Fig. 4: The upper row reveals the histograms for the 23rd representative feature (the 23rd feature has the highest positive weight value of 0.38 in the sigmoid prediction equation and one of the five representative features altered by BCR status) of our novel model after stratifying by the risk groups for one-vs.-other comparisons on the first external Validation set (CPCBN, Canada).

The target distribution is marked in red, whereas the other distribution is marked in green; the overlapped distribution is highlighted in brown. The histology images are patches selected based on the distribution patterns (dominant red range for low and high-risk group, the overlapped range for intermediate groups). Overall, the variance for feature distributions is differed by the risk groups. Specifically, the feature distribution is shifting between these risk groups. We identified a clear histopathological gradient for distortion of glandular architecture (e.g., disappearance of organized glandular architecture) by the risk groups based on these patches. p-value was estimated using the Levene Test and the two-sided significance level was set to ≤0.0001. Example histology images are captured at ×10 objective magnification (~330 × 330 µm). The supplementary section includes the entire feature distribution visualization and the access information to larger image sets representing these risk groups.