneously predicting survival, tumor subtype, etc.), the model captures a holistic representation of the disease state [ 77 ]. In evaluations on cancer datasets, such deep embeddings outperformed naive early integration in classification and survival prediction, likely because the network automatically denoises and emphasizes features most predictive of outcome [ 78 ]. Aside from deep learning, classical feature selection is also employed – for example, Karathanasis et al. applied Elastic Net regularization to gene expression data to select a subset of genes most associated with progression [ 79 ]. Similarly, the 36-gene MGUS signature by Sun et al. was identified through cross-validated selection of the top genes concordantly linked to progression in multiple training subsets. Such approaches reduce thousands of genes to a manageable signature that is both predictive and biologically interpretable (often enriched for pathways like cell cycle or DNA repair, as noted) [ 80 , 81 ]. In sum, effective dimensionality reduction, whether via algorithmic embedding or expert-driven feature curation, is foundational for building robust multi-omics models in this domain.

Machine Learning vs. Deep Learning Models: With a unified feature set in hand (or the ability to generate one on the fly), the next step is model training. A range of ML and DL methods have been explored for MM risk prediction. Traditional ML algorithms like random forests (RF), support vector machines (SVM), gradient boosting (e.g. XGBoost), and Elastic Net logistic regression have the advantage of working well on smaller datasets and providing some insight into feature importance. For instance, an RF model can rank variables by Gini importance or information gain, which can highlight which mutations, genes, or labs are most influential in predicting progression [ 82 ]. In practice, Santamaría et al. evaluated multiple classifiers (RF, SVM, k-NN, etc.) on routine clinical data of MGUS patients and found SVM performed best, achieving ~ 94% accuracy in identifying the few progressors in their cohort [ 83 ]. In Karathanasis’ multi-dataset gene expression study, Elastic Net (a regularized regression) consistently yielded high AUCs (~ 0.8–0.85) for MGUS progression prediction and had the benefit of automatic gene selection via its sparse coefficients. These classical ML models are relatively transparent and easier to implement, but they may not fully capture nonlinear interactions between dozens of omic features without substantial tuning [ 79 ].

Deep learning approaches, on the other hand, excel at learning complex functions from large data. Deep neural networks (DNNs) have been applied to myeloma multi-omics in various ways: from simple feed-forward networks that take concatenated features, to more specialized architectures. One cutting-edge example applied to myeloma is the geometric graph neural network (GGNN) developed by Zhu et al. [ 84 ]. This approach constructed a graph of genes/proteins connected by prior knowledge (protein–protein interactions, pathways) and computed network geometry features (like Ollivier-Ricci curvature) to feed into a GNN that predicts survival [ 84 ]. By incorporating the topology of molecular interactions, the model improved prognostic accuracy in the CoMMpass MM dataset and provided an interpretable framework highlighting key network hubs (e.g. identifying curvature changes around genes like CCND1 in high-risk patients) [ 84 ]. Another deep learning strategy uses attention mechanisms to integrate multi-omics: e.g. a cross-attention model can weight the contribution of each omic modality differently for each patient, essentially learning which data type is most informative in a given context (as proposed in CrossAttOmics for other cancers). Graph-based and attention-based DL models are powerful, but they typically require larger sample sizes and careful validation to avoid overfitting to noise [ 85 , 86 ].

Explainable AI: Regardless of model type, an essential consideration for clinical deployment is explainability. Clinicians and regulatory bodies will demand clarity on why an AI model flags a particular MGUS patient as high-risk. Tools like SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-Agnostic Explanations) have become standard for interpreting ML/DL predictions in biomedicine. These methods can rank features (e.g. specific gene expression levels, presence of a mutation, or a lab value) by their contribution to an individual prediction [ 87 ]. In the MGUS context, a SHAP summary might reveal, for instance, that a high risk score for a patient is primarily driven by an extremely aberrant FLC ratio combined with upregulation of cell-cycle genes and the presence of 1q gain, a combination that intuitively aligns with aggressive biology [ 88 ]. Explainability was highlighted in Karathanasis et al.’s work, where pathway enrichment analysis on the selected genes revealed involvement of PI3K-Akt, MAPK, Wnt, and mTOR pathways in progression [ 79 ], thus providing biological context to the black-box models. By using explainable AI, researchers turned the model’s “brain” inside out to some degree, reassuring that known myeloma-driving pathways were indeed being picked up by the algorithm. For deep models, integrated gradients and attention weight visualization are also used to interpret which features or sequence positions are most influential. Explainability not only aids clinician trust but can sometimes lead to discovery of novel biomarkers, for example, if an AI model consistently places heavy weight on a little-known gene, that gene may warrant further study as a potential player in progression [ 89 ].

Performance Evaluation: To ensure reliability, multi-omics models must be rigorously evaluated using proper validation frameworks. Given the limited size of many MGUS/SMM datasets, strategies like cross-validation and bootstrapping are commonly employed to estimate performance without overfitting. Ideally, models are trained on one cohort and tested on independent cohorts (external validation) to gauge generalizability. For instance, the PANGEA model was trained on a Dana-Farber cohort and prospectively validated on cohorts from Greece, UK, and Czech Republic [ 90 ]. Common performance metrics include the area under the ROC curve (AUROC) for classification (e.g. high- vs. low-risk), sensitivity at a fixed specificity (to see how many progressors can be caught with few false alarms), and for time-to-event predictions, the concordance index (C-index) or time-dependent AUC. In Sun et al.’s MGUS 36-gene model, a remarkable Harrell’s C-statistic of 0.928 was achieved for 10-year progression [ 80 ]. However, such high apparent performance in a single study must be viewed with caution until validated elsewhere, models may overfit or suffer from bias in patient selection. Therefore, the field is moving toward multi-institutional data pooling and consistent benchmark tasks to allow apples-to-apples comparisons of model performance. Ultimately, an AI risk score would need not just statistical significance but also clinical utility, for example, a high-risk group with, say, > 50% 5-year progression risk that clearly merits intervention versus a low-risk group with < 5% risk that can be safely observed. Maximizing the calibrated predictive value of these scores is the end goal of the methodological advances described above [ 91 ].

Representative case studies and emerging models

Gene expression signatures and hybrid models

One striking outcome from these studies is the emergence of gene expression profiling (GEP) as a strong predictor of progression risk when combined with traditional markers. Sun et al.’s 36-gene signature (GS36) derived at UAMS was able to identify a subset of MGUS patients with a more than 50% probability of progression within 10 years [ 92 ]. This is a dramatic enrichment compared to the ~ 10–20% risk in the overall MGUS population. GS36 alone stratified patients into high vs. low risk with a 25-fold difference in progression rates (54% vs. 2%) [ 93 ]. When the signature was combined with two routine labs (FLC ratio and immunoparesis status), an “ultra-high-risk” group was defined (having all factors abnormal) with 82% 10-year progression risk, effectively approaching the risk level of SMM and suggesting these MGUS patients might warrant SMM-like monitoring or even early intervention. Importantly, the genes in GS36 were not random: many are involved in cell cycle regulation, protein folding, and metabolism of unfolded proteins [ 94 ], aligning with known facets of myeloma biology. This provides face validity and suggests that the signature captures an underlying aggressive phenotype. Another independent effort by the Myeloma Institute used a 70-gene MyPRS GEP70 signature (originally developed for active myeloma prognosis) and showed it could predict which MGUS/SMM cases would progress faster [ 95 , 96 ]. Taken together, these point to the principle that high-risk precursor disease often already bears a transcriptional program resembling active myeloma, for instance, overexpression of MYC target genes or proliferation-associated genes, and that capturing this program via GEP can substantially refine risk stratification beyond clinical factors.

Multi-omics and machine learning models

In the context of precursor stages, multi-omics and AI approaches extend classical MGUS and SMM scores by combining clinical variables with genomic and other molecular features, using the integration and modeling strategies outlined in Sect. “ Multi-omics integration and AI modeling strategies .” Rather than introducing new algorithms, these studies test whether adding genomic lesions, mutational signatures or microenvironmental features can sharpen risk discrimination beyond purely clinical models, and they consistently show that precursor clones with “MM-like” genomic profiles carry a much higher short-term progression risk than biologically indolent counterparts. As such models are refined and externally validated, they provide a template for future AI-driven, clinically deployable multi-omics risk scores in MGUS and SMM [ 97 – 99 ].

