Public datasets such as MIMIC provide de-identified electronic health records (EHRs) that are essential for accessible and reproducible healthcare research. Yet even subtle decisions in the preprocessing pipeline can compromise the validity of a study. Temporal leakage, feature–task misalignment, biased imputation, data drift, and limited reproducibility can all lead to models that appear successful while failing to reflect the underlying clinical problem.
This session examines common pitfalls in preparing complex, real-world and multimodal EHR data and introduces practical approaches for preventing them. It is designed for both researchers with experience in medical data and participants who are new to EHR preprocessing.
What participants will do
Participants will analyse how preprocessing choices can affect clinical machine-learning results and examine failure modes such as information from the future entering the training data, features being poorly aligned with the prediction task, or missing values being handled in ways that introduce bias. They will also discuss how to recognise data drift and strengthen the reproducibility of preprocessing pipelines.
Through an interactive exchange, participants will contribute examples from their own work, formulate preventive measures, and help develop best-practice solutions for EHR data preparation. The discussion will focus not only on identifying problems but also on making preprocessing decisions explicit, testable, and easier to reproduce.
What to expect
Participants can expect a practical and discussion-oriented session led by two researchers working on reliable and reproducible AI in healthcare. The session combines methodological input with peer exchange and will give participants the opportunity to share their own experiences, challenges, and proposed solutions.
They will leave with a clearer understanding of how preprocessing can shape the reliability of healthcare AI research, a set of strategies for reducing common data-related risks, and new ideas for improving preprocessing standards. The session also offers an opportunity to connect with researchers interested in reliable and reproducible EHR-based machine learning.
No specific preparation is required. Optional background resources include an overview of the MIMIC-IV dataset and an example of planning a study while addressing preprocessing challenges: