De-noising and Feature Enhancement of fMRI Data Based on Naive Bayesian Link Prediction

The traditional functional connectivity network (FCN) is greatly affected by external factors such as head movements, resulting in inaccurate classification results. This paper proposes a link prediction algorithm based on Naive Bayes to reconstruct the FCN. The proposed Naive Bayes-based link prediction algorithm can de-noise and enhance the features of the original FCN. Specifically, the FCN is obtained by matching the AAL template with pre-processing and sliding time windows. Then the Bayesian algorithm is used to change the weight of each node and then expressed by Common Neighbors (CN), Adamic-Adar (AA) and Resource Allocation (RA) coefficients to obtain reconstructed functional connection network. We call the coefficients as Local Naive Bayes Common Neighbors (LNB-CN), Local Naive Bayes Adamic-Adar (LNB-AA) and Local Naive Bayes Resource Allocation (LNB-RA). Finally, the classifier is used to classify the reconstructed network to verify the effectiveness of the proposed algorithm. The use of schizophrenia data provided by the Olin Neuropsychiatry Research Center of Hartford in United States was used to validate the effect of the reconstructed FCN by link prediction. In the experiment, we compared the changes of network characteristics before and after FCN reconstruction, and the classification performance was evaluated using three indicators-accuracy, sensitivity and specificity. The experimental results show that compared with the original FCN, the reconstructed FCN can achieve better results. To some extent, it has helped the diagnosis of brain diseases.