Abstract
INTRODUCTION: Recent advances in artificial intelligence have created opportunities for medical anomaly detection through multimodal learning frameworks. However, traditional systems struggle to capture the complex temporal and semantic relationships in clinical data, limiting generalization and interpretability in real-world settings. METHODS: To address these challenges, we propose a novel framework that integrates symbolic representations, a graph-based neural model (PathoGraph), and a knowledge-guided refinement strategy. The approach leverages structured clinical records, temporally evolving symptom graphs, and medical ontologies to build semantically interpretable latent spaces. Our method enhances model robustness under sparse supervision and distributional shifts. RESULTS: Extensive experiments across electronic health records and diagnostic datasets show that our model outperforms existing baselines in detecting rare comorbidity patterns and abnormal treatment responses. DISCUSSION: Additionally, it improves interpretability and trustworthiness, which are critical for clinical deployment. By aligning domain knowledge with multimodal AI, our work contributes a generalizable and explainable solution to healthcare anomaly detection.