Exploratory identification of cardiological patient profiles using clustering of categorical clinical data

Standardised nursing documentation using codified terminologies supports consistent communication, reduces data fragmentation, and enables secondary data analysis and quality improvement. However, nursing documentation remains heterogeneous and longitudinal evidence is limited. In this context, a cardiology-focused healthcare organisation comprising two hospitals implemented the ICNP™ coding system for the adult population.

Explore whether number of nursing interventions and nursing diagnosis differ among clusters of categorical clinical and organisational variables. Methods A retrospective dataset of 4,800 hospital episodes recorded between 2024 and 2025 in Italy, was analysed using R. A descriptive cluster analysis was performed to identify homogeneous patient profiles based on categorical variables (care setting, diagnosis codes, and type of admission). Pairwise dissimilarities were computed using Gower distance, and clustering was performed using the k-modes algorithm with multiple random initialisations to reduce convergence to local optima. The optimal number of clusters was selected based on the average silhouette width for k values ranging from 2 to 10. Cluster profiles were described using frequency distributions. External validation was conducted using two variables not included in the clustering procedure: number of nursing interventions and number of diagnoses. Because variance homogeneity was violated, between-group differences were assessed using Kruskal–Wallis tests with Dunn–Bonferroni post hoc comparisons. Effect sizes were quantified using rank-based eta squared.

The silhouette profile showed a clear maximum at k = 3, indicating the best balance between within-cluster cohesion and between-cluster separation. The final solution included 632 cases in cluster 1 (13.2%), 3,058 in cluster 2 (63.7%), and 1,110 in cluster 3 (23.1%). Significant global differences were observed for both nursing interventions (χ² = 151.29, p < 0.001, eta²= 0.031) and number of diagnoses (χ² = 114.51, p < 0.001, eta²= 0.024), with small effect sizes. Post hoc analyses showed that cluster 3 (median = 30 nursing interventions; median = 5 diagnoses) differed significantly from clusters 1 (median = 18 nursing interventions; median = 3 diagnoses) and 2 (median = 14 nursing interventions; median = 3 diagnoses), whereas no significant differences were detected between clusters 1 and 2 for both nursing interventions and number of diagnoses. Overall, the clustering solution demonstrates supports an exploratory interpretation.

A clinically relevant subgroup showed higher nursing workload and diagnostic complexity, although effect sizes were small, indicating limited explanatory power. These findings support the feasibility of secondary use of codified nursing data for exploration profiling and highlight the need for further validation and longitudinal analyses.