Despite the excellent generalization capabilities of large-scale language models (LLMs), their severe limitations, such as illusions, lack of domain-specific knowledge, and ambiguity in the reasoning process, challenge their direct application to clinical decision support systems (CDSSs). To address these challenges, this study proposes a synergistic joint model that integrates knowledge graphs (KGs) and LLMs to enhance domain-specific knowledge and improve explainability in CDSSs. The proposed model leverages KGs to provide structured, domain-specific insights while utilizing LLMs’ generative capabilities to dynamically extract, refine, and expand medical knowledge. This bi-directional interaction ensures that CDSS recommendations remain both clinically accurate and contextually comprehensive. Performance evaluation of the joint model for mental health etiology, stress detection, and emotion recognition tasks of the CDSS showed up to a 12.0% increase in accuracy and an 8.6% increase in F1 score when compared to the standalone LLM model, with additional significant improvements when using the model with medical domain knowledge. Thus, the reliable and up-to-date domain knowledge obtained through the joint model not only improves the task performance of the CDSS, but also provides direct evidence of how such decisions were made. These findings validate the broad applicability and effectiveness of our KG–LLM joint model, highlighting its potential in real-world clinical decision support scenarios.

Forecasting research repeatedly finds that simple methods can match or beat complex ones out of sample. We test this in a safety-critical domain, near-real-time prediction of fire hazard state and structural damage, using a simulator-grounded benchmark: high-fidelity computational fluid dynamics (Fire Dynamics Simulator, FDS) provides reference data, an FDS-calibrated zone model (CFAST) generates a large corpus cheaply, and the temperature trajectories drive a finite-element model (OpenSees) and a HAZUS/Eurocode-informed damage rule. Under one protocol we compare simple, deep (PatchTST, TimesNet), and physics-informed (PINN, PIKAN) forecasters. Complex models do not dominate: a small corpus lets parsimonious models approach best accuracy, and physics helps mainly when data are scarce (crossover near twenty scenarios). We propose CD-PINN, which identifies a data-optimal reduced-order physics residual from the corpus by physics-guided regression over a candidate library and uses it as the physics constraint. This lifts a per-event physics model to the accuracy of corpus-trained forecasters while staying interpretable. On 26 laboratory-fire experiments, however, in-distribution rankings do not transfer: the large accuracy spread collapses to near-parity, so a leaderboard poorly predicts laboratory-fire accuracy. For downstream damage, we further show that the label definition, not the model class, sets the achievable ceiling.