Infantile Spasm Detection Based on Spatio-Temporal Attention Graphconv U-Net

Accurate detection of infantile spasms is crucial for timely intervention and treatment. The widespread use of electroencephalography (EEG) is limited, hindering the popularization of EEG-based epilepsy analysis in under-resourced medical settings. In response to these challenges, this study introduces a novel approach called spatio-temporal attention Graphconv U-Net (STA-GUNet) for detecting infantile spasms solely based on video data. The dataset comprises video clips collected from clinical records at China Medical University, consisting of 559 5-s clips depicting infantile spasms and 1031 5-s clips without infantile spasms. The videos are analyzed using the YOLOv7-Infant detection method, which tracks the dynamic changes in infants’ skeletal movements by examining key features such as position, velocity, and acceleration in a spatio-temporal context. The proposed model autonomously learns spatial and temporal patterns, leveraging attention mechanisms, graph convolutional networks, and a U-Net structure to capture intricate spatio-temporal dependencies, thereby enhancing the classification accuracy and robustness. To assess the model’s generalizability, a ten-fold cross-validation was conducted, demonstrating an overall accuracy of 93.14%, precision of 90.04%, $F1$ score of 90.28%, AUROC of 91.84%, and recall of 90.52%. Furthermore, an independent evaluation on a YouTube video dataset of infantile spasms demonstrates that STA-GUNet maintains strong detection performance, confirming its robustness and potential applicability in diverse real-world scenarios. To enhance the predictive performance of the model and validate its robustness, we completely separate the training set and the test set, and then proceed with the validation. These results support that the proposed detection method not only accurately identifies infantile spasms but also maintains a low false-positive rate, positioning it as a valuable adjunctive tool for clinical diagnosis.