A Reproducible Unsupervised Learning Workflow for Hippocratic Oath Profile Identification Using an Empirically Informed Synthetic Benchmark

Hananel A.; Calienes H. · CEUR Workshop Proceedings (2026)

Resumen

This paper presents a reproducible and stability-aware workflow for exploratory clustering of ordinal questionnaire data. The workflow is evaluated on an empirically informed synthetic benchmark containing n* = 150 observations and six ordinal analytical variables concerning bioethical education and perceptions related to the Hippocratic Oath. The methodological study was developed within an ongoing doctoral research program in which qualitative interviews informed the construction of an original 10-item Likert questionnaire. The associated empirical quantitative dataset comprises n0 = 38 complete questionnaire responses. The synthetic benchmark is used here as a computational test bed for developing and auditing the multivariate workflow; its n* = 150 observations are not interpreted as additional empirical participants or as increasing the inferential power of the doctoral sample. Standardized Pearson-based Principal Component Analysis (PCA) is used as the primary low-dimensional representation and is complemented by a polychoric-correlation sensitivity analysis. The first two Pearson components explain 61.94% and 29.54% of standardized variance, respectively (91.49% cumulatively), and both satisfy the eigenvalue-greater-than-one retention criterion. Cluster number is evaluated using Within-Cluster Sum of Squares (WSS), average silhouette width, and the Gap statistic, with all three criteria selecting k = 3. The resulting three-cluster solution has balanced sizes of 50/50/50, a mean silhouette width of 0.85470, and perfect agreement between k-means and Ward clustering (ARI = 1.00000). The same partition is recovered across 100 random seeds, leave-one-out perturbations, cluster-wise bootstrap stability assessment, PAM, and the polychoric sensitivity representation. A four-cluster solution, examined as a sensitivity analysis, shows a lower mean silhouette width (0.74379), a k-means-Ward ARI of 0.97879, and a smallest cluster of 16 observations. Because the results are obtained from a synthetic methodological benchmark, the recovered profiles are not interpreted as empirical subgroups of the doctoral participants or of the broader medical-student population. The contribution is methodological: an auditable workflow that integrates dimensionality reduction, cluster-number selection, cross-algorithm agreement, perturbation stability, measurement-scale sensitivity, and substantive characterization, providing a prespecified analytical framework for subsequent application to empirical questionnaire data.

Instituciones: Universidad Peruana Cayetano Heredia; Universidad Católica Santo Toribio de Mogrovejo