Abstract DP273: Feature Fusion of MRI Radiomics and Deep-Learned Representations Predicts One-Year Mental Health Outcome Post Pediatric Arterial Ischemic Stroke

Abstract

Background: Pediatric arterial ischemic stroke (PAIS) survivors face higher risks for adverse mental health outcomes such as anxiety, depression, and executive dysfunction. Early risk stratification could enable timely and targeted interventions. Methods: We retrospectively analyzed structural MRI from 174 children with PAIS (mean age at stroke onset 3.8 years, range 1 month to 17 years, 60% male (n = 105) and 40% female (n = 69)) enrolled in the institutional stroke registry. For each lesion, we extracted handcrafted radiomic features (1,024 per scan) and deep-learning features (128-dimensional embeddings) from a 3D convolutional neural network (CNN) ( Figure 1 ) trained on lesion-focused MRI patches. Three experiments were run: radiomics only, deep learning only, and feature fusion (radiomics plus deep learning). Class and modality balancing were applied during training. Four classifiers were evaluated (Gradient Boost, Random Forest, Bagging, AdaBoost). Mental health outcome at approximately one year was binarized (poor vs normal) using standardized behavioral assessments completed by parents and teachers (BASC-3 and BRIEF). A patient-level confidence ratio quantified inter-rater agreement and was used to weight evaluation. Primary metrics were area under the ROC curve (AUC) on an independent test set at both the stroke level and patient-level for patients with multiple strokes. Results: Models trained on fused features showed the highest discrimination at the stroke level (Gradient Boost AUC ≈ 0.79, 95% CI 0.71–0.85) ( Table 1 and Figure 2 ), outperforming radiomics only (Gradient Boost AUC ≈ 0.60) and deep learning only (CNN AUC 0.72). At the patient level, a combined decision rule improved sensitivity for identifying at-risk children without loss of specificity. Incorporating confidence ratios yielded a weighted AUC ≈ 0.79 for the best-fused model, comparable to stroke-level performance while accounting for label reliability. Conclusions: Integrating handcrafted radiomic features with deep-learned representations improves prediction of poor one-year mental health post PAIS compared with unimodal models. Confidence-aware evaluation provides a pragmatic way to handle subjective outcome labels. This multimodal approach could support earlier identification of children who need targeted mental health interventions and follow-up.

Ähnliche Arbeiten