INTRODUCTION
Risk stratification in malignant haematology aims to guide haematologists in making treatment decisions weighing individual disease biology and patient‐specific factors. Depending on the disease, there is a multitude of models that are currently considered the gold standard in clinical routine such as the European LeukemiaNet (ELN) 1 model in acute myeloid leukaemia (AML), the Revised International Prognostic Scoring System (IPSS‐R) 2 in myelodysplastic neoplasms (MDS) or the International Prognostic Index (IPI) 3 in diffuse large B‐cell lymphoma (DLBCL). These systems have proven their worth over decades. Their basis often is routine clinical or genetic data, guided by conventional statistical modelling and expert consensus. In clinical routine, they provide the means for pragmatic decision support as they essentially function as checklists or simple decision trees that can be memorized without assigning complex weights to their individual variables. Yet, haematological malignancies are inherently heterogeneous and often characterized by multiple concurrent aberrations. The current static approaches are likely unable to fully appreciate the complex interplay in disease biology. Given the increasing access to next‐generation sequencing or whole‐exome/genome and transcriptomic profiling, the dimensionality of patient‐specific data is going to rapidly outpace the interpretive capacity of conventional rule‐based systems. This introduces an urgent need for more flexible, data‐driven models capable of handling high‐dimensional datasets to integrate disease biology and patient‐specific factors.
Machine learning (ML) may address this challenge better than static rule‐based approaches. While initial applications of ML in haematology emerged over two decades ago, 4 , 5 , 6 , 7 early efforts were constrained by limited data integration and computational resources. In recent years, advances in algorithms, computational power, and access to large, multimodal datasets have brought AI‐guided clinical decision support systems within reach. 8 , 9 These systems can be trained to identify complex patterns and interactions in heterogeneous datasets and offer outcome predictions and risk stratification that may surpass traditional models both in accuracy and scale.
Importantly, these systems are not intended to replace clinical expertise but to augment it by supporting clinicians in evaluating vast and multifaceted patient data and informing personalized evidence‐based care. However, their implementation also introduces new challenges including model transparency, bias mitigation, validation, regulation and clinical integration. An understanding of the capabilities and limitations of AI‐driven decision support tools is essential for both clinicians and researchers in haematology. Ensuring that these technologies enhance rather than complicate patient care will depend on rigorous assessment, safe clinical integration and continued interdisciplinary dialogue between haematology and computational sciences.
In this review, we aim to equip haematologists with an understanding of both capabilities and limitations of ML models for risk prediction, highlight studies using ML‐based risk prediction across malignant haematology and discuss emerging trends and potential pitfalls for the clinical integration of these tools in the upcoming years.
A BRIEF INTRODUCTION TO MACHINE LEARNING FOR HAEMATOLOGISTS
The terms ‘artificial intelligence’ and ‘machine learning’ have become ubiquitous in healthcare. Despite widespread exposure, understanding of AI and ML varies significantly. A survey 10 showed that 21.9% of self‐reported ML‐literate respondents misidentified its definition, likely owing to the overlapping use of AI terminology in medicine.
AI broadly refers to digital problem‐solving techniques that would otherwise require human intelligence. 11 ML, a subset of AI, is focused on developing algorithms that learn from data to make predictions or uncover patterns, 11 setting it apart from traditional software, which operates based on explicit instructions. These models scale with their training data, which make them particularly suitable for handling the high‐dimensional datasets encountered in haematology, including molecular profiles, clinical variables and laboratory parameters. Within ML, deep learning (DL) utilizes neural networks that mimic information processing in the neocortex via interconnected artificial neurons (also termed perceptrons ). 12 Much like neural plasticity, the weights of the connections between these neurons are calibrated in the training stage, enabling neural networks to process large amounts of data such as images or text. 12 Convolutional neural networks (CNNs) were initially considered state of the art in image processing 13 ; however, the introduction of the transformer architecture (incorporating multi‐head attention) 14 revolutionized image and text processing, enabling the development of large language models and foundation models with increasing multimodal capabilities such as text, image or speech processing.
At the core of ML is model inference through methods like supervised learning, unsupervised learning and reinforcement learning. 11 Supervised learning, which pairs input data (e.g. genetic profile, differential count, pharmacogenomics data) with corresponding output labels (e.g. disease subtype, treatment outcome), is particularly suited to risk prediction tasks in haematology. It allows models to forecast specific outcomes based on patient data. Unsupervised learning, which does not rely on labelled outcomes, is invaluable for discovering new insights into disease mechanisms or patient stratification. In unsupervised learning, data are sorted by similarity or dissimilarity, allowing for cluster formation of patients with biologically related subgroups of a given entity. Reinforcement learning (RL) is an iterative method that links decision‐making to changes in outcomes or states of a simulated environment, thereby aiming at developing an optimal strategy to achieve a predefined goal. Since RL explicitly requires ML models to alter their environment, that is, make actual treatment decisions in the medical context, it is currently not used for direct risk prediction but can optimize treatment strategies or clinical pathways based on patient outcomes, for instance, in simulated environments. 15 The nature of rare data in medicine often requires pooling and harmonizing datasets from different institutions. If sharing data are not feasible, techniques like federated learning 16 where the model is shared between collaborators or swarm learning 17 where only model weights (learnable parameters) are shared offer alternatives that enable collaborative model development without sharing data.
Evaluating the performance of risk prediction models is crucial, particularly in haematology, where outcomes can significantly impact patient well‐being and survival. Sensitivity measures the proportion of actual high‐risk patients correctly identified, while specificity measures the proportion of truly low‐risk patients (or non‐events) correctly classified. High specificity reduces the number of false alarms, which can be important to avoid unnecessary treatment or interventions. The area under the receiver operating characteristic (AUC‐ROC) is particularly valuable for risk prediction models because it provides a comprehensive measure of the model's ability to distinguish between different outcomes (e.g. high risk vs. low risk) across all possible thresholds. A higher AUC indicates better model performance, with a score of 1 representing a perfect model and a score of 0.5 indicating no discriminative ability. Datasets in haematology are often imbalanced, with certain genetic or phenotypic characteristics being more common than others or certain events may occur rarely. In these cases, AUC‐ROC may be misleading and the precision–recall curve and its area under the curve (AUC‐PR) become important. These metrics focus on the model's ability to retrieve relevant instances (recall) ensuring that among patients predicted as high risk, a high proportion truly is high risk (precision). F1 score balances precision and recall as a harmonic mean. It is useful especially in settings where both missing high‐risk patients and overpredicting risk carry consequences. The concordance‐index (c‐index) evaluates the performance of survival models and any predictive model where the outcome is time until an event occurs. It provides an effective measure of a model's discriminative capacity in scenarios where the timing of events is critical. Beyond the discrimination of outcomes, clinical risk prediction models must also demonstrate good calibration, that is, the agreement between predicted and observed event probabilities. 18 A well‐calibrated model ensures that, for example, patients predicted to have a 20% probability of relapse experience relapse in approximately 20% of cases. Poor calibration can result in over‐ or underestimation of absolute risk and may therefore misguide clinical decisions even when discrimination performance is high. 19 Calibration can be evaluated, for example, using calibration plots or the Brier score 20 and can be improved via, for example, isotonic or Platt scaling. 21 Ideally, both discrimination performance as well as calibration should be reported for models intended to guide clinical decisions. In practice, no single metric suffices; however, combining several complementary metrics usually provides the most reliable picture of a model's performance.
A common pitfall of ML prediction models is overfitting, which describes the notion that a model does adequately learn patterns during training but rather memorizes the training data. Upon evaluating new previously unseen data, model performance subsequently plummets, and the model is not transferable or generalizable to other data within the scope it was trained for. Therefore, overfitting can be assumed if a model performs well on training data but performance drops in internal or external validation sets. Large, diverse datasets, independent validation sets, techniques such as cross‐validation, dropout, regularization, dimensionality reduction (i.e. reducing the number of variables relative to sample size) or alterations to model architecture may help mitigate or detect overfitting. 22 Table 1 summarizes key AI terminology.
