Artificial Intelligence in Haematologic Diagnostics: Current Applications and Future Perspectives

Abstract

Background

Clinical researchers and laboratory specialists are striving to explore artificial intelligence (AI) to facilitate and optimize haematological diagnostics in response to the growing demand for more efficient and accurate diagnoses.

Summary

This review summarizes current approaches integrating AI into blood and bone marrow cytomorphology, flow cytometry (FC), genetics, and haemostasis. Efforts include automated cell differentiation in peripheral blood and bone marrow aspirates, algorithms for identifying causes of anaemia, tools for rapid diagnosis of acute leukaemia, and other haematological entities. AI in FC may reduce subjectivity and variability, while in genomics, machine learning is increasingly implemented for processing high-throughput sequencing data and may enable automated detection of karyotypes in the future. In haemostasis, AI allows for automation in quality control, the establishment of personalized reference ranges, and potentially automated result interpretation. AI has, however, limitations such as cross-platform compatibility and often lacks sufficient validation. Ethical concerns include risks of bias and regulations are lagging behind the rapid developments.

Key Messages

AI shows promise for automating and improving many steps in haematological diagnostics, though final interpretation still needs expert haematologists.

Plain Language Summary

Haematology encompasses a wide variety of disorders affecting the blood and bone marrow. A diagnosis often requires multiple – and, with increased understanding of diseases, progressively complex – tests. Doctors and laboratory professionals are therefore exploring how artificial intelligence (AI) can help improve the way disorders of the blood and bone marrow are detected and understood, particularly how AI offers new ways to make this process faster and more accurate. This article reviews how AI is being used in several key areas of haematological laboratories. These include examining cells under a microscope, analysing patterns in cells using specialized machines (a method called Flow Cytometry), studying genetic changes, and measuring the ability of blood to form clots (a process called haemostasis). Some current uses of AI include automatically sorting different types of blood cells, identifying possible causes of anaemia (a condition where there are not enough healthy red blood cells), and quickly detecting certain types of leukaemia, a cancer of the blood. AI can also help reduce human error when analysing test results and speed up the processing of large genetic datasets. However, there are challenges. Many AI tools do not work well across different laboratories or have not been tested enough to be used in everyday practice. Currently and in the near future, AI is best used to support experts by handling some of the early steps in the testing process. Final decisions about a diagnosis still need to be made by trained professionals in blood diseases.

Introduction

Haematological diagnostics encompass an increasing range of methodologies that continually advance in technical intricacy and complexity of interpretation. The integration of multiple modalities has become crucial for diagnostic and prognostic assessment, posing a challenge amidst evolving requirements. Simultaneously, demographic shifts contribute to a rise in haematological conditions, intensifying workload pressures amid a decline in the availability of well-trained personnel. Consequently, clinical researchers and laboratory specialists are exploring how artificial intelligence (AI) can streamline and improve haematological diagnostics.

The term “artificial intelligence” (AI) emerged in the 1950s [1, 2] and is used for a wide range of existing and fictional, more or less sophisticated tasks performed by machines for problem-solving or imitation of human intelligence. After several waves of development, AI is currently transforming various fields of application. Recent achievements were driven by important progress in machine learning (ML), which was enabled by improved hardware and the extensive availability of large datasets. ML is a subset of AI, which involves the computational generation of rules or algorithms derived from a given dataset and a specified prediction model. Traditionally, ML takes one of these three learning approaches: supervised learning, unsupervised learning, and reinforcement learning [3]. While a comprehensive overview on general aspects of ML is beyond the scope of this review article focusing on the integration of AI in haematology diagnostics, a short summary is certainly helpful.

General Aspects of ML

In supervised learning, both input data (independent variables or predictors) and output data (dependent variable, the outcome or response) are required to optimize predictions. For this type of combined input and output data, the term “labelled” data has been coined. The basic tasks of supervised learning are regression and classification. Regression aims to optimize functions so as to best predict a continuous outcome variable based on a set of predictors, whereas classification focuses on optimizing a model to predict discrete outcome categories [3]. Unsupervised learning uses only input data without a corresponding outcome variable. The algorithm searches for patterns or structures in the input data. Most algorithms require certain assumptions about the data and might need careful optimization for the generation of useful results. Typical examples include principal component analysis and cluster analysis [3]. In reinforcement learning, the learning algorithm engages in a trial-and-error approach, interacting with the environment to receive specific rewards. Through this feedback loop, the algorithm continuously improves. This approach has enabled computers to master the game of Go, an abstract strategy board game [4], and, more recently, has made it possible for vehicles to drive autonomously [5].

