Speakers
Prof Frank Visseren, University Medical Centre in Utrecht
Cardiovascular risk modelling in clinical practice, success and Challenges
Frank Visseren (1965) is Professor of Medicine, Epidemiologist and Vascular Medicine specialist at the University Medical Center in Utrecht, The Netherlands. He is head of the Department of Vascular Medicine/Endocrinology/Diabetes and responsible for patient care, research, and teaching. His main research interest is in the field of (rare) lipid disorders especially familial dysbetalipoproteinemia, insulin resistance/DM2 and development of vascular complications, and in translating the results of large clinical trials and cohorts to individual patients by estimating CV risk and treatment benefit, in collaboration with (inter)national partners. He is (co-)author of >500 publications and supervised >50 thesis projects. He is the clinical director of the Second Manifestations of ARTerial disease (SMART) cohort at the UMC Utrecht. This is a cohort started in 1996 with currently >16,000 patients with vascular diseases and/or type 2 diabetes or severe hypertension or familial hyperlipidemia, growing with 800 patients annually. His research group has a track record in prediction research. He was member of several National guideline committees and Chaired the Task Force of the 2021 ESC Cardiovascular disease Prevention Guidelines and he is steering committee member of the ESC Cardiovascular Risk Collaboration.
Prof Luc Smits, Maastricht University
How many models make it? A quantitative look at the prediction pipeline
Luc Smits is a Professor of Clinical Epidemiology and Risk‑Based Care in the Faculty of Health, Medicine and Life Sciences at Maastricht University, The Netherlands. Much of his research is concerned with evaluating and implementing risk-based care strategies. With his recent textbook, Improving Health Care with Clinical Prediction Models — From Idea to Impact (2025), he provides a practical guide to developing prediction models that truly benefit patients and healthcare systems — aiming to reduce research waste and bridge the gap between model development and real-world impact.
His lecture examines the promise and limitations of clinical prediction models in healthcare. Although such models hold considerable potential — offering improvements in diagnosis, prognosis, cost-effectiveness, and operational efficiency — their real-world impact remains limited. By analyzing the life cycle of published models, he and his team have observed that only a small fraction are implemented in practice, and of those, few undergo thorough evaluation regarding their benefits or harms to patients and care delivery. He will outline strategies and methodological approaches aimed at increasing the success rate of prediction modelling initiatives and mitigating research waste.
Prof Matthew Sperrin, University of Manchester
Prediction under intervention : challenges and trade-offs
Matt Sperrin is Professor of Biostatistics and Health Data Science at University of Manchester. Matt is a statistician by background, with research interests at the intersection of prediction modelling and causal inference. He is especially interested in how causal inference can be used to allow prediction under intervention (i.e., what would be the risk IF...), and how causal principles can help to ensure prediction models are generalisable and fair. He programme directs the MSc Health Data Science, and co-leads CHAI (the causality in healthcare AI hub).
Causality and prediction are often two separate activities. In particular, prediction can be done in a way that is agnostic to underlying knowledge, mechanism or causal structure. However, it is very often useful to exploit any existing causal knowledge in the context of prediction. This is most directly the case when a decision is to be made on the basis of a prediction, where the decision will affect the risk itself (sometimes called performative prediction). In such cases, decisions are better supported by information about how the predicted risk reacts to those decisions: prediction under intervention. In this talk I will describe our attempts to build models that allow for prediction under intervention, and the inevitable challenges and trade-offs that arise. The motivating example is our recent development of a model that predicts cardiovascular risk under intervention, which is designed to be used to support decisions in primary prevention.
Dr Alicia Uiji, Amsterdam University Medical Centre
Making routine care data actionable: phenotyping and risk prediction in cardiovascular care. How can we move from large-scale EHR analyses to changes in care delivery?
This session will highlight approaches to phenotyping within routine care data and discuss strategies for embedding cardiovascular risk prediction into health systems. Practical considerations such as standardisation, federated analyse and computerised decision support systems will be highlighted as practical foundations for translating insight into care.
Prof Angela Wood, University of Cambridge
Dynamic landmark risk prediction models for electronic health records
Angela Wood is a Professor of Biostatistics and Health Data Science at the University of Cambridge and Research Director at Health Data Research UK. Her research focuses on the development and application of statistical methods for advancing epidemiological research and in utilising whole-population electronic health records. This includes developing statistical methodology for handling missing data and measurement error, multiple longitudinal risk factors, multiple imputation, risk prediction, and meta-analysis.
