New explainable DTM model predicts stroke outcomes with 81% AUC

A team of researchers from multiple institutions, led by Lisa Herzog, has introduced multimodal Deep Transformation Models (DTMs) for predicting stroke outcomes, combining the interpretability of statistical models with the representational power of neural networks. Their work, accepted at MICCAI 2026, addresses the critical need for explainability in clinical AI systems. The DTMs integrate diffusion-weighted MRI with structured clinical data from 407 patients, achieving an AUC of 0.81 [0.75, 0.87]—state-of-the-art for this task. Functional independence before stroke and stroke severity on admission emerged as the strongest tabular predictors, preserving interpretability for clinicians.

A key technical contribution is the adaptation of two popular xAI methods—Grad-CAM and Occlusion—to 3D CNN-based DTMs, generating explanation maps that highlight relevant brain regions. The maps consistently pointed to frontal lobe areas, which are known correlates of age. Intriguingly, when age was explicitly added as a tabular predictor, those frontal lobe signals disappeared, confirming the model’s ability to capture age-related patterns from imaging data alone. Similarity analyses of these maps revealed distinct spatial patterns linked to different stroke mechanisms, offering new avenues for systematic error analysis and hypothesis generation. This work represents a significant step toward trustworthy AI for stroke prognosis.