Enhancing Deep Learning Performance with Parallel Coordinates and Data Encoding for Thrombogenesis Prediction

Deep Neural Networks (DNNs) are driving significant advancements in the biomedical field, particularly in thrombogenesis detection, which has traditionally relied on computational fluid dynamics (CFD) analysis-a process that is inherently time-intensive, often taking from several days to several months. In this study, we employed DNN to significantly reduce the processing time to just a few minutes for thrombogenesis detection, while enhancing accuracy through hyperparameter tuning. An initial step involved implementing a parallel coordinate method to identify and address suspected outliers, which were subsequently removed to optimize performance. Furthermore, the study evaluated the performance of three major data processing techniques: data labeling, one-hot encoding, and latent space encoding. Comparative performance analysis revealed that while all three methods demonstrated strong efficacy, the latent space encoding method yielded the highest performance, achieving an R2 score of 99.02%. This work highlights the impact of advanced neural network architectures and optimized data processing strategies on improving the accuracy and efficiency of biomedical applications, offering valuable insights into the potential of DNN for thrombogenesis detection.