TABLE 1
Key terminology in artificial intelligence.
| Key term | Description |
| Artificial intelligence (AI) | Broad field of computer science aimed at creating systems capable of performing tasks that normally require human intelligence, such as pattern recognition, forecasting or decision‐making. |
| Machine learning (ML) | Subfield of AI where algorithms learn from data to make predictions or decisions without being explicitly programmed. |
| Deep learning (DL) | A form of ML using multilayer neural networks that can learn complex patterns from large datasets such as genomics or imaging data. |
| Supervised learning | Category of ML where a model is trained on labelled data (inputs with known outcomes) to predict outcomes in new cases. |
| Unsupervised learning | Category of ML where a model identifies structures or clusters in unlabelled data, revealing subgroups or new biological patterns. |
| Reinforcement learning (RL) | Category of ML where a model learns through trial‐and‐error interaction with a simulated environment to optimize decisions; may be used to test treatment strategies. |
| Neural network (NN) | Computational architecture inspired by the structure of the neocortex, composed of layers of interconnected artificial neurons (alternatively perceptrons ) with learnable weights (connections between the neurons). |
| Convolutional neural network (CNN) | Type of neural network excelling in image analysis (e.g. blood or bone marrow microscopy). |
| Transformer | Neural network architecture is designed to process sequential data (text, genomic sequences) using attention mechanisms rather than recurrence, being the foundation of modern large language models. |
| Large language model (LLM) | Transformer‐based model trained on massive text datasets to understand and generate human language. LLMs can be adapted to biomedical texts for literature synthesis or report generation. |
| Foundation model | Very large model trained on diverse, multimodal datasets (text, images, structured data) that can be fine‐tuned for a wide range of downstream tasks, including clinical prediction, image analysis or report summarization. |
| Overfitting | Instead of learning patterns from the data, the model essentially memorizes the training data, leading to high performance on training sets, but inadequate performance on (external) test data, indicating low generalizability. |
| Regularization | Mathematical strategies (e.g. L1/L2 penalties, dropout) to prevent overfitting by constraining model complexity. |
| Calibration | Agreement between predicted probabilities and actual outcomes which is essential for trustworthy clinical risk prediction. |
| Federated learning | Decentralized technique where models are trained across institutions without sharing raw patient data, preserving privacy. |
| Swarm learning | Decentralized technique where model weights are shared across institutions without sharing raw patient data or the model itself, preserving privacy. |
| Explainability (XAI) | Techniques that make model reasoning interpretable (e.g. SHAP values, attention maps), enabling clinicians to understand which features drive predictions. |
RISK PREDICTION WITH MACHINE LEARNING IN MYELOID NEOPLASMS
Myelodysplastic neoplasms
Early diagnosis of myelodysplastic neoplasms (MDS) enables physicians to decide about any optional treatment sooner, potentially delaying the progression of MDS and enhancing the patient's quality of life. Radhachandran et al. 23 developed an ML algorithm using 2 years of patient data from electronic health records (EHR) to predict MDS 1 year before diagnosis, achieving an AUROC of 0.87. The model identified the patient's age, haematocrit, red blood cell count and platelet count as the most significant predictors. Clichet et al. 24 also focused on predicting the risk of MDS but utilized multiparametric flow cytometry data instead. Their elasticnet algorithm, trained on data from 191 patients and externally validated on 89 patients, demonstrated an AUROC of 0.935, also distinguishing between high‐ and low‐risk MDS. Addressing MDS risk classification, Nazha et al. 25 introduced an unsupervised personalized prediction model that incorporates genetic alterations, laboratory results, WHO 2016 subtypes and demographic data. This model estimates the survival and the probability of leukaemia transformation. More recently, Mosquera‐Orgueira et al. 26 developed the Artificial Intelligence Prognostic Scoring System for MDS (AIPSS‐MDS), which relies on traditional clinical, haematological and cytogenetic variables. ML models have also shown promise in predicting treatment responses based on various genetic and clinical factors. Nazha et al. 27 applied Apriori market basket analysis to identify genomic biomarkers predictive of resistance to hypomethylating agents (HMA) in MDS patients. Expanding on this, Radakovich et al. 28 created an ML model that assesses the response to HMA by analysing serial blood counts during the first 90 days of treatment. After 3 months, this model could predict HMA response with an AUROC of 0.84 in an independent validation cohort. Similarly, Sharplin et al. 29 developed a random forest model to forecast early treatment failure in patients receiving azacitidine. Table 2 summarizes studies investigating ML for risk prediction in MDS.
TABLE 2
Studies investigating ML‐based outcome prediction in myelodysplastic neoplasms (MDS).
Abbreviations: AML, acute myeloid leukaemia; AUROC, area under the receiver operating characteristic; CMML, chronic myelomonocytic leukaemia; HMA, hypomethylating agents; LFS, leukaemia‐free survival; OS, overall survival; RF, random forest; RSF, random survival forest; XGB, extreme gradient boosting.
| References | Sample(s) | ML model(s) | Output | Validation |
| Radhachandran et al. 23 | MDS ( n = 790 470) | XGB | MDS risk prediction 1 year prior to diagnosis, AUROC 0.87 | Internal, hold‐out |
| Clichet et al. 24 | MDS ( n = 191) | Elasticnet | Risk of MDS | External ( n = 89) |
| Nazha et al. 25 | MDS ( n = 1471) | RSF | OS and risk of leukaemia transformation | External ( n = 465) |
| Mosquera‐Orgueira et al. 26 | MDS ( n = 7202) | RSF | OS and LFS | External ( n = 2957) |
| Nazha et al. 27 | MDS with HMA treatment ( n = 433) | Apriori market basket analysis | Risk of HMA resistance | External ( n = 113) |
| Radakovich et al. 28 | MDS ( n = 424) | Random Forest, gradient‐boosted decision trees | HMA response | External ( n = 90) |
| Sharplin et al. 29 | MDS, AML, CMML treated with azacitidine ( n = 273) | RSF | Risk of treatment failure, survival outcome | Internal, cross‐validation |
Myeloproliferative neoplasms
Myeloproliferative neoplasms (MPN) are phenotypically very heterogeneous, with an inherent risk of disease progression. Chronic myeloid leukaemia (CML) can be effectively treated with tyrosine kinase inhibitors (TKI). Nevertheless, TKI response and outcome prediction are paramount. Shanbehzadeh et al. 30 evaluated multiple ML methods and identified SVM as the most accurate method for predicting 5‐year survival in CML patients, based on 12 clinical features. In myelofibrosis, Mosquera‐Orgueira et al. 31 developed an Artificial Intelligence Prognostic Scoring System for Myelofibrosis (AIPSS‐MF) model, based on eight clinical variables at diagnosis, which achieved high accuracy in predicting overall survival (OS) and leukaemia‐free survival (LFS) for myelofibrosis patients. The grading of fibrosis is an important component of disease classification, prognostication and monitoring. Ryou et al. 32 developed an ML method to objectively assess bone marrow reticulin fibrosis, creating a Continuous Index of Fibrosis (CIF) using manually labelled fibrosis areas for training. This approach, which includes a quantitative fibrosis map and megakaryocyte analysis, accurately differentiates between ET and pre‐fibrotic MF (AUC = 0.94) and identifies MPN patients at risk of disease progression. With the incorporation of effective therapies for MF, accurately predicting outcomes after allogeneic haematopoietic cell transplantation (allo‐HCT) is crucial for determining the optimal timing for this procedure. Hernández‐Boluda et al. 33 created a random survival forest (RSF) model to identify patients at high risk for poor transplantation outcomes. Table 3 summarizes studies investigating ML for risk prediction in MPN.
TABLE 3
Studies investigating ML‐based outcome prediction in myeloproliferative neoplasms (MPN).
Abbreviations: CML, chronic myeloid leukaemia; CNN, convolutional neural network; LFS, leukaemia‐free survival; MF, myelofibrosis; OS, overall survival; RF, random forest; RSF, random survival forest; SVM, support vector machine.
| References | Sample(s) | ML model(s) | Output | Validation |
| Shanbehzadeh et al. 30 | CML ( n = 837) | SVM | CML 5‐year survival, accuracy 85.7% | Internal cross‐validation |
| Mosquera‐Orgueira et al. 31 | MF ( n = 1386) | RF | OS and LFS | Internal cross‐validation |
| Ryou et al. 32 | MPN, reactive marrow ( n = 107) | Ranking‐CNN, RF | MPN patients at risk of progression | ‐ |
| Hernández‐Boluda et al. 33 | MF ( n = 5183) | RSF | OS after transplant | ‐ |
Acute myeloid leukaemia
Acute myeloid leukaemia (AML) is a severe haematological malignancy characterized by its heterogeneity in clinical presentation, genetic landscape and prognosis. This variability presents a significant challenge in predicting patient outcomes, with some subgroups showing potential for cure while others face poor prognoses under current treatment standards. Shaik et al. 34 developed an ML‐based genomic‐driven prognostication model for AML patients with the RUNX1::RUNX1T1 fusion. This model demonstrated a strong correlation with measurable residual disease (MRD) and clinical outcomes, highlighting the potential of ML to uncover prognostic markers that are clinically relevant. Furthermore, the same group 35 introduced a supervised ML algorithm aimed at prognosticating AML with mutated NPM1 based on genetic profiles. Despite NPM1 mutations defining a distinct AML subgroup, these mutations do not always serve as the initiating event, underscoring the complexity of AML's genetic underpinnings. The concept of age‐related clonal haematopoiesis (ARCH) as a precursor to myeloid malignancies, including in the NPM1 AML subgroup, adds another layer of complexity to prognostication. Patkar et al. 35 employed a computational approach to decipher the clinical significance of numerous genetic variables. They identified that the type of NPM1 mutation, the corrected NPM1 variant allele fraction (VAF), the presence of DNMT3A R882 mutation, FLT3 internal tandem duplication VAF and IDH2 mutations were all clinically relevant. From these ML‐derived variables, a scoring system was developed, capable of classifying AML with mutated NPM1 into three prognostic classes with significantly different outcomes.