Very recent large-scale genomic studies have further refined our understanding of progression from MGUS and SMM to overt myeloma. Alberge et al. analyzed whole-genome and whole-exome data across precursor and MM cohorts and proposed an “MM-like” genomic score that places each case along a continuous axis of progression, from MGUS to SMM to MM [ 66 ]. Maura et al. introduced the concepts of “genomic MM” and “genomic MGUS,” demonstrating that a substantial fraction of SMM and a subset of MGUS already harbor MM-like genomic profiles with high progression risk, while most MGUS and a minority of SMM show a truly premalignant genomic pattern; incorporation of these genomic features into the IMWG 2/20/20 model significantly improved prediction of progression [ 65 ]. Together, these studies exemplify the shift from purely clinical staging to integrated genomic and multi-omics scores that map myeloma precursor conditions along a biology-defined continuum and provide a rational basis for future machine-learning–based risk models.

Models in active myeloma as analogues

In active myeloma, gene expression–based risk stratification (like the GEP70 and EMC92 signatures) is already a reality in some centers. These signatures identify high-risk MM patients who have significantly inferior survival. Their existence paves the way for analogous signatures in precursors. Additionally, multi-omics approaches in MM have yielded insights [ 100 ]. While aimed at survival rather than progression, the use of multi-omics plus a knowledge-based graph is highly relevant. It demonstrated that integrating gene expression, mutations, and copy number data on a network improved outcome prediction beyond using any single data type [ 101 ]. For example, patients whose tumor had a high “curvature” in certain PPI sub-networks had worse survival. In principle, a similar graph could be constructed for MGUS vs. MM progression risk, linking multi-omic features of MGUS clones to known oncogenic network topology [ 102 ]. Another development in MM is the emergence of minimal residual disease (MRD) detection by next-generation sequencing; while not an omics integration per se, it highlights the prognostic value of extremely sensitive molecular techniques [ 103 ]. One could imagine in the future a MGUS risk model that incorporates an “MRD-like” measurement (i.e. detecting an extremely low level of malignant cells with specific mutations), if such could be measured, it might indicate an MGUS that is essentially a nascent myeloma.

In summary, recent studies validate that multi-omics and AI models can markedly enhance risk predictions. By combining data types (gene expression + clinical, or multiple omics together) and by using advanced algorithms, these models achieve both improved performance (higher C-statistics/AUCs) and yield mechanistic insights through the importance of certain genes/pathways. The case studies reviewed here form a proof-of-concept that moving “from MGUS to myeloma” risk scoring into the era of multi-omics and machine learning is not only feasible but already demonstrating superior results. The next sections will tackle how to translate these research advances into clinical tools, and what challenges must be navigated along the way.

Translational challenges and model deployment

Despite promising results from research studies, significant barriers remain before multi-omics AI risk scores can be deployed in routine clinical practice. These challenges span technical issues with data and models, as well as practical considerations in the healthcare setting.

Heterogeneity and standardization of data

Multi-omics data are inherently complex and can vary widely depending on the platforms and protocols used. For example, one center’s MGUS gene expression data might come from RNA-sequencing, while another’s come from microarrays, directly combining these without careful batch correction can lead to spurious results. Likewise, genomic variant calling can differ by pipeline, and even something as “simple” as a serum FLC level can vary by assay kit. To build robust models, data standardization is critical [ 104 , 105 ]. This includes technical normalization (removing batch effects, scaling features) and biological harmonization (ensuring consistent definitions, e.g. what threshold defines immunoparesis). The PANGEA study confronted this by using common data elements (M-spike, FLC, etc.) that are routinely measured everywhere [ 106 ], but for omics, one approach is to adopt reference pipelines or use transfer learning to adjust models trained on one dataset to another [ 107 ]. For instance, an MGUS progression model trained on microarray data might need retraining or calibration on RNA-seq data. The heterogeneity of patient populations is also a factor, models might perform differently in different ethnic groups or with different comorbidities [ 107 ]. Large, diverse training cohorts and external validations across geographic populations are necessary to ensure generalizability. We may also need consensus on which omics are most useful to include for a “clinically deployable” test; a parsimonious model that uses 2–3 key modalities might translate better than one requiring whole-genome sequencing, transcriptomics, proteomics and metabolomics for every patient.

Experience from existing molecular diagnostics provides a practical template for this process [ 108 ]. In solid tumors and hematological malignancies, broad discovery-phase omics profiling has often been followed by locked, analytically validated panels that measure a restricted set of biomarkers under standardized pre-analytical and analytical conditions (defined tissue or sample type, minimum tumor or plasma cell content, centralized laboratories and predefined quality metrics) [ 109 , 110 ]. In multiple myeloma, for example, gene expression–based signatures such as GEP70/EMC92 and targeted next-generation sequencing panels for high-risk lesions are already implemented in CLIA-accredited reference laboratories and in clinical trials, with harmonized protocols for sample collection, bioinformatics pipelines and reporting formats [ 111 , 112 ]. These experiences illustrate how an initially complex multi-omics pipeline can be distilled into a reproducible “test kit,” and suggest that MGUS/SMM multi-omics scores could follow a similar trajectory from research-grade analyses to standardized assays suitable for routine use [ 108 , 111 ].

Cost and logistics

Implementing a multi-omics risk score means patients would need to undergo additional tests beyond standard of care. Currently, MGUS work-up is relatively minimal (blood tests, maybe a marrow biopsy in some) [ 113 ]. Requiring, say, whole-genome sequencing and RNA-seq on every MGUS patient’s plasma cells would be cost-prohibitive in many settings and may not be covered by insurers. Thus, there is a tension between the richness of data and practicality. One solution is to identify surrogate markers or smaller targeted panels that capture the most predictive features. For example, if a 36-gene signature is validated, one could design a targeted RNA-seq or RT-PCR panel for those genes rather than doing a full transcriptome profile [ 114 ]. Similarly, instead of whole-genome sequencing, a targeted DNA panel could check for high-risk mutations (RAS, TP53, etc.) and copy number changes (1q gain, 17p loss) known to be relevant. The cost of omics is continually dropping, but the test turnaround time and requirement for viable sample (e.g. enough bone marrow plasma cells for RNA) also come into play [ 115 ]. In a community clinic, sending samples for multi-omics analysis might not be feasible unless centralized services exist. Therefore, a stepwise deployment could start in academic centers or referral clinics, where infrastructure exists, and as evidence and cost-effectiveness accrue, expand outward. Health economic analyses will be needed to show that improved risk stratification and, presumably, better targeted early interventions justify the cost of testing [ 116 ]. In practical terms, such analyses should quantify not only the per-test cost but also downstream metrics such as cost per quality-adjusted life-year (QALY) gained, the number needed to test or treat to prevent one progression to symptomatic myeloma, and the proportion of patients who are meaningfully reclassified into higher- or lower-risk categories compared with standard clinical scores [ 117 , 118 ]. Simple budget-impact models, for example comparing a strategy in which only patients with very high clinical risk (e.g. IMWG 2/20/20 high-risk SMM) receive multi-omics testing versus broader testing of all referrals to a myeloma center, can help clinicians and payers decide where the test adds enough value to warrant implementation [ 119 , 120 ]. These quantitative considerations are crucial to determine whether multi-omics panels are initially confined to tertiary centers and clinical trials or can realistically diffuse into community practice.

Interpretability and clinician trust

Regardless of an AI model’s accuracy, clinicians may be reluctant to rely on its predictions without understanding the underlying rationale. This is particularly true for critical decisions, such as prescribing early therapy to an asymptomatic MGUS patient, a choice that, as of now, lacks standardized clinical guidelines. To gain the trust of healthcare providers, it is essential to incorporate explainability into the AI tool’s output [ 121 ]. For example, rather than simply presenting a “High-risk score = 0.85,” the report could include the specific factors contributing to this risk, such as “High-risk due to: IgA isotype, very high sFLC ratio, 1q21 amplification, and elevated expression of proliferation genes”. This level of detail would resonate with clinical reasoning and facilitate discussion with the patient [ 122 ]. Furthermore, visualization tools could enhance understanding; for instance, a physician could be shown where the patient’s risk score positions them in relation to others, or identify which molecular profile cluster the patient belongs to, such as “This patient clusters with profiles of known progressors” [ 123 ]. Education also plays a key role, as many hematologists will need to familiarize themselves with genomics and AI concepts to confidently integrate them into clinical practice [ 124 ]. One effective way to foster trust is by involving clinicians in the model development process, such as co-designing the output reports and selecting actionable insights [ 125 ]. Additionally, “white-box” models, like rule-based classifiers or simplified scoring systems derived from the machine learning model, could serve as a bridge. For example, even if the backend ML model uses more complex calculations, presenting a simplified scoring system based on the most frequently used features could make the output more accessible [ 126 ]. Real-world deployments of AI-based clinical decision support systems, such as sepsis prediction scores embedded in electronic health records or explainable imaging triage tools, suggest that adoption is highest when models provide concise explanations aligned with clinical reasoning and can be interrogated through simple “what-if” scenarios (for example, how the risk would change if M-protein or free light chain ratio improved), rather than functioning as opaque, non-interactive alerts [ 127 , 128 ].

Regulatory and validation hurdles

From a regulatory perspective, a multi-omics AI risk score would likely be classified as a medical device, specifically as software as a medical device (SaMD), by agencies such as the FDA [ 129 ]. Gaining approval would necessitate evidence of three key criteria: analytical validity (i.e., whether the test accurately and reliably measures what it is intended to), clinical validity (i.e., whether the score correlates with patient outcomes), and clinical utility (i.e., whether its use improves patient management or outcomes) [ 130 ]. Achieving regulatory approval for an AI model is a complex process, and it may require prospective trials where clinical management is guided by the AI score to demonstrate its benefit [ 131 ]. For example, a trial might compare the treatment of high-risk SMM patients, as identified by the AI model, to a cohort of patients who are observed without intervention, in order to assess whether the AI-guided treatment results in improved outcomes [ 132 ].

Additionally, the model would need a clearly defined intended use and target population. Regulators would also closely examine the risks associated with false positives (e.g., unnecessary treatment for patients incorrectly classified as high-risk) and false negatives (e.g., missing a high-risk patient who subsequently progresses). Furthermore, because AI models can evolve through retraining on new data, regulatory approval is typically specific to a particular algorithm version [ 133 ]. Any updates or changes to the model may require re-approval or the use of adaptive “locked” models to ensure regulatory compliance. Finally, the deployment plan must address how the model will be updated as new data becomes available without disrupting the existing regulatory framework. For laboratory-developed tests (LDTs), which many genomics tests fall under, regulation may be less centralized, though performance must still be validated in CLIA-certified laboratories [ 134 ].

Beyond classical concepts of analytical and clinical validity, several standardized frameworks now provide practical guidance on how to develop, report and evaluate prediction models and AI tools. For instance, the TRIPOD and TRIPOD-AI guidelines specify minimum reporting standards for multivariable prediction models, while tools such as PROBAST help to systematically assess risk of bias [ 135 , 136 ]. For interventional studies where management is guided by an AI-derived score, the SPIRIT-AI and CONSORT-AI extensions outline key elements that should be captured in protocols and trial reports. Designing MGUS/SMM multi-omics risk models and their prospective evaluations in accordance with these frameworks will make the resulting tools easier for regulators, payers and clinicians to appraise and compare with existing risk scores [ 137 ].

Privacy and data sharing

The use of multi-omics data, particularly genomic information, raises significant concerns regarding patient privacy. Since a patient’s DNA sequence is both unique and identifiable, incorporating such data into AI models, especially those hosted on cloud-based platforms, introduces critical security considerations [ 138 ]. Ensuring robust data protection requires the implementation of strong encryption measures and strict adherence to health data privacy regulations, such as HIPAA and GDPR. This challenge has also contributed to the growing interest in federated learning, which is explored further in the next Sect [ 139 ]. Federated learning allows models to be trained on decentralized data without transferring the raw data beyond the confines of hospital firewalls. In practice, this approach could involve installing software locally at each healthcare site, with only the model parameters, rather than the raw patient data, being transmitted. This method helps mitigate privacy risks while still enabling the use of AI models in healthcare settings [ 140 ].

Integration with workflow

Finally, any AI tool deployed in a clinical setting must integrate smoothly into the existing clinical workflow. This requires seamless integration with electronic health records (EHR) and laboratory information systems. Ideally, when a clinician orders an “MGUS AI risk panel,” the necessary data collection would be automated [ 141 ]. For example, the patient might undergo blood sampling for omics analysis, the laboratory would generate the relevant data, and the AI software would then process both the new omics data and the existing laboratory results from the EHR to compute the risk score. The output would be delivered as a report within the EHR, and the entire process should have a reasonable turnaround time, preferably within days, not months. Achieving this streamlined process demands close coordination between data providers, analysts, and clinicians. Additionally, confirmatory tests may be required; for instance, if the AI identifies a specific mutation, clinical-grade sequencing might be necessary to validate the finding. All of these procedural steps must be carefully defined and structured [ 142 ].

In additions to data- and model-related limitations, the performance and clinical utility of predictive models are strongly influenced by the therapeutic context in which they are applied. Most existing risk scores for MGUS, SMM, and MM were developed in cohorts treated with historical standards of care or, in the case of precursor conditions, with observation alone [ 7 , 30 , 31 , 143 ]. As treatment paradigms evolve, for example, with earlier intervention in high-risk SMM, the widespread use of triplet or quadruplet induction, novel antibodies, bispecifics, and cellular therapies, the natural history of disease and the baseline hazards on which these models were trained may change substantially [ 144 – 146 ]. Consequently, models need to be regularly revalidated and, when necessary, recalibrated in contemporary, treatment-reflective cohorts [ 147 ]. Ideally, future multi-omics and AI-based tools will explicitly incorporate therapeutic exposures and response dynamics as time-dependent variables, so that risk estimates are understood as treatment-conditional and can be updated as patients transition from observation to different lines of therapy [ 148 , 149 ].

Impact of evolving therapy on model validity

Most MGUS and SMM risk models have been derived in cohorts in which patients were either observed without intervention or treated according to historical standards of care. As early treatment strategies for high-risk SMM and genomically defined high-risk states become increasingly common, those models may no longer describe the actual clinical trajectory of contemporary patients. Trials such as I-PRISM, which evaluated triplet therapy in high-risk SMM, demonstrate that early therapeutic intervention can change progression patterns and highlight the importance of response depth and minimal residual disease in predicting outcome in this setting [ 150 , 151 ].

In active myeloma, recent individualized risk-prediction tools explicitly integrate treatment variables. The IRMMa model, for instance, uses multi-state modeling to account for high-dose melphalan with autologous stem-cell transplantation and for maintenance therapy when estimating patient-specific survival probabilities [ 148 ]. This approach illustrates that risk is inherently treatment-conditional: the same genomic profile may have a different prognosis depending on whether the patient receives intensive induction, transplant, and maintenance, or more conservative therapy.

For MGUS and SMM, similar therapy-aware models will be needed as early-intervention strategies, immunotherapies, and bispecific antibodies move into the precursor space [ 152 ]. Two complementary strategies are likely to be required. First, models should be recalibrated in contemporary cohorts that reflect current treatment paradigms, rather than simply extrapolating hazard estimates from untreated historical populations. Second, future multi-omics and AI tools will probably need to explicitly include therapeutic exposures and response dynamics (depth, duration, MRD status) as features, enabling truly dynamic risk prediction that can be updated as patients move from observation to intervention.

Finally, consensus efforts such as the recent IMS-IMWG genomic definition of high-risk MM underline that risk categories themselves are evolving as new molecular data and outcome correlations emerge [ 153 ]. Any multi-omics risk system for MGUS and SMM will therefore have to be flexible enough to adapt to changing treatment standards and updated genomic staging frameworks.

