Unsupervised Phenotyping Explorer
Cluster the kidney cohort in radiomic-plus-clinical space and ask whether the resulting phenotypes mean anything.
- k-means
- PCA
- phenotypes
- unsupervised
Mean silhouette width
0.175
Patients
180.000
Features used
4.000
k-means clusters in PCA space
Silhouette width by patient
Points below zero are likely misclassified.
Cluster profiles (feature medians) and malignancy breakdown
| Cluster | N | age_at_nephrectomy | body_mass_index | radiographic_size_cm | preop_egfr | % malignant |
|---|---|---|---|---|---|---|
| Cluster 1 | 72.000 | 52.596 | 24.418 | 4.372 | 73.715 | 0.0% |
| Cluster 2 | 51.000 | 66.878 | 30.255 | 4.391 | 87.389 | 2.0% |
| Cluster 3 | 57.000 | 64.834 | 31.153 | 3.176 | 61.550 | 3.5% |
Structure vs. label
Unsupervised structure reflects whatever dominates the variance of the chosen features, which need not be malignancy. A cluster with a higher malignant fraction is a post-hoc, exploratory observation — not evidence of a validated phenotype until it replicates in new data.