Comprehensive Summary
This paper reviews and analyzes studies that applied deep learning models to predict hematoma expansion in patients with intracerebral hemorrhage (ICH) using computed tomography (CT) scans. The authors searched four major databases and evaluated study quality using QUADAS-2 and METRICS tools, then put diagnostic metrics through meta-analysis. In total, 22 studies were included, with 11 contributing to exclusive deep learning analyses and 6 to combined deep learning plus clinical imaging feature models. Exclusive deep learning models showed strong accuracy, and combined models that included both deep learning and clinical features performed even better. This suggested that performance varied depending on segmentation techniques and study quality, with manual or semi-automated segmentation yielding higher sensitivity. Overall, deep learning based models showed reliable predictive accuracy and robustness, particularly when integrated with clinical features.
Outcomes and Implications
Accurate prediction of hematoma expansion is critical, since expansion occurs in many patients and is linked to worse outcomes. This review shows that deep learning (DL) tools using CT scans can help doctors identify high-risk patients earlier, allowing faster treatments like lowering blood pressure or giving clot-stopping drugs. Compared to traditional CT signs, which are not very reliable, DL models were much more accurate. Still, most of the studies were small or done at single hospitals, so larger, multi-center studies are needed before these tools can be widely used. Despite this, the strong results suggest DL models may soon become an important part of stroke and ICH management.