Traditional statistical learning models require well-structured input data, organized into specific variables, posing a significant challenge for relatively unstructured data. Converting unstructured primary data into a structured format involves a process commonly known as feature extraction. This entails isolating or transforming specific elements of the data, assigning them to distinct variables. Feature extraction, as performed by humans, is often cumbersome and time-consuming. Recently, deep learning (DL) models, which rely on multiple layers of artificial neural networks (ANNs), have gained popularity. One advantage of these models is their ability to process nonlinear and relatively unstructured data such as images, text, speech, or videos [6]. Therefore, the data preparation step of feature extraction can often be omitted or at least reduced to a minimum [7]. DL may outperform traditional methods for some types of tasks, while more traditional, feature-based methods may be better suited for others [7]. Further advances in artificial neuronal networks have led to the development of large language models (LLMs) such as ChatGPT (Chat Generative Pre-Trained Transformer) and others for the processing of natural language [8–10]. Recently, these models have been further developed into large multimodal models that integrate text and image processing capabilities [11–13].

Extensive reviews of AI in medicine or haematology have been published elsewhere [14–18]. In this review, we briefly summarize aspects of the application of AI in haematology diagnostics at present and in the near future and aim to provide a concise overview of common terms and key concepts. There are numerous challenges in the field: firstly, haematology covers a wide range of diseases, including many rare ones with incidence rates of 1 in 1,000,000 or lower. Integrating these rare entities into most ML models poses significant challenges due to the need for large datasets. Consequently, rare diseases are often underrepresented or display skewed frequency distribution, leading to biased model development. In the worst-case scenario, prediction models may simply miss such rare entities. Secondly, many diagnoses rely on criteria that combine information from multiple diagnostic methods, such as medical history, clinical examinations, laboratory tests, and imaging. Each contributes unique data types, necessitating comprehensive models capable of processing diverse data types. Thirdly, some fields in haematology are not well standardized and the interpretation of certain biomarkers may be strongly user-dependent, e.g., detection of dysplasia for the diagnosis of myelodysplastic syndromes (MDS). This results in significant inter-observer variability, hindering the creation of reliable reference data. Finally, many complex diagnostic procedures lack standardization across platforms, laboratories, and sites, e.g., staining and pre-processing of samples for flow cytometry (FC). This severely limits generalizability of prediction models, either because platform-related variability exceeds biological variability, rendering any consistent prediction impossible, or because the training dataset lacks representative data from multiple platforms and sites, leading to overfitted models. Such single platform models may perform excellently on their original platform or site but may fail when applied to others.

Blood Count and Cytomorphology

Peripheral blood cell counters are part of the basic equipment of every haematology laboratory. The results of a complete blood count (CBC) are parameterized and are therefore easy to use in ML. Preselecting certain parameters can aid in investigating specific questions [19, 20]. For example, it can help diagnose anaemia and distinguish its causes, such as iron deficiency anaemia (IDA) versus thalassaemia based on haemoglobin, erythrocyte count, red blood cells (RBCs), and reticulocytes indices [21, 22]. Urrechaga et al. [23] developed an index that combines MicroR – the percentage of microcytic RBCs with a volume less than 60 fL – and HYPO-He, which represents the percentage of hypochromic RBCs with a haemoglobin content below 17 pg (1.055 fmol), along with the red cell distribution width (RDW). This so-called “M-H-RDW index” provided a sensitivity of 100% and a specificity of 9,293% for β-thalassaemia screening at a cut-off value of −7.6. Schoorl et al. [24] developed six algorithms – three for IDA and three for β-thalassaemia, which can be used to differentiate IDA and β-thalassaemia. Similarly, some information from CBCs can be integrated into ML algorithms to identify the causes of thrombocytopaenia, as reviewed by Elshoeibi et al. [25].

Malignant diseases that have a relevant expression in the blood have been a major focus of AI research. In fact, malignant disorders such as acute lymphoblastic leukaemia (ALL) [26], chronic lymphocytic leukaemia [27], and myeloproliferative neoplasms [28] may be rapidly suggested or even diagnosed based on CBC results. However, in all of these examples, independent validation is still required before implementation. Given the relevance of cell counts in many haematological diseases and the potential for their use in AI, integration with other AI programs in a single platform such as morphology, FC and molecular information will be highly relevant [29].

