XSRD-Net: Explainable Stroke Relapse Detection

• XSRD-Net is a multimodal deep learning framework that combines 3D CTA imaging and tabular clinical data to predict stroke relapse risk.
• It employs both binary classification and relapse-free survival regression to deliver targeted predictions using metrics like AUC, F1, and c-index.
• The model offers interpretability by analyzing modality contributions, highlighting the significant roles of carotid artery imaging and heart disease markers.

XSRD-Net, short for EXplainable Stroke Relapse Detection, is a multimodal deep learning approach for predicting stroke recurrence risk from 3D intracranial CT angiography (CTA) and tabular clinical variables comprising age, gender, coronary heart disease (CHD), and peripheral artery disease (PAD). It was introduced for two related objectives: binary relapse detection and relapse-free survival (RFS) time prediction followed by classification. A central characteristic of the method is that it is explicitly framed as explainable: it seeks not only to estimate recurrence risk, but also to quantify modality contribution and to identify clinically relevant image regions and tabular features that drive the prediction, with the stated aim of supporting secondary prevention and therapy planning (Gapp et al., 9 Sep 2025).

1. Clinical problem and study cohort

Stroke is described as the second most frequent cause of death world wide with an annual mortality of around 5.5 million. The reported recurrence rate is between 5 and 25% in the first year, and mortality for relapses is stated as 40%. Within that framing, XSRD-Net addresses early identification of patients at risk of recurrence.

The study collected data from 119 patients with at least one ischemic cerebral event (ICE) between 2010 and 2024. The cohort comprised 78 non-relapse patients and 41 relapse patients. The reported mean age was 69.10 ± 10.18. Men accounted for 66.4% of the cohort and women for 33.6%. With respect to vascular comorbidity, 20.2% had CHD only, 16.8% had PAD only, 5.0% had CHD + PAD, and 58.0% had neither of those heart-disease indicators.

The train-test split was 95 patients for training, including 32 relapses, and 24 patients for testing, including 9 relapses. The imaging modality consisted of 3D intracranial CTA. These images were registered to a common reference image using affine registration via SimpleElastix. The tabular modality comprised age, gender, CHD, and PAD (Gapp et al., 9 Sep 2025).

The tabular encoding was explicit. Age was z-normalized as

$$\hat{y}_i = \frac{y_i - \mu}{\sigma}, \quad \mu = \frac{1}{N}\sum_{i=1}^N y_i,\quad \sigma = \sqrt{\frac{1}{N-1}\sum_{i=1}^N(y_i-\mu)^2}.$$

Gender was encoded as Female = $[0,1]$ and Male = $[1,0]$. Heart disease was encoded as No disease = $[0,0]$, CHD = $[1,0]$, PAD = $[0,1]$, and CHD + PAD = $[1,1]$. These design choices matter because the model’s interpretability analyses are reported at the level of both full modalities and individual tabular attributes.

2. Prediction tasks and decision rules

The work studied two tasks. Task 1 is a binary classification problem with class 0 = non-relapse and class 1 = relapse. The model output is a score in $[0,1]$, and thresholding is used to obtain the predicted class. The decision rule is stated as follows: given a continuous model output $s \in [0,1]$, predict relapse if $s > \theta$, and predict non-relapse if $[0,1]$0. Threshold selection used the F-score, and the paper distinguishes $[0,1]$1 as recall-weighted and $[0,1]$2 as precision-weighted.

Task 2 is the main task. It first performs regression to predict RFS time, defined as the number of days from the initial event until relapse, or censoring/no relapse, and then performs a subsequent classification. The classification rule is based on predicted RFS time $[0,1]$3: relapse if $[0,1]$4, non-relapse if $[0,1]$5. The threshold $[0,1]$6 is selected by maximizing F1 on the training data. The paper states that $[0,1]$7 was chosen between $[0,1]$8 days and $[0,1]$9 days, and it also mentions a hard cut-off at 2555 days for non-relapse cases (Gapp et al., 9 Sep 2025).

A further methodological element concerns incomplete labels. For some relapse patients, the exact RFS time was unknown. The training procedure therefore used a dynamic estimation strategy: the current predicted mean RFS of known relapse patients was used to initialize and update missing relapse RFS values epoch by epoch. This indicates that the regression setting was not a straightforward fully observed survival problem, but a partially observed target-estimation procedure embedded into model training.

3. Empirical performance across classification and RFS prediction

For Task 1 on the test set, the reported results were AUC 0.84 and F1 0.69 for the tabular-only model, AUC 0.67 and F1 0.53 for the vision-only model, and AUC 0.82 and F1 0.69 for the multimodal model. For the tabular model, sensitivity and specificity depended on threshold: at $[1,0]$0, Sens 1.00 and Spec 0.50; at $[1,0]$1, Sens 1.00 and Spec 0.50; and at $[1,0]$2, Sens 0.78 and Spec 0.81.

For Task 2 regression on the test set, the reported c-index values were 0.45 for tabular-only, 0.59 for vision-only, and 0.56 for XSRD-Net. When restricted to relapse cases only, the c-index values were 0.38, 0.64, and 0.68, respectively. After converting the regression outputs to class predictions, the derived classification results were AUC 0.34 and F1 0.53 for tabular-only, AUC 0.72 and F1 0.60 for vision-only, and AUC 0.71 and F1 0.63 for XSRD-Net (Gapp et al., 9 Sep 2025).

