ML-Based Approach to Predict Carotid Arterial Blood Flow Dynamics

Abstract

In the current study, a numerical model has been developed to simulate the blood flow characteristics in the human carotid artery. The data thus generated is analyzed to understand the blood flow variations and predict the flow characteristics using Machine Learning techniques. In developing the numerical model, the key features of the system, namely, the blood, is modeled as an incompressible Newtonian fluid, and the artery is an elastic pipe. This model is simulated using COMSOL software by varying the material properties of the artery. Univariate analysis was performed to gain insight into the features' behaviour and target variables. Subsequently, machine-learning regression models were trained using the data generated from the idealized human carotid artery. Furthermore, the validity of the data was ensured by comparing it with flow division ratios available in the literature. The evaluation of these models was conducted by calculating the Mean Absolute Error values for the test dataset, resulting in the following values: polynomial regressor (0.0106), hyper-tuned support vector regressor (0.0487), decision tree regressor (0.000), random forest regressor (0.0156), Adaboost (0.0508), gradient-boosting (0.0044), and XGboost (0.0043). A quantile loss function was employed to assess the prediction uncertainty. According to the theory of loss function, models with low loss values are considered good predictors. The prediction uncertainty was measured by applying quantile loss function, and it identified that the random forest regressor as the best predictor model for the data, followed by the polynomial regression of degree 3. Prediction intervals for the target variable were computed by leveraging the random forest quantile regressor model. Moreover, the developed polynomial model was utilized to investigate the presence of stenosis in the artery.