ICH PrognoSys: Predicting intracranial hemorrhage prognosis using multimodal data

Abstract

Intracranial hemorrhage (ICH) is a life-threatening condition that requires a timely and accurate prognosis to guide critical clinical decisions. Traditionally, radiologists and clinicians assess prognosis through the evaluation of CT scans and patient history; however, time constraints and inter-observer variability can limit this manual process. In this study, convolutional neural networks (CNNs), including ResNet-18, DenseNet-121, and VGG-16, were employed to analyze CT scan slices. At the same time, tree-based machine learning models random forest and extreme gradient boosting (XGBoost) were used to process clinical tabular data. To improve transparency and trust in model predictions, explainable AI (XAI) methods SHapley Additive exPlanations (SHAP) for clinical data and Gradient-weighted Class Activation Mapping (Grad-CAM) for CT images were applied. The ResNet- 18 model achieved the highest performance among image-based models, while the random forest model with recursive feature elimination (RFE) led among tabular models. A web application was developed to enable clinicians to upload CT scans and clinical data, receiving prognosis predictions along with visual explanations, thereby serving as an accessible decision-support tool in clinical settings.