For the multimodal regression model, additional test-set thresholds were reported. At $[1,0]$3, the model achieved Sens 0.56 and Spec 0.88. At $[1,0]$4, it again achieved Sens 0.56 and Spec 0.88. At $[1,0]$5, it achieved Sens 0.56 and Spec 0.75.

The paper’s interpretation is task-dependent. For Task 1, it concludes that tabular data alone was sufficient to separate relapse from non-relapse fairly well. For Task 2, it states that vision mattered more than tabular data, and that the multimodal model was more balanced than the tabular-only model. The authors also emphasize that the regression task benefited from late fusion and that the multimodal setup may generalize better. A plausible implication is that the utility of multimodal fusion depends strongly on whether the target is immediate relapse discrimination or ordering of RFS trajectories.

4. Modality contribution and multimodal behavior

A distinctive element of XSRD-Net is its use of modality contribution measures. The reported modality contribution variables satisfy $[1,0]$6 and $[1,0]$7. This provides a direct account of how much each modality, or each tabular attribute within the multimodal setting, contributes to the prediction.

For the multimodal regression task, the paper reports an overall contribution ratio of vision:tabular = 0.68:0.32. In the detailed decomposition for Task 2, the contributions were Vision (3D CTA) 0.68, Age 0.03, Gender 0.20, CHD 0.03, and PAD 0.06. For multimodal classification in Task 1, the contributions were Vision (3D CTA) 0.01, Age 0.47, Gender 0.35, CHD 0.07, and PAD 0.10 (Gapp et al., 9 Sep 2025).

These values underpin two different multimodal regimes. In Task 1, the near-zero image contribution indicates what the paper describes as “unimodal collapse” toward the stronger tabular stream. Age and gender dominated the prediction, while CHD + PAD together contributed 0.17. In Task 2, by contrast, imaging became the main source of signal, while gender was the strongest tabular attribute and CHD and PAD remained relevant markers.

This contrast is important for interpreting the model family. XSRD-Net is not presented as a uniformly superior multimodal classifier; rather, its behavior depends on the target definition. For binary relapse detection, the tabular stream appears to capture most of the separative information available in the cohort. For RFS prediction, imaging contributes the dominant signal, and the multimodal model appears to recover complementary information instead of collapsing completely onto the tabular features.

5. Interpretability findings and anatomical-clinical linkage

The interpretability analysis combines modality contribution analysis with occlusion sensitivity and saliency-style localization. A major reported finding is that heart disease variables matter across tasks, especially CHD and PAD. The discussion explicitly notes that heart diseases significantly influence outcomes. In the regression setting, the contributions were CHD 0.03 and PAD 0.06; in the classification setting, they were CHD 0.07 and PAD 0.10.

On the imaging side, the occlusion-sensitivity analysis found that the most relevant region for RFS prediction was the supra-aortic region up to the upper neck, especially the arteria carotis communis region. The saliency maps likewise highlighted importance around the carotid arteries (Gapp et al., 9 Sep 2025).

The paper’s central interpretability conclusion is therefore a cross-modal linkage: heart disease variables in the tabular stream and carotid artery structures in the CTA stream together appear to provide the strongest signal for relapse detection and RFS prediction. The authors connect the carotid findings to carotid plaques, vascular course/shape, and possible dissections or atherosclerotic changes, and they specifically note that carotid artery disease can lead to stroke events.

This suggests a clinically coherent picture in which cardiovascular comorbidity reflects systemic vascular burden, while carotid anatomy captures a local vascular substrate for recurrent ischemic events. The paper treats this link as a central outcome and states that ongoing data collection and model retraining are intended to strengthen it further. In practical terms, the stated clinical message is that XSRD-Net may help identify patients at higher relapse risk early, support therapy planning, prioritize closer follow-up or more aggressive secondary prevention, and guide attention toward carotid pathology and cardiac or peripheral vascular disease.

6. Terminology, limitations, and related naming

The name XSRD-Net refers specifically to EXplainable Stroke Relapse Detection. It is a stroke-recurrence prediction framework built on multimodal clinical and imaging data. It should not be conflated with RadiX-Net, a distinct family of structured sparse neural network topologies based on mixed-radix numeral systems and Kronecker products. The RadiX-Net paper explicitly notes that it does not use the term “XSRD-net” and instead concerns deterministic construction of sparse DNN topologies rather than stroke relapse modeling (Robinett et al., 2019).

Several limitations are directly implied in the reported study design and discussion. The cohort is small, with 119 patients. The dataset is described as single-center / retrospective. There are missing RFS times for some relapse patients, requiring dynamic target estimation during training. The class distribution is imbalanced, with 41 relapse versus 78 non-relapse patients. The paper also states that the interpretability findings are promising but still need stronger validation (Gapp et al., 9 Sep 2025).

The stated future directions are correspondingly concrete: continued data collection, model retraining, further investigation of the carotid–stroke relationship, and a search for additional biomarkers for recurrence prediction. Taken together, these points position XSRD-Net as a multimodal, explainable prototype for stroke relapse risk modeling whose principal empirical strengths lie in tabular discrimination for binary relapse detection and in image-informed multimodal prediction for RFS-related ranking and derived classification.