Going in a similar direction, Eckardt et al. 36 explored how age influences the prognostic significance of genetic mutations in AML. They developed ML models that effectively predicted complete remission and 2‐year overall survival, with AUC scores of 0.801 and 0.791 respectively. By employing Shapley (SHAP) values, the study quantified each variable's contribution across different age groups, revealing significant variability in the prognostic value of genetic alterations with age. This finding challenges the uniform application of risk stratification models and underscores the necessity for age‐adjusted treatment strategies in AML. Eckardt et al. 37 also utilized nine ML models to predict complete remission and 2‐year overall survival in intensively treated AML patients, achieving high accuracies for both internal testing and external validation. The achievement of complete remission is a crucial milestone in AML therapy, as refractory disease is associated with dismal outcomes. Thus, accurately identifying patients at risk is essential for tailoring treatment concepts to individual disease biology. Using unsupervised learning, the same group 38 identified four distinct biological clusters in AML re‐stratifying patients regarding ELN2017 risk assignment. Clusters differed significantly regarding both molecular alterations present per cluster and outcomes and were validated in an external cohort. Gal et al. 39 applied three different ML algorithms using pretreatment gene expression data to predict complete remission in intensively treated AML patients. Pathway enrichment analysis for the most significantly deregulated genes indicated the guanosine diphosphate fucose biosynthesis pathway as the most significant, suggesting a crucial role for N‐glycosylation in AML. Wagner et al. 40 identified a three‐gene signature ( CALCRL, CD109, LSP1 ) that distinctively predicted event‐free survival (EFS) and overall survival (OS) in AML patients. This signature, validated in public AML cohorts comprising 1050 patients (adult and paediatric), was found to be uniquely predictive of survival in AML among 39 distinct malignancies, further establishing ML's potential in identifying novel prognostic markers. In a smaller study, Karami et al. 41 utilized six different ML algorithms to predict OS in AML patients, identifying the gradient boosted tree as the most reliable method. Most recently, Jia et al. 42 employed 11 ML models to predict survival status at 1, 2 and 3 years after AML diagnosis based on epidemiological characteristics in an overarching study that included 87 090 patients, showcasing ML's scalability and adaptability in handling large datasets for survival prediction. Table 4 summarizes studies investigating ML for risk prediction in AML.
TABLE 4
Studies investigating ML‐based outcome prediction in acute myeloid leukaemia (AML).
Abbreviations: ANN, artificial neural net; DT, decision tree; GBT, gradient boosting tree; KNN, k nearest neighbour; LDA, linear discriminant analysis; LR, logistic regression; NN, neural network; pSVM, polynomial support vector machine; QDA, quadratic discriminant analysis; RBF‐SVM, radial basis kernel function support vector machine; RF, random forest; SVM, support vector machine; W‐Bayes Net, weighted Bayesian Network; XGB, extreme gradient boosting.
| References | Sample(s) | ML model(s) | Output | Validation |
| Shaik et al. 34 | AML RUNX1::RUNX1T1 ( n = 131) | LR | Two risk classes with divergent OS & RSF | Internal, validation cohort |
| Patkar et al. 35 | AML NPM1 ( n = 110) | LR | Three AML NPM1 risk groups with divergent OS | ‐ |
| Eckardt et al. 36 | AML ( n = 3062) | RF, LR, XGB | Likelihood of CR, 2‐year OS, age‐stratified prognostic variables | Internal |
| Eckardt et al. 37 | AML ( n = 1383) | SVM, RF, LR, RBF‐SVM, ANN, adaptive boosting, pSVM, gradient boosting, KNN | Likelihood of CR, 2‐year OS | External ( n = 664) |
| Eckardt et al. 38 | AML ( n = 1383) | Meta‐clustering of 11 different clustering algorithms | Risk assignment based on molecular and cytogenetics, comparison to ELN 2017 | External ( n = 664) |
| Gal et al. 39 | AML ( n = 473) | KNN, RF, SVM | Likelihood of CR | Internal, cross‐validation |
| Wagner et al. 40 | AML ( n = 593) | ANN | EFS, OS, 3‐gene expression signature | External, public cohorts ( n = 1050) |
| Karami et al. 41 | AML ( n = 249) | RF, DT, LR, Naive Bayes, W‐Bayes Net, GBT | OS | Internal, cross‐validation |
| Jia et al. 42 | AML ( n = 87 090) | XGB, glmnet, KNN, LR, Naive Bayes, NN, RF, SVM, LDA, QDA, DT | 1‐,2‐,3‐year OS | Internal, cross‐validation |
RISK PREDICTION WITH MACHINE LEARNING IN LYMPHOID NEOPLASMS
Diffuse large B‐cell lymphoma
In diffuse large B‐cell lymphoma (DLBCL), multiple studies investigated divergent biological subgroups beyond the conventional cell‐of‐origin classification (COO). 43 Early studies by Shipp et al. 44 delineated follicular lymphoma (FL) and DLBCL and predicted response to CHOP‐based regimens using a supervised approach and identified two groups with distinct 5‐year OS rates. In comprehensive genomic analyses, both Chapuy et al. 45 and Schmitz et al. 46 identified five and four (partially overlapping) genetic driver subgroups with impact on 5‐year OS rates, including favourable risk ABC and poor‐risk GCB subtypes, thereby highlighting the need for genetic stratification beyond COO. Zhuang et al. 47 used unsupervised clustering to identify two subgroups with divergent survival (immune‐enriched and cell cycle‐enriched) in a cohort of 414 DLBCL patients using comprehensive gene expression data and subsequently constructed a classifier that sorted patients into their proposed subgroups using only seven divergent genes. Similarly, He et al. 48 used consensus clustering on gene expression data and identified two clusters differing by their expression of programmed cell death‐related genes, identifying SCARB2 , GBP1 and STOM as highly relevant genes across models predicting expression patterns. Comparing a stacked model‐based outcome prediction with routine scores, Biccler et al. 49 used data from the Danish and Swedish lymphoma registries (including clinicopathological data, treatments and outcomes) and reported superior performance of their model (validation set c‐index 0.74) outperforming both IPI (c‐index 0.66) and NCCN‐IPI (c‐index 0.68) in predicting patient survival. Mosqueira Orgueira et al. 50 used identified two clusters delineated by their differential expression of TNFRSF9, BIRC3, BCL2L1 and G3BP2 . Subsequently, they developed a random survival forest‐based classifier including clinical information, COO classification, their proposed four‐gene expression cluster assignment and gene expression data, reporting a test set c‐index of 0.79 for survival prediction. Furthermore, using demographic and clinical information, Detrait et al. 51 evaluated five ML classifiers in predicting primary refractory disease, among which naïve Bayes (NB) achieved the highest classification performance with an F1‐score of 0.82. Features with the highest importance for model decisions were IPI score, age > 80 years, high body mass index, time to treatment and PET‐CT scan performed after two cycles of treatment. As comprehensive genomic analyses are often not available in clinical routine, several studies investigated COO prediction in DLBCL with a reduced set of input variables. Zhao et al. 52 identified eight genes ( MYBL1, LMO2, BCL6, MME, IRF4, NFKBIZ, PDE4B and SLA ) to predict either GCB or non‐GCB subtypes with 91.0–94.4% validation accuracy. Similarly, Viswanathan et al. 53 used a multilayer perceptron‐based model to predict COO using gene expression data of 20 genes, achieving a test set and external validation accuracy of 94.7% and 95.5%, respectively, demonstrating close alignment between survival predictions of their classifier and conventional COO subtyping. Table 5 summarizes studies investigating ML for risk prediction in DLBCL.
TABLE 5
Studies investigating ML‐based outcome prediction in diffuse large B‐cell lymphoma (DLBCL).
Abbreviations: AB, adaptive boosting; DT, decision tree; GBM, gradient boosting machine; KNN, k nearest neighbour; LDA, linear discriminant analysis; LR, logistic regression; MLP, multi‐layer perceptron; NB, naïve Bayes; NN, nearest neighbour; RF, random forest; RSF, random survival forest; SVM, support vector machine; XGB, extreme gradient boosting.