A related but more challenging field is the application of AI in cytomorphology. In recent years, many laboratories worldwide have introduced digital microscopy on peripheral blood films and developed algorithms aiming to reduce the number of peripheral blood films that require manual evaluation. One example is CellaVision (Ideon Science Park 223, 70 Lund Sweden) which allows automation of the haematology workflow, creating a streamlined process. The system includes a microscope with a digital camera and a sophisticated software. Slides are automatically identified and placed under the microscope. Individual cells are located and digital images of leukocytes, erythrocytes, and thrombocytes are captured. A ML-derived algorithm provisionally classifies leukocytes and erythrocytes. The pre-classification/pre-characterization is then reviewed and verified by a qualified lab technician. The system is particularly agile in the classification of normal cells; its photographic archive represents an easy-to-use instrument for review and characterization of abnormal cells.

In contrast, bone marrow cell morphology in routine diagnostics still largely depends on manual evaluation. The development of AI in this field is hampered by the fact that its results are poorly parameterized, and when they are, they lack good standardization. Bone marrow contains a wider variety of cells with subtle differences, whose interpretation may often be context dependent. The multilayer composition of bone marrow particles poses an additional challenge for automated microscopy, which can be overcome by various technical solutions. As an example, Systems by Scopio Labs (Tel Aviv, Israel) combine a series of images into a reconstructed high-resolution image, followed by automated recognition and classification of cells and/or manual analysis [30]. Manual microscopic evaluation of bone marrow requires considerable amounts of time, even by experienced morphologists, and is subject to significant inter-observer variability. Reliable AI-assistance would therefore have a great impact. In this optic, several studies have aimed to develop digital evaluation of bone marrow smears. There are two main approaches: the first focuses primarily on accurately identifying specific cell types. Matek et al. [31] employed ANN-based approaches for the automated identification of cell types in bone marrow samples [32]. This pilot study showed promising results though only limited validation was performed. Similarly, Wang et al. [32] used an ANN-based method for the automated identification of cell types in bone marrow samples, reporting a recall and accuracy of 0.842 and 0.988 respectively in an internal but independent dataset. Currently, haematological centres around the world are developing algorithms to integrate the first commercial digital microscopy tools into their diagnostic workflows. These tools support cell differentiation and assess both the quantity and quality of the different haematologic lineages in bone marrow aspirates.

A second approach is to focus on improving morphological diagnosis for specific clinical questions. Tang et al. [33] published a method for the automated distinction of abnormal from normal lymphocytes based on colour, texture and geometrical cytological characteristics. The authors reported a positive predictive value of 99.04% for the identification of abnormal lymphocytes, though in a very small test set. Eckardt et al. [34] used DL to distinguish acute myeloid leukaemia (AML) from healthy bone marrow smears, achieving a sensitivity of 87% and a specificity of 89% in an internal validation set. The same team used DL to specifically detect acute promyelocytic leukaemia (APL), an AML subtype characterized by specific genetic abnormalities, in, which early initiation of specific treatment with all-trans retinoic acid is required to improve patient outcomes. This study was however designed as feasibility study, lacking an independent test set [35].

The diagnosis of MDS relies heavily on morphological assessments of dysplastic changes in cell lineages, which are subject to high inter-observer discrepancies [36]. Accordingly, several ML approaches were developed to detect signs of dysplasia in individual cells, either within a single cell lineage [37–39] or across multiple lineages. As an example of the latter, Lee et al. [40] used DL to detect dysplasia in granulocytes, megakaryocytes and erythrocytes. These studies highlight yet another potential application of AI in enhancing the objectivity and reproducibility of bone marrow smear evaluation.

