Electronic health records (EHRs) contain many kinds of multimodal data. This data includes demographics, laboratory results, diagnoses, treatments, and clinical notes. Accurate retrieval of similar patients is essential. It provides the base for clinical decision support systems, cohort identification, and personalized medicine. Traditional methods that use global embeddings often add noise from features that are not relevant to the task. These methods also lack interpretability. A cluster-guided embedding framework is proposed in this work. Features are first grouped into clusters that have clear clinical meaning. Examples of such clusters are ophthalmic examinations, systemic laboratory tests, and medication history. Specialized embeddings are then learned for each cluster. These embeddings are combined through fusion. The framework builds on representation learning techniques that are already established. The main contribution is the clustering strategy that uses clinical knowledge. This strategy is designed for ophthalmology workflows in a large regional EHR system. Systematic evaluation is performed for retrieval accuracy and for interpretability. The framework is applied to approximately 5000 ophthalmology patient records. The records come from BioArc, which is the largest EHR platform in western Iran. Results are obtained. Cluster-specific embeddings perform better than global baselines. The improvement appears in retrieval precision, recall, and clinical relevance. The difference is clear and consistent across metrics.
