Deep Learning-Driven Thrombogenesis Prediction For Covid-19 Patients Using Cfd Data

During the global pandemic, particularly for Covid-19 patients who suffer greatly from cardiovascular issues, these issues lead to a higher mortality rate. This study focuses on analyzing Computational Fluid Dynamics (CFD) simulation files and to develop a predictive model to determine percent of Lagrangian particles that washout in one cardiac cycle. The proposed model relies on a simple arterial ultrasound procedure, making it non-invasive. The ability to predict thrombogenesis at a faster rate would dramatically improve patient care efficiency and facilitate prompt treatment decisions which could lead to saving lives. This transition to Deep Learning (DL) allows us to explore more intricate relationships within cardiovascular data, enabling the model to extract and learn complex patterns, potentially leading to significantly improved prediction accuracy. Our research proposes an implementation of DL using Ray Tune where we have a pipeline which tunes the hyper-parameters like the learning rate and epochs on a large set of model structures. The proposed system does this based on a range of each of the variables and optimizes for the best Mean Squared Error (MSE) results. The model was evaluated on CFD data, achieving high accuracy with an R² of up to 92.8% and 94.8% on two of the three files. This demonstrates the model's ability to capture complex cardiovascular patterns and suggests potential for rapid, non-invasive thrombogenesis prediction in clinical settings. Our results indicate that further improvements could be achieved with larger datasets and more extensive hyperparameter tuning This implementation and comprehensive optimization process aims to provide a system to detect blood clots using deep learning which can ultimately enhance patient care and outcomes by improving thrombogenesis prediction for Covid-19 patients and other individuals at risk.