Flow Cytometry

FC is a technique that analyses and quantifies characteristics of cells in fluid suspension as they pass through a laser beam. It utilizes fluorescent-labelled antibodies to specifically target and identify cell surface or intracellular markers, allowing complex analysis of cell populations from heterogeneous samples. FC is critical for the diagnosis and follow-up of haematological diseases such as chronic and acute leukaemia, multiple myeloma, and lymphomas. The clinical interpretation of complex FC data relies on specialized software for the selection of the cells of interest (“gating”) and the study of their antigen expression [41]. In traditional gating, a series of two-dimensional plots are used to visualize the data, and hierarchical gates are manually set to select specific cell populations [42]. The quality of the process strongly depends on the judgement and experience of the operator. The process is hard to generalize, standardize, and transfer between laboratories and between FC instruments designed by different manufacturers. Manual gating allows for the incorporation of biological and clinical knowledge into the analysis. However, the two-dimensional plots in manual gating often fail to show the complex underlying structure of the data, becoming severely limited with the use of large panels, which may include up to 20–50 parameters for each cell. Moreover, the (human) operator is at risk of introducing bias or errors into the analysis, with high interpersonal variability. This leads to limited reproducibility and objectivity, and can strongly affect outcome prediction [43, 44].

Therefore, FC seems to be an ideal candidate for the implementation of AI-augmented diagnostics. Over more than a decade, a plethora of ML-tools and packages have been developed and found application in research but have failed thus far to be implemented in routine diagnostics [42, 44–50]. Most tools are based on clustering algorithms [51, 52], where cells are first grouped based on similarity and then assigned specific identities in a second step. With additional tweaking, these models can directly label cells in new samples [53], and even detect abnormal cells that were not included in the training set [54].

Beyond the identification of cell clusters, several groups have developed ML algorithms for the diagnosis of lymphoid neoplasia. One model published in 2012 by Zare et al. [49] was able to distinguish mantle cell lymphoma from small lymphocytic lymphoma with 100% sensitivity and 92% specificity. More recently, Zhao et al. [55] published a CNN-based algorithm trained to distinguish the most frequent leukaemic lymphoid neoplasms with circulating cells (chronic lymphoid leukaemia, marginal zone lymphoma, mantle cell lymphoma, prolymphocytic leukaemia, follicular lymphoma, hairy cell leukaemia, and lymphoplasmacytic lymphoma). The model’s performance was promising in an independent dataset, though precise indications on sensitivity and specificity are lacking. Another study by Ng and Zuromski [56] claimed greater than 95% sensitivity in detecting any B-cell malignancy, though low specificity would necessitate human verification of all positive cases.

Parallel to developments in cytomorphology, Monaghan et al. [57] published a method for the rapid diagnosis of acute leukaemia using FC data. Their algorithm was trained to automatically classify cytopenia samples into AML, APL, ALL, or non-neoplastic cytopenias. However, the method was developed on a single platform, with strictly standardized antibodies [57], severely limiting its applicability to other laboratories. In addition, the reported performances of the models are a result of cross validation, and were not tested on independent test data. Until independent validation – preferably across multiple platforms, the true performance is therefore unknown.

Minimal or measurable residual disease (MRD) detection for the follow-up of acute leukaemia is a critical application of FC. In 2016, an ML-based algorithm capable of detecting MRD in AML was published, but was restricted to AML with a specific aberrant expression pattern [58]. In 2019, Li et al. [59] published an ML-based algorithm detecting MRD in AML and MDS. While both authors cited high performances, independent validation was omitted. In a study published in 2018 by Ko et al. [50], a labelled ML approach led to the automated distinction between AML, MDS, and normal samples with high accuracy across two different platforms. The team used diagnostic FC data for training and validation, while post-induction data was used for testing the model. Interestingly, the model showed high prognostic power when applied to post-treatment FC data from AML patients [50]. Reiter et al. [60] developed their own ML approach and compared it to other existing ML algorithms for the detection of MRD in paediatric B-ALL using FC data provided by three different laboratories, highlighting the importance of benchmarking and cross-platform compatibility. A more modest yet highly practical approach involves the automated generation of radar plots that “pre-gate” cells, requiring subsequent manual evaluation. This strategy does not eliminate the need for human expertise but could serve to streamline and standardize the assessment of MRD [61].

ML-based algorithms in haematological FC hold great potential. However, cross-platform compatibility imposes a major limitation on both the establishment and validation of any method. While some approaches have been proposed to enable better cross-platform compatibility [62, 63], most published ML-based algorithms are trained on a single platform. As a result, many AI-based algorithms initially show spectacular results but fail when expanded to other laboratories or broader patient groups.

In general, the relatively small cohort of patients used for the demonstration of most models’ capabilities does not allow for an adequate exploration of their sensitivity and specificity, especially when taking into account numerous important factors, such as disease subtype, patient age, and general health status. Thus, while fully automated FC interpretation is unrealistic, automated gating may be a potential aid in the intermediate future [61].

