HieraMed: A hierarchical representation framework for next-visit diagnosis prediction from longitudinal electronic health records.

Next-visit diagnosis prediction from electronic health records (EHRs) is a challenging multi-label prediction task because predictive information is distributed across medical-code semantics, heterogeneous visit contents, and longitudinal patient history. This article presents HieraMed, a hierarchical method for multi-label next-visit diagnosis prediction from structured EHR data. HieraMed first converts standardized medical codes into semantically enriched embeddings using natural-language descriptions. It then integrates diagnoses, procedures, and medications within each encounter into visit-level representations, and models visit sequences with a Liquid Neural Network to capture patient-state evolution over time. The method was evaluated on MIMIC-III and MIMIC-IV. HieraMed achieved a visit-level precision@20 of 67.80% and a code-level accuracy@20 of 63.39% on MIMIC-III, and 71.65% and 65.64% on MIMIC-IV. Compared with the best baseline, it improved visit-level precision@20 by 3.02 percentage points on MIMIC-III and 2.66 percentage points on MIMIC-IV. These results indicate that organizing EHR data through a code-visit-patient hierarchy yields consistent improvements in next-visit diagnosis prediction under the evaluated benchmark settings.•HieraMed provides a hierarchical modelling pipeline for next-visit diagnosis prediction from structured EHR data.•The method integrates code-level semantic augmentation, visit-level heterogeneous event aggregation, and patient-level temporal modelling.•Validation on MIMIC-III and MIMIC-IV shows consistent improvements over strong baseline models.