Multi-Layered Deep Graph Convulutional Neural Network in predicting and classifying symptoms hematoma and hemorrhage patients

Hemorrhages and hematomas are characterized as medical conditions that involve the escape of blood into or around the brain tissue. Predicting and classifying symptoms in these hematoma presenting patients in a timely and critical manner is essential in guiding neurosurgeons toward optimal treatment methods. Our study introduces a novel graph convolutional network for processing CAT scan images. This network extracts key features and classifies hematoma symptoms based on those features, enhancing the diagnostic process. It is built around a 13-layer convolutional neural network that is fed inputs in graph data, where nodes are defined as pixels within the segmented hematoma region and edges as spatial proximity in connecting neighboring pixels. To maintain a robust model and prevent overfitting, our team employed ReLu activation/normalization layers, L2 regularization, and 80-20 train/test split of the data. Our experimental results indicate that our model achieved up to 86% classification accuracy for loss of consciousness symptoms, 81% for aphasia, and 82% for headaches. Given the complexity of classifying hematoma and hemorrhage symptoms, future research should focus on integrating past patient history data with CAT scan images through feature-level fusion, creating a unified data representation for the machine learning model.