Molecular Genetics

The advent of massive parallel sequencing approaches over the last decades has boosted our understanding of haematological malignancies as well as non-malignant, mutation-driven haematological disorders. However, the comprehensive genomic perspective on haematological diseases has highlighted the complexity of accurate data analysis and interpretation, which is essential for routine clinical diagnostics. The shift from individual marker analysis (e.g., Sanger sequencing) to high-throughput technologies – such as large next-generation sequencing panels, whole exome sequencing, whole genome sequencing, long read sequencing, and ATAC sequencing – has greatly increased the complexity of data processing, variant calling, and interpretation in clinical settings, to the extent that manual handling is no longer feasible. AI approaches are therefore anticipated to facilitate the accurate evaluation of complex molecular data, to enhance efforts for timely diagnosis, prediction of clinical outcomes, monitoring of MRD, and to support clinical decision-making. Data analysis from next-generation sequencing applications in haematology includes several sequential steps. Firstly, individual sequencing reads are aligned to a reference genome, a process integral to most sequencing platforms. Secondly, variants are identified by “variant calling,” either performed on the same platform (e.g., Ion Reporter for Ion Torrent systems, Dragen for Illumina systems, CLC Workbench for QIAGEN) or by independent applications. Solutions offered by the sequencing platforms may be best suited to specific technical challenges but are not always tailored to the requirements of particular sample types. This is especially important in follow-up detection of variants at low allele frequency (VAF), as, for example, in MRD analyses. These special requirements distinguish haematological genetic diagnostics from constitutional genetics. A particular challenge is the characterization of mutational markers at VAF 40–60%, which may either relate to somatic mutations characteristic of haematologic malignancies or to a germline variant not suitable for follow-up assessment. Recently, AI, including ML and DL methods, has been increasingly implemented to process the large datasets generated by high-throughput sequencing [64]. Furthermore, advancements such as biomarker identification, liquid biopsy approaches, and the characterization of relationships between molecular profiles and therapeutic responses will benefit from the integration of AI approaches into current diagnostic workflows in molecular genetics and cytogenetics in haematology [64–66]. Various commercial platforms offering AI components are currently available and under constant refinement, either as part of sequencing platforms or standalone applications.

Equally, AI solutions are progressively implemented in cytogenetics. For over a decade, laboratories have routinely used AI-assisted fluorescence in situ hybridization analysis platforms (e.g., MetaSystems MetaCyte, Applied Spectral Imaging) to automate scanning, signal counting, and preliminary classification, which markedly reduces manual counting time [67–71]. In recent years, several laboratories have implemented semi-automated karyotyping software that can separate and group chromosomes, pre-assign karyotypes, and flag numerical abnormalities for review by an analyst, although structural rearrangements still require full manual interpretation. Currently, software companies in the field of karyotyping strive to resolve these challenges [72–75]. For microarray-based analysis and optical genome mapping, AI-assisted algorithms are applied mainly for quality control, aberration detection, and preliminary copy number variant classification, before expert review and clinical correlation. Microarray-based analysis and optical genome mapping face significant challenges in automation due to varying analysis parameters based on indication and data quality which may differ between runs. For example, the significance of a small copy number aberration depends on whether it includes at least one gene pertinent to the specific malignancy. Across all modalities, final interpretation remains the responsibility of a board-certified haematologist or geneticist, in line with accreditation standards.

Despite notable progress in recent years, several challenges need to be addressed to unlock the full potential of AI in genetic haematology diagnostics. In molecular genetics, high-throughput sequencing of haematological malignancies produces heterogeneous datasets with variable depth, quality, and complexity, influenced by tumour heterogeneity, clonal evolution, and frequent low VAF. AI-based variant calling and annotation tools can aid in data reduction, but they are limited by the need for curated training datasets and cannot yet reliably distinguish between clinically relevant low-VAF somatic variants and benign artefacts or germline variants without expert oversight.

In cytogenetics, the diversity of structural chromosome rearrangements in haematological malignancies – often involving complex and cryptic abnormalities – limits the accuracy of automated recognition. AI can reliably perform numerical chromosome sorting and detect supernumerary chromosomes, but current algorithms have insufficient sensitivity for balanced translocations, small insertions, or subtle structural aberrations without manual verification. Furthermore, AI-based disease classification for haematological malignancy karyotypes is currently hindered by limited standardization of data, complexity of abnormalities, integration requirements, regulatory constraints, and the interpretability gap. Future progress will depend on multicentre, large-scale, standardized image datasets paired with rich clinical metadata, along with AI models designed for explainability and regulatory compliance.

