Clinical meaningfulness of unsupervised clustering groupings

Determine whether patient groupings produced by unsupervised clustering can be interpreted as clinically meaningful when the clustering solutions are unstable across bootstrap samples.

Background

The paper evaluates k-medoids, agglomerative, and spectral clustering of 4,509 patients in the CENTER-TBI cohort under different distance metrics and cluster-number selection criteria. The resulting solutions differ substantially across analytic choices and are often not reproduced in bootstrap samples, indicating that apparent patient subgroups may depend strongly on methodological decisions and the particular observed sample.

Against this backdrop, the authors explicitly question whether an unstable grouping can represent a genuine clinical pattern in the broader patient population. The paper does not resolve whether any such grouping is clinically meaningful; instead, it concludes that clustering results should be treated cautiously and not overinterpreted in clinical research.