| References | Sample(s) | ML model(s) | Output | Validation |
| Shipp et al. 44 | DLBCL ( n = 58) FL ( n = 19) | Weighted voting (supervised) | Identified two patient groups with 70% vs. 12% 5‐y OS | Internal cross‐validation |
| Chapuy et al. 45 | DLBCL ( n = 304) | Non‐negative matrix factorization consensus clustering | Identified 5 clusters with driver genetics and divergent outcomes | ‐ |
| Schmitz et al. 46 | DLBCL ( n = 574) | RF, SVM, GenClass | Identified 4 with driver genetics and divergent outcomes | Internal cross‐validation |
| Zhuang et al. 47 | DLBCL ( n = 414 training, n = 729 validation) | Consensus clustering, NB, DT, RF AB, XGB, SVM, Fisher discriminant analysis | Identified immune‐enriched and cell cycle‐enriched subtypes | External validation sets |
| He et al. 48 | DLBCL ( n = 95 training, n = 72 validation) | Consensus clustering, SVM, Lasso, GBM, ridge regression, KNN, LR, LDA, DT, XGB, NB, elastic net regression, stepwise regression | Identified two subgroups based on programmed cell death‐related genes | External validation sets |
| Biccler et al. 49 | DLBCL ( n = 2759 training, n = 2414 validation) | Stacked model | Developed a stacked model outperforming IPI and NCCN‐IPI in survival prediction | External validation set |
| Mosqueira Orgueira et al. 50 | DLBCL ( n = 233 training, n = 64 validation) | mclust, RSF | Identified two clusters based on a 4‐gene signature with divergent survival rates | External validation set |
| Detrait et al. 51 | DLBCL ( n = 130) | SVM, RF, LR, NB, XGB | F1‐score of 0.82 in predicting primary refractory patients | Internal cross‐validation |
| Zhao et al. 52 | DLBCL ( n = 414 training, n = 855 validation) | DT, RF, SVM, NN, AB, Fisher discriminant analysis, Bagging | Identified 8 genes to predict COO with 91.0–94.4% accuracy | External validation sets |
| Viswanathan et al. 53 | DLBCL ( n = 381) | MLP | Identified 20 genes to predict COO with 94.7%–95.9% accuracy | External validation set |
Follicular lymphoma and mantle cell lymphoma
Zha et al. 54 developed an extreme gradient boosting (XGB) model (named FLIPI‐C) to predict POD24 using clinical and laboratory data of 1938 FL patients, achieving an AUC of 0.70 in an external validation set, reportedly outperforming both FLIPI and FLIPI2. The most predictive features in their model were lymphocyte/monocyte ratio, LDH, haemoglobin, β2‐microglobulin, PET SUV max and nodal involvement. Using electronic health record (EHR) data from the Veterans Affairs Cancer Registry System, Li et al. 55 developed a random survival forest (RSF) model integrating clinical and laboratory data among which albumin levels, age and erythrocyte count were the most important variables. Yet, their model was unable to outperform a Cox model based on POD24. Mosqueira Orgueira et al. 56 used gene expression data of 184 FL patients to predict outcomes with an RSF‐based model named IAC‐FL. Best results were obtained when integrating their model with age and the m7‐FLIPI, achieving a c‐index of 0.79, reportedly outperforming m7‐FLIPI alone (c‐index 0.74). Using a neural net‐based approach, Carreras et al. 57 identified 43 among 22 215 genes that were associated with favourable or unfavourable outcomes in FL. Highest ranking genes for favourable prognosis were TDRD12 and ZNF230 and EVA1B, KRT19, BTN2A3P, KLK10, TRPM4, TMEM70 and SLC24A2 for unfavourable prognosis. The same group conducted a similar study in mantle cell lymphoma (MCL), 58 where their model identified KIF18A, YBX3, PEMT, GCNA and POGLUT3 to be associated with poor outcomes while SELENOP, AMOTL2, IGFBP7, KCTD12 and ADGRG2 were associated with favourable outcomes. The authors report these alterations to be also prognostically relevant in external datasets of DLBCL and a multi‐entity cohort derived from The Cancer Genome Atlas. Using clinicopathological, genetic and laboratory data of 862 MCL patients, Hill et al. 59 developed a gradient‐boosted model incorporating 20 features (named ‘integrative MIPI’ or iMIPI) with an ROCAUC of 0.82 and a simplified model (iMIPIs) with 10 features, achieving an ROCAUC of 0.81 in predicting relapse or progression after front‐line treatment with LDH, Ki67 and platelet counts being the top prognostic variables. Zhang et al. 60 used recursive feature elimination to identify four immune‐related genes ( CD247, CD3E, CD4, GATA3 ) delineating two clusters in 152 MCL patients with divergent outcomes. Table 6 summarizes studies investigating ML for risk prediction in FL and MCL.
TABLE 6
Studies investigating ML‐based outcome prediction in follicular lymphoma (FL) and mantle cell lymphoma (MCL).
Abbreviations: DT, decision tree; MLP, multi‐layer perceptron; OS, overall survival; RF, random forest; ROCAUC, receiver‐operating‐characteristic area‐under‐the‐curve; RSF, random survival forest; SVM, support vector machine; XGB, extreme gradient boosting.
| References | Sample(s) | ML model(s) | Performance | Validation |
| Zha et al. 54 | FL ( n = 1938 training, n = 1145 validation) | XGB, DT, RF, ridge classification | POD24 prediction, 0.70 ROCAUC in external validation | External validation set |
| Li et al. 55 | FL ( n = 523) | RSF | Survival prediction, ROCAUC 0.73 (RSF did not outperform Cox model) | Internal cross‐validation |
| Mosqueira Orgueira et al. 56 | FL ( n = 184 training, n = 138 validation) | RSF‐based model (named IAC‐FL) | Survival prediction, c‐index 0.79 when integrating IAC‐FL, m7‐FLIPI, and age | External validation set |
| Carreras et al. 57 | FL ( n = 184) | MLP, radial basis function neural networks | Identified 25 favourable and 18 unfavourable prognostic genes | ‐ |
| Carreras et al. 58 | MCL ( n = 123) | MLP, radial basis function neural networks | Identified 10 affecting prognosis also in DLBCL and a multi‐entity cohort | External validation sets of DLBCL ( n = 414), multi‐entity ( n = 7289) |
| Hill et al. 59 | MCL ( n = 862) | XGB, RF | ROCAUC 0.81–0.82 | Internal cross‐validation |
| Zhang et al. 60 | MCL ( n = 152) | SVM, RF, Lasso, k‐means clustering | Identification of a 4‐gene signature defining two clusters with divergent OS | Internal cross‐validation |
Hodgkin lymphoma
A study by Rask Kragh Jørgensen et al. 61 used a stacked model developed on data from the Danish National Lymphoma Register and validated on Norwegian and Swedish registers to predict PFS and OS in Hodgkin lymphoma (HL), achieving a c‐index of 0.665 and 0.749, respectively, thereby outperforming both IPS‐3 and IPS‐7 while being on par with A‐HIPI. DeAndrés‐Galiana et al. 62 predicted treatment response (CR and PR) versus PD as well as CR versus PR in a cohort of 263 Spanish HL patients, obtaining accuracies of 95.8% and 92%, respectively, based on clinicopathological and laboratory data. Parodi et al. 63 compared five ML models in predicting outcomes of 130 HL patients with the SVM model outperforming its competitors. Among selected predictive variables, they highlight an overexpression of XIST to be enriched in female patients and being associated with favourable outcomes. Patricio et al. 64 used gene expression data from 106 HL patients to predict response after two cycles of ABVD, achieving an ROCAUC of 0.97 with an SVM classifier and identifying a set of 14 predictive genes. Similarly, Kuang et al. 65 identified nine genes ( LAMP1, STAT1, MMP9, C1QB, ICAM1, CD274, CCL19, HCK and LILRB2 ) associated with unfavourable prognosis using gene expression data from 130 HL patients, resulting in a predictive performance of ROCAUC 0.785 for 5‐year OS. Table 7 summarizes studies investigating ML for risk prediction in HL.
TABLE 7
Studies investigating ML‐based outcome prediction in Hodgkin's lymphoma (HL).
Abbreviations: DT, decision tree; KNN, k nearest neighbour; LR, logistic regression; NB, naïve Bayes; RF, random forest; ROCAUC, receiver‐operating‐characteristic area under the curve; SVM, support vector machine; XGB, extreme gradient boosting.
| References | Sample(s) | ML model(s) | Performance | Validation |
| Rask Kragh Jørgensen et al. 61 | HL ( n = 707 training, n = 760 validation) | Stacked model | Validation cohort: c‐index 0.749 (OS) and 0.665 (PFS) | External validation sets |
| deAndrés‐Galiana et al. 62 | HL ( n = 263) | KNN | CR+PR vs. PD accuracy 95.8% CR vs. PR accuracy 92% | Internal cross‐validation |
| Parodi et al. 63 | HL ( n = 130) | KNN, SVM, DT, neural net, Logic Learning Machine | Best‐performing model SVM, accuracy 81.5% | Internal cross‐validation |
| Patricio et al. 64 | HL ( n = 106) | KNN, SVM, RF, XGB, DT, NB | Best‐performing model SVM, ROCAUC 0.97 | Internal cross‐validation |
| Kuang et al. 65 | HL ( n = 130) | Lasso, RF, LR | ROCAUC 0.785 (5‐year OS) | Internal cross‐validation |
Chronic lymphocytic leukaemia
In chronic lymphocytic leukaemia (CLL), multiple studies investigated ML models to predict time‐to‐first treatment (TTFT). 66 , 67 , 68 , 69 Chen et al. 66 used clinical data, laboratory values and standard genetic markers to predict TTFT, reporting high performance for both RF and RSF, albeit that no model outperformed a standard Cox model. Meiseles et al. 67 developed a model trained on a reduced set of clinical and laboratory data and found that the addition of genetic and flow cytometry data did not substantially boost TTFT predictive performance in their sample. Morabito et al. 68 used autoencoders on gene expression data and identified IGF1R and QTRT1 to be significantly associated with TTFT in multivariable analysis, resulting in a c‐index of 0.79 for their final model also including CLL‐IPI relevant variables. Similarly, Mosquera Orgueira et al. 69 identified 19 genes that clustered CLL patients into two groups with different TTFT and developed a tree‐based classifier predicting the need for treatment within 5 years after diagnosis with 89% accuracy. Parviz et al. 70 reported that the addition of clinical variables to CLL‐IPI improved outcome prediction for treatment initiation, occurrence of infection or death, while the addition of recurrent genetic lesions did not improve performance. In an ensemble learning approach, Agius et al. 71 developed CLL‐treatment infection model (CLL‐TIM) on data from the Danish National CLL Registry, predicting treatment initiation or infection within 2 years with a precision and recall of 0.72 and 0.75 respectively. Coombes et al. 72 used unsupervised learning to identify patient clusters with divergent survival and reported mutated immunoglobulin heavy‐chain variable region gene ( IGHV ) status, absent Zap 70 expression, female sex and younger age to be associated with favourable outcomes. Using flow cytometry data, Hoffmann et al. 73 identified 17 distinctive cell populations in 157 CLL patients, of which a population enriched for CD4+ cells had the highest classification performance (ROCAUC 0.83) for outcome prediction. Table 8 summarizes studies investigating ML for risk prediction in CLL.
