Toward universal representations of the living: Physiological invariance for transportable medical AI.

Medical artificial intelligence (AI) models often degrade when deployed beyond their training environment, suggesting reliance on context-specific correlations rather than stable physiological structure. This article considers whether improved transportability may require representation learning strategies aligned with biological mechanisms expected to persist across populations, devices, and care pathways. Physiological invariance is introduced as the hypothesis that outcome-relevant predictive relationships may be mediated by latent physiological processes that are more stable across environments than observed measurements shaped by workflows or data acquisition. Multimodal self-supervised learning combined with mechanism-informed regularization may help identify such environment-stable structure, although empirical validation remains limited. Physiological invariance is not proposed as a sufficient or necessary condition for generalization, but as a candidate structural explanation for transportability in domains where shared biological mechanisms exist.