Out-of-distribution detection for reliable medical AI

This thesis investigates the impact of distribution shift on clinical prediction models and examines how out-of-distribution (OOD) detection can improve the reliability and safety of medical artificial intelligence systems in clinical practice. Clinical prediction models are typically developed using specific datasets, yet their performance may deteriorate when applied in different healthcare settings due to variations in patient populations, institutional practices, data collection procedures, and temporal changes in care delivery. Such shifts can result in unreliable predictions and affect clinical decision-making.
To address this challenge, the thesis establishes a methodological foundation for applying OOD detection to medical data and demonstrates its clinical relevance across several studies. A benchmarking study evaluated existing OOD detection methods on medical tabular data under multiple distribution-shift scenarios, showing that although current methods can detect obvious OOD cases, they struggle when faced by more subtle shifts.
The thesis further demonstrates the clinical consequences of distribution shift in ICU benchmarking and fall-risk prediction. In both applications, predictive performance deteriorated for OOD patients, leading to unreliable patient-level predictions, potentially affecting downstream assessments such as ICU quality evaluation and fall prevention strategies. These findings highlight the importance of integrating OOD detection into clinical workflows.
In addition, this thesis proposes methodological improvements, including the Capturing Extreme Activations (CEA) approach and a framework for improving external validation by identifying distribution-shifted cases. Overall, this work demonstrates that strong average model performance alone does not guarantee reliable real-world deployment. By explicitly accounting for distribution shift, OOD detection can help identify unreliable predictions, improve the interpretation of model performance, and support safer, more trustworthy implementation of AI systems in healthcare.