In conclusion, while the science behind multi-omics risk models is progressing rapidly, the transition to clinical deployment requires careful attention to standardization, validation, and the usability of these tools. The solutions developed to address these challenges will likely serve as benchmarks for the adoption of AI applications in oncology. In the next section, we outline a conceptual roadmap for implementing such a risk score system in practice, while considering the aforementioned factors.

Toward a clinically deployable AI-based risk scoring framework

Data acquisition and processing

Patients with MGUS/SMM would undergo a standardized multi-omics profiling at diagnosis (and potentially at intervals thereafter). This could involve obtaining bone marrow aspirate for genomic and transcriptomic analysis of clonal plasma cells, and peripheral blood for proteomic/metabolomic assays or immune cell profiling. Given practical constraints, one might start with a focused panel of assays, for example: targeted DNA sequencing of a myeloma gene panel, a 50-gene expression panel (such as a qPCR kit covering the signature genes), and measurement of key cytokines or metabolites that research has linked to progression [ 154 ]. These raw data would be processed through a bioinformatics pipeline: quality control, normalization (e.g. TPM normalization for RNA, variant calling for DNA, etc.), and feature extraction (like computing pathway scores or immune cell subsets from raw data) [ 155 , 156 ]. Alongside, standard clinical data (M-protein level, FLC ratio, marrow plasma cell % from pathology, cytogenetics, etc.) are pulled from the electronic health record. All these features are fed into the AI model’s input layer in a predefined format. Automation is crucial here – as much as possible, the data flow from lab instruments to model input should be hands-free to avoid errors and delays. For instance, a local software could automatically parse the sequencing results and EHR data to populate the model’s input variables [ 157 ].

AI model inference

The core of the system is the predictive model, which, as discussed, could be an ensemble of sub-models or a unified network [ 158 ]. In a deployable tool, one might opt for a simpler model if it offers similar performance, for example, a gradient-boosted trees model might be preferred over a very deep neural network if it’s easier to validate and explain. Nonetheless, the model would take the processed multi-omics features and output a risk score, often between 0 and 1, representing the probability of progression to MM within a certain time frame (e.g. 2 years, 5 years) [ 159 ]. Some models might output a risk category (low/intermediate/high). In parallel, the model can generate uncertainty estimates, for instance via Monte Carlo dropout in a neural net or through an ensemble of models, to flag cases where the prediction might be less reliable due to out-of-distribution inputs or conflicting data [ 160 ]. The inference step should be quick (sub-second once data is in hand), so that results can be available soon after lab results are ready [ 161 ].

Explainability and reporting

Crucially, the model doesn’t act alone, it is coupled with an explainability module that translates the raw prediction into a human-friendly explanation. For example, if the model is a tree ensemble, the tool might compute SHAP values for the top 10 features. If it’s a neural network, it might use an attribute method to quantify feature importance [ 162 ]. The output report might say: “Progression Risk Score = 0.82 (High). Top contributors: Bone marrow plasma cells 15% (high risk), Serum FLC ratio 45 (high risk), MAPK pathway gene expression signature (high), 17p deletion (present).” Each of those factors could be accompanied by a brief note or visual indicator (like a red up arrow for high risk features, green down arrow for protective features) [ 163 ]. The report might also place the patient in context: e.g. “This risk corresponds to an approximate 60% chance of progression at 5 years. Patients in this risk category historically had a median time to progression of X years under observation.” Such framing helps clinicians and patients understand the implications [ 164 ]. The language and format of the report should be developed with user feedback – too much detail can overwhelm, too little can be uninformative. Ideally, it fits on one page in the patient chart. Importantly, the report should emphasize that the score is a tool to aid (not replace) clinical judgment, including a disclaimer per regulatory requirements [ 162 ].

Integration into clinical workflow

The system would be integrated with hospital information systems to facilitate ordering and resulting. A physician could order the “MGUS AI Risk Panel” just like ordering a laboratory test. The order triggers the necessary lab workflows (e.g. sample collection for sequencing). Once data are generated and the model produces a result, the risk report appears in the EHR, possibly in a dedicated section or as a PDF attachment [ 165 ]. Alerts could also be configured, for instance, a very high-risk result might trigger a flag in the system, suggesting referral to a specialist or consideration of therapy. Conversely, a very low-risk result might suggest extended interval follow-up. The model’s recommendations can be linked to clinical guidelines. For example, if high-risk SMM is identified, the report might include: “Recommendation: Consider early treatment per [guideline] or clinical trial enrollment” [ 166 ]. Integration also involves periodic recalculation if new data arrives, for example, if at a 6-month visit, the patient has new lab results or even a repeat marrow, the model could be re-run to update the risk (this is analogous to how some cardiovascular risk calculators get updated with new cholesterol readings) [ 167 ].

Continuous learning and updates

A deployed model should have a mechanism for ongoing validation and improvement. All predictions and outcomes should ideally be logged (with patient consent and privacy safeguards) to see how accurate the model’s predictions prove over time [ 168 ]. If approved by regulators, one could implement a learning health system where the model is periodically retrained on new accumulated data from all sites. In practice, due to regulatory constraints, the initial deployment might be a locked model, but updated versions can be released as more evidence is gathered [ 169 ]. This ties into the concept of federated learning for multi-center data: each clinic using the tool could contribute to a secure aggregation of model gradients without sharing raw patient data, enabling improvement of the global model. In the near term, continuous learning might simply be manual: model developers gather new data, retrain/improve the model, and go through an update approval cycle annually or as needed [ 170 ].

Validation and feedback loop

Even after deployment, ongoing post-market validation is critical. This might include prospective studies where the model’s recommendations are followed to see if patient outcomes improve (e.g. early treatment of those flagged high-risk leads to fewer cases of end-organ damage or better survival). Also, user feedback from clinicians should be collected, are they finding the tool helpful? Are there cases where it gave an obviously wrong prediction? This feedback can highlight issues like certain subpopulations where the model is less accurate (maybe it underestimates risk in young patients, for instance). The feedback loop allows refinement not just of the model but of the entire workflow (maybe clinicians want the report to also suggest which trials a high-risk patient could enroll in, that could be added) [ 169 ].

In building this system, leveraging cross-disciplinary expertise is crucial. Laboratory scientists, data scientists, IT specialists, clinicians, and patient representatives should all be involved. For example, patient advocates might guide how to communicate risk in a compassionate yet clear way. Clinicians will ensure that the output aligns with decision-making needs. Data scientists will maintain the model’s accuracy and guard against algorithmic biases.

By following this framework, the hope is to create a risk scoring tool that is not only accurate but also clinically meaningful and actionable. One concrete vision is that in a future not far off, every patient diagnosed with SMM would get an AI-derived risk assessment that tells them and their doctor the likely trajectory of their disease, guiding whether they should start therapy or can be safely observed. Likewise, MGUS patients would be stratified such that those with negligible risk could be spared unnecessary frequent testing, whereas those with worrisome molecular profiles could be watched more closely or even offered interceptive therapy in clinical trials. In essence, the goal is to personalize the management of precursor plasma cell disorders using the best of modern technology, thereby improving patient outcomes and optimizing use of healthcare resources (Fig. 1 ).

Future directions

The future of risk stratification in MGUS and SMM lies in the integration of advanced technologies such as single-cell RNA sequencing, spatial transcriptomics, and microbiome analysis. These innovations promise to provide deeper insights into disease progression, offering the potential to identify high-risk subclones and track changes within the tumor microenvironment [ 171 ]. Additionally, the application of federated learning could facilitate collaborative, multi-center studies while protecting patient privacy, thus overcoming challenges associated with fragmented datasets. Longitudinal, dynamic risk models will enable real-time updates to patient risk profiles, allowing for more personalized and timely treatment interventions [ 172 ].

