Beyond Accuracy: Why Replacing Traditional Scores with AI in Thrombosis Care Creates New Ethical and Regulatory...

Clinical prediction scores have been central to thrombosis and hemostasis practice for decades, providing transparent, interpretable, and validated approaches to risk assessment. Increasingly, artificial intelligence (AI)-based predictive models are being developed to improve risk stratification by integrating large, complex datasets and identifying patterns beyond conventional statistical models. However, improved predictive performance alone does not determine clinical value, and AI introduces new ethical, regulatory, and governance challenges. This narrative review compares traditional thrombosis risk scores and AI-based predictive models, focusing on methodological differences, ethical considerations, regulatory requirements, and the competencies clinicians require for responsible implementation. Traditional scores benefit from simplicity, transparency, and ease of clinical interpretation but have important limitations, including modest predictive performance, limited adaptability, potential bias, and uncertain evidence that score-guided management improves outcomes compared with clinical judgment alone. Many of the concerns currently associated with AI-based risk prediction, including bias, generalizability, validation, and monitoring, also exist within traditional prediction models but have historically received less regulatory attention. AI-based models offer potential advantages through dynamic risk prediction, integration of multimodal data, and improved predictive accuracy; however, they introduce additional challenges related to explainability, data governance, privacy, cybersecurity, accountability, reproducibility, and postdevelopment surveillance. AI should be viewed as an evolution of clinical decision support rather than a replacement for established approaches or clinical expertise. Safe implementation requires robust validation, continuous monitoring, human oversight, transparent governance frameworks, and clinician AI literacy.