Искусственный интеллект, видоизменяющий парадигму гематологического диагноза и лечения злокачественных заболеваний: от статической оценки до динамического управления точностью.

Introduction

Hematologic malignancies are highly heterogeneous, with their development driven by genetic and epigenetic alterations, tumor microenvironment remodeling, and clonal evolution, resulting in significant prognostic variations and high recurrence risks [ 1 , 2 ]. Although targeted, immunological and cellular therapies are advancing [ 2 ], there is still a considerable gap in the existing system for key aspects such as diagnosis, prognosis evaluation and treatment follow-up, and for precise, whole-process management [ 2 , 3 ].

Diagnosis and classification require comprehensive evaluation of morphological and immunophenotypic findings, yet current methods remain heavily reliant on manual interpretation, which is inefficient and prone to subjective bias and inter-center variability [ 3 , 4 ]. Prognosis assessment predominantly depends on static baseline metrics and staging systems, failing to capture the dynamic evolution of tumor microenvironment and clonal architecture [ 1 , 2 ]. In therapeutic decision-making and efficacy monitoring, biomarkers such as minimal residual disease (MRD) lack cross-platform standardized and transferable criteria, while research on drug resistance mechanisms and novel targets has yet to establish actionable clinical pathways [ 5 , 6 ].

Artificial intelligence (AI) is being increasingly applied to the diagnosis and treatment of hematologic malignancies through four key applications: First, automated identification of cellular and histological patterns for classifying peripheral blood smears, bone marrow morphology, and pathological Sects. [ 7 – 9 ]; Second, integration of multi-omics and multimodal data using machine learning or deep learning to build diagnostic and prognostic models incorporating genomic, transcriptomic, and radiomics data [ 10 ]; Third, dynamic prediction based on longitudinal data, employing time-series or joint models to assess survival outcomes, recurrence risks, and treatment efficacy [ 11 , 12 ], Fourth, embedding decision support systems into electronic health records (EHRs) to provide real-time risk alerts and treatment recommendations within clinical workflows [ 13 , 14 ].

AI has shown potential in extracting patterns from high-dimensional, multimodal data that are difficult for human eyes to discern, particularly in enhancing diagnostic consistency, refining prognostic stratification, and supporting personalized treatment adjustments. However, its development remains constrained by challenges such as biased training datasets, insufficient cross-center generalization, weak “black box” explainability, and inadequate ethical oversight [ 15 – 17 ]. Relying solely on AI is insufficient to reshape clinical pathways; it must be advanced in tandem with multi-center data governance, interpretable algorithms, human-machine collaboration, and prospective clinical trials [ 15 , 17 – 19 ].

Unlike previous reviews that focused on single diseases or technical approaches, this study systematically examines AI advancements in diagnosing leukemia, lymphoma, multiple myeloma, and myelodysplastic syndromes, analyzing their diagnostic classification, prognosis assessment, and treatment decision-making capabilities while evaluating their technical strengths and limitations. Building on this foundation, we propose an AI-driven dynamic diagnostic-therapeutic system and its integration framework with electronic health records, providing insights for developing and optimizing intelligent clinical pathways for hematologic malignancies.

Development of artificial intelligence in the field of hematologic malignancies

The application of artificial intelligence in hematologic malignancies has evolved through distinct phases: the traditional machine learning and early omics era, the high-dimensional machine learning and deep learning exploration phase, the multimodal learning and longitudinal modeling stage, and finally the large language model and clinical decision support phase, as illustrated in Fig. 1 .

Traditional machine learning and early omics phase (1999–2014)

This phase primarily utilizes high-throughput yet single-modal early omics data, such as gene expression chips, with traditional machine learning (ML) algorithms like support vector machines (SVM) and random forests (RF) being the main applications [ 20 – 23 ]. The landmark achievement came in 1999 when Golub et al. pioneered the use of DNA microarrays and machine learning to automatically distinguish acute myeloid leukemia (AML) from acute lymphoblastic leukemia (ALL) [ 20 ], demonstrating for the first time that leukemia subtypes could be identified solely through gene expression profiles. Bullinger et al. utilized gene expression profiles to establish prognostic subtypes in human AML, achieving further stratification of risk populations that were previously unidentifiable through conventional cytogenetics [ 22 ]. This phase has initially achieved automatic disease classification and risk stratification but relies on manual feature screening, making it difficult to fully exploit high-dimensional, nonlinear omics data.

High-dimensional machine learning and deep learning exploration phase (2015–2019)

With advances in computing power and data scale, high-dimensional machine learning and deep learning (DL) have been progressively applied to hematologic oncology research. At this stage, most deep learning applications remain focused on specific tasks and small-sample cohorts, with the field still in its methodological exploration phase.

Omics: High-dimensional molecular signature-driven prognostic stratification

In omics analysis, researchers integrate high-dimensional molecular features, such as mutation profiles and transcriptomes, into prognostic models to refine risk stratification for hematologic malignancies like AML [ 24 – 26 ]. Subsequent studies in solid tumor cohorts demonstrated deep learning’s superiority over traditional linear models in high-dimensional data integration and survival subgroup classification, providing a replicable technical paradigm for developing multi-omics deep learning models in hematologic oncology [ 27 ].

Image: Deep learning-driven analysis of blood pathology images

In image analysis, convolutional neural networks (CNNs) have been applied to blood pathology imaging for automated cell morphology recognition and classification [ 7 , 28 – 30 ]. A landmark study by Matek et al. demonstrated that their CNN model, developed from AML peripheral blood smears, achieved near-human expert performance in leukemia cell identification [ 7 ]. These advancements have significantly enhanced the objectivity, consistency, and efficiency of cell morphology diagnosis.

Multimodal learning and longitudinal modeling phase (2020–present)

With the accumulation of multi-omics and clinical data, research has shifted focus from single-indicator or single-modal analysis to integrating clinical features and multi-omics information to construct comprehensive prognostic models, while exploring dynamic risk profiles of disease progression [ 31 , 32 ]. For instance, studies have developed predictive models outperforming traditional methods by combining clinical parameters and RNA sequencing data from multiple myeloma (MM) patients [ 32 ]. Meanwhile, AI is beginning to analyze time-series data in EHRs to predict disease recurrence or treatment response early, enabling more precise individualized treatment decisions [ 33 ].

The phase of large language models and clinical decision support (2024–present)

Large language models (LLMs), exemplified by GPT-4, are being explored as potentially useful tools for clinical decision support. They are well suited to processing unstructured text and synthesizing information from medical literature and guidelines for complex cases, demonstrating significant utility in differential diagnosis and the systematic comparison of medical oncology guidelines [ 34 , 35 ]. Recent studies have also explored combining LLMs with specialized medical tools to support more structured task execution; however, these systems remain at an early stage, and their reliability, calibration, safety, and real-world clinical utility in hematologic practice remain to be established [ 36 , 37 ].

Application of artificial intelligence in the diagnosis and treatment of hematologic malignancies

Leukemia

Leukemia is a highly heterogeneous malignant tumor with multiple subtypes originating from hematopoietic stem cells [ 38 ]. Diagnosis primarily follows the MICM integrated pathway (morphology, immunology, cytogenetics, and molecular biology), which heavily relies on manual interpretation and lacks standardization [ 39 , 40 ]. Furthermore, significant genetic and phenotypic variation in leukemia cells poses substantial challenges for prognostic assessment and personalized treatment decisions. The development of AI holds the promise of addressing these bottlenecks.

Diagnosis and disease classification

The application of AI in leukemia diagnosis and classification is evolving from traditional ML that relies on manually designed features to DL with end-to-end feature learning, and further advancing toward multimodal data fusion. ML models that depend on expert-defined features offer strong interpretability and practical value in scenarios with limited data resources. A representative work is the ALLSorts hierarchical classifier proposed by Schmidt et al., which uses RNA sequencing data to identify molecular subtypes of B-cell acute lymphoblastic leukemia (B-ALL), achieving 84%−92% accuracy across a multicenter cohort encompassing 18 subtypes [ 41 ]. The innovation of this model lies in its meta-subtype hierarchical architecture, which can assign cases with ambiguous classifications to higher-level categories, thereby enhancing clinical applicability. However, its performance is limited by the quality of RNA-seq data, and its ability to identify rare subtypes remains insufficient due to sample scarcity.

In contrast, DL models employ convolutional neural networks to achieve end-to-end feature learning, enabling automatic extraction of complex disease-related patterns from raw data, demonstrating significant potential [ 42 ]. For instance, Yadav et al.’s 3SNet model achieved 84% to 93.8% accuracy in leukemia cell morphology classification by integrating grayscale, gradient direction histogram, and local binary pattern features from cellular microscopy images [ 43 ]. However, the clinical utility of morphological AI is often constrained by pre-analytical variability across laboratories, because subtle differences in fixation, staining intensity, and scanning settings can introduce domain shift that disproportionately degrades reliability for rare or dysplastic cells [ 44 ].

Following the single-modal approaches in ML and DL, current research is shifting toward multi-source data fusion. A notable example is the work by Hybel et al., who integrated imaging flow cytometry (IFC) with DL by employing CNNs to jointly model bright-field images and DNA morphology of leukemia stem cells and normal hematopoietic stem cells in AML patients, achieving a discriminant accuracy of up to 93.42% [ 45 ]. This demonstrates the value of multi-channel imaging information fusion in precise cell-level classification. However, the study’s limited sample size and inherent cellular heterogeneity remain critical challenges for its robust clinical application.

Beyond imaging flow cytometry, machine learning has also been applied to conventional multicolor flow cytometric datasets to alleviate the time-consuming and operator-dependent nature of manual MRD interpretation. In B-ALL, an AI-enhanced pipeline reduced manual MRD review time to approximately 1 min per case while maintaining excellent agreement with conventional expert analysis, highlighting a practical route to improve the accessibility of flow-based MRD testing [ 46 ]. However, the clinical readiness of flow cytometry is limited less by model design than by cross-platform non-equivalence, as variation in antibody panels, compensation, instrument settings, and gating conventions can cause data drift that requires continuous quality control and expert adjudication [ 47 ].

Prognosis assessment and treatment decision

Most current leukemia prognosis models still rely on static pre-treatment metrics. For instance, Eckardt et al. conducted a multicenter study involving 1,383 AML patients, systematically comparing nine supervised learning algorithms that integrated multidimensional data to predict complete response rates and 2-year overall survival (OS) [ 48 ]. While the study quantified the importance of multi-gene features, its retrospective data nature limited direct clinical translation. Subsequent research by the team [ 49 ] further used interpretable AI frameworks, such as Shapley Additive Explanations (SHAP), for age-stratified analysis, revealing that the prognostic weights of key genes, such as NPM1 and TP53, dynamically change with age. This highlights the need to consider age covariates in risk assessment.

Beyond baseline prognostic modeling, prospective evidence suggests that functional drug profiling may support treatment selection in AML. The SMARTrial showed that ex vivo drug response reports could be returned within 7 days for most participants, and in an AML validation cohort, incorporation of drug response profiles improved ELN-2022 risk stratification for adverse-risk patients [ 50 ]. However, translating ex vivo sensitivity signals into in vivo efficacy remains challenged by a microenvironment gap, as stromal protection and immune interactions within the bone marrow niche can reshape drug vulnerabilities and thereby decouple predicted sensitivity from real-world response [ 51 ].

The transformation in prognostic assessment stems from incorporating dynamic biomarkers such as MRD into predictive models. Jeha et al.’s study exemplifies this approach, integrating multi-timepoint MRD monitoring with detailed molecular subtypes in a cohort of 598 pediatric ALL patients, establishing a truly dynamic biomarker-based framework for prognostic evaluation and treatment decision-making. The study measured MRD levels in bone marrow and peripheral blood multiple times during early induction and consolidation phases, using Kaplan-Meier survival analysis and multivariate Cox regression to assess event-free survival and overall survival outcomes [ 52 ]. This model enables risk stratification based on real-time MRD updates during treatment, facilitating risk-adaptive therapeutic adjustments.

AI applications also demonstrate significant potential in emerging therapies. The deep learning model RCMNet, developed by Zhang et al., achieved 99.63% accuracy in CAR-T cell morphology recognition by integrating attention mechanisms, providing an automated tool for quality control in treatment processes [ 53 ].

AI applications in leukemia are currently shifting the focus of diagnosis and treatment from static baseline assessments to dynamic risk evaluation and personalized decision-making, driven by longitudinal indicators such as MRD. While technological advancements continue to be implemented in both morphological and flow cytometry domains, cross-center consistency and standardization remain critical gaps. To realize clinical value, more rigorous external validation and workflow-level integration are essential. We have compiled representative studies from recent years, as shown in Table 1 .

Lymphoma

Lymphoma is a highly heterogeneous malignant tumor originating from lymphocytes [ 58 ]. This disease encompasses over 100 subtypes, and its accurate diagnosis requires integrating clinical, morphological, immunophenotypic, and molecular genetic information, with pathological interpretation being crucial [ 59 ]. Given the significant variations among these subtypes, stratification of prognosis and selection of treatment strategies becomes more complex and challenging. AI can identify microscopic patterns that are difficult for human eyes to discern, thereby improving process efficiency and aiding in decision-making for complex cases [ 60 ].

Diagnosis and disease classification

In the field of pathological image analysis, DL has achieved automated feature extraction and classification of H&E-stained slides while identifying morphological patterns associated with specific molecular alterations. For instance, the image-based DL algorithm developed by Perry et al. can directly identify double/triple-hit lymphoma (DHL/THL) from H&E slides with a sensitivity of 100% and an area under the curve (AUC) of 0.95, outperforming traditional fluorescence in situ hybridization (FISH) detection [ 61 ]. The Carreras team used a ResNet-based CNN model to automatically distinguish follicular lymphoma (FL) from reactive hyperplastic tissue in conventional H&E whole-slide images, achieving a model accuracy of 99.80% through rigorous data segmentation and case-level validation [ 62 ]. These achievements highlight the efficiency potential of automated histomorphological analysis, though most studies remain based on single-center, small-sample data. External validation in large, prospective, multicenter cohorts is urgently needed to confirm their generalizability and clinical applicability.

In molecular subtype identification, the standardized interpretation of genomic data remains a core challenge for precise lymphoma genotyping [ 63 ]. Unlike solid tumors, molecular data from lymphomas are primarily used for classification rather than targeted therapy, leading to inconsistent interpretation frameworks and significant interobserver variability. AI technologies show promise in providing standardized solutions. For instance, Shen et al. developed a simplified 38-gene classifier named “LymphPlex” by integrating genome sequencing, RNA sequencing, and FISH techniques. This classifier successfully classified diffuse large B-cell lymphoma (DLBCL) into seven molecular subtypes with distinct prognostic characteristics, validated across over 1,000 patient cases [ 64 ]. Bobee et al. established a classification system for seven B-cell non-Hodgkin lymphoma (B-NHL) subtypes using expression profiles of approximately 130 genes and RF [ 65 ]. Genomic classifiers can be fragile to batch effects and cohort shift, because differences in library preparation, sequencing depth, and variant calling pipelines may change feature distributions, and performance may deteriorate when deployed in underrepresented patient groups without rigorous external calibration [ 66 ].

Prognosis assessment and treatment decision

In the development of prognostic models, multimodal data fusion is emerging as a key trend. Lee et al. developed a deep learning model integrating digital pathology and clinical data to predict immune chemotherapy responses in diffuse large B-cell lymphoma (DLBCL) patients, achieving a validation AUC of approximately 0.86 [ 67 ]. The model’s innovation lies in enhancing robustness and interpretability through knowledge distillation and attention mechanisms. However, given its limited sample size and retrospective design, comprehensive validation incorporating genomic data is required.

In radiomics, Jiang’s team developed a multicenter, interpretable deep learning system named DeepENKTCL. By automatically segmenting PET/CT images of extranodal NK/T-cell lymphoma (ENKTCL), the system extracts radiomics and topological features, then provides explanations using the SHAP framework. The resulting FusionScore outperforms traditional models in stratification for progression-free survival (PFS) and OS [ 68 ].

Beyond prognostic radiomics, PET/CT-based AI has also been developed to automate lymphoma staging under established clinical systems. Li et al. developed a multi-view deep learning model based on baseline 18 F-FDG PET/CT maximum-intensity-projection (MIP) images to classify limited versus advanced stage, achieving accuracy comparable to early-career nuclear medicine physicians while substantially reducing interpretation time [ 69 ]. Hasanabadi et al. further showed that radiomics-based machine learning could reproduce Lugano staging in a multicenter externally validated cohort, with extranodal features improving discrimination and survival analysis supporting its clinical relevance [ 70 ].

In contrast, traditional machine learning based on large-scale real-world cohorts has demonstrated robust performance. Zha et al. developed the FLIPI-C model using XGBoost to predict 24-month progression of disease (POD24) from data of 1,938 FL patients across 17 centers. The model achieved AUC values of 0.76 in internal validation and 0.70 in external validation, outperforming existing benchmarks and providing a potential tool for early intervention in high-risk patients [ 71 ]. However, the model does not incorporate genomic information and relies on retrospective data, and further prospective studies are needed to confirm its clinical utility.

In the field of CAR-T therapy prediction, Raj et al. proposed the unsupervised InflaMix model for predicting inflammatory responses. The INFLAmmation MIXture model successfully predicted 6-month PFS in three independent cohorts by integrating 14 pre-infusion laboratory and cytokine markers [ 72 ]. Its advantages include a single blood draw and simplification to 6 routine indicators, demonstrating strong clinical implementation potential.