However, the successful deployment of these technologies requires addressing significant challenges in data integration, particularly across diverse omics platforms. Ensuring the standardization and harmonization of multi-omics data will be crucial for building robust predictive models [ 173 ]. Ethical considerations, including the mitigation of algorithmic bias and equitable access to AI-based tools, must also be prioritized to avoid exacerbating healthcare disparities. As AI models evolve, regulatory frameworks will need to adapt to accommodate continuously learning systems, ensuring their safety and efficacy in clinical settings. These advancements hold the potential to transform the management of plasma cell disorders, but careful attention to both technical and ethical concerns is essential for their clinical implementation [ 174 ]. Emerging genomic scores such as the MM-like score in precursor conditions and individualized treatment-aware models such as IRMMa in active myeloma together point towards a future in which MGUS and SMM risk prediction is both multi-omics-driven and therapy-conditional, with models that are continuously updated as new treatments and clinical trial data become available [ 148 , 175 , 176 ].

Conclusion

In conclusion, this manuscript provides a comprehensive overview of the evolving landscape of risk stratification in MGUS and SMM, emphasizing the transformative potential of multi-omics integration and AI-driven models. By highlighting the promise of advanced technologies such as single-cell RNA sequencing, spatial transcriptomics, and federated learning, the work outlines a clear vision for more personalized and dynamic approaches to disease management. However, successful clinical implementation will require overcoming significant challenges, including data integration across diverse platforms, addressing ethical concerns, and adapting regulatory frameworks to accommodate continuously evolving AI systems. With careful attention to these technical and ethical issues, the integration of these innovations could substantially enhance the precision and efficacy of risk assessment in plasma cell disorders.

Acknowledgements

This work was supported by the Intelligent Management of Chronic Disease Technology Team of Longyan City.

Author contributions

Yanyun Wu, Dongliang Zhang, Jingyao Jiang, Linghui Zheng, Zhiming Zhou wrote and carried out investigation and visualization for the manuscript. Zhenxing Zhang, Sina Nouri contributed conceptualization and revised the manuscript. All authors reviewed and approved the final version. Yanyun Wu, Dongliang Zhang contributed equally to this work and share first authorship.

Funding

This research did not receive any financial support from public, commercial, or nonprofit organizations.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Competing interests

The authors declare that they have no competing interests.

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Clinical trial number

Not applicable.

Footnotes

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Yanyun Wu and Dongliang Zhang contributed equally to this work and share first authorship.

Contributor Information

Zhenxing Zhang, Email: MarilynNoelle16508@outlook.com.

Sina Nouri, Email: 1sinanouri@gmail.com.

References

1. Korde N, Kristinsson SY, Landgren O. Monoclonal gammopathy of undetermined significance (MGUS) and smoldering multiple myeloma (SMM): novel biological insights and development of early treatment strategies. Blood J Am Soc Hematol. 2011;117(21):5573–81.

2. Wallington-Beddoe CT, Mynott RL. Prognostic and predictive biomarker developments in multiple myeloma. J Hematol Oncol. 2021;14(1):151.

3. Dar MA, et al. Multiomics technologies: role in disease biomarker discoveries and therapeutics. Brief Funct Genomics. 2023;22(2):76–96.

4. Yetgin A. Revolutionizing multi-omics analysis with artificial intelligence and data processing. Quant Biology. 2025;13(3):e70002.

5. Kumar A, Metta D. AI-driven precision oncology: predictive biomarker discovery and personalized treatment optimization using genomic data. Int J Adv Res Publ Rev. 2024;1(3):21–38.

6. Pérez-Persona E, et al. New criteria to identify risk of progression in monoclonal gammopathy of uncertain significance and smoldering multiple myeloma based on multiparameter flow cytometry analysis of bone marrow plasma cells. Blood. 2007;110(7):2586–92.

7. Rajkumar SV, et al. Serum free light chain ratio is an independent risk factor for progression in monoclonal gammopathy of undetermined significance. Blood. 2005;106(3):812–7.

8. Lee DD, Davies FE, Morgan GJ. Evolution of the classification and management of smoldering multiple myeloma. J Cell Immunol. 2024;6(3):125–39.

9. Jovanović KK, et al. Deregulation and targeting of TP53 pathway in multiple myeloma. Front Oncol. 2019;8:665.

10. Avigan ZM, Mitsiades CS, Laganà A. The role of 1q abnormalities in multiple myeloma: genomic insights, clinical implications, and therapeutic challenges. Seminars in hematology. Elsevier; 2025.

11. Rajkumar SV. Multiple myeloma: 2024 update on diagnosis, risk-stratification, and management. Am J Hematol. 2024;99(9):1802–24.

12. Visram A, Cook J, Warsame R. Smoldering multiple myeloma: evolving diagnostic criteria and treatment strategies. Hematology. 2021;2021(1):673–81.

13. Mellgard G, et al. Risk stratification models overestimate progression risk in contemporary patients with smoldering multiple myeloma. HemaSphere. 2024;8(3):e61.

14. Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019;25(1):44–56.

15. Kyle RA, Rajkumar SV. Monoclonal gammopathy of undetermined significance. Neoplast Dis Blood, 2012: pp. 751–85.

16. Corsale AM. Prediction of progression to symptomatic multiple myeloma in patients with high-risk plasma cell dyscrasias. 2025.

17. Fend F, Dogan A, Cook JR. Plasma cell neoplasms and related entities—evolution in diagnosis and classification. Virchows Arch. 2023;482(1):163–77.

18. Moritz DY. Multiple myeloma and other plasma cell diseases. University of Zagreb. School of Medicine. Department of Internal Medicine; 2019.

19. Halvarsson B-M. Genetic predisposition to sporadic and Familial multiple myeloma. Lund University; 2022.

20. Chen LY, et al. Monoclonal gammopathy of increasing significance: time to screen? Haematologica. 2022;108(6):1476.

21. Li C et al. Global patterns and trends in multiple myeloma incidence: Age, period and birth cohort analysis. 2022.

22. Dhodapkar MV. MGUS to myeloma: a mysterious gammopathy of underexplored significance. Blood J Am Soc Hematol. 2016;128(23):2599–606.

23. van Nieuwenhuijzen N, et al. From MGUS to multiple myeloma, a paradigm for clonal evolution of premalignant cells. Cancer Res. 2018;78(10):2449–56.

24. Ho M, et al. Changing paradigms in diagnosis and treatment of monoclonal gammopathy of undetermined significance (MGUS) and smoldering multiple myeloma (SMM). Leukemia. 2020;34(12):3111–25.

25. Dalla Palma AB. Risk factors of short-term progression in patients with monoclonal gammopathies. 2021.

26. Rajkumar SV, Dispenzieri A, Kyle RA. Monoclonal gammopathy of undetermined significance, Waldenström macroglobulinemia, AL amyloidosis, and related plasma cell disorders: diagnosis and treatment . in Mayo Clinic Proceedings . 2006. Elsevier.

27. González-Calle V, et al. High‐risk cytogenetic abnormalities in multiple myeloma: PETHEMA‐GEM experience. HemaSphere. 2024;8(12):e70031.

28. Kyle RA, et al. Long-term follow-up of monoclonal gammopathy of undetermined significance. N Engl J Med. 2018;378(3):241–9.

29. Cherry BM, et al. Modeling progression risk for smoldering multiple myeloma: results from a prospective clinical study. Leuk Lymphoma. 2013;54(10):2215–8.

30. Mateos MV, et al. International myeloma working group risk stratification model for smoldering multiple myeloma (SMM). Blood Cancer J. 2020;10(10):102.

31. Lakshman A, et al. Risk stratification of smoldering multiple myeloma incorporating revised IMWG diagnostic criteria. Blood Cancer J. 2018;8(6):59.

32. Zanwar S, Rajkumar SV. Current risk stratification and staging of multiple myeloma and related clonal plasma cell disorders: MULTIPLE MYELOMA, GAMMOPATHIES. Leukemia, 2025: pp. 1–8.

33. Köse MC. Understanding Disease Progression and Target Identification for Multiple Myeloma . 2023, Universite de Liege (Belgium).

34. Abdar Esfahani M, et al. The epigenetic revolution in hematology: from benchside breakthroughs to clinical transformations. Clin Experimental Med. 2025;25(1):230.

