Stability of feature attribution: Contrasting supervised and unsupervised selection for radiopathomics and TCGA...

Interpretable machine learning is increasingly used in oncology, yet feature attributions from supervised models (e.g., Random Forest, XGBoost) can be unstable and bias-prone when grounded solely in SHAP explanations. We contrast target-prediction accuracy with the reliability of feature-importance estimates and assess stability via feature-elimination tests on TCGA data (705 samples, 1936 features). While supervised models achieved modest gains (Random Forest ROC-AUC: 0.8851 to 0.8865; XGBoost: 0.8681 to 0.8695), their feature-selection stability was low (6/10 and 3/10 retained, respectively). Unsupervised and non-target methods were markedly more stable: Feature Agglomeration, Highly Variable Gene Selection, and Spearman retained 10/10 features with unchanged performance (0.8823, 0.8823, 0.8766). We recommend combining unsupervised criteria with causal design and external validation to mitigate model-specific biases.