Current AI research in lymphoma primarily focuses on common subtypes such as diffuse DLBCL and FL, where PET/CT radiomics continues to demonstrate value in predicting treatment efficacy and stratifying risk. The core challenge lies in the limited generalization capacity of models due to insufficient samples of long-tail rare subtypes, which will require solutions through multi-center data collaboration and other technical approaches. We have compiled representative studies from recent years, as shown in Table 2 .

Multiple myeloma

MM, a malignant tumor arising from clonal proliferation of plasma cells in the bone marrow, predominantly affects elderly patients. Its classic CRAB symptoms (hypercalcemia, renal impairment, anemia, and osteoporosis) severely compromise patients’ quality of life [ 75 ]. The challenges in early diagnosis, risk stratification, and treatment decision-making stem from the often insidious nature of bone lesions, uneven distribution of myeloma cells, and their molecular and clinical heterogeneity. AI offers new possibilities to overcome these limitations.

Diagnosis and disease classification

As the disease progresses, plasma cells in the bone marrow can enter the peripheral blood, forming circulating plasma cells (CPCs). Elevated CPC levels not only indicate an increased risk of extramedullary infiltration but also serve as a key indicator for prognosis assessment and MRD monitoring [ 76 ]. Chen et al. developed the Morphogo CNN morphological analysis system, trained on over 300,000 cell images, achieving 99.6% accuracy in CPC detection [ 76 ]. Notably, the model threshold can be adjusted according to clinical needs. At a lower threshold of 0.60, sensitivity for CPC detection increased to 96.2%, but specificity decreased to 78.0%, raising the risk of false positives. This threshold was empirically chosen to illustrate high-sensitivity screening rather than determined through formal optimization. Moreover, model performance was affected by staining quality and image resolution, and the lack of multicenter prospective validation limits its generalizability. Meanwhile, Andrade’s team built a real-time plasma cell detection model using You Only Look Once v7 (YOLOv7), achieving over 0.80 accuracy on public datasets, providing a viable solution for rapid screening in resource-limited environments [ 77 ].

Radiomics and automatic segmentation algorithms have demonstrated significant value in the quantitative assessment of MM lesions. Key research focuses include radiomics-based diagnostic models [ 78 , 79 ] and DL-driven automatic segmentation algorithms [ 80 – 82 ]. Sachpekidis et al. developed a DL-based automated workflow that achieved CT-based automatic segmentation of whole-body skeletal structures and mapped them to PET SUV maps, using multi-threshold strategies to calculate total tumor volume (MTV) and total lesion glycolysis (TLG) [ 82 ]. While this method provides an automated, reproducible quantitative approach to tumor burden assessment, its retrospective design and limited sample size limit direct clinical generalizability.

Prognosis assessment and treatment decision

Current prognostic stratification for MM relies on invasive bone marrow biopsies and cytogenetic analyses [ 83 ]. To address limited access to cytogenetic testing in real-world practice, Hao et al. developed a lumbar-spine MRI radiomics model for predicting ISS/R-ISS categories in newly diagnosed MM. A fusion model incorporating peripheral blood biomarkers further improved performance, suggesting a practical non-invasive supplement for baseline risk stratification when molecular profiling is unavailable [ 84 ]. The Mosquera team developed three molecular risk scores for MM by integrating clinical trial data from over 14,000 newly diagnosed cases (NDMM) into the EMN–HARMONY platform. These scores were validated against PFS and OS through internal cross-validation and external cohort testing, based on bone marrow cytogenetics and baseline clinical parameters [ 85 ]. In MM, risk stratification tools remain difficult to operationalize unless they align with established staging logic and real-world heterogeneity, because transplant eligibility, frailty, and extramedullary disease can shift outcomes independently of baseline biomarkers and may limit portability across treatment settings [ 86 ].

In recent years, the prognostic evaluation of MM has transitioned from static stratification to dynamic risk modeling. The SCOPE model, proposed by Hussain et al., is a time-series multi-task machine learning framework based on the transformer architecture. It can simultaneously perform event prediction, biomarker trajectory inference, and counterfactual treatment effect assessment [ 12 ]. Designed for NDMM and relapsed/refractory multiple myeloma (RRMM), this method provides time-updated, personalized, dynamic prognostic and treatment-selection criteria, demonstrating superior discrimination and calibration compared to static systems. However, its potential indicator bias and out-of-domain generalization capability remain unresolved challenges. The Murie team constructed the mmSYGNAL transcriptional program network based on Systems Genetic Network Analysis (SYGNAL), deriving program activity from multi-omics data and training cell genetics subtype-specific ML models accordingly, forming a unified framework for personalized dynamic risk assessment and treatment selection [ 87 ]. This framework performed excellently in multi-cohort external validation, with high consistency in drug response prediction and in vitro drug sensitivity. However, its clinical application still relies on high-quality omics data and prospective validation.

Current AI research in MM has shifted focus from manual feature extraction to end-to-end frameworks. These systems utilize PET/CT and whole-body MRI for lesion detection, bone marrow burden quantification, and relapse prediction, while incorporating time-series models to transform prognosis assessment into an updatable dynamic system. Clinical challenges include inconsistent data acquisition and post-processing protocols, as well as insufficient prospective evidence. The next step involves establishing standardized imaging protocols and conducting multicenter clinical trials to validate their generalizability and clinical benefits. Representative studies in recent years are summarized in Table 3 .

Myelodysplastic syndromes

Myelodysplastic syndromes (MDS) are myeloid clonal disorders characterized by hematopoietic stem cell dysfunction, dysplasia, and ineffective hematopoiesis [ 92 ]. In early MDS, increased apoptosis of bone marrow precursors is closely associated with ineffective hematopoiesis and contributes to peripheral cytopenias [ 93 ]. They exhibit diverse manifestations in peripheral blood and bone marrow, with a high risk of progression to AML [ 92 ]. Its diagnosis should be distinguished from other bone marrow failure disorders with overlapping clinical manifestations, particularly aplastic anemia (AA) and paroxysmal nocturnal hemoglobinuria (PNH), which are distinct entities but may similarly present with peripheral cytopenias [ 94 ]. This complexity poses a serious challenge to early diagnosis and precise stratification of MDS, but it also provides a broad space for AI applications.

Diagnosis and disease classification

Current AI-based diagnostic methods primarily rely on bone marrow smears, peripheral blood smears, and multi-parameter flow cytometry (MFC) data [ 95 ]. Among these, analysis of bone marrow smear images represents a key research direction. The cell classification algorithm for bone marrow smears developed by Lee et al. can automatically identify hematopoietic cell lineages in patients with MDS and detect abnormally growing cells in bone marrow smears. The AUC for distinguishing pathological and normal hematopoietic cells ranged from 0.945 to 0.996, and the accuracy ranged from 0.912 to 0.993, enabling automatic quantification of morphological characteristics [ 96 ].

In blood sample analysis, Zhu et al. optimized the MDS-CBC score (e-MDS-CBC) using ML by integrating the immature platelet fraction (IPF) with erythroid and granulocyte parameters, significantly improving MDS screening efficacy. The advantage lies in enabling primary screening with routine blood parameters alone. However, its generalizability depends on the distribution of white blood cell populations (CPD) and requires validation in broader populations [ 97 ].

Furthermore, AI models based on flow cytometry have demonstrated significant potential. The Clichet team developed an MFC-based MDS diagnostic model using the Elastic Net al.gorithm, identifying five key features from flow cytometry data. The model achieved an AUC of 0.935, 91.8% sensitivity, and 92.5% specificity on the validation set, outperforming the conventional Ogata score. However, its high data preprocessing requirements necessitate broader cohort validation for robustness [ 98 ].

Prognosis assessment and treatment decision

The prognosis assessment for MDS has traditionally relied on scoring systems such as the International Prognostic Scoring System (IPSS) and its revised version, IPSS-R. The recently proposed molecular IPSS (IPSS-Mol) integrates the mutation status of 31 genes, achieving a discriminant C-index of 0.75 and significantly outperforming IPSS-R, emerging as a more valuable clinical tool [ 99 ]. Mosquera Orgueira et al. utilized the Rasch-Schwarz-Fischer (RSF) algorithm to incorporate eight basic clinical and cytogenetic variables, developing an MDS prognostic scoring system with C-indexes of 0.78 for OS and 0.84 for leukemia-free survival (LFS). Its external validation demonstrated comparable performance to IPSS-Mol without requiring genetic testing, though heterogeneity in data entry across centers may affect its long-term robustness [ 100 ].

Beyond static baseline scores, dynamic risk modeling using longitudinal clinical data is emerging in MDS. Bobak et al. developed a gradient-boosted decision-tree model that repeatedly estimates 1-year mortality risk using follow-up blood values plus diagnosis-based features, trained on the Düsseldorf MDS registry and externally validated in independent cohorts from Heidelberg and Mannheim, achieving AUROC values around 0.8 and outperforming a diagnosis-only baseline model [ 101 ].

