HEIP: a hybrid entropy–ontology framework for uncertainty‑aware multilabel medical image classification

The Hybrid Entropy–Ontology–Integrated Pipeline (HEIP) is introduced as an uncertainty-aware framework for multilabel chest radiograph classification. In high-stakes clinical settings, overconfident misclassification can undermine trust and compromise patient safety. HEIP addresses this challenge through three complementary reliability mechanisms. First, a per-sample entropy-regulated gating module converts predictive uncertainty into a learned gating vector, implementing a data-dependent information bottleneck that attenuates intermediate feature activations for uncertain inputs. Second, a learned ontological co-occurrence buffer captures empirical label dependencies and introduces a soft structural regularizer that discourages clinically implausible label combinations. Third, a path-integral–inspired latent regularization term penalizes abrupt embedding deviations across training iterations, promoting stable feature-space evolution. Experiments on multilabel chest radiograph datasets demonstrate that predictive entropy correlates negatively with per-sample accuracy, supporting its role as an effective uncertainty proxy. The learned co-occurrence structure aligns with medically plausible label relationships, and calibration analysis indicates improved reliability without degradation of discriminative performance. By jointly integrating entropy modulation, ontological grounding, and latent smoothness, HEIP provides a unified reliability-oriented extension to deep medical image classifiers, emphasizing trustworthiness and semantic consistency alongside conventional performance metrics.