Deep Neural Networks for the Early Prediction of Abnormal Blood Flow: A Systematic Review of Techniques, Clinical Validation, and Future Directions

Blood flow dynamics and their abnormalities are pivotal in both cardiovascular and cerebrovascular diseases. Detecting these early is challenging, primarily because traditional methods are often invasive, resource-intensive, or inadequate in prediction. Deep neural networks (DNNs) offer a promising alternative, capable of interpreting complex, nonlinear patterns in high-dimensional medical imaging and physiological data, thus providing a non-invasive means to anticipate potential issues. In our systematic review, we thoroughly examined existing DNN methods aimed at early prediction of these problematic flow states. We focused on aspects such as model architectures, data types, testing methodologies, and their applicability in real-world scenarios. Our search through major biomedical and engineering databases identified 192 studies that met our criteria, encompassing a variety of architectures, including convolutional, recurrent, transformer, and physics-informed models. The field is predominantly led by CNN-based models, which constitute 34.9% of the studies, primarily targeting image-based tasks. In contrast, recurrent and transformer models appear in only 2.1% and 6.3% of the studies, despite their proficiency in handling temporal flow. Physics-informed models serve as a bridge, offering enhanced interpretability and physiological alignment. However, 44.3% of studies fail to specify their DNN architecture. While these models demonstrate strong internal performance, external validation is limited to just 16.7% of studies. Real-world clinical evaluation is notably scarce, occurring in only 6.3% of the cases. The challenges are substantial, involving dataset diversity, inconsistent reporting standards, and complex model interpretability. Although DNNs are promising for non-invasive hemodynamic predictions their transition from laboratory to clinical practice is hindered by insufficient external validation reporting inconsistencies and interpretability issues. Future efforts should focus on multi-center prospective studies, the utilization of explainable AI, and the development of standardized datasets to enable DNNs to realize their potential in clinical applications.