In treatment decision-making, Radakovich et al. demonstrated that machine learning model based on continuous blood routine changes within 90 days after hypomethylating agent (HMA) therapy can predict treatment response early, providing a novel approach for cost-effective dynamic monitoring. However, current evidence remains limited to retrospective single-center validation [ 102 ]. Allogeneic hematopoietic stem cell transplantation (Allo-HSCT) remains the only potentially curative approach for MDS [ 103 ]. The selection of pre-transplant conditioning regimens is critical. Shimomura’s team developed a machine learning model using 2,567 transplantation cases that could distinguish patient populations that benefited from myeloablative conditioning (MAC) from those that benefited from reduced-intensity conditioning (RIC). This data-driven support for personalized transplantation strategies highlights the value of AI in complex treatment decisions, though its conclusions require confirmation through prospective multicenter studies [ 104 ].

Current AI applications in MDS primarily focus on two domains: streamlined multi-parameter flow models and deep learning for bone marrow smears. The prognostic aspect is transitioning from static baseline analysis to data fusion and temporalization. However, its widespread adoption faces challenges, including noise in labeled data, difficulty distinguishing critical disease cases, limited sample sizes, and model generalization issues due to data domain shifts. We have summarized representative related studies in recent years, as shown in Table 4 .

Challenges and strategies in clinical translation

While AI research in hematologic malignancies continues to gain momentum, its transition from algorithmic innovation to clinical implementation faces significant challenges. To transform models that excel in retrospective data into reliable real-world decision-support tools, we must systematically bridge the gaps in data, algorithms, and ethical oversight (Fig. 2 ).

Data aspects

High-quality, large-scale training data is fundamental to ensuring AI’s strong generalization capabilities [ 15 ]. In hematologic malignancies, dataset complexity extends beyond sample size: clinically meaningful labels are hierarchical and periodically revised, which can complicate retrospective harmonization across cohorts [ 39 , 107 , 108 ]. Moreover, clinically relevant endpoints are often time-varying and assay-dependent, increasing the need for standardized reporting and improved cross-platform harmonization, with broader multicenter external validation still needed [ 109 , 110 ]. Several open or access-controlled resources now support benchmarking and validation, including CRDC/GDC [ 111 ], MMRF CoMMpass [ 112 ], TARGET [ 113 ], Beat AML [ 114 ], St. Jude Cloud [ 115 ], FlowRepository [ 116 ], and EGA [ 117 ], while Scientific Data [ 118 ] and Zenodo [ 119 ] are expanding reusable cytomorphology and cytometry benchmarks. However, the inherent multimodal nature of hematologic tumor data and clinical realities pose a primary challenge. Diagnostic data—spanning morphology, immunology, cytogenetics, and molecular biology—vary not only in origin and format but are also scattered across heterogeneous information systems across different medical institutions, creating severe “data silos” [ 120 ]. Furthermore, differences in sample processing, staining protocols, and scanning equipment among centers introduce uncontrollable biases, severely limiting the models’ out-of-domain generalization performance [ 16 ].

To overcome data bottlenecks and achieve clinically actionable outcomes, a phased approach of “data construction—standardization—integration” should be implemented. The first step involves establishing a multi-center shared database that systematically aggregates multimodal raw data (e.g., morphological and imaging) while simultaneously recording interoperable metadata [ 18 ]. This metadata facilitates the identification of domain and time drift and provides structured support for external validation and generalization evaluation. Secondly, establishing unified data collection and quality control standards can reduce technical variations at the source [ 19 ], improve cross-center consistency and comparability, and emphasize practicality and implementability. Finally, promote multimodal alignment and hierarchical fusion [ 121 ], adopt traceable, interpretable fusion strategies, and seamlessly integrate with EHR workflows. This measure aims to systematically reduce data heterogeneity, enhance the model’s robustness and out-of-domain generalization in a multi-center environment, and support continuous drift monitoring and model governance.

Algorithmic aspects

In algorithmic development, three core challenges persist: generalization, interpretability, and fairness. First, inadequate generalization capability remains a critical barrier to clinical translation. Models trained on specific datasets often exhibit “domain shift” phenomena, making them ill-suited for diverse populations, testing platforms, or evolving clinical practices. This directly results in significant performance degradation of AI tools during cross-center validation [ 17 , 19 ]. Second, the “black box” decision-making inherent in complex models like deep learning severely hinders clinical adoption. The opaque decision-making processes make it difficult for physicians to understand and trust model outputs, creating a fundamental obstacle in life-critical medical decisions [ 122 ]. Third, algorithmic bias and fairness issues demand attention. When training data inadequately represent diverse racial, age, and gender groups, models may amplify existing healthcare inequalities, leading to systematic diagnostic or prognostic biases against minority populations [ 123 ].

To address algorithmic challenges, a self-supervised pre-training and task-adaptive architecture can be adopted. This framework first constructs a general representation with strong generalization capability through cross-modal alignment and contrastive learning using large-scale unlabeled multimodal medical data [ 124 ]. Subsequently, it performs fine-tuning on downstream tasks with limited labeled samples to adapt to scenarios such as diagnostic classification and prognosis evaluation [ 125 , 126 ]. By incorporating causal annotations and expert prior knowledge, the model’s decision-making becomes more interpretable and clinically credible at the mechanistic level. Its output also extends from traditional classification probability to event risk and individualized intervention benefit, aligning more closely with the logic of clinical decision-making [ 126 ] (Fig. 3 ).

However, the introduction of such complex models also makes their robustness and reliability critical prerequisites for clinical translation. Therefore, rigorous validation and interpretive techniques must be implemented. First, models must undergo large-scale, multicenter external validation and adhere strictly to the TRIPOD + AI Declaration for complete and transparent reporting [ 15 ]. Second, explainable AI (XAI) techniques such as SHAP, LIME, and attention mechanisms should be integrated to visualize the decision-making basis of models, transforming them from “black boxes” into “transparent systems” [ 122 ]. Finally, fairness audits throughout the entire model development cycle—actively detecting and correcting performance disparities across subgroups—are ethically imperative to ensure health equity [ 127 ]. These measures aim to enhance AI model performance, enabling them to better serve clinical practice.

Ethical and regulatory aspects

The clinical deployment of AI must ensure patient safety and public welfare, yet this process faces multiple challenges. The foremost issue lies in liability delineation and accountability. When AI-assisted decision-making errors occur, the legal responsibility among developers, deployment institutions, and clinicians remains an unresolved gray area [ 124 ]. Second, regulatory science faces challenges. The traditional medical device approval process struggles to keep pace with the rapid iteration of AI software. Regulatory agencies need to establish a more adaptive review framework to balance innovation, safety, and efficiency [ 103 ]. Additionally, broad social and ethical risks cannot be overlooked. Beyond algorithmic bias, issues such as data misuse and patient consent must be proactively considered [ 123 , 128 , 129 ].

To address ethical and regulatory challenges and establish a trustworthy clinical AI system, systematic progress must be made through institutional design and technological synergy. At the institutional level, promoting transparent research reporting and trial pre-registration forms the foundation of trust. Transparent reporting requires full disclosure of the entire research process, in accordance with international standards such as TRIPOD + AI [ 15 ], ensuring procedural transparency and reproducible results. Trial pre-registration helps clarify research hypotheses and evaluation metrics, preventing selection bias at the source and enhancing evidentiary reliability [ 106 ]. Regarding technological implementation, advocating human-machine collaborative evaluation and decision-making paradigms is essential [ 17 ]. It is crucial to define the primary role in critical decisions alongside AI’s supportive function, ensuring AI outputs effectively assist rather than interfere with clinical judgment.

Future directions and summary

In the future, the diagnosis and treatment of hematologic cancers may increasingly move toward multi-level, multimodal data integration and more coordinated AI-assisted decision support to enable dynamic optimization of clinical care. Based on the AI clinical deployment process, this system can be divided into several tiers: the infrastructure and input layer, the model processing layer, the analysis layer, and the decision-making and evaluation layer, corresponding to the information flow shown in Fig. 4 .

The realization of this vision depends on the deep integration of AI with EHR systems. For instance, generative models like Foresight have demonstrated the feasibility of building universal temporal models based on large-scale real-world EHRs [ 130 ]. In such systems, AI could function as a monitoring and decision-support layer that helps identify evolving patient risks and trigger clinician review when necessary. These alerts are then reviewed by physicians to guide interventions, thereby driving the automation and dynamism of the diagnostic and treatment process (Fig. 5 ).

AI has shown potential in hematologic malignancy diagnosis and treatment, with emerging applications spanning disease diagnosis and classification, prognosis assessment, and treatment decision support. To realize this vision clinically, rigorous validation, regulatory approvals, robust IT infrastructure, and ongoing professional oversight are essential.

