A systematic evaluation of explainable AI methods for high-dimensional transcriptome-based cancer survival prediction

Explainable Artificial Intelligence (XAI) holds the promise to compensate for the "black-box" nature of deep learning which impedes transcriptome-based cancer survival prediction. However, there is a lack of systematic benchmarking XAI frameworks tailored for high-dimensional survival data. To bridge this gap, we systematically evaluated six representative XAI methods in three main categories: gradient-based, propagation-based, and perturbation-based approaches by using a Self-Normalizing Neural Network (SNN) as the baseline survival model. 6,248 samples across 15 cancer types from The Cancer Genome Atlas (TCGA) was analysed in this evaluation with a unified framework we developed. The evaluation metrics encompassed three key dimensions: prognostic factor enrichment (univariate Cox regression significance), biological consistency (supported by four authoritative databases, including OpenTargets), and explanation stability (Kuncheva Index). Among the six XAI methods, we find that DeepSHAP achieved the best overall performance, identifying the highest number of statistically significant prognostic factors while maintaining superior explanation stability; LRP (Layer-wise Relevance Propagation) showed slightly lower prognostic specificity but the highest consensus with biological databases in capturing general cancer genes, making it suitable for validating biological plausibility. In contrast, the perturbation-based method, PFI (Permutation Feature Importance) exhibited systematic failure and extremely low stability due to its inability to handle feature collinearity in high-dimensional transcriptomic data. Furthermore, we identified explanation stability as a robust proxy for the biological validity of the XAI. Collectively, this study establishes an empirical framework for selecting trustworthy AI explanation tools for precision medicine.