35. Ram M, et al. The genetic and molecular drivers of multiple myeloma: current insights, clinical implications, and the path forward. Pharmacogenomics Person Med. 2024. 10.2147/PGPM.S350238.

36. Christina YY. Computational network mining in High-Risk patients with multiple myeloma. The Ohio State University; 2020.

37. Pirouzbakht M, Hamzeh S, Soleimani H, Samarkhazan. Beyond single biomarkers: multi-omics strategies to predict immunotherapy outcomes in blood cancers. Clin Exp Med. 2025;25(1):355.

38. Sutanto H, et al. Genetic, epigenetic, and molecular determinants of multiple myeloma and precursor plasma cell disorders: a pathophysiological overview. Med Oncol. 2025;42(7):234.

39. Xu L, Su Y. Genetic pathogenesis of immunoglobulin light chain amyloidosis: basic characteristics and clinical applications. Exp Hematol Oncol. 2021;10(1):43.

40. Goldman-Mazur S, Vesole DH, Jurczyszyn A. Clinical implications of cytogenetic and molecular aberrations in multiple myeloma. Acta Haematol Pol. 2021;52(1):18–28.

41. Farswan A, et al. Characterizing the mutational landscape of MM and its precursor MGUS. Am J Cancer Res. 2022;12(4):1919–p.

42. Janz S, et al. Germline risk contribution to genomic instability in multiple myeloma. Front Genet. 2019;10:424.

43. Aksenova AY, et al. Genome instability in multiple myeloma: facts and factors. Cancers. 2021;13(23):5949.

44. Walker BA, et al. Aberrant global methylation patterns affect the molecular pathogenesis and prognosis of multiple myeloma. Blood. J Am Soc Hematol. 2011;117(2):553–62.

45. Mattioli M, et al. Gene expression profiling of plasma cell dyscrasias reveals molecular patterns associated with distinct IGH translocations in multiple myeloma. Oncogene. 2005;24(15):2461–73.

46. Huang M, et al. Epigenetic mechanisms and next-gen editing platforms in hematology: from molecular basis to therapeutic frontiers. Crit Rev Oncol Hematol. 2025;215:104916.

47. Bhawe K, Roy D. Interplay between NRF1, E2F4 and MYC transcription factors regulating common target genes contributes to cancer development and progression. Cell Oncol. 2018;41(5):465–84.

48. Chan KI, et al. MYC oncogene: a druggable target for treating cancers with natural products. Aging Disease. 2024;15(2):640.

49. Li F, et al. A multi-omics approach to reveal critical mechanisms of activator protein 1 (AP-1). Biomed Pharmacother. 2024;178:117225.

50. Gong L, Qiu L, Hao M. Novel insights into the initiation, evolution, and progression of multiple myeloma by multi-omics investigation. Cancers. 2024;16(3):498.

51. Aljabban J, et al. Characterization of monoclonal gammopathy of undetermined significance progression to multiple myeloma through meta-analysis of GEO data. Heliyon. 2023. 10.1016/j.heliyon.2023.e17298.

52. Nissom PM, et al. Transcriptome and proteome profiling to Understanding the biology of high productivity CHO cells. Mol Biotechnol. 2006;34(2):125–40.

53. Giannakoulas N, Ntanasis-Stathopoulos I, Terpos E. The role of marrow microenvironment in the growth and development of malignant plasma cells in multiple myeloma. Int J Mol Sci. 2021;22(9):4462.

54. Schinke C, et al. Characterizing the role of the immune microenvironment in multiple myeloma progression at a single-cell level. Blood Adv. 2022;6(22):5873–83.

55. Botta C, et al. Mechanisms of immune evasion in multiple myeloma: open questions and therapeutic opportunities. Cancers. 2021;13(13):3213.

56. Chahin M, et al. Clinical considerations for immunoparesis in multiple myeloma. Cancers. 2022;14(9):2278.

57. Maiso P, et al. Bone marrow mesenchymal stromal cells in multiple myeloma: their role as active contributors to myeloma progression. Cancers. 2021;13(11):2542.

58. Solimando AG, et al. The route of the malignant plasma cell in its survival niche: exploring multiple myelomas. Cancers. 2022;14(13):3271.

59. Li L et al. Circulating immune cells and risk of osteosarcoma: a Mendelian randomization analysis. Front Immunol, 2024. Volume 15–2024.

60. Trudu M. INTEGRATED ROLE OF MITOCHONDRIA IN MULTIPLE MYELOMA. 2024.

61. Nair R, Gupta P, Shanmugam M. Mitochondrial metabolic determinants of multiple myeloma growth, survival, and therapy efficacy. Front Oncol. 2022;12:1000106.

62. Jiménez C, et al. Characterization of human B cell hematological malignancies using protein-based approaches. Int J Mol Sci. 2024;25(9):4644.

63. Schmidt TM, Fonseca R, Usmani SZ. Chromosome 1q21 abnormalities in multiple myeloma. Blood Cancer J. 2021;11(4):83.

64. Ram M, et al. The genetic and molecular drivers of multiple myeloma: current Insights, clinical Implications, and the path forward. Pharmgenomics Pers Med. 2024;17:573–609.

65. Maura F et al. Genomics define malignant transformation in myeloma precursor conditions. J Clin Oncol. 2025; Jco2501733.

66. Alberge J-B et al. Genomic landscape of multiple myeloma and its precursor conditions. Nat Genet. 2025; 1–11.

67. Mukherjee A, et al. From data to cure: a comprehensive exploration of multi-omics data analysis for targeted therapies. Mol Biotechnol. 2025;67(4):1269–89.

68. Valous NA, et al. Graph machine learning for integrated multi-omics analysis. Br J Cancer. 2024;131(2):205–11.

69. Schulte-Sasse R, et al. Integration of multiomics data with graph convolutional networks to identify new cancer genes and their associated molecular mechanisms. Nat Mach Intell. 2021;3(6):513–26.

70. Xiao S, et al. Graph neural networks with multiple prior knowledge for multi-omics data analysis. IEEE J Biomedical Health Inf. 2023;27(9):4591–600.

71. Agamah FE, et al. Computational approaches for network-based integrative multi-omics analysis. Front Mol Biosci. 2022;9:967205.

72. Li R, et al. Graph signal processing, graph neural network and graph learning on biological data: a systematic review. IEEE Rev Biomed Eng. 2021;16:109–35.

73. Ouyang D, et al. Integration of multi-omics data using adaptive graph learning and attention mechanism for patient classification and biomarker identification. Comput Biol Med. 2023;164:107303.

74. He B, et al. DGHNE: network enhancement-based method in identifying disease-causing genes through a heterogeneous biomedical network. Brief Bioinform. 2022. 10.1093/bib/bbac405.

75. Zhang X, et al. OmiEmbed: a unified multi-task deep learning framework for multi-omics data. Cancers. 2021;13(12):3047.

76. Withnell E, et al. XOmiVAE: an interpretable deep learning model for cancer classification using high-dimensional omics data. Brief Bioinform. 2021;22(6):bbab315.

77. Xiao Y, et al. Dual alignment feature embedding network for multi-omics data clustering. Knowl Based Syst. 2025;309:112774.

78. Zhang C, et al. Deep latent space fusion for adaptive representation of heterogeneous multi-omics data. Brief Bioinform. 2022;23(2):bbab600.

79. Karathanasis N, Spyrou GM. Predicting the progression from asymptomatic to symptomatic multiple myeloma and stage classification using gene expression data. Cancers. 2025;17(2):332.

80. Sun F, et al. A gene signature can predict risk of MGUS progressing to multiple myeloma. J Hematol Oncol. 2023;16(1):70.

81. He B, et al. Gene coexpression network and module analysis across 52 human tissues. Biomed Res Int. 2020;2020:p6782046.

82. Santamaría A, et al. Machine learning-based predictive model for risk stratification of multiple myeloma from monoclonal gammopathy of undetermined significance. Electronics. 2025;14(15):3014.

83. Ryu AJ, et al. Artificial intelligence-enabled screening strategy for drug repurposing in monoclonal gammopathy of undetermined significance. Blood Cancer J. 2023;13(1):28.

