Description This paper examines a largely overlooked source of bias in health artificial intelligence: the assumption that disease states are static, linear, and classifiable at single points in time. Drawing on clinical practice, systems physiology, and early-warning theory, it argues that many chronic and complex diseases behave as dynamic processes characterized by fluctuation, delay, instability, and phase transitions. When AI systems are developed, validated, and implemented using static disease representations, systematic errors can be introduced upstream—before any algorithm is trained. These errors may manifest as biased predictions, mistimed interventions, safety risks, and implementation failure, even in technically well-performing models. The paper situates static disease modeling as a conceptual risk factor for bias and patient harm, and outlines why safe and equitable health AI requires disease models that account for temporal dynamics, system stability, and change over time. It is intended as a conceptual and clinical contribution to discussions on bias mitigation, patient safety, and responsible AI implementation in healthcare.
Anita Domargård (Fri,) studied this question.