This review has several limitations. It is a narrative rather than a systematic review, and selection bias cannot be excluded. The included studies are highly heterogeneous in disease subtype, data modality, study design, and evaluation metrics, limiting direct comparison across studies. Many cited AI studies remain retrospective, single-center, or proof-of-concept, with limited prospective validation and uncertain generalizability. In addition, improved model performance does not necessarily translate into clinical utility or patient benefit. Discussions on LLM-assisted decision support and EHR-embedded closed-loop systems are forward-looking and should be interpreted as conceptual directions rather than established standards of care.

While challenges remain, continued methodological refinement, accumulation of higher-quality clinical evidence, and deeper interdisciplinary collaboration may expand the role of AI in hematologic malignancies. Such progress may support more dynamic and individualized medical care, but its clinical translation will still require rigorous validation, governance, and responsible implementation.

Acknowledgements

The authors thank their colleagues for constructive comments and technical support.

Author contributions

Z.L., J.Z., L.H., and X.C. conducted the literature search and study screening, extracted and synthesized the evidence, and drafted the manuscript. J.W. and W.T. conceived the scope and structure of the review, provided supervision and critical revisions for important intellectual content, and finalized the manuscript. All authors reviewed and approved the final version for submission.

Funding

This work was supported by the Fundamental Research Program of Shanxi Province (Grant No. 202303021211224; Weiwei Tian) and the Open Fund of the Key Laboratory of Molecular Diagnosis and Treatment of Hematological Diseases of Shanxi Province (Grant No. KLMDT202402; Weiwei Tian).

Data availability

No datasets were generated or analysed during the current study.

Declarations

Competing interests

The authors declare no competing interests.

Footnotes

Publisher’s note

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

Zhujin Li and Jie Zhao contributed equally to this work and share first authorship.

References

1. Takahashi K, Tanaka T (2023) Clonal evolution and hierarchy in myeloid malignancies. Trends Cancer 9:707–715. 10.1016/j.trecan.2023.05.004

2. Khoury R, Raffoul C, Khater C, Hanna C (2025) Precision medicine in hematologic malignancies: evolving concepts and clinical applications. Biomedicines 13:1654. 10.3390/biomedicines13071654

3. Nazha A, Elemento O, Ahuja S, Lam B, Miles M, Shouval R et al (2025) Artificial intelligence in hematology. Blood 146:2283–2292. 10.1182/blood.2025029876

4. Radakovich N, Nagy M, Nazha A (2020) Machine learning in haematological malignancies. Lancet Haematol 7:e541–e550. 10.1016/S2352-3026(20)30121-6

5. Chandhok NS, Sekeres MA (2025) Measurable residual disease in hematologic malignancies: a biomarker in search of a standard. EClinicalMedicine 86:103348. 10.1016/j.eclinm.2025.103348

6. El Alaoui Y, Elomri A, Qaraqe M, Padmanabhan R, Yasin Taha R, El Omri H et al (2022) A review of artificial intelligence applications in hematology management: current practices and future prospects. J Med Internet Res 24:e36490. 10.2196/36490

7. Matek C, Schwarz S, Spiekermann K, Marr C (2019) Human-level recognition of blast cells in acute myeloid leukaemia with convolutional neural networks. Nat Mach Intell 1:538–544. 10.1038/s42256-019-0101-9

8. Sun S, Yin Z, Van Cleave JG, Wang L, Fried B, Bilal KH et al (2025) DeepHeme, a high-performance, generalizable deep ensemble for bone marrow morphometry and hematologic diagnosis. Sci Transl Med 17:eadq2162. 10.1126/scitranslmed.adq2162

9. Eckardt J-N, Middeke JM, Riechert S, Schmittmann T, Sulaiman AS, Kramer M et al (2022) Deep learning detects acute myeloid leukemia and predicts NPM1 mutation status from bone marrow smears. Leukemia 36:111–118. 10.1038/s41375-021-01408-w

10. Kosvyra Α, Karadimitris Α, Papaioannou Μ, Chouvarda I (2024) Machine learning and integrative multi-omics network analysis for survival prediction in acute myeloid leukemia. Comput Biol Med 178:108735. 10.1016/j.compbiomed.2024.108735

11. Zhou Y, Smith J, Keerthi D, Li C, Sun Y, Mothi SS et al (2024) Longitudinal clinical data improve survival prediction after hematopoietic cell transplantation using machine learning. Blood Adv 8:686–698. 10.1182/bloodadvances.2023011752

12. Hussain Z, De Brouwer E, Boiarsky R, Setty S, Gupta N, Liu G et al (2024) Joint AI-driven event prediction and longitudinal modeling in newly diagnosed and relapsed multiple myeloma. Npj Digit Med 7:200. 10.1038/s41746-024-01189-3

13. Agius R, Riis-Jensen AC, Wimmer B, Cunha-Bang C, Murray DD, Poulsen CB et al (2024) Deployment and validation of the CLL treatment infection model adjoined to an EHR system. NPJ Digit Med 7:147. 10.1038/s41746-024-01132-6

14. Castro GA, Almeida JM, Machado-Neto JA, Almeida TA (2024) A decision support system to recommend appropriate therapy protocol for AML patients. Front Artif Intell 7:1343447. 10.3389/frai.2024.1343447

15. Collins GS, Moons KGM, Dhiman P, Riley RD, Beam AL, Van Calster B et al (2024) TRIPOD + AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ 385:e078378. 10.1136/bmj-2023-078378

16. Asadi-Aghbolaghi M, Darbandsari A, Zhang A, Contreras-Sanz A, Boschman J, Ahmadvand P et al (2024) Learning generalizable AI models for multi-center histopathology image classification. Npj Precis Oncol 8:151. 10.1038/s41698-024-00652-4

17. Vasey B, Nagendran M, Campbell B, Clifton DA, Collins GS, Denaxas S et al (2022) Reporting guideline for the early-stage clinical evaluation of decision support systems driven by artificial intelligence: DECIDE-AI. Nat Med 28:924–933. 10.1038/s41591-022-01772-9

18. Touré V, Krauss P, Gnodtke K, Buchhorn J, Unni D, Horki P et al (2023) Fairification of health-related data using semantic web technologies in the Swiss Personalized Health Network. Sci Data 10:127. 10.1038/s41597-023-02028-y

19. Rajagopal A, Ayanian S, Ryu AJ, Qian R, Legler SR, Peeler EA et al (2024) Machine learning operations in health care: a scoping review. Mayo Clin Proc Digit Health 2:421–437. 10.1016/j.mcpdig.2024.06.009

20. Golub TR, Slonim DK, Tamayo P, Huard C, Gaasenbeek M, Mesirov JP et al (1999) Molecular classification of cancer: class discovery and class prediction by gene expression monitoring. Science 286:531–537. 10.1126/science.286.5439.531

21. Shipp MA, Ross KN, Tamayo P, Weng AP, Kutok JL, Aguiar RCT et al (2002) Diffuse large B-cell lymphoma outcome prediction by gene-expression profiling and supervised machine learning. Nat Med 8:68–74. 10.1038/nm0102-68

22. Bullinger L, Döhner K, Bair E, Fröhling S, Schlenk RF, Tibshirani R et al (2004) Use of gene-expression profiling to identify prognostic subclasses in adult acute myeloid leukemia. N Engl J Med 350:1605–1616. 10.1056/NEJMoa031046

23. Haferlach T, Kohlmann A, Wieczorek L, Basso G, Kronnie GT, Béné M-C et al (2010) Clinical utility of microarray-based gene expression profiling in the diagnosis and subclassification of leukemia: Report from the International Microarray Innovations in Leukemia Study Group. J Clin Oncol 28:2529–2537. 10.1200/JCO.2009.23.4732

24. Papaemmanuil E, Gerstung M, Bullinger L, Gaidzik VI, Paschka P, Roberts ND et al (2016) Genomic classification and prognosis in acute myeloid leukemia. N Engl J Med 374:2209–2221. 10.1056/NEJMoa1516192

25. Botta C, Di Martino MT, Ciliberto D, Cucè M, Correale P, Rossi M et al (2016) A gene expression inflammatory signature specifically predicts multiple myeloma evolution and patients survival. Blood Cancer J 6:e511. 10.1038/bcj.2016.118

26. Wang M, Lindberg J, Klevebring D, Nilsson C, Lehmann S, Grönberg H et al (2018) Development and validation of a novel RNA sequencing-based prognostic score for acute myeloid leukemia. J Natl Cancer Inst 110:1094–1101. 10.1093/jnci/djy021

27. Chaudhary K, Poirion OB, Lu L, Garmire LX (2018) Deep learning-based multi-omics integration robustly predicts survival in liver cancer. Clin Cancer Res 24:1248–1259. 10.1158/1078-0432.CCR-17-0853

28. Jha KK, Dutta HS (2019) Mutual information based hybrid model and deep learning for acute lymphocytic leukemia detection in single cell blood smear images. Comput Methods Programs Biomed 179:104987. 10.1016/j.cmpb.2019.104987

29. Kimura K, Tabe Y, Ai T, Takehara I, Fukuda H, Takahashi H et al (2019) A novel automated image analysis system using deep convolutional neural networks can assist to differentiate MDS and AA. Sci Rep 9:13385. 10.1038/s41598-019-49942-z

30. Shahin AI, Guo Y, Amin KM, Sharawi AA (2019) White blood cells identification system based on convolutional deep neural learning networks. Comput Methods Programs Biomed 168:69–80. 10.1016/j.cmpb.2017.11.015

31. Maura F, Rajanna AR, Ziccheddu B, Poos AM, Derkach A, Maclachlan K et al (2024) Genomic classification and individualized prognosis in multiple myeloma. J Clin Oncol 42:1229–1240. 10.1200/JCO.23.01277

32. Mosquera Orgueira A, González Pérez MS, Díaz Arias JÁ, Antelo Rodríguez B, Alonso Vence N, Bendaña López Á et al (2021) Survival prediction and treatment optimization of multiple myeloma patients using machine-learning models based on clinical and gene expression data. Leukemia 35:2924–2935. 10.1038/s41375-021-01286-2

33. Mataraso SJ, Espinosa CA, Seong D, Reincke SM, Berson E, Reiss JD et al (2025) A machine learning approach to leveraging electronic health records for enhanced omics analysis. Nat Mach Intell 7:293–306. 10.1038/s42256-024-00974-9

34. Eriksen AV, Möller S, Ryg J (2024) Use of GPT-4 to diagnose complex clinical cases. NEJM AI 1:AIp2300031. 10.1056/AIp2300031

35. Ferber D, Wiest IC, Wölflein G, Ebert MP, Beutel G, Eckardt J-N et al (2024) GPT-4 for information retrieval and comparison of medical oncology guidelines. NEJM AI 1:AIcs2300235. 10.1056/AIcs2300235

36. Ferber D, El Nahhas OSM, Wölflein G, Wiest IC, Clusmann J, Leßmann M-E et al (2025) Development and validation of an autonomous artificial intelligence agent for clinical decision-making in oncology. Nat Cancer 6:1337–1349. 10.1038/s43018-025-00991-6

37. Goodell AJ, Chu SN, Rouholiman D, Chu LF (2025) Large language model agents can use tools to perform clinical calculations. NPJ Digit Med 8:163. 10.1038/s41746-025-01475-8

38. Choi H-S, Kim BS, Yoon S, Oh S-O, Lee D (2024) Leukemic stem cells and hematological malignancies. Int J Mol Sci 25:6639. 10.3390/ijms25126639

39. Arber DA, Orazi A, Hasserjian RP, Borowitz MJ, Calvo KR, Kvasnicka H-M et al (2022) International consensus classification of myeloid neoplasms and acute leukemias: integrating morphologic, clinical, and genomic data. Blood 140:1200–1228. 10.1182/blood.2022015850

40. Aakash F, Gisriel SD, Zeidan AM, Bennett JM, Bejar R, Bewersdorf JP et al (2024) Contemporary approach to the diagnosis and classification of myelodysplastic neoplasms/syndromes-recommendations from the International Consortium for Myelodysplastic Neoplasms/Syndromes (MDS [icMDS]). Mod Pathol 37:100615. 10.1016/j.modpat.2024.100615

41. Schmidt B, Brown LM, Ryland GL, Lonsdale A, Kosasih HJ, Ludlow LE et al (2022) ALLSorts: an RNA-Seq subtype classifier for B-cell acute lymphoblastic leukemia. Blood Adv 6:4093–4097. 10.1182/bloodadvances.2021005894

42. Zhou SK, Le HN, Luu K, V Nguyen H, Ayache N (2021) Deep reinforcement learning in medical imaging: a literature review. Med Image Anal 73:102193. 10.1016/j.media.2021.102193

43. Yadav DP, Kumar D, Jalal AS, Kumar A, Singh KU, Shah MA (2023) Morphological diagnosis of hematologic malignancy using feature fusion-based deep convolutional neural network. Sci Rep 13:16988. 10.1038/s41598-023-44210-7

44. Fan BE, Yong BSJ, Li R, Wang SSYW, Aw MYN, Chia MF, Chen DTY, Neo YS, Occhipinti B, Ling RR, Ramanathan K, Ong YX, Lim KG, Wong WYK, Lim SP, Latiff STBA, Shanmugam H, Wong MS, Ponnudurai K, Winkler S (2024) From microscope to micropixels: a rapid review of artificial intelligence for the peripheral blood film. Blood Rev 64:101144. 10.1016/j.blre.2023.101144

45. Hybel TE, Jensen SH, Rodrigues MA, Hybel TE, Pedersen MN, Qvick SH et al (2024) Imaging flow cytometry and convolutional neural network-based classification enable discrimination of hematopoietic and leukemic stem cells in acute myeloid leukemia. Int J Mol Sci 25:6465. 10.3390/ijms25126465

46. Seheult JN, Otteson GE, Timm MM, Weybrighe MJ, Shi M, OITEANU H, Jevremovic D, Chen C, Chiu A, Horna P (2026) Artificial intelligence accelerates the interpretation of measurable residual B lymphoblastic leukemia by flow cytometry. Blood Adv 10:58–69. 10.1182/bloodadvances.2025016126

47. Vahedi A, Royaei M, Madani T, Mercier F, Poopak B (2025) A machine learning approach for the rapid identification of measurable residual disease in acute myeloid leukemia. Haematologica 110:1854. 10.3324/haematol.2024.286019

48. Eckardt J-N, Röllig C, Metzeler K, Kramer M, Stasik S, Georgi J-A et al (2023) Prediction of complete remission and survival in acute myeloid leukemia using supervised machine learning. Haematologica 108:690–704. 10.3324/haematol.2021.280027

49. Eckardt J, Hahn W, Ries RE, Chrost SD, Winter S, Stasik S et al (2025) Age-stratified machine learning identifies divergent prognostic significance of molecular alterations in AML. Hemasphere 9:e70132. 10.1002/hem3.70132

50. Liebers N, Bruch PM, Terzer T, Hernandez MH, Paramasivam N, Fitzgerald D et al (2023) Ex vivo drug response profiling for response and outcome prediction in hematologic malignancies: the prospective non-interventional SMARTrial. Nat Cancer 4:1648–1659. 10.1038/s43018-023-00645-5

51. Cheung HL, Wong YH, Li YY, Yang XX, Ko LH, Kabigting FET, Chan KC, Leung AYH, Chan BP (2025) Microenvironment matters: In vitro 3D bone marrow niches differentially modulate survival, phenotype and drug responses of acute myeloid leukemia (AML) cells. Biomaterials 312:122719

52. Jeha S, Choi J, Roberts KG, Pei D, Coustan-Smith E, Inaba H et al (2021) Clinical significance of novel subtypes of acute lymphoblastic leukemia in the context of minimal residual disease-directed therapy. Blood Cancer Discov 2:326–337. 10.1158/2643-3230.BCD-20-0229

53. Zhang R, Han X, Lei Z, Jiang C, Gul I, Hu Q et al (2022) RCMNet: a deep learning model assists CAR-T therapy for leukemia. Comput Biol Med 150:106084. 10.1016/j.compbiomed.2022.106084

54. Alabdulqader EA, Alarfaj AA, Umer M, Eshmawi AA, Alsubai S, Kim T et al (2024) Improving prediction of blood cancer using leukemia microarray gene data and Chi2 features with weighted convolutional neural network. Sci Rep 14:15625. 10.1038/s41598-024-65315-7

55. Abbasi EY, Deng Z, Ali Q, Khan A, Shaikh A, Reshan MSA et al (2024) A machine learning and deep learning-based integrated multi-omics technique for leukemia prediction. Heliyon 10:e25369. 10.1016/j.heliyon.2024.e25369

56. Tang M, Antić Ž, Fardzadeh P, Pietzsch S, Schröder C, Eberhardt A et al (2024) An artificial intelligence-assisted clinical framework to facilitate diagnostics and translational discovery in hematologic neoplasia. eBioMedicine 104:105171. 10.1016/j.ebiom.2024.105171

57. Shehta AI, Nasr M, El Ghazali AEDM (2025) Blood cancer prediction model based on deep learning technique. Sci Rep 15:1889. 10.1038/s41598-024-84475-0

58. Campo E, Jaffe ES, Cook JR, Quintanilla-Martinez L, Swerdlow SH, Anderson KC et al (2022) The international consensus classification of mature lymphoid neoplasms: a report from the Clinical Advisory Committee. Blood 140:1229–1253. 10.1182/blood.2022015851

59. UmaMaheswaran SK, Brahmaiah VP, Suresh C, Balaji SR, Ahmad F (2025) Artificial intelligence based malignant lymphoma type prediction using enhanced super resolution image and hybrid feature extraction algorithm. Comput Biol Med 193:110335. 10.1016/j.compbiomed.2025.110335

60. Syrykh C, Van Den Brand M, Kather JN, Laurent C (2025) Role of artificial intelligence in haematolymphoid diagnostics. Histopathology 86:58–68. 10.1111/his.15327

