BenchXAI: Comprehensive benchmarking of post-hoc explainable AI methods on multi-modal biomedical data.

Area of Science:

• Biomedical data analysis
• Artificial intelligence in medicine
• Explainable AI (XAI)

Background:

• Digitalization and AI offer predictive modeling opportunities in medicine.
• Deep learning models excel but suffer from a 'black-box' problem, hindering trust.
• Existing explainable AI (XAI) methods are often evaluated on single data types, limiting cross-modal applicability.

Purpose of the Study:

• To develop and introduce BenchXAI, a novel benchmarking package for evaluating XAI methods in biomedical contexts.
• To assess the robustness, suitability, and limitations of fifteen XAI methods across diverse biomedical data modalities.
• To provide a framework for the statistical evaluation and visualization of XAI method performance and robustness.

Main Methods:

• Development of BenchXAI, a package for comprehensive XAI method evaluation.
• Application of BenchXAI to three common biomedical tasks: clinical, imaging/signal, and biomolecular data.
• Implementation of a novel sample-wise normalization for post-hoc XAI methods to enable statistical analysis.

Main Results:

• Four XAI methods (Integrated Gradients, DeepLift, DeepLiftShap, GradientShap) demonstrated strong performance across all evaluated biomedical tasks.
• Certain XAI methods (Deconvolution, Guided Backpropagation, LRP-α1-β0) showed limitations on specific tasks.
• Benchmarking revealed performance variations among XAI methods depending on the data modality and task.

Conclusions:

• BenchXAI provides a vital resource for comparing and validating XAI methods in medicine.
• The study highlights the need for modality-aware XAI evaluation for reliable medical AI.
• Findings support the increasing necessity of explainable AI in the biomedical domain, especially with emerging regulations like the EU AI Act.