Multimorbidity is a growing global health challenge that complicates patient management because of heterogeneous disease combinations and complex clinical trajectories. Existing multimorbidity phenotyping approaches often rely on predefined disease groupings or conventional clustering techniques that inadequately capture latent relationships among patients, limiting the discovery of clinically meaningful phenotypes. We developed an interpretable artificial intelligence framework to discover clinically meaningful multimorbidity phenotypes from routine electronic health records by integrating structured clinical data, disease representation learning, network analysis, statistical validation, and explainable artificial intelligence. The framework was evaluated using 15,732 patient encounters and identified five distinct multimorbidity phenotypes with significantly different demographic and clinical characteristics. Statistical validation demonstrated significant differences in age (ANOVA, F = 89.27, p < 0.001), waiting time (ANOVA, F = 129.84, p < 0.001), healthcare setting (χ² = 1192.84, p < 0.001), and ICD-10 disease distributions (χ² = 777.02, p < 0.001), confirming the robustness and clinical distinctiveness of the identified phenotypes. Explainable artificial intelligence further revealed that prior healthcare utilization, laboratory investigations, medication-related variables, and disease-network embeddings were the principal determinants of phenotype membership, enabling transparent interpretation of patient stratification. By combining representation learning with disease network modelling and explainable AI, the proposed framework transforms routinely collected electronic health records into clinically interpretable multimorbidity phenotypes, providing a scalable foundation for precision medicine, clinical decision support, and population health management.