Multimodal explainable AI approach to pediatric appendicitis detection: integrating clinical data and ultrasound images

Pediatric appendicitis is a surgical emergency in which early and accurate diagnosis is critical to prevent severe complications. Clinical evaluation is often challenging because symptoms overlap with other abdominal conditions, and routine laboratory tests or imaging findings may be inconclusive. Therefore, improved diagnostic approaches are needed to enhance reliability and consistency. A multimodal attention fusion framework was developed to integrate patient symptoms, laboratory indicators, and ultrasound images for appendicitis detection. The workflow included data preprocessing, exploratory analysis, model development, training, and validation. Ultrasound images were standardized through resizing and pixel-intensity normalization, before feature extraction using a convolutional neural network, while clinical variables were analyzed as structured inputs. A cross-attention mechanism was employed to fuse multimodal representations and enable final classification. The model was trained and evaluated using the Regensburg Pediatric Appendicitis Dataset. The proposed framework achieved strong diagnostic performance, with 97.3% accuracy, 96.5% precision, 97.8% sensitivity, 96.2% specificity, 97.1% F1-score, and 96.8% AUROC. The multimodal attention-based approach improves diagnostic accuracy and consistency in pediatric appendicitis. By integrating imaging and clinical data and highlighting clinically relevant features, the framework improves predictive performance and aids in interpretability, demonstrating its usefulness as a tool for clinical evaluation.