Integrating Clinical and Imaging Data: A Review of Hybrid Machine Learning and Deep Learning Approaches for Stroke Prognosis

Abstract

The global death rate from brain strokes can be reduced through early and precise prediction, but existing AI diagnostic systems use only unimodal data. These systems are unable to build complete patient profiles because they use machine learning (ML) techniques for handling clinical records and deep learning (DL) methods for neuroimaging analysis. The paper presents a systematic review which investigates AI-driven stroke prognosis research through three different modalities: clinical data only, MRI data solely, and emerging hybrid (clinical + MRI) frameworks. The study evaluates the prediction pipelines used in research through an extensive examination of their methods for missing value processing and their approaches to feature selection and their prediction accuracy evaluation. The comparison shows each modality predicts accurate results while also demonstrating their limitations which arise from public dataset restrictions and the need for advanced computing to process images. The analysis shows multimodal data fusion has become essential for clinical decision support system development by providing valuable insights about real-time system implementation.