TABLE 8
Studies investigating ML‐based outcome prediction in chronic lymphocytic leukaemia (CLL).
Abbreviations: AB, adaptive boosting; GBM, gradient boosting machine; LR, logistic regression; RF, random forest; RSF, random survival forest; SVM, support vector machine; XGB, extreme gradient boosting.
| References | Sample(s) | ML model(s) | Output | Validation |
| Chen et al. 66 | CLL ( n = 737) | RF, RSF, LR, SVM, GBM | TTFT prediction c‐index RSF 0.93 (1 year) RF 0.92 (5 years) | Internal cross‐validation |
| Meiseles et al. 67 | CLL ( n = 109) | GBM | TTFT prediction (2 years) AUPRC 0.78 (extensive) 0.77 (reduced) | Internal cross‐validation |
| Morabito et al. 68 | CLL ( n = 217) | Autoencoder | Identified IGF1R and QTRT1 as predictors of TTFT, c‐index 0.79 | Internal cross‐validation |
| Mosquera Ogueira et al. 69 | CLL ( n = 196 training, n = 98 validation) | mclust, boosted trees | Identified 19 genes as predictors of TTFT, accuracy 89% | External validation |
| Parviz et al. 70 | CLL ( n = 314) | AB, XGB, LightGBM, CatBoost, RF, ExtraTrees, SVM | Prediction of treatment initiation, infection, or death with CLL‐IPI, clinical variables; no added benefit of genetics | Internal cross‐validation |
| Agius et al. 71 | CLL ( n = 3270) | Ensemble model named CLL‐treatment‐infection‐model (‐TIM) | Prediction of treatment initiation or infection, precision 0.72 recall 0.75 | External validation |
| Coombes et al. 72 | CLL ( n = 247) | k medoids | Clusters with divergent survival, favourable outcomes associated with mutated IGHV, absent ZAP70 expression, female sex, younger age | ‐ |
| Hoffmann et al. 73 | CLL ( n = 157) | Algorithmic population descriptions (ALPODS) | Identified 17 populations in multi‐colour flow cytometry, ROCAUC 0.83 for a CD4+ enriched population | Internal cross‐validation |
Multiple myeloma
To predict undetectable measurable residual disease (MRD) in multiple myeloma (MM), Guerrero et al. 74 used cytogenetics, tumour burden and immune biomarkers, achieving an accuracy of 0.72 in external validation. Using a transformer‐based model, Hussain et al. 75 predicted PFS, OS and occurrence of adverse events (AE) outperforming the International Staging System (ISS) and with comparable performance to an RSF‐based model and DynamicDeepHit (a neural network) and predicted longitudinal biomarker trajectories, outperforming other tested neural network models. Using clinical, standard genetic and laboratory data from 338 elderly MM patients, Bao et al. 76 compared neural networks—DeepSurv and DeepHit—to an RSF and standard Cox model, reporting comparable performances in predicting OS. Using standard prognostic laboratory values, Farswan et al. 77 propose a model with alternative thresholds for albumin, calcium, glomerular filtration rate (GFR) and haemoglobin using k‐adaptive partitioning, reportedly outperforming RISS scores. Mosquera Orgueira et al. 78 used clinical and gene expression data to develop a 50‐variable RF model, predicting OS with a c‐index of 0.78, incorporating age, ISS stage, β2‐microglobulin, first‐line treatment and expression of 46 genes, prominently involving cell cycle regulation and ubiquitin conjugation. In another study by the same group, 79 unsupervised clustering was applied to clinical, laboratory and standard cytogenetic data of 730 MM patients, identifying two clusters with divergent outcomes restratifying patients from ISS and R‐ISS categories while only slightly outperforming both. Venezian Povoa et al. 80 used gene expression data and cytogenetics to predict treatment sensitivity, reporting an AUC of 0.69 for their proposed multi‐learning training (MuLT) model. Borisov et al. 81 used gene expression data of patients receiving bortezomib‐based regimens to identify FGFR3 , MAF , IGHA2 , IGHV1‐69 and GRB14 upregulation as markers of favourable response. Katsenou et al. 82 used proteomic profiles of 39 MM patients and unsupervised clustering to screen for drug vulnerabilities, reporting CD9, ITB3 and F13A to be the top three most significant markers. To predict severe infections in daratumumab‐based regimens based on routine clinical and laboratory data, Mikulski et al. 83 identified haemoglobin level and ECOG to be significantly associated with severe infections in multivariable analyses, reporting an ROCAUF of 0.87 for gradient boosting and RF respectively. Table 9 summarizes studies investigating ML for risk prediction in MM.
TABLE 9
Studies investigating ML‐based outcome prediction in multiple myeloma (MM).
Abbreviations: BIRCH, balanced iterative reducing and clustering using hierarchies; GBM, gradient boosting machine; LR, logistic regression; MLP, multi‐layer perceptron; RF, random forest; RR, ridge regression; RSF, random survival forest; SVM, support vector machine; XGB, extreme gradient boosting.
| References | Sample(s) | ML model(s) | Output | Validation |
| Guerrero et al. 74 | MM ( n = 301 training, n = 186 validation) | LR | Prediction of undetectable MRD, accuracy 0.72 | External validation |
| Hussain et al. 75 | MM ( n = 703 training, n = 720 validation) | Transformer model (named SCOPE), DynamicDeepHit, RSF | Prediction of PFS, OS, adverse events, biomarker trajectories | External validation |
| Bao et al. 76 | MM ( n = 338) | RSF, DeepSurv, DeepHit | OS prediction, c‐indexes 0.77–0.80 | Internal cross‐validation |
| Farswan et al. 77 | MM ( n = 716 training, n = 1254 validation) | k adaptive partitioning, BIRCH clustering, J48 classifier | Modified Risk Staging (MRS) model, introduces alternative thresholds for prognostic laboratory values | External validation |
| Mosquera Orgueira et al. 78 | MM ( n = 730) | mclust, RF | Prediction of OS, c‐index 0.78 | Internal validation |
| Mosquera Orgueira et al. 79 | MM ( n = 752) | mclust | Identification of two risk clusters, c‐indexes 0.59–0.65 stratified by age | Internal cross‐validation |
| Venezian Povoa et al. | MM ( n = 1525) | LightGBM, MLP, SVM, Multi Learning Training (MuLT) | Prediction of treatment sensitivity, AUC 0.69 | Internal cross‐validation |
| Borisov et al. 81 | MM ( n = 53) | SVM, RR, RF, NB, MLP | Identified FGFR3 , MAF , IGHA2 , IGHV1‐69 and GRB14 upregulation as prognostically favourable for bortezomib‐based regimens | Internal cross‐validation |
| Katsenou et al. 82 | MM ( n = 39) | k means, LR, RF, SVM | Screening proteomic profiles for drug vulnerabilities | Internal validation |
| Mikulski et al. 83 | MM ( n = 139) | Gradient boosting, RF, J48 classifier | Prediction of severe infections, ROCAUC 0.87 | Internal cross‐validation |
DISCUSSION AND OUTLOOK
ML models may soon enable data‐driven personalized outcome predictions using complex and high‐dimensional data in haematology. Yet, several challenges remain hindering widespread routine use. Any AI risk prediction model should be benchmarked against existing models in clinical routine as AI models not necessarily outperform conventional statistical models. Common limitations of current ML‐based outcome classifiers in haematology include their predominantly retrospective design and reliance on single institution training data with small sample size and high class imbalance. They often incorporate only one ML model, therefore lacking comparisons to other ML models or conventional statistical models, often include only a single data modality and lack external or prospective validation and standardization in reporting. Furthermore, training data and model parameters are seldom made publicly accessible, hindering independent verification and clinical translation. To mitigate bias and ensure robust generalizability, models should be trained on large, diverse datasets and validated on external cohorts. However, a lack of digitization, disincentives to data sharing and privacy concerns currently hinder the development of models on a large scale. These may be overcome using co‐operative learning strategies such as federated 84 or swarm learning, 17 centralized data collection efforts such as the Harmony Alliance 85 or alternatively by sharing synthetic samples that mimic the behaviour of real patient cohorts without compromising privacy. 86 Additionally, training sets should include real‐world data and provide adequate representation of minority groups and patients with rare subtypes to mitigate bias and enable generalization to a broader patient population in routine clinical settings rather than being tailored to clinical trial collectives many models are currently trained on. 87 , 88 , 89
Models trained with the explicit intent of acting as a clinical decision support tool should be rigorously assessed and evaluated. Therefore, transparent quality control and standardization of reporting on such models in medical literature are essential. Guidance can be found in the transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD), initially developed for multivariate statistical models 90 and recently updated for machine learning 91 and LLMs 92 ; TRIPOD provides a comprehensive checklist streamlining transparent reporting including data, analytical methods, model specifications, limitations and intended use as well as stakeholder involvement and open science considerations. Technological risk mitigation should be guided by incorporating ‘Good Machine Learning Practices’ as recently proposed by the Food and Drug Administration (FDA). 93
Apart from outcome prediction, AI models have been successfully implemented in haematological diagnostics including evaluation of peripheral blood 94 and bone marrow smears, 95 histopathology 96 and radiological imaging, 97 , 98 identifying genotype–phenotype links, 99 , 100 automating flow cytometry, 101 supporting karyotype analysis 102 and interpretation of genetic variants. 103 Hence, diagnostic AI models may complement each other and support decision‐making models in clinical routine in the near future. As AI permeates the diagnostic and therapeutic workflow in haematology, clinical integration requires a digital infrastructure and a workforce trained in AI applications. The first necessitates seamless integration in laboratory and clinical information systems to enhance and not disrupt existing workflows. AI models should aggregate information in an existing digital environment to provide comprehensive and explainable decision support in a user‐friendly interface without increasing workload by requiring haematologists to enter redundant information into external software. Haematologists should be trained in interpreting model outputs and be aware of model‐dependent limitations. This requires incorporating AI literacy in medical school and residency, ideally interwoven with concrete use cases as they arise during medical training. Furthermore, incorporating AI literacy into continuing medical education is essential, given that rapid technological advances demand ongoing updates in user competencies.