84. Zhu J, et al. Geometric graph neural networks on multi-omics data to predict cancer survival outcomes. Comput Biol Med. 2023;163:107117.

85. Zhang Y, et al. Local augmented graph neural network for multi-omics cancer prognosis prediction and analysis. Methods. 2023;213:1–9.

86. Jiang R, et al. A transformer-based weakly supervised computational pathology method for clinical-grade diagnosis and molecular marker discovery of gliomas. Nat Mach Intell. 2024;6(8):876–91.

87. Timón-Reina S, Rincón M, Martínez-Tomás R. An overview of graph databases and their applications in the biomedical domain. Database. 2021;2021:pbaab026.

88. Shah V. Genetics of high risk myeloma. 2021.

89. Claussen CM et al. Gene expression profiling predicts clinical outcomes in newly diagnosed multiple myeloma patients in a standard of care setting . 2016, American Society of Hematology Washington, DC.

90. Cowan A, et al. Personalised progression prediction in patients with monoclonal gammopathy of undetermined significance or smouldering multiple myeloma (PANGEA): a retrospective, multicohort study. Lancet Haematol. 2023;10(3):e203–12.

91. Zhan F, et al. Gene-expression signature of benign monoclonal gammopathy evident in multiple myeloma is linked to good prognosis. Blood. 2007;109(4):1692–700.

92. Mailankody S, et al. Molecular and biologic markers of progression in monoclonal gammopathy of undetermined significance to multiple myeloma. Leuk Lymphoma. 2010;51(12):2159–70.

93. Seth S, et al. Monoclonal gammopathy of undetermined significance: current concepts and future prospects. Curr Hematol Malig Rep. 2020;15(2):45–55.

94. Anguiano A, et al. Gene expression profiles of tumor biology provide a novel approach to prognosis and May guide the selection of therapeutic targets in multiple myeloma. J Clin Oncol. 2009;27(25):4197–203.

95. van Laar R, et al. The 70-Gene MyPRSR prognostic risk score signature predicts increased risk of progression from MGUS to multiple myeloma requring treatment. Blood. 2016;128(22):3275.

96. Li JR, Wang C, Cheng C. Identifying high-risk multiple myeloma patients: a novel approach using a clonal gene signature. Int J Cancer. 2024;155(9):1684–95.

97. Alhamrani SQ, et al. Machine learning for multi-omics characterization of blood cancers: a systematic review. Cells. 2025. 10.3390/cells14171385.

98. Dou Y, Mirzaei G. MO-GCAN: multi-omics integration based on graph convolutional and attention networks. Bioinformatics. 2025. 10.1093/bioinformatics/btaf405.

99. Spandana C, Kumar R, Gautam A. Cancer Detect Hum Blood Samples Using Mach Learn Techniques.

100. Bauer M, et al. A detailed exploration of using RNA-Seq data in established multiple myeloma gene expression profile microarray based risk scores. Clin Lymphoma Myeloma Leuk. 2019;19(10):e57–8.

101. Alharbi F et al. Comparative analysis of multi-omics integration using advanced graph neural networks for cancer classification . arXiv preprint arXiv:2410.05325, 2024.

102. Wang YH, et al. Immune signatures of bone marrow cells can independently predict prognosis in patients with myelodysplastic syndrome. Br J Haematol. 2022;196(1):156–68.

103. Ferla V, et al. Minimal residual disease detection by next-generation sequencing in multiple myeloma: promise and challenges for response-adapted therapy. Front Oncol. 2022;12:932852.

104. Al-Zoghby AM, et al. A comprehensive review of multimodal deep learning for enhanced medical diagnostics. Computers, Materials & Continua. 2025. 10.32604/cmc.2025.065571.

105. He B, et al. Assessing the impact of data preprocessing on analyzing next generation sequencing data. Front Bioeng Biotechnol. 2020;8:817.

106. Landgren O, et al. Association of immune marker changes with progression of monoclonal gammopathy of undetermined significance to multiple myeloma. JAMA Oncol. 2019;5(9):1293–301.

107. Agarwal A, Ghobrial IM. Monoclonal gammopathy of undetermined significance and smoldering multiple myeloma: a review of the current understanding of epidemiology, biology, risk stratification, and management of myeloma precursor disease. Clin Cancer Res. 2013;19(5):985–94.

108. Licata L, et al. Multigene signatures for early breast cancer in clinical practice: a report of the Lombardy genomic assays for breast cancer working group. Front Oncol. 2023;13:1081885.

109. Slodkowska EA, Ross JS. MammaPrint 70-gene signature: another milestone in personalized medical care for breast cancer patients. Expert Rev Mol Diagn. 2009;9(5):417–22.

110. Sinn P, et al. Multigene assays for classification, prognosis, and prediction in breast cancer: a critical review on the background and clinical utility. Geburtshilfe Frauenheilkd. 2013;73(09):932–40.

111. van Laar R, et al. Translating a gene expression signature for multiple myeloma prognosis into a robust high-throughput assay for clinical use. BMC Med Genomics. 2014;7:25.

112. Cerchione C, et al. Gene expression profiling in multiple myeloma: redefining the paradigm of risk-adapted treatment. Front Oncol. 2022;12:820768.

113. Leonardos D, et al. Applications of multiparameter flow cytometry in the diagnosis, prognosis, and monitoring of multiple myeloma patients. Diseases. 2025;13(10):320.

114. Li J-r, et al. Enhancing prognostic power in multiple myeloma using a plasma cell signature derived from single-cell RNA sequencing. Blood Cancer J. 2024;14(1):38.

115. Xiang Y, et al. Single-cell sequencing reveals the mechanisms of multiple myeloma progression: clarity or confusion? Annals of Hematology; 2025. pp. 1–18.

116. Adamiak G. Methods for the economic evaluation of health care programmes. BMJ Publishing Group Ltd; 2006.

117. Seefat MR, et al. A systematic review of cost-effectiveness analyses of novel agents in the treatment of multiple myeloma. Cancers. 2021;13(22):5606.

118. Asrar MM, et al. A systematic review of economic evaluations of treatment regimens in multiple myeloma. Expert Rev PharmacoEcon Outcomes Res. 2021;21(4):799–809.

119. Owens CE, et al. The economic costs of precision medicine for clinical translational research among children with high-risk cancer. NPJ Precision Oncol. 2024;8(1):224.

120. Guzauskas GF, et al. Population genomic screening for three common hereditary conditions: a cost-effectiveness analysis. Ann Intern Med. 2023;176(5):585–95.

121. Payne VL, et al. Clinician perspectives on how situational context and augmented intelligence design features impact perceived usefulness of sepsis prediction scores embedded within a simulated electronic health record. J Am Med Inf Assoc. 2024;31(6):1331–40.

122. Kostick-Quenet K, et al. Trust criteria for artificial intelligence in health: normative and epistemic considerations. J Med Ethics. 2024;50(8):544–51.

123. Ackerhans S, et al. Perceived trust and professional identity threat in AI-based clinical decision support systems: scenario-based experimental study on AI process design features. JMIR Form Res. 2025;9:e64266.

124. Osonuga A, et al. Bridging the digital divide: artificial intelligence as a catalyst for health equity in primary care settings. Int J Med Inf. 2025;204:106051.

125. Canfell OJ, et al. Artificial intelligence after the bedside: co-design of AI-based clinical informatics workflows to routinely analyse patient-reported experience measures in hospitals. BMJ Health Care Inform. 2024. 10.1136/bmjhci-2024-101124.

126. Borys K, et al. Explainable AI in medical imaging: an overview for clinical practitioners - Saliency-based XAI approaches. Eur J Radiol. 2023;162:110787.

127. Abràmoff MD, et al. Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care offices. NPJ Digit Med. 2018;1:39.

128. Wong A, et al. External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients. JAMA Intern Med. 2021;181(8):1065–70.

129. Marco-Ruiz L, et al. A multinational study on artificial intelligence adoption: clinical implementers’ perspectives. Int J Med Inf. 2024;184:105377.

