Skip to main content

Table 2 A comparison of the cluster quality metrics for each loss, including the highest and lowest Silhouette scores and the highest and lowest log-rank P-values across ten runs for each loss function

From: Novel deep learning-based solution for identification of prognostic subgroups in liver cancer (Hepatocellular carcinoma)

Method

Loss Function

Log-rank P-value (lowest)

Log-rank P-value (highest)

Silhouette score (highest)

Silhouette score (lowest)

Baseline-binary cross entropy loss

BCE

6.68E−04

1.38E−01

0.29

0.2

Baseline-mean squared error loss

MSE/LR

4.07E−04

1.51E−01

0.31

0.18

Clustering loss

LRC

9.11E−02

3.83E−01

0.92

0.59

Survival loss

LRS

1.89E−96

6.70E−62

0.77

0.62

Combined survival and Clustering loss-hybrid model

LRSC

1.55E−77

2.62E−61

0.7

0.59

  1. A summary of the highest and lowest Silhouette scores and log-rank P-values across 10 runs for each loss function. The log-rank P-values for MSE and BCE varied between significant and non-significant, whereas the survival-based losses produced the lowest log-rank P-values. Silhouette scores for MSE and BCE remained below 0.4, indicating low confidence in group assignments whereas the clustering-based loss LRC was able to produce the highest Silhouette scores, indicating high confidence in group assignments