61. Perry C, Greenberg O, Haberman S, Herskovitz N, Gazy I, Avinoam A et al (2023) Image-based deep learning detection of high-grade B-cell lymphomas directly from hematoxylin and eosin images. Cancers 15:5205. 10.3390/cancers15215205

62. Carreras J, Ikoma H, Kikuti YY, Nagase S, Ito A, Orita M et al (2025) Histological image classification between follicular lymphoma and reactive lymphoid tissue using deep learning and explainable artificial intelligence (XAI). Cancers 17:2428. 10.3390/cancers17152428

63. Zhang W, Ao Q, Guan Y, Zhu Z, Kuang D, Li MMQ et al (2022) A novel diagnostic approach for the classification of small B-cell lymphoid neoplasms based on the NanoString platform. Mod Pathol 35:632–639. 10.1038/s41379-021-00954-z

64. Shen R, Fu D, Dong L, Zhang M-C, Shi Q, Shi Z-Y et al (2023) Simplified algorithm for genetic subtyping in diffuse large B-cell lymphoma. Signal Transduct Target Ther 8:145. 10.1038/s41392-023-01358-y

65. Bobée V, Drieux F, Marchand V, Sater V, Veresezan L, Picquenot J-M et al (2020) Combining gene expression profiling and machine learning to diagnose B-cell non-Hodgkin lymphoma. Blood Cancer J 10:59. 10.1038/s41408-020-0322-5

66. Reardon B, Culhane AC, Van Allen EM (2026) Convergence of machine learning and genomics for precision oncology. Nat Rev 26:21–229. 10.1038/s41568-025-00897-6

67. Lee JH, Song G, Lee J, Kang S, Moon KM, Choi Y et al (2024) Prediction of immunochemotherapy response for diffuse large B -cell lymphoma using artificial intelligence digital pathology. J Pathol Clin Res 10:e12370. 10.1002/2056-4538.12370

68. Jiang C, Jiang Z, Zhang X, Qu L, Fu K, Teng Y et al (2025) Robust and interpretable deep learning system for prognostic stratification of extranodal natural killer/T-cell lymphoma. Eur J Nucl Med Mol Imaging 52:1739–1750. 10.1007/s00259-024-07024-x

69. Li D, Li R, Ma Y, Huang SY, Han M, Gao YZ et al (2026) Multi-view deep learning for automated lymphoma staging from 18F-FDG PET/CT: physician-level accuracy with high-throughput workflow. EJNMMI Res. 10.1186/s13550-025-01357-w

70. Hasanabadi S, Aghamiri SMR, Abin AA, Cheraghi M, Bakhshayesh KM, Vosoughi H et al (2025) Automatic Lugano staging for risk stratification in lymphoma: a multicenter PET radiomics and machine learning study with survival analysis[J]. Nucl Med Commun 46:1200–1211. 10.1097/MNM.0000000000002046

71. Zha J, Chen Q, Zhang W, Jing H, Ye J, Liu H et al (2025) A machine learning-based model to predict POD24 in follicular lymphoma: a study by the Chinese workshop on follicular lymphoma. Biomark Res 13:2. 10.1186/s40364-024-00716-4

72. Raj SS, Fei T, Fried S, Ip A, Fein JA, Leslie LA et al (2025) An inflammatory biomarker signature of response to CAR-T cell therapy in non-Hodgkin lymphoma. Nat Med 31:1183–1194. 10.1038/s41591-025-03532-x

73. Zhang X, Zhao Z, Wang R, Chen H, Zheng X, Liu L et al (2024) A multicenter proof-of-concept study on deep learning-based intraoperative discrimination of primary central nervous system lymphoma. Nat Commun 15:3768. 10.1038/s41467-024-48171-x

74. Swiderska-Chadaj Z, Hebeda KM, Van Den Brand M, Litjens G (2021) Artificial intelligence to detect MYC translocation in slides of diffuse large B-cell lymphoma. Virchows Arch 479:617–621. 10.1007/s00428-020-02931-4

75. Malard F, Neri P, Bahlis NJ, Terpos E, Moukalled N, Hungria VTM et al (2024) Multiple myeloma. Nat Rev Dis Primers 10:45. 10.1038/s41572-024-00529-7

76. Chen P, Zhang L, Cao X, Jin X, Chen N, Zhang L et al (2024) Detection of circulating plasma cells in peripheral blood using deep learning-based morphological analysis. Cancer 130:1884–1893. 10.1002/cncr.35202

77. Andrade CLB, Ferreira MV, Alencar BM, Junior AMA, Lopes TJS, Dos Santos AS et al (2024) Enhancing diagnostic accuracy of multiple myeloma through ML-driven analysis of hematological slides: new dataset and identification model to support hematologists. Sci Rep 14:11176. 10.1038/s41598-024-61420-9

78. Zhang S, Liu M, Li S, Cui J, Zhang G, Wang X (2023) An MRI-based radiomics nomogram for differentiating spinal metastases from multiple myeloma. Cancer Imaging 23:72. 10.1186/s40644-023-00585-4

79. Cao J, Li Q, Zhang H, Wu Y, Wang X, Ding S et al (2024) Radiomics model based on MRI to differentiate spinal multiple myeloma from metastases: a two-center study. J Bone Oncol 45:100599. 10.1016/j.jbo.2024.100599

80. Bauer F, Kächele J, Bernhard J, Hajiyianni M, Weinhold N, Sauer S et al (2025) Advanced automated model for robust bone marrow segmentation in whole-body MRI. Acad Radiol 32:2824–2835. 10.1016/j.acra.2024.12.060

81. Wennmann M, Neher P, Stanczyk N, Kahl K-C, Kächele J, Weru V et al (2023) Deep learning for automatic bone marrow apparent diffusion coefficient measurements from whole-body magnetic resonance imaging in patients with multiple myeloma: a retrospective multicenter study. Invest Radiol 58:273–282. 10.1097/rli.0000000000000932

82. Sachpekidis C, Enqvist O, Ulén J, Kopp-Schneider A, Pan L, Mai EK et al (2024) Artificial intelligence–based, volumetric assessment of the bone marrow metabolic activity in [18F]FDG PET/CT predicts survival in multiple myeloma. Eur J Nucl Med Mol Imaging 51:2293–2307. 10.1007/s00259-024-06668-z

83. Rajkumar SV (2024) Multiple myeloma: 2024 update on diagnosis, risk-stratification, and management. Am J Hematol 99:1802–1824. 10.1002/ajh.27422

84. Hao W, Zheng F, Gou X, Zhang D, Yin P, Zhang CC, Hong N (2026) Prediction of ISS and R-ISS stratification in newly diagnosed multiple myeloma using lumbar spine MRI radiomics model: A two-center multimodal study. J Imaging Inf Med 1–13. 10.1007/s10278-025-01833-8

85. Mosquera Orgueira A, Gonzalez Perez MS, D’Agostino M, Cairns DA, Larocca A, Palacios JJL et al (2025) Machine learning risk stratification strategy for multiple myeloma: insights from the EMN-HARMONY alliance platform. HemaSphere 9:e70228. 10.1002/hem3.70228

86. Bonello F, Cani L, D’Agostino M (2022) Risk stratification before and during treatment in newly diagnosed multiple myeloma: from clinical trials to the real-world setting. Front Oncol 12:830922. 10.3389/fonc.2022.830922

87. Murie C, Turkarslan S, Patel AP, Coffey DG, Becker PS, Baliga NS (2025) Individualized dynamic risk assessment and treatment selection for multiple myeloma. Br J Cancer 132:922–936. 10.1038/s41416-025-02987-6

88. Ye Z, Lv J, Wu X, Li H, Niu X, Han X et al (2025) <article-title update="added">AI-quantified ¹¹C-MET PET/CT bone marrow metabolic activity for prognostic assessment in newly diagnosed multiple myeloma. Eur J Nucl Med Mol Imaging. 10.1007/s00259-025-07594-4

89. Baffour FI, Huber NR, Ferrero A, Rajendran K, Glazebrook KN, Larson NB et al (2023) Photon-counting detector CT with deep learning noise reduction to detect multiple myeloma. Radiology 306:229–236. 10.1148/radiol.220311

90. Van Den Berghe T, Verberckmoes B, Kint N, Wallaert S, De Vos N, Algoet C et al (2024) Predicting cytogenetic risk in multiple myeloma using conventional whole-body MRI, spinal dynamic contrast-enhanced MRI, and spinal diffusion-weighted imaging. Insights Imaging 15:106. 10.1186/s13244-024-01672-1

91. Ren L, Xu B, Xu J, Li J, Jiang J, Ren Y et al (2023) A machine learning model to predict survival and therapeutic responses in multiple myeloma. Int J Mol Sci 24:6683. 10.3390/ijms24076683