130. Jafri L, et al. Insights from semi-structured interviews on integrating artificial intelligence in clinical chemistry laboratory practices. BMC Med Educ. 2024;24(1):170.

131. Hou H, Zhang R, Li J. Artificial intelligence in the clinical laboratory. Clin Chim Acta. 2024;559:119724.

132. Herman DS, et al. Artificial intelligence and mapping a new direction in laboratory medicine: A review. Clin Chem. 2021;67(11):1466–82.

133. Aung YYM, Wong DCS, Ting DSW. The promise of artificial intelligence: a review of the opportunities and challenges of artificial intelligence in healthcare. Br Med Bull. 2021;139(1):4–15.

134. Tazawa Y. Perspective for the development of companion diagnostics and regulatory landscape to encourage personalized medicine in Japan. Breast Cancer. 2016;23(1):19–23.

135. Collins GS, et al. TRIPOD + AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:e078378.

136. Moons KG, et al. PROBAST + AI: an updated quality, risk of bias, and applicability assessment tool for prediction models using regression or artificial intelligence methods. BMJ. 2025. 10.1136/bmj-2024-082505.

137. Wolff RF, et al. PROBAST: a tool to assess the risk of bias and applicability of prediction model studies. Ann Intern Med. 2019;170(1):51–8.

138. Zafar I, et al. Molecular biology in the exabyte era: taming the data deluge for biological revelation and clinical transformation. Comput Biol Chem. 2025;119:108535.

139. Androulakis IP, et al. The dawn of a new era: can machine learning and large language models reshape QSP modeling? J Pharmacokinet Pharmacodyn. 2025;52(4):36.

140. Yuan R, et al. Harmony-based data integration for distributed single-cell multi-omics data. PLoS Comput Biol. 2025;21(9):e1013526.

141. Magrabi F, Lyell D, Coiera E. Automation in contemporary clinical information systems: a survey of AI in healthcare settings. Yearb Med Inf. 2023;32(1):115–26.

142. Ye J, et al. The role of artificial intelligence for the application of integrating electronic health records and patient-generated data in clinical decision support. AMIA Jt Summits Transl Sci Proc. 2024;2024:459–67.

143. Visram A, Cook J, Warsame R. Smoldering multiple myeloma: evolving diagnostic criteria and treatment strategies. Hematol Am Soc Hematol Educ Program. 2021;2021(1):673–81.

144. Lonial S, et al. Randomized trial of Lenalidomide versus observation in smoldering multiple myeloma. J Clin Oncol. 2020;38(11):1126–37.

145. Mateos MV, et al. Lenalidomide plus dexamethasone for high-risk smoldering multiple myeloma. N Engl J Med. 2013;369(5):438–47.

146. Devasia AJ, Chari A, Lancman G. Bispecific antibodies in the treatment of multiple myeloma. Blood Cancer J. 2024;14(1):158.

147. CARDIOL RA. Validation and Updating of Clinical Prediction Models: Why and How?.

148. Maura F, et al. Genomic classification and individualized prognosis in multiple myeloma. J Clin Oncol. 2024;42(11):1229–40.

149. Maffini E, et al. Long-term follow up of tandem autologous-allogeneic hematopoietic cell transplantation for multiple myeloma. Haematologica. 2018;104(2):380.

150. Nadeem O, et al. Deeper response predicts better outcomes in high-risk-smoldering-myeloma: results of the I-PRISM phase II clinical trial. Nat Commun. 2025;16(1):358.

151. Zhang SC, Ballas LK. Radiation for multiple myeloma In the era of novel agents: Indications, Safety, and dose Selection . In Seminars In radiation oncology . Elsevier; 2025.

152. Zhang S, et al. The benefits and safety of monoclonal antibodies: implications for cancer immunotherapy. J Inflamm Res. 2025;18:4335–57.

153. Avet-Loiseau H, et al. International myeloma Society/International myeloma working group consensus recommendations on the definition of High-Risk multiple myeloma. J Clin Oncol. 2025;43(24):2739–51.

154. Li X, et al. Unveiling the cellular landscape: insights from single-cell RNA sequencing in multiple myeloma. Front Immunol. 2024;15:1458638.

155. Pilcher W, et al. Cross center single-cell RNA sequencing study of the immune microenvironment in rapid progressing multiple myeloma. NPJ Genomic Med. 2023;8(1):3.

156. Kou J, et al. Flexible assistance strategy of lower limb rehabilitation exoskeleton based on admittance model. Sci China Technological Sci. 2024;67(3):823–34.

157. Jang JS, et al. Molecular signatures of multiple myeloma progression through single cell RNA-Seq. Blood Cancer J. 2019;9(1):2.

158. Nguyen D, et al. A comparison of Monte Carlo dropout and bootstrap aggregation on the performance and uncertainty Estimation in radiation therapy dose prediction with deep learning neural networks. Phys Med Biol. 2021;66(5):054002.

159. Fuse Y, et al. Artificial intelligence in clinical data analysis: a review of large language models, foundation models, digital twins, and allergy applications. Allergol Int. 2025. 10.1016/j.alit.2025.06.005.

160. Hadweh P, et al. Machine learning and artificial intelligence in intensive care medicine: critical recalibrations from rule-based systems to frontier models. J Clin Med. 2025. 10.3390/jcm14124026.

161. Seoni S, et al. Application of uncertainty quantification to artificial intelligence in healthcare: A review of last decade (2013–2023). Comput Biol Med. 2023;165:107441.

162. Ponce-Bobadilla AV, et al. Practical guide to SHAP analysis: explaining supervised machine learning model predictions in drug development. Clin Transl Sci. 2024;17(11):e70056.

163. Tarabanis C, et al. Explainable SHAP-XGBoost models for in-hospital mortality after myocardial infarction. Cardiovasc Digit Health J. 2023;4(4):126–32.

164. Hu C, et al. Explainable artificial intelligence model for mortality risk prediction in the intensive care unit: a derivation and validation study. Postgrad Med J. 2024;100(1182):219–27.

165. Elhaddad M, Hamam S. AI-driven clinical decision support systems: an ongoing pursuit of potential. Cureus. 2024;16(4):e57728.

166. Bagheri M et al. AI-driven decision-making in healthcare information systems: a comprehensive review. 2024.

167. Sargam GS. AI-DRIVEN CLINICAL DECISION SUPPORT: ENHANCING CARE QUALITY THROUGH EHR-INTEGRATED INTELLIGENCE. International Journal of Advanced Research in Computer Science. 2025. 10.26483/ijarcs.v16i3.7262.

168. Yoo J, et al. Review learning: real world validation of privacy preserving continual learning across medical institutions. Comput Biol Med. 2025;192:110239.

169. Rosenthal JT, Beecy A, Sabuncu MR. Rethinking clinical trials for medical AI with dynamic deployments of adaptive systems. Npj Digit Med. 2025;8(1):252.

170. Rieke N, et al. The future of digital health with federated learning. NPJ Digit Med. 2020;3(1):119.

171. Sklavenitis-Pistofidis R, et al. Single-cell RNA sequencing defines distinct disease subtypes and reveals hypo-responsiveness to interferon in asymptomatic waldenstrom’s macroglobulinemia. Nat Commun. 2025;16(1):1480.

172. Hagos YB, et al. Deep learning enables Spatial mapping of the mosaic microenvironment of myeloma bone marrow Trephine biopsies. Cancer Res. 2024;84(3):493–508.

173. Yahiaoui ME, et al. Federated learning with privacy preserving for multi- institutional three-dimensional brain tumor segmentation. Diagnostics. 2024. 10.3390/diagnostics14242891.

174. Zorlu T, et al. Challenges, difficulties, and delayed diagnosis of multiple myeloma. Diagnostics (Basel). 2025. 10.3390/diagnostics15131708.

175. Alberge JB, et al. Genomic landscape of multiple myeloma and its precursor conditions. Nat Genet. 2025;57(6):1493–503.

176. Bayani A, et al. Aptamer-based approaches in leukemia: a paradigm shift in targeted therapy. Clin Exp Med. 2025;25(1):186.

Associated Data

Data Availability Statement

No datasets were generated or analysed during the current study.