Perceived vs. True Emergence: A Cognitive Account of Generalization in Clinical Time Series Models

Understanding how deep learning models form high-level states is a central challenge and the cognitive science of emergence provides a promising framework for interpreting these internal processes. We investigate how a neural network can learn to perceive emergent states from complex clinical time series using an information-theoretic objective $\Psi$ that balances temporal predictability and abstraction. We introduce a framework that distinguishes perceived emergence, a model’s ability to identify emergent patterns within its training environment ($\Psi>0$; In-Distribution), from true emergence, the persistence of these patterns under distribution shift ($\Psi>0$; Out-Of-Distribution). We evaluate this framework by reciprocal training and verification across two large critical care datasets, MIMIC-IV and eICU, comprising 63 harmonized variables. Our experiments demonstrate that models trained with $\Psi$ capture perceived emergence within their training environments and also exhibit true emergence across datasets, indicating robust generalization. We provide a processing account of this generalization by analyzing the internal mechanics of the learned representations and the stability of their mutual information under distributional shift, thereby contributing to a clearer understanding of how such models may achieve out-of-distribution generalization in clinical settings.