Given the increasing complexity of data in haematology, data‐driven ML models may be well equipped to address the complex biological interactions present in haematological malignancies. As these models mature, they have the potential to provide data‐driven disease classification systems and risk‐adapted treatment support. Bridging the gap between science and clinical routine will depend on sustained collaboration, robust validation and regulatory guidance.
AUTHOR CONTRIBUTIONS
J‐NE conceptualized the study. J‐NE and WW performed the literature search and wrote the initial draft of the manuscript. J‐NE created visualizations. JMM and TH critically reviewed and edited the manuscript. All authors approved the final version of the manuscript.
FUNDING INFORMATION
No funding was received for this study.
CONFLICT OF INTEREST STATEMENT
J‐NE declares consulting services for AstraZeneca, Novartis, Janssen. Furthermore, he holds shares in Cancilico; has received an institutional research grant by Novartis; and has received honoraria by Amgen, AstraZeneca, Janssen, Novartis and Pfizer. JMM declares consulting services for Janssen, Roche, Gilead, Abbvie, Jazz, Pfizer, Astellas, Novartis, AstraZeneca, Clycostem. Furthermore, he holds shares in Cancilico and Synagen; has received institutional research grants by Janssen, Jazz and Novartis; and has received honoraria by Novartis, Roche, Janssen, Abbvie, Pfizer, Sanofi, Astellas and Beigene. TH is part owner of MLL Munich Leukemia Laboratory. WW is employed by MLL Munich Leukemia Laboratory.
ACKNOWLEDGEMENTS
Open Access funding enabled and organized by Projekt DEAL.
Eckardt J‐N, Walter W, Middeke JM, Haferlach T. Artificial intelligence for risk assessment and outcome prediction in malignant haematology. Br J Haematol. 2026;208(1):25–38. 10.1111/bjh.70231
DATA AVAILABILITY STATEMENT
Data sharing not applicable to this article as no datasets were generated or analysed during the current study.
REFERENCES
1. Döhner H, Wei AH, Appelbaum FR, Craddock C, DiNardo CD, Dombret H, et al. Diagnosis and management of AML in adults: 2022 recommendations from an international expert panel on behalf of the ELN. Blood. 2022;140(12):1345–1377. 10.1182/blood.2022016867
2. Greenberg PL, Tuechler H, Schanz J, Sanz G, Garcia‐Manero G, Solé F, et al. Revised international prognostic scoring system for myelodysplastic syndromes. Blood. 2012;120(12):2454–2465.
3. International Non‐Hodgkin's Lymphoma Prognostic Factors Project . A predictive model for aggressive non‐Hodgkin's lymphoma. N Engl J Med. 1993;329(14):987–994.
4. Kothari R, Cualing H, Balachander T. Neural network analysis of flow cytometry immunophenotype data. IEEE Trans Biomed Eng. 1996;43(8):803–810.
5. Zini G, d'Onofrio G. Neural network in hematopoietic malignancies. Clin Chim Acta. 2003;333(2):195–201.
6. Masić N, Gagro A, Rabatić S, Sabioncello A, Dasić G, Jaksić B, et al. Decision‐tree approach to the immunophenotype‐based prognosis of the B‐cell chronic lymphocytic leukemia. Am J Hematol. 1998;59(2):143–148.
7. Lyons‐Weiler J, Patel S, Bhattacharya S. A classification‐based machine learning approach for the analysis of genome‐wide expression data. Genome Res. 2003;13(3):503–512.
8. Topol EJ. High‐performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019;25(1):44–56.
9. He J, Baxter SL, Xu J, Xu J, Zhou X, Zhang K. The practical implementation of artificial intelligence technologies in medicine. Nat Med. 2019;25(1):30–36.
10. Mezzalira E, D'Agnolo M, Boaro MP, Reggiani G, Cellini A, Boaro A, et al. European Hematology Association 2025 Congress. [cited 2025 Aug 20]. Assessing Knowledge Gaps and Attitudes Towards Artificial Intelligence Among Healthcare Professionals Working in Haematology. Available from: https://library.ehaweb.org/eha/2025/eha2025‐congress/4162910/elisabetta.mezzalira.assessing.knowledge.gaps.and.attitudes.towards.artificiai.html?f=menu%3D6%2Abrowseby%3D8%2Asortby%3D2%2Amedia%3D3%2Ace_id%3D2882%2Aot_id%3D31588%2Amarker%3D5845%2Afeatured%3D19595
11. Russell S, Norvig P. Artificial intelligence, global edition. London: Pearson; 2021. Available from: https://elibrary.pearson.de/book/99.150005/9781292401171
12. Goodfellow IJ, Bengio Y, Courville A. Deep learning. Cambridge, MA, USA: MIT Press; 2016. [cited 2023 Dec 9]. Available from: https://mitpress.mit.edu/9780262035613/deep‐learning/
13. Krizhevsky A, Sutskever I, Hinton GE. ImageNet classification with deep convolutional neural networks. Advances in neural information processing systems. NIPS'12: Proceedings of the 26th International Conference on Neural Information Processing Systems – Volume 1. Curran Associates, Inc; 2012. p. 1097–1110 [cited 2023 Dec 9]. Available from: https://papers.nips.cc/paper_files/paper/2012/hash/c399862d3b9d6b76c8436e924a68c45b‐Abstract.html
14. Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez AN, et al. Attention is all you need. 2017. arXiv.org. [cited 2025 Apr 23]. Available from: https://arxiv.org/abs/1706.03762v7
15. Eckardt JN, Wendt K, Bornhäuser M, Middeke JM. Reinforcement learning for precision oncology. Cancers (Basel). 2021;13(18):4624.
16. Li L, Fan Y, Tse M, Lin KY. A review of applications in federated learning. Comput Ind Eng. 2020;149:106854.
17. Warnat‐Herresthal S, Schultze H, Shastry KL, Manamohan S, Mukherjee S, Garg V, et al. Swarm learning for decentralized and confidential clinical machine learning. Nature. 2021;594(7862):265–270.
18. Alba AC, Agoritsas T, Walsh M, Hanna S, Iorio A, Devereaux PJ, et al. Discrimination and calibration of clinical prediction models: users' guides to the medical literature. JAMA. 2017;318(14):1377–1384.
19. Van Calster B, McLernon DJ, van Smeden M, Wynants L, Steyerberg EW, Bossuyt P, et al. Calibration: the Achilles heel of predictive analytics. BMC Med. 2019;17(1):230.
20. Brier GW. Verification of forecasts expressed in terms of probability. Mon Weather Rev. 1950;78(1):1–3.
21. Guo C, Pleiss G, Sun Y, Weinberger KQ. On calibration of modern neural networks [Internet]. arXiv 2017. [cited 2025 Oct 8]. Available from: http://arxiv.org/abs/1706.04599
22. Bejani MM, Ghatee M. A systematic review on overfitting control in shallow and deep neural networks. Artif Intell Rev. 2021;54(8):6391–6438.
23. Radhachandran A, Garikipati A, Iqbal Z, Siefkas A, Barnes G, Hoffman J, et al. A machine learning approach to predicting risk of myelodysplastic syndrome. Leuk Res. 2021;109:106639.
24. Clichet V, Lebon D, Chapuis N, Zhu J, Bardet V, Marolleau JP, et al. Artificial intelligence to empower diagnosis of myelodysplastic syndromes by multiparametric flow cytometry. Haematologica. 2023;108(9):2435–2443.
25. Nazha A, Komrokji R, Meggendorfer M, Jia X, Radakovich N, Shreve J, et al. Personalized prediction model to risk stratify patients with myelodysplastic syndromes. JCO. 2021;39(33):3737–3746.
26. Mosquera Orgueira A, Perez Encinas MM, Diaz Varela NA, Mora E, Díaz‐Beyá M, Montoro MJ, et al. Machine learning improves risk stratification in myelodysplastic neoplasms: an analysis of the Spanish Group of Myelodysplastic Syndromes. Hema. 2023;7(10):e961.
27. Nazha A, Sekeres MA, Bejar R, Rauh MJ, Othus M, Komrokji RS, et al. Genomic biomarkers to predict resistance to hypomethylating agents in patients with myelodysplastic syndromes using artificial intelligence. JCO Precis Oncol. 2019;3: PO.19.00119. https://pubmed.ncbi.nlm.nih.gov/31663066/
28. Radakovich N, Sallman DA, Buckstein R, Brunner A, Dezern A, Mukerjee S, et al. A machine learning model of response to hypomethylating agents in myelodysplastic syndromes. iScience. 2022;25(10):104931.
29. Sharplin K, Proudman W, Chhetri R, Tran ENH, Choong J, Kutyna M, et al. A personalized risk model for azacitidine outcome in myelodysplastic syndrome and other myeloid neoplasms identified by machine learning model utilizing real‐world data. Cancers (Basel). 2023;15(16):4019.
