Interpretable prediction from longitudinal clinical data requires more than accurate risk estimates: a model should preserve meaningful domain structure, reveal how risk evolves over time, and remain auditable when transferred across heterogeneous healthcare settings. We address this challenge with a Knowledge-Guided Temporal Fuzzy Neural Network (KG-TFNN) that integrates clinically anchored yet trainable fuzzy states, explicit temporal fuzzy-risk and observation-process descriptors, parsimonious additive and cross-feature Sugeno inference, and attention-based temporal aggregation within a single auditable prediction pathway. Unlike approaches that rely primarily on latent representations or post-hoc explanations, KG-TFNN exposes membership evolution, feature-level risk contributions, native rule firing, temporal contributions, and parameter drift for direct structural analysis. We evaluated the framework for predicting a ≧ 2-point increase in Sequential Organ Failure Assessment (SOFA) score within 6 h using patient-disjoint MIMIC-IV development and locked testing, followed by fully frozen transport to the multicenter eICU-CRD database comprising 80,239 patients across 205 hospitals. Explicit temporal representation produced the largest controlled ablation effect on discrimination (ΔAUROC = 0.0510), exceeding the contribution of knowledge-guided initialization alone (ΔAUROC = 0.0061), while structural analyses demonstrated reproducible global behavior and identifiable membership and rule evolution. Although KG-TFNN did not exceed tuned LightGBM in discrimination, it enabled forms of temporal and structural auditing unavailable from predictive performance alone and revealed important limitations in cross-database calibration and patient-specific explanation. These findings establish KG-TFNN as a rigorously evaluated framework for studying the trade-off between predictive performance, semantic interpretability, temporal dynamics, and transportability, providing a reusable design principle for trustworthy longitudinal machine learning rather than a task-specific clinical alarm.