Erick Almeida de Souza, Bruno Hebling Vieira, Carlos Ernesto Garrido Salmon
{"title":"利用深度卷积-递归模型,可从基于任务的动态功能连接中预测个体认知特征。","authors":"Erick Almeida de Souza, Bruno Hebling Vieira, Carlos Ernesto Garrido Salmon","doi":"10.1093/cercor/bhae412","DOIUrl":null,"url":null,"abstract":"<p><p>There has been increased interest in understanding the neural substrates of intelligence and several human traits from neuroimaging data. Deep learning can be used to predict different cognitive measures, such as general and fluid intelligence, from different functional magnetic resonance imaging experiments providing information about the main brain areas involved in these predictions. Using neuroimaging and behavioral data from 874 subjects provided by the Human Connectome Project, we predicted various cognitive scores using dynamic functional connectivity derived from language and working memory functional magnetic resonance imaging task states, using a 360-region multimodal atlas. The deep model joins multiscale convolutional and long short-term memory layers and was trained under a 10-fold stratified cross-validation. We removed the confounding effects of gender, age, total brain volume, motion and the multiband reconstruction algorithm using multiple linear regression. We can explain 17.1% and 16% of general intelligence variance for working memory and language tasks, respectively. We showed that task-based dynamic functional connectivity has more predictive power than resting-state dynamic functional connectivity when compared to the literature and that removing confounders significantly reduces the prediction performance. No specific cortical network showed significant relevance in the prediction of general and fluid intelligence, suggesting a spatial homogeneous distribution of the intelligence construct in the brain.</p>","PeriodicalId":9715,"journal":{"name":"Cerebral cortex","volume":null,"pages":null},"PeriodicalIF":2.9000,"publicationDate":"2024-10-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"Individual cognitive traits can be predicted from task-based dynamic functional connectivity with a deep convolutional-recurrent model.\",\"authors\":\"Erick Almeida de Souza, Bruno Hebling Vieira, Carlos Ernesto Garrido Salmon\",\"doi\":\"10.1093/cercor/bhae412\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<p><p>There has been increased interest in understanding the neural substrates of intelligence and several human traits from neuroimaging data. Deep learning can be used to predict different cognitive measures, such as general and fluid intelligence, from different functional magnetic resonance imaging experiments providing information about the main brain areas involved in these predictions. Using neuroimaging and behavioral data from 874 subjects provided by the Human Connectome Project, we predicted various cognitive scores using dynamic functional connectivity derived from language and working memory functional magnetic resonance imaging task states, using a 360-region multimodal atlas. The deep model joins multiscale convolutional and long short-term memory layers and was trained under a 10-fold stratified cross-validation. We removed the confounding effects of gender, age, total brain volume, motion and the multiband reconstruction algorithm using multiple linear regression. We can explain 17.1% and 16% of general intelligence variance for working memory and language tasks, respectively. We showed that task-based dynamic functional connectivity has more predictive power than resting-state dynamic functional connectivity when compared to the literature and that removing confounders significantly reduces the prediction performance. No specific cortical network showed significant relevance in the prediction of general and fluid intelligence, suggesting a spatial homogeneous distribution of the intelligence construct in the brain.</p>\",\"PeriodicalId\":9715,\"journal\":{\"name\":\"Cerebral cortex\",\"volume\":null,\"pages\":null},\"PeriodicalIF\":2.9000,\"publicationDate\":\"2024-10-03\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Cerebral cortex\",\"FirstCategoryId\":\"3\",\"ListUrlMain\":\"https://doi.org/10.1093/cercor/bhae412\",\"RegionNum\":2,\"RegionCategory\":\"医学\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q2\",\"JCRName\":\"NEUROSCIENCES\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Cerebral cortex","FirstCategoryId":"3","ListUrlMain":"https://doi.org/10.1093/cercor/bhae412","RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"NEUROSCIENCES","Score":null,"Total":0}
Individual cognitive traits can be predicted from task-based dynamic functional connectivity with a deep convolutional-recurrent model.
There has been increased interest in understanding the neural substrates of intelligence and several human traits from neuroimaging data. Deep learning can be used to predict different cognitive measures, such as general and fluid intelligence, from different functional magnetic resonance imaging experiments providing information about the main brain areas involved in these predictions. Using neuroimaging and behavioral data from 874 subjects provided by the Human Connectome Project, we predicted various cognitive scores using dynamic functional connectivity derived from language and working memory functional magnetic resonance imaging task states, using a 360-region multimodal atlas. The deep model joins multiscale convolutional and long short-term memory layers and was trained under a 10-fold stratified cross-validation. We removed the confounding effects of gender, age, total brain volume, motion and the multiband reconstruction algorithm using multiple linear regression. We can explain 17.1% and 16% of general intelligence variance for working memory and language tasks, respectively. We showed that task-based dynamic functional connectivity has more predictive power than resting-state dynamic functional connectivity when compared to the literature and that removing confounders significantly reduces the prediction performance. No specific cortical network showed significant relevance in the prediction of general and fluid intelligence, suggesting a spatial homogeneous distribution of the intelligence construct in the brain.
期刊介绍:
Cerebral Cortex publishes papers on the development, organization, plasticity, and function of the cerebral cortex, including the hippocampus. Studies with clear relevance to the cerebral cortex, such as the thalamocortical relationship or cortico-subcortical interactions, are also included.
The journal is multidisciplinary and covers the large variety of modern neurobiological and neuropsychological techniques, including anatomy, biochemistry, molecular neurobiology, electrophysiology, behavior, artificial intelligence, and theoretical modeling. In addition to research articles, special features such as brief reviews, book reviews, and commentaries are included.