30. Shanbehzadeh M, Afrash MR, Mirani N, Kazemi‐Arpanahi H. Comparing machine learning algorithms to predict 5‐year survival in patients with chronic myeloid leukemia. BMC Med Inform Decis Mak. 2022;22(6):236.
31. Mosquera‐Orgueira A, Pérez‐Encinas M, Hernández‐Sánchez A, González‐Martínez T, Arellano‐Rodrigo E, Martínez‐Elicegui J, et al. Machine learning improves risk stratification in myelofibrosis: an analysis of the Spanish registry of myelofibrosis. Hema. 2023;7(1):e818.
32. Ryou H, Sirinukunwattana K, Aberdeen A, Grindstaff G, Stolz BJ, Byrne H, et al. Continuous indexing of fibrosis (CIF): improving the assessment and classification of MPN patients. Leukemia. 2023;37(2):348–358.
33. Hernández‐Boluda JC, Mosquera‐Orgueira A, Gras L, Koster L, Tuffnell J, Kröger N, et al. Use of machine learning techniques to predict poor survival after hematopoietic cell transplantation for myelofibrosis. Blood. 2025;145(26):3139–3152.
34. Shaikh AF, Kakirde C, Dhamne C, Bhanshe P, Joshi S, Chaudhary S, et al. Machine learning derived genomics driven prognostication for acute myeloid leukemia with RUNX1‐RUNX1T1. Leuk Lymphoma. 2020;61(13):3154–3160. 10.1080/10428194.2020.1798951
35. Patkar N, Shaikh AF, Kakirde C, Nathany S, Ramesh H, Bhanshe P, et al. A novel machine‐learning‐derived genetic score correlates with measurable residual disease and is highly predictive of outcome in acute myeloid leukemia with mutated NPM1. Blood Cancer J. 2019;9(10): Available from: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6773777/
36. Eckardt JN, Hahn W, Ries RE, Chrost SD, Winter S, Stasik S, et al. Age‐stratified machine learning identifies divergent prognostic significance of molecular alterations in AML. Hema. 2025;9(5):e70132.
37. Eckardt JN, Röllig C, Metzeler K, Kramer M, Stasik S, Georgi JA, et al. Prediction of complete remission and survival in acute myeloid leukemia using supervised machine learning. Haematologica. 2023;108(3):690–704.
38. Eckardt JN, Röllig C, Metzeler K, Heisig P, Stasik S, Georgi JA, et al. Unsupervised meta‐clustering identifies risk clusters in acute myeloid leukemia based on clinical and genetic profiles. Commun Med. 2023;3(1):1–13.
39. Gal O, Auslander N, Fan Y, Meerzaman D. Predicting complete remission of acute myeloid leukemia: machine learning applied to gene expression. Cancer Informat. 2019;18:1176935119835544.
40. Wagner S, Vadakekolathu J, Tasian SK, Altmann H, Bornhäuser M, Pockley AG, et al. A parsimonious 3‐gene signature predicts clinical outcomes in an acute myeloid leukemia multicohort study. Blood Adv. 2019;3(8):1330–1346.
41. Karami K, Akbari M, Moradi MT, Soleymani B, Fallahi H. Survival prognostic factors in patients with acute myeloid leukemia using machine learning techniques. PLoS One. 2021;16(7):e0254976.
42. Jia Z y, Abulimiti M, Wu Y, Ma L n, Li X y, Wang J. A novel perspective on survival prediction for AML patients: integration of machine learning in SEER database applications. Heliyon. 2025;11(2):e42030.
43. Alizadeh AA, Eisen MB, Davis RE, Ma C, Lossos IS, Rosenwald A, et al. Distinct types of diffuse large B‐cell lymphoma identified by gene expression profiling. Nature. 2000;403(6769):503–511.
44. Shipp MA, Ross KN, Tamayo P, Weng AP, Kutok JL, Aguiar RCT, et al. Diffuse large B‐cell lymphoma outcome prediction by gene‐expression profiling and supervised machine learning. Nat Med. 2002;8(1):68–74.
45. Chapuy B, Stewart C, Dunford AJ, Kim J, Kamburov A, Redd RA, et al. Molecular subtypes of diffuse large B cell lymphoma are associated with distinct pathogenic mechanisms and outcomes. Nat Med. 2018;24(5):679–690.
46. Schmitz R, Wright GW, Huang DW, Johnson CA, Phelan JD, Wang JQ, et al. Genetics and pathogenesis of diffuse large B‐cell lymphoma. N Engl J Med. 2018;378(15):1396–1407.
47. Zhuang X, Liu B, Long J, Wang H, Yu J, Ji X, et al. Machine‐learning‐based classification of diffuse large B‐cell lymphoma patients by a 7‐mRNA signature enriched with immune infiltration and cell cycle. Clin Transl Oncol. 2024;26(4):936–950.
48. He T, Geng J, Hou C, Li H, Zhang H, Zhao P, et al. Predictive biomarkers and molecular subtypes in DLBCL: insights from PCD gene expression and machine learning. Discov Oncol. 2025;16(1):542.
49. Biccler JL, Eloranta S, de Nully Brown P, Frederiksen H, Jerkeman M, Jørgensen J, et al. Optimizing outcome prediction in diffuse large B‐cell lymphoma by use of machine learning and Nationwide lymphoma registries: a Nordic lymphoma group study. JCO Clin Cancer Inform. 2018;2:1–13.
50. Mosquera Orgueira A, Díaz Arias JÁ, Cid López M, Peleteiro Raíndo A, Antelo Rodríguez B, Aliste Santos C, et al. Improved personalized survival prediction of patients with diffuse large B‐cell lymphoma using gene expression profiling. BMC Cancer. 2020;20:1017.
51. Detrait MY, Warnon S, Lagasse R, Dumont L, Prophétis SD, Hansenne A, et al. A machine learning approach in a monocentric cohort for predicting primary refractory disease in diffuse large B‐cell lymphoma patients. PLoS One. 2024;19(10):e0311261.
52. Zhao S, Dong X, Shen W, Ye Z, Xiang R. Machine learning‐based classification of diffuse large B‐cell lymphoma patients by eight gene expression profiles. Cancer Med. 2016;5(5):837–852.
53. Viswanathan A, Kundal K, Sengupta A, Kumar A, Kumar KV, Holmes AB, et al. Deep learning‐based classifier of diffuse large B‐cell lymphoma cell‐of‐origin with clinical outcome. Brief Funct Genomics. 2023;22(1):42–48.
54. Zha J, Chen Q, Zhang W, Jing H, Ye J, Liu H, et al. A machine learning‐based model to predict POD24 in follicular lymphoma: a study by the Chinese workshop on follicular lymphoma. Biomark Res. 2025;13(1):2.
55. Li C, Patil V, Rasmussen KM, Yong C, Chien HC, Morreall D, et al. Predicting survival in veterans with follicular lymphoma using structured electronic health record information and machine learning. Int J Environ Res Public Health. 2021;18(5):2679.
56. Mosquera Orgueira A, Cid López M, Peleteiro Raíndo A, Abuín Blanco A, Díaz Arias JÁ, González Pérez MS, et al. Personally tailored survival prediction of patients with follicular lymphoma using machine learning transcriptome‐based models. Front Oncol. 2022;11. 10.3389/fonc.2021.705010/full
57. Carreras J, Kikuti YY, Miyaoka M, Hiraiwa S, Tomita S, Ikoma H, et al. Artificial intelligence analysis of the gene expression of follicular lymphoma predicted the overall survival and correlated with the immune microenvironment response signatures. Mach Learn Know Extr. 2020;2(4):647–671.
58. Carreras J, Nakamura N, Hamoudi R. Artificial intelligence analysis of gene expression predicted the overall survival of mantle cell lymphoma and a large pan‐cancer series. Healthcare (Basel). 2022;10(1):155.
59. Hill HA, Jain P, Ok CY, Sasaki K, Chen H, Wang ML, et al. Integrative prognostic machine learning models in mantle cell lymphoma. Cancer Res Commun. 2023;3(8):1435–1446.
60. Zhang W, Shi JN, Wang HN, Zhang T, Zhou X, Zhang HM, et al. Identification of immune‐related genes and development of a prognostic model in mantle cell lymphoma. Ann Transl Med. 2022;10(24):1323.
61. Rask Kragh Jørgensen R, Bergström F, Eloranta S, Tang Severinsen M, Bjøro Smeland K, Fosså A, et al. Machine learning‐based survival prediction models for progression‐free and overall survival in advanced‐stage Hodgkin lymphoma. JCO Clin Cancer Inform. 2024;8:e2300255.
62. deAndrés‐Galiana EJ, Fernández‐Martínez JL, Luaces O, Del Coz JJ, Fernández R, Solano J, et al. On the prediction of Hodgkin lymphoma treatment response. Clin Transl Oncol. 2015;17(8):612–619.
63. Parodi S, Manneschi C, Verda D, Ferrari E, Muselli M. Logic learning machine and standard supervised methods for Hodgkin's lymphoma prognosis using gene expression data and clinical variables. Health Informatics J. 2018;24(1):54–65.
64. Patrício A, Costa RS, Henriques R. On the challenges of predicting treatment response in Hodgkin's lymphoma using transcriptomic data. BMC Med Genet. 2023;16(1):170.
