A novel approach for brain connectivity using recurrent neural networks and integrated gradients

Brain connectivity is an important tool for understanding the cognitive and perceptive neural mechanisms in the neuroimaging field. Many methods for estimating effective connectivity have relied on the linear regressive model. However, the linear regression approach might fail to account for the complexity inherent in brain connectivity. Due to the recent success of deep neural networks (DNNs), regressive data are able to be predicted with high accuracy. This study aimed to develop a connectivity method using the prediction performance of a DNN model and the parameters of the model. To this end, a method is proposed that utilizes integrated gradients in a recurrent neural network model. It is an extended application of explainable artificial intelligence in the multivariate autoregressive DNN model. It would be advantageous compared to the methods using the parameters of the linear regressive model or Granger's approach referring to the difference in error between the models. The performance of the connectivity estimation was tested by simulated datasets with various conditions. The overall performance was good on multiple metrics including recall (0.94), precision (0.90), F1-score (0.92), and accuracy (0.97). Compared with other conventional methods, the proposed method is robust and precise. The proposed method also demonstrates that it can be applied to estimate the actual brain connectivity in a magnetoencephalography study. In conclusion, the connectivity method based on integrated gradients provides an accurate estimation of brain connectivity by effectively capturing complex interactions, which is validated through high performance metrics such as recall, precision, F1-score, and accuracy across multiple simulated datasets. It introduces a novel framework to combine DNN and integrated gradients and to estimate effective connectivity by the explainable AI.