Can artificial intelligence revolutionise haemovigilance?

A number of studies suggest that we must be wary of the risks of Gen AI use at its present state of development4. Gen AI models are known to “hallucinate,” to fabricate answers that “sound” authoritative yet are incorrect. Gen AI programs are rapidly improving but require human-in-the-loop validation. For example, misclassification in haemovigilance support tools can affect diagnosis of serious AEs or create clinician alert fatigue when every patient with a fever is reported as having an AE. Another concern with the use of AI is systemic bias. AI models are trained on existing datasets, which may inherently reflect racial, gender, or rural bias, as Obermeyer et al. reported when Black patients’ needs were underestimated due to the use of healthcare spending as a proxy for illness burden8.

Donor and recipient privacy is another challenge, since these AI models require large datasets for training and fine-tuning. Approaches, called federated learning, have been developed to decentralize the use of healthcare big data in order to maintain local control and confidentiality while models are trained on individual datasets9. Data can be analyzed locally without requiring physical data transfer, subsequently the results of the local analysis and model training can be uploaded to a central site for aggregation and further analysis. Development of data standards is another critical step in the process of building robust haemovigilance across hospitals, systems, regions, and countries.

Let us return to the question, “Can AI revolutionize haemovigilance?” I believe the answer is yes, AI can, and over time will change the way hemovigilance is conducted. AI promises to make haemovigilance tasks easier, from detection of transfusion related AEs through the use of Gen AI and onto the use of agentic AI, which can plan and execute multi-step tasks, speeding the aggregation of event reports and reporting into anonymized and centralized haemovigilance systems (ideally with human-in-the-loop review) where analysis of huge datasets can occur, and ultimately to generate widespread communication of findings to enhance patient safety. For donor haemovigilance, ML models that can predict donor AEs (e.g., vasovagal reactions) are already in use10 and AI is poised to assist with donor recruitment and retention. Models that identify AEs and flag them for clinician attention and centralization can be expected as well. From an educational perspective, Gen AI Chatbots will evolve to synthesize and serve up more accurate donation and transfusion information to donors and patients.

I expect the AI changes will occur over a period of a few years. Each new application will feel quite revolutionary as AI is applied to new tasks and provides services that were previously unavailable. Incorporation of these tools into everyday use, into the EMR for example, will require new regulatory pathways varying by country/region. Ideally, the use of AI will assist low-and-middle-income countries to engage in advanced haemovigilance processes, using smart phone apps and tools for detecting and reporting to centralized systems where further aggregation, analytics, and benchmarking can be possible. The future of haemovigilance is bright and AI will evolve to speed the process.