65. Kuang Z, Tu J, Li X. Combined identification of novel markers for diagnosis and prognostic of classic Hodgkin lymphoma. Int J Gen Med. 2021;14:9951–9963.
66. Chen D, Goyal G, Go RS, Parikh SA, Ngufor CG. Improved interpretability of machine learning model using unsupervised clustering: predicting time to first treatment in chronic lymphocytic leukemia. JCO Clin Cancer Inform. 2019;3:1–11. 10.1200/CCI.18.00137
67. Meiseles A, Paley D, Ziv M, Hadid Y, Rokach L, Tadmor T. Explainable machine learning for chronic lymphocytic leukemia treatment prediction using only inexpensive tests. Comput Biol Med. 2022;145:105490.
68. Morabito F, Adornetto C, Monti P, Amaro A, Reggiani F, Colombo M, et al. Genes selection using deep learning and explainable artificial intelligence for chronic lymphocytic leukemia predicting the need and time to therapy. Front Oncol. 2023;13. 10.3389/fonc.2023.1198992/full
69. Mosquera Orgueira A, Antelo Rodríguez B, Alonso Vence N, Bendaña López Á, Díaz Arias JÁ, Díaz Varela N, et al. Time to treatment prediction in chronic lymphocytic leukemia based on new transcriptional patterns. Front Oncol. 2019;9. 10.3389/fonc.2019.00079/full
70. Parviz M, Brieghel C, Agius R, Niemann CU. Prediction of clinical outcome in CLL based on recurrent gene mutations, CLL‐IPI variables, and (para)clinical data. Blood Adv. 2022;6(12):3716–3728.
71. Agius R, Brieghel C, Andersen MA, Pearson AT, Ledergerber B, Cozzi‐Lepri A, et al. Machine learning can identify newly diagnosed patients with CLL at high risk of infection. Nat Commun. 2020;11(1):363.
72. Coombes CE, Abrams ZB, Li S, Abruzzo LV, Coombes KR. Unsupervised machine learning and prognostic factors of survival in chronic lymphocytic leukemia. J Am Med Inform Assoc. 2020;27(7):1019–1027.
73. Hoffmann J, Eminovic S, Wilhelm C, Krause SW, Neubauer A, Thrun MC, et al. Prediction of clinical outcomes with explainable artificial intelligence in patients with chronic lymphocytic leukemia. Curr Oncol. 2023;30(2):1903–1915.
74. Guerrero C, Puig N, Cedena MT, Goicoechea I, Perez C, Garcés JJ, et al. A machine learning model based on tumor and immune biomarkers to predict undetectable MRD and survival outcomes in multiple myeloma. Clin Cancer Res. 2022;28(12):2598–2609.
75. Hussain Z, De Brouwer E, Boiarsky R, Setty S, Gupta N, Liu G, et al. Joint AI‐driven event prediction and longitudinal modeling in newly diagnosed and relapsed multiple myeloma. NPJ Digit Med. 2024;7(1):200.
76. Bao L, Wang YT, Zhuang JL, Liu AJ, Dong YJ, Chu B, et al. Machine learning‐based overall survival prediction of elderly patients with multiple myeloma from multicentre real‐life data. Front Oncol. 2022;12:922039.
77. Farswan A, Gupta A, Gupta R, Hazra S, Khan S, Kumar L, et al. AI‐supported modified risk staging for multiple myeloma cancer useful in real‐world scenario. Transl Oncol. 2021;14(9):101157.
78. Mosquera Orgueira A, González Pérez MS, Díaz Arias JÁ, Antelo Rodríguez B, Alonso Vence N, Bendaña López Á, et al. Survival prediction and treatment optimization of multiple myeloma patients using machine‐learning models based on clinical and gene expression data. Leukemia. 2021;35(10):2924–2935.
79. Mosquera Orgueira A, González Pérez MS, Diaz Arias J, Rosiñol L, Oriol A, Teruel AI, et al. Unsupervised machine learning improves risk stratification in newly diagnosed multiple myeloma: an analysis of the Spanish myeloma group. Blood Cancer J. 2022;12(4):76.
80. Venezian Povoa L, Ribeiro CHC, da Silva IT. Machine learning predicts treatment sensitivity in multiple myeloma based on molecular and clinical information coupled with drug response. PLoS One. 2021;16(7):e0254596.
81. Borisov N, Sergeeva A, Suntsova M, Raevskiy M, Gaifullin N, Mendeleeva L, et al. Machine learning applicability for classification of PAD/VCD chemotherapy response using 53 multiple myeloma RNA sequencing profiles. Front Oncol. 2021;11:652063.
82. Katsenou A, O'Farrell R, Dowling P, Heckman CA, O'Gorman P, Bazou D. Using proteomics data to identify personalized treatments in multiple myeloma: a machine learning approach. Int J Mol Sci. 2023;24(21):15570.
83. Mikulski D, Kędzior MK, Mirocha G, Jerzmanowska‐Piechota K, Witas Ż, Woźniak Ł, et al. Predictors and profile of severe infectious complications in multiple myeloma patients treated with daratumumab‐based regimens: a machine learning model for pneumonia risk. Cancers (Basel). 2024;16(21):3709.
84. Rieke N, Hancox J, Li W, Milletarì F, Roth HR, Albarqouni S, et al. The future of digital health with federated learning. npj Digit Med. 2020;3(1):1–7.
85. Home ‐ HARMONY Healthcare Alliance [Internet]. [cited 2025 June 16]. Available from: https://www.harmony‐alliance.eu/
86. Eckardt JN, Hahn W, Röllig C, Stasik S, Platzbecker U, Müller‐Tidow C, et al. Mimicking clinical trials with synthetic acute myeloid leukemia patients using generative artificial intelligence. npj Digit Med. 2024;7(1):1–11.
87. Futoma J, Simons M, Panch T, Doshi‐Velez F, Celi LA. The myth of generalisability in clinical research and machine learning in health care. Lancet Digit Health. 2020;2(9):e489–e492.
88. Norori N, Hu Q, Aellen FM, Faraci FD, Tzovara A. Addressing bias in big data and AI for health care: a call for open science. Patterns (N Y). 2021;2(10):100347.
89. Yang J, Dung NT, Thach PN, Phong NT, Phu VD, Phu KD, et al. Generalizability assessment of AI models across hospitals in a low‐middle and high income country. Nat Commun. 2024;15(1):8270.
90. Collins GS, Reitsma JB, Altman DG, Moons KGM. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. Br J Cancer. 2015;112(2):251–259.
91. Collins GS, Moons KGM, Dhiman P, Riley RD, Beam AL, Calster BV, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:e078378.
92. Gallifant J, Afshar M, Ameen S, Aphinyanaphongs Y, Chen S, Cacciamani G, et al. The TRIPOD‐LLM reporting guideline for studies using large language models. Nat Med. 2025;31(1):60–69.
93. Food and Drug Administration . Good machine learning practice for medical device development: guiding principles. Silver Spring, MD, USA: FDA; 2025. Available from: https://www.fda.gov/medical‐devices/software‐medical‐device‐samd/good‐machine‐learning‐practice‐medical‐device‐development‐guiding‐principles
94. Matek C, Schwarz S, Spiekermann K, Marr C. Human‐level recognition of blast cells in acute myeloid leukaemia with convolutional neural networks. Nat Mach Intell. 2019;1(11):538–544.
95. Matek C, Krappe S, Münzenmayer C, Haferlach T, Marr C. Highly accurate differentiation of bone marrow cell morphologies using deep neural networks on a large image data set. Blood. 2021;138(20):1917–1927.
96. Pescia C, Sozanska AM, Thomas E, Cooper RA. Artificial intelligence in haematopathology: current perspective and future directions. Diagn Histopathol. 2025;31(5):267–276.
97. Bai A, Si M, Xue P, Qu Y, Jiang Y. Artificial intelligence performance in detecting lymphoma from medical imaging: a systematic review and meta‐analysis. BMC Med Inform Decis Mak. 2024;24(1):13.
98. Cairns J, Frood R, Patel C, Scarsbrook A. The role of AI in lymphoma: an update. Semin Nucl Med. 2025;55(3):377–386.
99. Eckardt JN, Middeke JM, Riechert S, Schmittmann T, Sulaiman AS, Kramer M, et al. Deep learning detects acute myeloid leukemia and predicts NPM1 mutation status from bone marrow smears. Leukemia. 2022;36:111–118. 10.1038/s41375-021-01408-w
100. Eckardt JN, Schmittmann T, Riechert S, Kramer M, Sulaiman AS, Sockel K, et al. Deep learning identifies acute promyelocytic leukemia in bone marrow smears. BMC Cancer. 2022;22(1):201.
101. Ng DP, Simonson PD, Tarnok A, Lucas F, Kern W, Rolf N, et al. Recommendations for using artificial intelligence in clinical flow cytometry. Cytometry B Clin Cytom. 2024;106(4):228–238.
102. Rosenblum LS, Holmes J, Taghiyev AF. The emergence of artificial intelligence‐guided karyotyping: a review and reflection. Genes. 2025;16(6):685.
103. Novakovsky G, Dexter N, Libbrecht MW, Wasserman WW, Mostafavi S. Obtaining genetics insights from deep learning via explainable artificial intelligence. Nat Rev Genet. 2023;24(2):125–137.
Associated Data
Data Availability Statement
Data sharing not applicable to this article as no datasets were generated or analysed during the current study.