The field of hematology is embracing the transformative power of artificial intelligence (AI) in both diagnostics and therapeutics. However, even for the 21st century hematologist who has tried to keep up with the pace of AI applications in medicine, the technology is evolving at a dizzying pace. With this remarkable and powerful tool comes a unique set of considerations especially as it applies to its use in clinical research and therapeutics in hematology.
As AI technologies such as deep neural networks and machine learning advance rapidly, they bring a wave of innovation to hematology, particularly in areas such as disease diagnostics, prognosis prediction, and personalized medicine incorporating genome sequencing for cancer therapy.
This year’s ASH annual meeting will shine a light on AI’s role in hematology, providing a comprehensive exploration of both its significant opportunities and critical risks. The potential application of AI to both malignant and classical hematology will be showcased at the sessions summarized below. Additionally, the Special Interest Session, The Role of Artificial Intelligence in Hematology Practice: The Good, the Bad, and the Ugly will consider the restrictions, limitations, and ethical concerns associated with using AI in research, guideline development, and patient privacy protection.
Malignant Hematology
One poster abstract to watch for is Generation of Multimodal Longitudinal Synthetic Data by Artificial Intelligence to Improve Personalized Medicine in Hematology, a project conducted by GenoMed4All and Synthema. The consortia’s project leveraged generative AI to create multimodal synthetic data from real-world data of patients with myeloid neoplasms. Specifically, the study explored the critical challenges of using real-world data such as clinical information, cytogenetics, somatic mutations, and transcriptomics to develop more effective predictive models and improve patient outcomes. By generating synthetic data through advanced machine learning frameworks such as the conditional generative adversarial network (GAN), tabular variational autoencoder (VAE), and tabular generative pre-training transformer (GPT) architectures, the researchers were able to replicate the statistical properties and complexity of real-world patient data while preserving privacy.
Novel diagnostic approaches like these can revolutionize research and clinical care, enabling hypothesis testing, model validation, and the potential acceleration of clinical trials — all without compromising patient data. Importantly, the study found high fidelity between the synthetic and real data sets, as well as demonstrated nearly identical transcriptomic signatures and survival outcomes between the two, indicating the reliability of AI-generated synthetic data for clinical research. This inspires confidence in algorithmic models that leverage data analysis in furthering clinical trial design and implementation, potentially saving many human-hours through increased efficiency.
However, the potential risks of using AI-generated data are not insignificant. Despite the impressive fidelity of the synthetic data, the challenges of ensuring clinical accuracy and preserving the intrinsic relationships between data layers remain significant. While generative AI offers an innovative way to overcome the limitations of real-world data access, particularly for rare diseases, the complexity and variability of hematologic conditions present significant challenges for the widespread application of AI-generated data in clinical practice. Moreover, the ethical considerations of using this synthesized data, such as privacy concerns and the potential for bias, cannot be overlooked.
Classical Hematology
A session on thrombosis risk and outcomes will feature the abstract, The AI-Mayo PE (AIM PE) Study: Validation of an Artificial Intelligence Algorithm Using Electrocardiograms to Predict Pulmonary Emboli. Conducted by a team of researchers from Mayo Clinic in Rochester, Minnesota, the AIM-PE study employed an AI-based algorithm to predict the presence of pulmonary emboli (PE) through electrocardiograms (ECGs) in more than 18,000 patients. While the traditional approach to diagnosing PE relies heavily on imaging such as computed tomography (CT) scans, the AI-ECG algorithm demonstrated a notable predictive capacity by incorporating machine learning techniques, combining ECG data with D-dimer levels to refine risk stratification, achieving an area under the curve of 0.93 in predicting PE. This approach could significantly reduce the need for imaging, particularly for patients with elevated risk but negative D-dimer results, enhancing both efficiency and patient safety by reducing unnecessary exposure to radiation and contrast agents.
Another example of leveraging AI in hematology will be showcased with the abstract, Can Machine Learning Supplant the Plasmic Score in Improving the Early Diagnosis of Thrombotic Thrombocytopenic Purpura? It explores the accuracy of an algorithmic approach to predicting disease risk/probability. For their study, researchers employed a machine learning model that was ultimately found to achieve improved predictive accuracy, specificity, and sensitivity based on such variables as hematologic metricizes, lactate dehydrogenase, creatinine, cancer status, indirect bilirubin, neurologic deficit, age, and history of transplant.
While the potential benefits of AI in hematology are undeniable, the inherent risks are also evident in these studies. The reliance on AI for diagnostic and prognostic purposes brings forward concerns about model transparency, interpretability, and the potential for algorithms to perpetuate existing biases in health care data. AI enhances diagnostic accuracy, but there remains a need for human oversight to ensure that algorithms do not override clinical judgment. The risks of over-reliance on AI, particularly in critical care settings, are real, and safeguards must be in place to mitigate the potential for errors or misdiagnosis. It would be timely and prudent to have a regulatory body oversee the development of AI models for machine learning, diagnostic algorithms, and patient privacy protection.
As AI continues to gain momentum in hematology, its dual potential as both a tool for innovation and a source of risk must be carefully managed. The studies to be presented at #ASH24 highlight the remarkable promise of AI-driven advances (in, for example, personalized medicine and disease diagnosis), as well as the challenges of ensuring that these technologies are applied safely, ethically, and effectively. AI in hematology is undoubtedly a double-edged sword — capable of cutting through the complexity of disease, but only when wielded with care and caution.
