A machine learning approach to predict cerebral perfusion status based on internal carotid artery blood flow

Cerebral blood flow (CBF), or perfusion, is a prerequisite for maintaining brain metabolism and normal physiological functions. Diagnosing and evaluating cerebral perfusion status is crucial to managing brain disease. However, cerebral perfusion imaging devices are complicated to operate, should be controlled by specialized technicians, are often large, and are usually installed in fixed places such as hospitals. It is significantly difficult for clinicians to obtain the cerebral perfusion status in time. Considering that CBF is mainly supplied by the internal carotid artery (ICA), this study proposes a cerebral perfusion status prediction model that can automatically quantify the level of cerebral perfusion in patients by modeling the association between ICA blood flow and cerebral perfusion.

Forty-eight participants were enrolled in the study after screening. We collected participants' ICA ultrasound and brain magnetic resonance imaging (MRI) data before and after dobutamine injection based on a rigorous experimental paradigm and built an ICA-cerebral perfusion datasetdd. Support vector machine (SVM), k-nearest neighbor (KNN), decision tree (DT), random forest (RF), gradient boosting decision tree (GBDT), and extreme gradient boosting (XGBOOST) were used for early prediction of cerebral perfusion status. The SHAP analysis was adopted to reveal the impact of interpretable predictions for each feature.

The XGBOOST model demonstrated the best overall classification performance with an accuracy of 78.01%, sensitivity of 96.67%, specificity of 98.23%, F1 score of 74.57%, Matthews correlation coefficient (MCC) of 62.17%, and area under the receiver operating characteristic curve (AUC) of 87.08%. Accelerated speed, peak systolic flow velocity, and resistance index of ICA blood flow are important factors for cerebral perfusion prediction.

The proposed method paves a new avenue for the study of predicting cerebral perfusion status automatically and providesv a noninvasive, real-time, and low-cost alternative to brain perfusion imaging. Moreover, this analysis identifies highly predictive features for the cerebral perfusion status and gives clinicians an intuitive understanding of the influence of key features. The prediction models can serve as an early warning tool that offers sufficient time for clinicians to take early intervention measures.