Dr Kym Snell, University of Birmingham
A Critical view on the use of electronic health records for clinical prediction modelling
Dr Kym Snell is an Associate Professor in Biostatistics at the University of Birmingham, with research interests in both applied and methodological aspects of prediction modelling. This includes development, validation and updating of clinical prediction models as well as the use of electronic health records and individual participant data from multiple sources for the purpose of prediction modelling.
Electronic health records (EHR) from primary and secondary care provide a valuable resource for healthcare research, offering easy and low-cost access to large amounts of patient data. This makes it an attractive and increasingly popular data source for developing and validating clinical prediction models compared to collecting prospective data cohorts. Examples of clinical prediction models developed using EHR include QRISK3, QDiabetes and eFalls, however there are limitations to using EHR, particularly from primary care records, that can lead to inaccurate clinical prediction models.
This presentation will cast a critical eye on the rising popularity of using EHR for clinical prediction models, discussing the advantages, limitations and opportunities for methodological development. Examples will be used to illustrate points drawing from experience in developing prediction models using CPRD. These will include models for predicting cardiovascular disease risk in postpartum women and models aiming to identify rheumatoid arthritis earlier in primary care. Some common issues encountered when using EHR for prediction modelling include inconsistent coding, missing predictor information and under or over diagnosis of conditions of interest. Additional challenges arise when outcomes are relatively rare. These data issues can all lead to bias and additional uncertainty in predictions, impacting the performance and reliability of a prediction model. Finally, opportunities to quantify the impact of some of the data issues will be discussed along with how we can assess whether a model may still offer some value in clinical decision making.
Dr Shishir Rao
Artificial intelligence for electronic health records: a focus on cardiovascular diseases
Dr Shishir Rao is a senior AI research scientist in Professor Kazem Rahimi's Deep Medicine research group at the University of Oxford, specialising in deep learning for multimodal healthcare data. He holds a DPhil which focussed on deep learning for electronic health records and comes from a background of mathematics and computer science.
Dr Rao co-leads AI-driven healthcare initiatives focusing on cardiovascular disease, perinatal risk assessment, musculoskeletal conditions, and heart failure research. He has co-pioneered the first foundation Transformer model for EHR - the BEHRT model. His research includes developing extensions to these Transformer-based model for spectrum of analyses from phenotyping to risk prediction to causal inference. Ultimately, Dr Rao's research focuses on bridging advanced computational methods with practical clinical applications to improve patient care through innovative AI solutions.
This talk introduces Transformer foundation models for electronic health records (EHR) and their application in cardiovascular disease risk prediction. The presentation focuses on developing AI models that can assess patients across the entire risk spectrum, from low-risk populations to high-risk niche groups such as those with heart failure. These advanced models leverage multimodal EHR data to provide comprehensive risk stratification for cardiovascular conditions. Beyond prediction, the presentation will also share how these models can be utilised for discovery research.
Dr David Kent, Tufts University
Guidance for Unbiased predictive Information for healthcare Decision making and Equity (GUIDE): considerations when race may be a prognostic factor
David M. Kent, MD, MS, is the Founder and Director of the Tufts Predictive Analytics and Comparative Effectiveness (PACE) Center at Tufts Medical Center (TMC), and Director of the Clinical and Translational Science (CTS) MS/PhD Program at Tufts University. He is a Professor of Medicine, Neurology, and CTS and works as a hospitalist at TMC. Dr. Kent’s expertise lies in clinical epidemiology, focusing on predictive analytics and comparative effectiveness, particularly in cardiovascular disease and stroke. He has received over $35 million in continuous funding from the NIH and PCORI since 2003 and has authored over 250 publications including in NEJM, JAMA, and Lancet. His research has been instrumental in developing methods for predicting heterogeneous treatment effects in clinical trials, establishing frameworks for evaluations of clinical prediction models, understanding patient selection for PFO closure after stroke (which have changed practice and guidelines worldwide), and studying covert cerebrovascular disease in routine care detected by natural language processing. He has served on multiple committees, including the NIH Steering Committee on Personalized Health Care and Prevention and the Optum Labs Scientific Advisory Board. Dr. Kent is a dedicated mentor, having guided over 60 junior faculty, fellows, and students, many of whom have moved into leadership roles in academia, industry, and government. His expertise in predictive analytics directly informs his work on The Guidance for Unbiased predictive Information for healthcare Decision-making and Equity (GUIDE), a framework to address bias in clinical prediction models. Clinical prediction models (CPMs) support medical decision-making but may contribute to racial disparities. GUIDE offers expert recommendations on evaluating and mitigating algorithmic bias, with a focus on applications in which race has prognostic importance. Developed via a 5-round Delphi process with a diverse expert panel, GUIDE affirms race as a social construct and emphasizes decisional context, competing fairness principles, and trade-offs in race-aware vs. race-unaware models.