Haemostasis

Since 2018, there has been a notable increase in the number of AI-related publications in the field of haemostasis and thrombosis [76] with figures reaching over 280 publications in 2024. AI and ML have found various applications in haemostasis laboratories, such as automation, quality assurance/control, interpreting test results, establishing personalized reference ranges, and developing diagnostic pathways using decision trees [76–79]. In the area of pre-analytical haemostasis, supervised ML approaches and backpropagation algorithms have been used to differentiate between clotted and non-clotted samples [80]. A significant number of studies have applied ML techniques to predict coagulation disorders and their severity. AI in the form of an ANN can be used to distinguishing between healthy individuals and patients with antiphospholipid syndrome (APS), based on the results of thrombin generation. Its high sensitivity makes it very effective in screening for APS [81]. ANNs achieve high accuracy in predicting the risk of pulmonary embolism/deep vein thrombosis based on clinical and laboratory factors, with models requiring between 15 and 39 factors for good accuracy [82, 83]. In a study on venous thromboembolism (VTE), an ANN integrated 64 clinical parameters and 62 genetic variants, emphasizing three key genetic factors: MTHFR C677T gene polymorphism, factor V Leiden, and intracellular adhesion molecule 1 [84]. However, a large randomized trial has already disproved the association of MTHFR C677T polymorphism with increased VTE, putting the reliability of this ANN study into question [85]. Moreover, AI has been employed to predict platelet phenotypes using microfluidic models and to classify morphological changes in activated platelets through FC [86]. However, it is currently unclear whether the sensitivity of this method in diagnosing platelet functional disorders or monitoring antiplatelet drugs is sufficient.

Overall, AI applications in haemostasis face significant challenges due to the intricate nature of the field, which may explain the relative scarcity of ML studies in this field. Haemostasis involves proteins, cells, and fluid dynamics, which pose unique challenges. In addition, haemostasis tests are subject to significant variability due to factors such as genetic differences, dietary habits, and the use of antithrombotic medication, all of which must be accounted for in developing a robust AI-based diagnostic tool. Moreover, in the field of haemostasis, datasets often lack good representativeness or external validation and sometimes even internal testing sets [76, 87, 88]. These limitations must be addressed before routine implementation can be considered.

Clinical Decision Support Systems

Ideally, any AI-based approach to diagnostics in haematology should not be limited to one modality but should integrate information from all possible modalities (see also Fig. 1). Systems capable of integrating information from multiple modalities and aiding in diagnosis or clinical decisions are collectively called “clinical decision support systems” or CDSS. A CDSS, as understood in the majority of current literature, is the informatics implementation of a decision algorithm in a specific clinical setting. Modern clinical informatics systems have the ability to seamlessly integrate CDSS. These are not limited to diagnostics, though several CDSS aim to improve diagnostic processes, such as pop-up warnings when ordering an inappropriate combination of iron studies [89] or when unnecessary repetition CBC differentials [90] is detected. The second example, published by Mahowald et al. [90], led to quantifiable cost savings and improvement in turnaround times for adequately ordered CBC differentials. In another example, Westbrook et al. [91] published a CDSS for the automated detection of heparin-induced thrombocytopaenia, based on the dynamics of thrombocyte counts. In this study, however, the CDSS failed to improve appropriate HIT testing due to both low sensitivity and low specificity and increased alert fatigue [91]. In the examples above, CDSS was based on explicit, evidence-based or expert-written algorithms. However, CDSS may serve as a means to implement decisions based on statistical models, including ML [92, 93]. Complex CDSS, containing one or several implementations of ML, could facilitate the integration of genomic data for cancer patients [91], where ML is used for statistical modelling and clinical prediction, but also for continuous and automated inclusion of newly published data. Recent advancements have positioned LLMs as powerful and versatile tools for automating the integration of diverse data and unstructured modalities, including unstructured non-English text [94, 95], with the possibility of automated updating. Current versions of general LLMs demonstrate remarkable diagnostic “reasoning” despite not being explicitly for this purpose [96]. In a recent comparison between two expert-based CDSS and two LLMs, the CDSS still outperformed the LLMs [96]. However, various ways of enhancing the performance of an LLM in a specific domain (for example, advanced prompting, fine tuning, or specific retraining), as well as the possibility of hybrid approaches, may lead to further improvements in the short term [97]. As an example, RDguru is an “agent” built on multiple LLMs, designed to assist in the diagnostic process of rare diseases. The proposed tool is able to process natural language descriptions, suggest differential diagnoses, answer questions on rare diseases, and provide at least some information on the underlying reasoning or sources [98]. Conversely, limited repeatability and the potential for the random generation of false content (“AI hallucination”), as well as the lack of reliable explanations, pose limitations to the widespread adoption of LLMs in a medical diagnostic setting [96]. In addition, LLM-based CDSS may be highly reliable in some medical domains while showing significant weaknesses in other domains [97]. Theoretically, LLMs have a strong capacity for processing and interacting in languages other than English. However, due to bias in the training data, true performance in non-English languages is often inferior [97]. Furthermore, LLMs are prone to biased prompt formulation and may repeat or even amplify biases or discriminating tropes historically present in the training data [97]. Finally, any CDSS deployed in clinical routine should add a tangible benefit to patients. A meta-analysis found that most controlled trials studying an intervention with a CDSS only led to small improvements in the targeted process of care [99]. This may be improved with careful selection of the target process. The highest benefit is likely to be found in a process where there is low guideline adherence and potentially strong impact on clinical outcome [99].

