Why the model decides what it decides
SHAP (SHapley Additive exPlanations) decomposes each prediction into per-feature contributions. Across the cohort, the model has learned a clean story — one feature dominates, and it has a threshold.
Top predictor
Tumour Area (cm²)
SHAP 0.141
Clinical threshold
1.5 cm²
ellipse tumour area
High-confidence acc
91.4%
when |p−0.5| ≥ 0.15
Global importance
Mean absolute SHAP value, computed across the 82-case held-out test set. Values here are from our retrained ensemble — rankings match the manuscript's Figure 4A.
How to read these bars
- Tumour area dominates — as large as the next four features combined. This is consistent across every tree-based model in the study.
- Size X and Y are highly correlated with area but carry independent signal because elongated tumours behave differently from round ones.
- Anatomical unit matters more than surgeon experience in this cohort — a function of how tumours distribute across body parts, not a causal effect of the anatomy itself.
- See-and-do and experience contribute almost nothing — the paper flags this as likely selection bias, not lack of skill effect.
Tumour-area dependence
The model learned a threshold at ~1.5 cm². Below it the prediction stays under 50%; above it the probability of ≥13 sections climbs steeply. Toggle SCC / Recurrent / Aggressive histology to see the whole curve shift.
Per-case waterfall
Starting from the cohort's base rate, each feature adds or subtracts probability. Pick a case below to see its full breakdown.
Age
69
Type
BCC
Size
4×5 mm
Unit
NOSE
Baseline (cohort prior)
47.9%
Final prediction
28.3%