Cognitive Impairment Assistant Diagnosis Based On Graph Neural Network Clustering Algorithm

Cognitive impairment refers to the impairment of the body’s cognitive functions and can be diagnosed as Mild Cognitive Impairment(MCI)or Severe Dementia depending on the degree of functional impairment.Cognitive impairment is prevalent in the elderly due to physical and neurological deterioration of the brain.According to scientific statistics,cognitive impairment is difficult to cure and affects about 10 million people each year,so effective and accurate diagnosis has attracted widespread attention.In recent years,combining 3D imaging results from advanced equipment and clinical cognitive impairment test scale results,doctors can analyze and get the exact type of disease.However,due to the high cost of the equipment,it is not possible to satisfy all hospitals and thus bottlenecks are encountered when trying to fully rely on 3D imaging equipment to diagnose the degree of cognitive level.Therefore,an unbiased,rapid and accurate diagnosis of cognitive dysfunction is still being explored.Unlike 3D imaging results,the use of the Clinical Cognitive Impairment Test Scale is the most common and simple way to detect cognitive impairment,and its results are equally easy to obtain,and this paper expects to make a preliminary diagnosis of cognitive level from clinical scale data.In this paper,we propose a method called K-Nearest-Neighbour Graph-Deep Modularity Networks(KG-DMoN)for distinguishing the degree or stage of cognitive impairment by clinical cognitive impairment test scales.This deep graph clustering approach proposed in this paper aims to bridge the gap between clustering targets and graph neural networks in an end-to-end manner.Firstly,a graph network-style construction of the data structure is performed by the K-Nearest-Neighbour(KNN)algorithm,and secondly,the connection between graph clustering and graph pooling is drawn.Next,a new loss function is designed in this paper to optimize the clustering algorithm.In the prediction process,since graph convolutional neural network is a kind of transductive learning,which cannot be predicted by training a good model like inductive learning,the spatial distribution is used to find the n nearest neighbors of the new sample and the original sample to determine the category to which this sample belongs.After experimenting with 1398 samples,the results show that the algorithm in this paper achieves high confidence in classifying normal or mild cognitive impairment(Normal & MCI)and dementia.Compared with clinicians’ diagnostic strategies,KGDMoN performed more than 4% in accuracy(ACC).Together with five other popular clustering algorithms,KG-DMoN performed more than 6% in Area Under Curve(AUC),more than 2% in precision.Normalized Mutual Information(NMI)and Adjusted Rand index(ARI)both exceeded 16% and 19%.In order to show the clustering effect more intuitively,this paper uses t-Distributed Stochastic Neighbor Embedding(t-SNE)to present the results,and the data distribution is downscaled.In the two-dimensional plot of the data result distribution,KG-DMoN shows a more realistic and KG-DMoN shows a more realistic and reasonable distribution than the other methods.Together,these results show that KG-DMoN can effectively help primary clinicians in primary hospitals to make preliminary classification of the degree of cognitive impairment.