A standard blood test misses four out of five cases of iron deficiency, detecting the condition only once it is already advanced. This matters because untreated iron deficiency is not a minor inconvenience—it reduces vaccine effectiveness, raises maternal mortality, impairs child development, and worsens outcomes for heart attack, stroke, and cancer patients. In the UK, 30% of people with acute heart failure have iron deficiency, yet current screening using the full blood count catches only advanced disease. The delay causes avoidable harm and drives up healthcare costs; anaemia alone costs the German surgical system over €1 billion annually. The researchers have built machine-learning models that analyse individual-cell data from the same routine blood test, boosting diagnostic sensitivity from 20% to over 80% without sacrificing accuracy. They now need to confirm the models work in hospital patients, not just blood donors. If successful, the approach would require only a software update to existing lab infrastructure, enabling real-time iron deficiency diagnosis at the first blood test. That shift—from late detection to early, automated screening—could reduce hospital stays, lower mortality, and cut long-term costs across the health system.
View original technical description
Iron deficiency (ID) is a major healthcare challenge, impacting quality of life, increasing mortality, and costing global health systems billions annually. Despite being inexpensive to treat and recognised for its potential to improve life quality and reduce healthcare costs when addressed early, ID remains systematically under-diagnosed and under-treated. Tackling ID presents a significant opportunity for preventative healthcare. Context: ID has serious health and economic consequences. It reduces vaccine effectiveness, raises maternal mortality, and impairs child development. It has major detrimental impacts on conditions like heart attack, stroke and cancer leading to longer hospital stays and higher mortality. In the UK, 30% of patients with acute heart failure have ID, despite high quality evidence demonstrating that treatment improves quality of life and reduces mortality by 40-60%. Costs of ID to healthcare are huge due to ID’s compounding effects on a wide range of other conditions. The cost of anaemia to surgical care alone, when assessed in the German system, was estimated in excess of €1 billion annually. Despite these well-documented impacts, ID remains common, affecting over 5% of the UK population and up to 30% internationally. Challenge: A critical challenge is the failure of current methods to diagnose ID early. We show that current screening using the full blood count (FBC) misses 80% of cases, only effectively detecting advanced disease, and resulting in a strong bias towards delayed diagnosis. This delay causes significant, avoidable risk. Our research demonstrates this can be overcome using a novel approach to the FBC. Solution: We are developing innovative tools in collaboration with clinicians, scientists, and mathematicians to detect ID early. By applying machine learning to the high-dimensional FBC (HD-FBC) (individual-cell level data collected during FBC analysis) we have created models in blood donors that increase diagnostic sensitivity from 20% to over 80% while maintaining specificity. Our approach augments an existing ubiquitous test, making it scalable without significant investment. We are uniquely positioned to deliver, having curated extensive HD-FBC datasets (usually deleted post-analysis) and implemented a secure federated learning (FL) network allowing efficient model training across institutions. This ensures performance across diverse populations without compromising data privacy. The Gap: Ensuring we can translate model performance from blood donors to the clinical setting. Aims: To improve healthcare outcomes and reduce morbidity and mortality due to underdiagnosis of ID by improving early detection. Gap Fund Objectives: Expand our ID detection model, trained in blood donors, to clinical cohorts at Barts Health NHS Trust. Improve stability across diverse populations by using FL to train simultaneously across large clinical and donor datasets (BartsHealth, University College London, INTERVAL trial) and validate within unrelated datasets (Amsterdam UMC, STRIDES trial) Develop protocols for prospective studies These steps will confirm performance in clinical datasets, allowing us to apply for DPFS to fund prospective impact studies Applications and benefits: The developed model will automate real-time screening of HD-FBC samples, replacing current, ineffective screening methods. It will provide rapid diagnosis at initial blood test, reducing diagnostic delay and the need for follow-up testing. It is rapidly scalable, requiring only the addition of software to current infrastructure. Improved ID identification will enhance patient care and reduce long-term costs associated with untreated ID and its complications. By addressing this significant gap in ID management, our approach offers a practical and impactful advancement in the drive towards preventative care.
Plain English summaries and category classifications on this site are generated by AI and may not perfectly reflect the original research.
Is something wrong? Let us know