Explainability in mulimodal deep transformation models for stroke outcome prediction

Interpretability can coexist with strong predictive performance: The models achieved state-of-the-art accuracy (AUC 0.81) for predicting stroke recovery while remaining interpretable, demonstrating that explainability doesn't require sacrificing clinical effectiveness.

Combining different data types requires careful interpretation: When clinical tabular data (like patient age) was added to imaging data, the explanation maps changed significantly—highlighting that AI systems may identify proxy patterns in images rather than direct causal relationships, which is crucial for clinicians to understand.

Explanation methods should be validated against known medical knowledge: The fact that explanation maps consistently highlighted frontal lobe regions known to correlate with age suggests the xAI methods (Grad-CAM and Occlusion) are capturing medically meaningful patterns, providing confidence in their reliability for clinical use.

Multiple explanation approaches strengthen trust in AI decisions: Using two different explanation methods that produced consistent results across the same regions provides stronger evidence that the model's reasoning is sound, rather than relying on a single interpretation technique.

AI can generate new clinical insights: Beyond prediction, the explanation maps revealed distinct spatial patterns that could support hypothesis generation about stroke pathophysiology and help identify systematic errors, suggesting explainable AI has value beyond individual patient decisions.