92. Mittelman M (2025) Introduction to the review series on myelodysplasia. Haematologica 110:281–282. 10.3324/haematol.2023.284948

93. Raddi MG, Bencini S, Peruzzi B, Mattiuz G, Pourcq SD, Tanturli M et al (2024) Flow cytometric analysis of erythroid precursors and mutational signatures of lower risk myelodysplastic syndromes identify responders to erythroid stimulating agents. Blood Cancer J 14:127. 10.1038/s41408-024-01112-9

94. Oster HS, Van de Loosdrecht AA, Mittelman M (2025) Diagnosis of myelodysplastic syndromes: the classic and the novel. Haematologica 110:300–311. 10.3324/haematol.2023.284937

95. Elshoeibi AM, Badr A, Elsayed B, Metwally O, Elshoeibi R, Elhadary MR et al (2023) Integrating AI and ML in myelodysplastic syndrome diagnosis: state-of-the-art and future prospects. Cancers 16:65. 10.3390/cancers16010065

96. Lee N, Jeong S, Park M-J, Song W (2022) Deep learning application of the discrimination of bone marrow aspiration cells in patients with myelodysplastic syndromes. Sci Rep 12:18677. 10.1038/s41598-022-21887-w

97. Zhu J, Lemaire P, Mathis S, Ronez E, Clauser S, Jondeau K et al (2022) Machine learning-based improvement of MDS-CBC score brings platelets into the limelight to optimize smear review in the hematology laboratory. BMC Cancer 22:972. 10.1186/s12885-022-10059-8

98. Clichet V, Lebon D, Chapuis N, Zhu J, Bardet V, Marolleau J-P et al (2023) Artificial intelligence to empower diagnosis of myelodysplastic syndromes by multiparametric flow cytometry. Haematologica. 10.3324/haematol.2022.282370

99. Bernard E, Tuechler H, Greenberg PL, Hasserjian RP, Arango Ossa JE, Nannya Y et al (2022) Molecular international prognostic scoring system for myelodysplastic syndromes. NEJM Evid 1:EVIDoa2200008. 10.1056/EVIDoa2200008

100. Mosquera Orgueira A, Perez Encinas MM, Diaz Varela NA, Mora E, Díaz-Beyá M, Montoro MJ et al (2023) Machine learning improves risk stratification in myelodysplastic neoplasms: an analysis of the Spanish Group of Myelodysplastic Syndromes. Hemasphere 7:e961. 10.1097/hs9.0000000000000961

101. Bobak J, Spohr P, Richter S, Streuer A, Schulz FI, Strupp C et al (2025) Dynamic Mortality Risk Prediction in Myelodysplastic Syndromes Using Longitudinal Clinical Data. JCO Clin Cancer Inf 9:e2500236. 10.1200/CCI-25-0023

102. Radakovich N, Sallman DA, Buckstein R, Brunner A, Dezern A, Mukerjee S et al (2022) A machine learning model of response to hypomethylating agents in myelodysplastic syndromes. Iscience 25:104931. 10.1016/j.isci.2022.104931

103. Hellström-Lindberg ES, Kröger N (2023) Clinical decision-making and treatment of myelodysplastic syndromes. Blood 142:2268–2281. 10.1182/blood.2023020079

104. Shimomura Y, Komukai S, Kitamura T, Sobue T, Kurosawa S, Doki N et al (2023) Identifying the optimal conditioning intensity for stem cell transplantation in patients with myelodysplastic syndrome: a machine learning analysis. Bone Marrow Transpl 58:186–194. 10.1038/s41409-022-01871-8

105. Romano NH, Ruiz C, Schlaepfer P, Balabanov S, Habringer S, Widmer CC (2025) AI-based detection of neutrophil dysplasia: an accessible and sensitive model for MDS diagnosis from peripheral blood. Ann Hematol 104:4477–4486. 10.1007/s00277-025-06533-5

106. De Almeida JG, Gudgin E, Besser M, Dunn WG, Cooper J, Haferlach T et al (2023) Computational analysis of peripheral blood smears detects disease-associated cytomorphologies. Nat Commun 14:4378. 10.1038/s41467-023-39676-y

107. Alaggio R, Amador C, Anagnostopoulos I, Attygalle AD, Araujo IBDO, Berti E et al (2022) The 5th edition of the World Health Organization classification of haematolymphoid tumours: lymphoid neoplasms. Leukemia 36:1720–1748. 10.1038/s41375-022-01620-2

108. Khoury JD, Solary E, Abla O, Akkari Y, Alaggio R, Apperley JF et al (2022) The 5th edition of the World Health Organization Classification of Haematolymphoid Tumours: Myeloid and Histiocytic/Dendritic Neoplasms. Leukemia 36:1703–1719. 10.1038/s41375-022-01613-1

109. Heuser M, Freeman SD, Ossenkoppele GJ, Buccisano F, Hourigan CS, Ngai LL et al (2021) 2021 update on MRD in acute myeloid leukemia: a consensus document from the European LeukemiaNet MRD Working Party. Blood 138(26):2753–2767. 10.1182/blood.2021013626

110. Chandhok NS, Sekeres MA (2025) Measurable residual disease in hematologic malignancies: a biomarker in search of a standard. EClinicalMedicine. 10.1016/j.eclinm.2025.103348

111. CRDC/GDC. https://datacommons.cancer.gov/repository/genomic-data-commons . Accessed 10 Mar 2026

112. MMRF CoMMpass. https://themmrf.org/finding-a-cure/personalized-treatment-approaches/s . Accessed 10 Mar 2026

113. TARGET. https://dbgap.ncbi.nlm.nih.gov/home/ . Accessed 10 Mar 2026

114. Beat AML. https://bloodcancerunited.org/research/beat-aml-bringing-precision-medicine-patients-clinical-trial . Accessed 10 Mar 2026

115. St J, Cloud. https://www.stjude.cloud/ . Accessed 10 Mar 2026

116. FlowRepository. http://flowrepository.org/ . Accessed 10 Mar 2026

117. The European Genome-phenome Archive (EGA). https://ega-archive.org/ . Accessed 10 Mar 2026

118. Scientific Data. https://www.nature.com/sdata/ . Accessed 10 Mar 2026

119. Zenodo. https://zenodo.org/ . Accessed 10 Mar 2026

120. Beauchamp NJ, Bryan RN, Bui MM, Krestin GP, McGinty GB, Meltzer CC et al (2023) Integrative diagnostics: the time is now—a report from the International Society for Strategic Studies in Radiology. J Am Coll Radiol 20:455–466. 10.1016/j.jacr.2022.11.015

121. Yang H, Yang M, Chen J, Yao G, Zou Q, Jia L (2024) Multimodal deep learning approaches for precision oncology: a comprehensive review. Brief Bioinform 26:bbae699. 10.1093/bib/bbae699

122. Plass M, Kargl M, Kiehl T, Regitnig P, Geißler C, Evans T et al (2023) Explainability and causability in digital pathology. J Pathol Clin Res 9:251–260. 10.1002/cjp2.322

123. Ning Y, Teixayavong S, Shang Y, Savulescu J, Nagaraj V, Miao D et al (2024) Generative artificial intelligence and ethical considerations in health care: a scoping review and ethics checklist. Lancet Digit Health 6:e848–e856. 10.1016/S2589-7500(24)00143-2

124. Mello MM, Guha N (2024) Understanding liability risk from using health care artificial intelligence tools. N Engl J Med 390:271–278. 10.1056/NEJMhle2308901

125. Huang S-C, Pareek A, Jensen M, Lungren MP, Yeung S, Chaudhari AS (2023) Self-supervised learning for medical image classification: a systematic review and implementation guidelines. Npj Digit Med 6:74. 10.1038/s41746-023-00811-0

126. Krones F, Marikkar U, Parsons G, Szmul A, Mahdi A (2025) Review of multimodal machine learning approaches in healthcare. Inf Fusion 114:102690. 10.1016/j.inffus.2024.102690

127. Lu J, Sattler A, Wang S, Khaki AR, Callahan A, Fleming S et al (2022) Considerations in the reliability and fairness audits of predictive models for advance care planning. Front Digit Health 4:943768. 10.3389/fdgth.2022.943768

128. Ziller A, Mueller TT, Stieger S, Feiner LF, Brandt J, Braren R et al (2024) Reconciling privacy and accuracy in AI for medical imaging. Nat Mach Intell 6:764–774. 10.1038/s42256-024-00858-y

129. Park HJ (2024) Patient perspectives on informed consent for medical AI: a web-based experiment. Digit Health 10:20552076241247938. 10.1177/20552076241247938

130. Kraljevic Z, Bean D, Shek A, Bendayan R, Hemingway H, Yeung JA et al (2024) Foresight-A generative pretrained transformer for modelling of patient timelines using electronic health records: A retrospective modelling study. Lancet Digit Health 6:e281–e290. 10.1016/S2589-7500(24)00025-6

Associated Data

Data Availability Statement

No datasets were generated or analysed during the current study.