Fig. 1.

Conventional workup: clinicians aggregate individual diagnostic results including clinical information and all laboratory values. Clinicians then weigh and integrate this information and decide on the most likely diagnosis and prognosis, allowing them to communicate the diagnosis and necessary next steps to the patient. AI-enhanced workup: results from single or multiple modalities are aggregated and, where applicable, categorized. Abnormal elements or patterns and risk constellations are flagged. In some cases, possible diagnoses are suggested. These outputs are subject to verification by laboratory staff before being reviewed by clinicians, who refine and contextualize the information prior to patient communication.

In general, the systematic evaluation of LLMs in clinical diagnostics is lagging behind rapid developments. Consequently, tools for the systematic evaluation of studies on AI-based CDSS are as relevant as their initial development [100, 101].

Ethical and Regulatory Considerations

Although AI presents substantial potential in diagnostic applications, it equally introduces significant risks and ethical concerns. Major risks are bias and discrimination in AI-based algorithms. Such bias potentially leads to missed or incorrect diagnoses [102]. This is especially true for minority groups and may exacerbate inequalities [103]. While all AI tools are susceptible to bias, those based on LLMs are especially prone to reproducing, and even amplifying, existing biases and stereotypes [104]. Examples come from the field of radiology, where research indicates that AI tends to exacerbate biases present in training data [97, 103]. Biases can be introduced at every step of development, validation, implementation, and utilization of an AI diagnostic tool. The potentially harmful consequences of these biases can only be avoided by active mitigation at every step of an AI-based application’s life cycle – from initial conception to deployment and regular updating [103]. Algorithmic bias is further exacerbated by a lack of explainability. Technologies based on DL are inherently inexplainable, a limitation often termed the “black box problem.” A model may outperform humans in the detection of specific diseases, but the underlying reasoning cannot be ascertained [105]. This leads to increased difficulties in detecting potential medical errors [105]. Furthermore, lack of explainability leads to difficulties in assigning accountability and liability [106]. Physicians are morally responsible for their decisions and actions and are therefore morally accountable for any harm resulting from diagnostic mistakes. However, with limited control or even understanding of the mechanisms underlying an AI-based diagnostic tool, the assignment of accountability and liability is challenging [102, 106, 107]. Another concern with moral and legal implications is privacy and data protection. Any software receiving patient data requires robust security to protect it from hacking [102, 107, 108]. AI solutions that rely on remote data processing may not adequately protect confidential data. Furthermore, the use of LLM in the medical field may lead to inadvertent leakage of patient information into the LLM’s database and thus the public domain [109]. According to the principle of autonomy, all patients should be given complete information on the risks of a diagnostic procedure, which includes potential risks due to the involvement of an AI-based system [108, 110]. However, with progressive integration of AI into clinical routine [110, 111], this aim seems increasingly impossible. Transparent and clear communication with the general public is therefore paramount. Another concern is unequal access. AI-based medical applications are not accessible equally across society [107, 108], leading to a further increase in social gaps. Finally, the widespread application of AI in diagnostics may lead to deskilling [112–114]. Any new technology leads to some skills becoming obsolete, which is not always detrimental. However, overreliance on AI further exacerbates the risks of bias, as well as the difficulties of accountability and liability [112, 113]. Deskilling not only leads to vulnerability in the event of unexpected unavailability of an AI diagnostic aid but also reduces the chances that clinicians will detect or correctly identify new or unusual presentations [113]. Adequate training of future haematologists and laboratory workers will therefore be needed to avoid loss of critical skills. On a broader scale concerning the risks of AI, Slattery et al. [115] proposed a repository and system for classification of AI risks, currently available as preprint.

