Revisiting AI Interpretability in Precision Oncology: Why Predictive Accuracy Does Not Ensure Stable Feature...

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• Artificial intelligence (AI) is increasingly vital in oncology for risk prediction, treatment planning, and biomarker discovery.
• Current AI evaluations often equate high predictive accuracy with reliable interpretation, potentially compromising reproducibility and clinical decision-making.
• This study addresses the need for robust interpretability metrics in AI for oncology.

Purpose of the Study:

• To reassess AI interpretability in oncology by introducing feature ranking order consistency as a stability-focused metric.
• To evaluate how AI model explanations respond to minimal input perturbations.
• To ensure AI models provide trustworthy and clinically actionable insights.

Main Methods:

• Compared supervised models (Linear Regression, LASSO, Random Forest, XGBoost) with unsupervised/statistical methods (PCA, Highly Variable Gene Selection, Spearman's rank correlation) using The Cancer Genome Atlas (TCGA) breast cancer multi-omics data.
• Assessed feature ranking stability by testing consistency after removing the top-ranked feature (<0.1% perturbation).
• Evaluated predictive performance using a Random Forest classifier with 10-fold cross-validation.

Main Results:

• Supervised models demonstrated unstable feature importance rankings even with minimal perturbations, indicating potentially fragile or misleading explanations despite high predictive accuracy.
• Unsupervised methods, specifically Highly Variable Gene Selection and Spearman's rank correlation, consistently produced stable and biologically coherent feature sets.
• These stable methods maintained competitive predictive performance compared to supervised approaches.

Conclusions:

• Interpretive instability is a significant limitation hindering the clinical application of many machine learning models in oncology.
• Integrating stability-based criteria, like feature ranking consistency, into AI evaluation frameworks is crucial for reproducible and trustworthy results.
• Prioritizing interpretability alongside accuracy is essential for the responsible and effective deployment of AI in precision oncology.