Researcher at the ADAPT Centre, Dr Ramisa Hamed, recently presented her paper at the 23rd International Conference on Informatics, Management, and Technology in Healthcare (ICIMTH 2025). Titled “Data Mapping Challenges in Reproducibility of ML for Acute Kidney Injury Prediction”, the paper explores the real-world barriers that hinder the replication of machine learning models in clinical environments. The study focuses on efforts to reproduce an acute kidney injury (AKI) prediction model within the Electronic Health Record (EHR) system at St. James’s Hospital (SJH), Dublin.
Reproducibility is essential in Machine Learning for Healthcare (ML4H) research, underpinning the generalisability and clinical relevance of predictive models. However, as Hamed’s work reveals, transplanting even a well-performing model into a new healthcare setting is far from straightforward. Her research identified a range of issues that complicate this process, including differences in how EHR systems are structured, the way data is formatted, and how clinical concepts are recorded or defined. Adding to these technical challenges are the legal and regulatory constraints that shape how patient data can be used or accessed.
To bridge these gaps, Hamed and her team combined expert knowledge with natural language processing (NLP) techniques and standardised medical terminologies to map and align predictor variables from the original model with the local data. Despite these efforts, they still encountered inconsistencies such as missing data and differences in measurement units, which forced them to adapt their approach to feature selection and data conversion.
Speaking about the research, Dr Hamed said: “Our findings highlight how challenging it is to reproduce ML models across healthcare systems, not just technically, but also from regulatory and semantic standpoints. These barriers must be tackled systematically if we are to ensure machine learning can be safely and reliably scaled in healthcare settings.”
The research illustrates how important it is to bring together technical expertise, clinical insight, and robust data frameworks when attempting to deploy machine learning across institutions. It also serves as a reminder that even the most promising models need to be validated and re-evaluated in each new context to ensure their accuracy and usefulness.
The research was supervised by Prof. Lucy Hederman, Prof. Gaye Stephens, Prof. Donal Sexton, and Prof. Mark Little, and supported by the ADAPT SFI Research Centre at Trinity College Dublin along with the SPARK Innovation Fund at St. James’s Hospital.
The full paper is published in the ICIMTH 2025 proceedings, Global Healthcare Transformation in the Era of Artificial Intelligence and Informatics, and is available via IOS Press Ebooks – Global Healthcare Transformation in the Era of Artificial Intelligence and Informatics
