A Method to improve the Reliability of Saliency Scores applied to Graph Neural Network Models in Patient Populations

Graph Neural Networks (GNN), a novel method to recognize features in heterogeneous information structures, have been recently used to model patients with similar diagnoses, extract relevant features and in this way predict for instance medical procedures and therapies. For applications in a medical field is relevant to leverage the interpretability of GNNs and evaluate which model inputs are involved in the computation of the model outputs, which is a useful information to analyze correlations between diagnoses and therapies from large datasets. We present in this work a method to sample the saliency scores of GNNs models computed with three different methods, gradient, integrated gradients, and DeepLIFT. The final sample of scores informs the customers if they are reliable if and only if all of them are convergent. This method will be relevant to inform customers which is the degree of confidence and interpretability of the computed predictions obtained with GNNs models.