The editorial text is listed below.
Morphological assessment of blood and bone marrow smears by expert pathologists using manual microscopy is the foundation of diagnosing many haematological diseases. In the past decade, artificial intelligence (AI) and particularly its subcategory of machine learning, has set to revolutionise this important initial stage of a patient's journey. November, 2021, saw the refinement of a network algorithm previously used in the classification of peripheral blood smears to automate the evaluation of bone marrow cell cytomorphological characteristics. In a study published in Blood, Matek and colleagues used the largest dataset in the literature to date, including more than 170 000 expert-annotated bone marrow microscopy images from 961 patients with a variety of haematological diseases, to train two distinct machine learning classifiers. They showed that the classifiers outperformed a feature-based approach to cell classification and achieved excellent accuracy for many cytomorphological features of clinical relevance. External validation was done in an independent dataset, although this process was compromised by the scarcity of large, annotated bone marrow image sets.
These findings represent the introduction of support tools that will soon aid diagnostic investigation in digitalised haematology laboratories.
Beyond the qualitative assessment of cellular morphology in images, automation can facilitate the quantification of features and aid the preselection of patients for molecular testing. Eckardt and colleagues showed that a deep learning algorithm could identify acute myeloid leukaemia and predict NPM1 mutation status from bone marrow smears. In the future, image-based classification tools trained on outcome data are expected to help clinical stratification and optimal treatment selection too. Automation could also help make the diagnostic process faster and more consistent, improving cost-effectiveness, and although the gold standard is predicted to remain diagnosis by haematologists, AI could help reduce the impact of shortages of trained experts. A long-lasting issue in low-income and middle-income countries, specialists are predicted to be in short supply in the USA too, according to projections by the Association of American Medical Colleges, due to challenges attracting and retaining physicians. In our field, lack of mentorship, and little exposure to haematology patients and research experience in the field during training have been discussed as having an impact on the shortage of classical haematologists. Against this backdrop, although it seems unlikely AI could ever substitute human judgement, its potential to complement the workforce and reduce workloads is very attractive.
However, many challenges remain before AI can become truly transformative. Cases used to train the model will not only determine its diagnostic accuracy but could also introduce bias and lead to wrong associations when the assumption that the study cohort represents the target population is incorrect. Additionally, many AI models are not entirely data driven; they can incorporate knowledge from guidelines and experts’ preferences that can further limit its applicability to different settings. Most haematological conditions are rare; therefore, a scarcity of large datasets to reliably train models and an abundance of expert consensus filling the evidence gaps compound these issues. As cautioned by The Lancet and Financial Times Commission exploring the convergence of digital health, artificial intelligence, and other emerging technologies with health futures, there are also important issues regarding data governance that will need to be addressed to allow the potential of these digital transformations to unfold.
In haematology, there is an ingrained appreciation of pathology as part of the diagnostic process, and it is exciting to witness the evolution of the AI-aided era in this field. The attainment of large datasets from representative populations will be key to ensure models can be applied to the target patients. There are already some international platforms being built, such as HARMONY, a European public-private partnership for big data in the haematology-oncology space, which could be used to develop more reliable AI models in the near future. However, governance issues surrounding data collection, storage, ownership, protection, and secondary use allowance will need to be carefully regulated to ensure data solidarity and the use of health data as a global public good in the imminent digital future. Matek and colleagues’ study sets the way forward by advocating for an open use of their unique data.
#bigdataforbloodcancer: Accelerating Better and Faster Treatment for Patients with Hematologic Malignancies.
The HARMONY Alliance (HARMONY and HARMONY PLUS) is a public-private European Network of Excellence for Big Data in Hematology. Our mission is to unlock and spread valuable knowledge on hematologic malignancies (blood cancers) among a large number of stakeholders, with the goal to harness and mine Big Data to speed up the development of improved treatments for patients and more effective treatment strategies.
Receive the latest news. Click here to subscribe!