In light of the numerous potential ethical issues emerging with the use of AI in health, numerous governmental and nongovernmental entities, including the US Food and Drug Administration and the WHO, are working to provide ethical guidelines for the development and deployment of AI in health [103, 111, 116–122]. Recently, the WHO has published specific guidelines for the use of LLM in health [123]. However, in the absence of legally binding regulations and governmental oversight, widespread application of these principles, especially by commercial enterprises, is unlikely [123]. In most countries, there is currently very limited regulation on AI in health [124]. In the USA, AI-based health technologies fall under the regulatory scope of software as a medical device [117]. Importantly, software providing clinical support or recommendations for healthcare professionals is not a software as a medical device and therefore does not fall under these regulations [117]. In the European Union, the current medical regulations, under which medical software falls, do not address AI in detail [125]. A new set of regulations is expected to be adopted soon in the EU, with the intention of regulating the use of AI in high-risk domains, including health [124]. While being praised for pioneering a legal framework for AI, the text is also criticized for leaving large gaps for interpretation [126]. Overall, governments are struggling to keep up with the rapid technological evolution [126, 127], and understaffing of regulatory bodies undermines effective enforcement where clear legislation exists. Regulations are therefore often limited to guidelines, voluntary standards, and codes of conduct [117]. As such, a high responsibility falls on the developer and the user [117].

Conclusion

AI-based tools are increasingly applied across haematologic diagnostics, including peripheral blood counts, cytomorphology, FC, genomics, and haemostasis. In addition, CDSS have been developed to help integrate multimodal information and thus assist diagnostic process. However, the level of maturity of these tools varies widely. Automated analysis of blood films is already in routine use with human oversight, while applications for bone marrow smears, FC, and haemostasis remain limited by technical, validation, and interoperability challenges. Across all domains, rare diseases, underrepresented data, and lack of external validation risk bias and reduce generalizability, highlighting the need for large, multicentre collaborations, robust datasets, and standardized clinical information integration. In the foreseeable future, AI will most likely serve as an assistive technology, speeding up analyses and improving reproducibility, while expert haematologists retain the final interpretative role.

Conflict of Interest Statement

A.S.K. received a research grant from Gilead. M.A. has received honoraria from AbbVie, AstraZeneca, BeiGene, Lilly, and Stemline and travel support from Johnson & Johnson, Lilly, and Roche. A.A.S. received research grants from Silence Therapeutics support for haematologic congress participation from Lilly, Novo Nordisk, and SOBI and is involved in patents received/pending (Silence Therapeutics, University of Bern). S.C.M. received reports grants from Swiss National Science Foundation and from the Foundation for the Fight Against Cancer and has consulted for and received honoraria from Novartis, Celgene/BMS, GSK, Ajax Therapeutics Inc., OrPha Swiss GmbH, AbbVie AG, and Amgen; in addition, S.C. Meyer has a patent for PAT058952-US-PSP pending and a patent for PAT058953-US-PSP pending. U.B. was a member of the journal’s Editorial Board at the time of submission. B.M., G.W., J.T., N.K., N.P., and A.R. have no conflicts of interest to declare.

Funding Sources

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Author Contributions

A.S.K., U.B., A.R., M.A., G.W., N.P., B.M., N.K., J.T., S.C.M., and A.A.S. made substantial contributions to the conception and design of the work, participated in the acquisition and interpretation of the literature, and were involved in drafting and critically revising the manuscript for important intellectual content. A.S.K. was provided the figure. All authors approved the final version to be published and agreed to be accountable for all aspects of the work, ensuring its accuracy and integrity.

Funding Statement

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.