Harmonized Multinational Quantitative EEG(HarMNqEEG) Norms Study

Quantifying individual brain developmental changes in patients against standard references is critical for clinical disease diagnosis,and this conception explains the idea of precision medicine initiatives.EEG age-dependent brain developmental trajectories provide a reliable description of brain activity and,in particular,one can detect brain dysfunction based on the degree of deviation from normal developmental trajectories.Accordingly,the study of brain disease diagnosis based on the standard developmental trajectory of the EEG is the well-known "quantitative electroencephalography"(q EEG)analysis.However,q EEG studies have focused on univariate spectra,lacking descriptions of brain functional connectivity,and q EEG studies have only been carried out in a single region or a single site without forming an international unified standard q EEG norm which hinders the development of global health care.Based on the traditional q EEG norm analysis,this dissertation proposed a multinational q EEG norm analysis,which mainly realizes two significant breakthroughs for the first time: extended the frequency domain q EEG method to multivariate analysis;removed the "batch effect" of multinational data based on the Harmonization Multinational q EEG(Har MNq EEG)model and relies on the developed high-dimensional fast kernel regression construct the human EEG brain development norm for the first time.These two breakthroughs have significantly improved the accuracy of clinical data diagnosis of brain diseases.This dissertation follows the principles of open science,from data collection to clinical data diagnosis,establishes and openly shares a set of comprehensive and efficient multinational q EEG norm datasets and platforms.Its main contributions include:(1)Based on the Riemannian manifold theory,this dissertation extended the traditional spectral q EEG norm analysis to the multivariate cross spectra,increasing the interpretation of the q EEG norm for brain functional connectivity in disease diagnosis.The cross spectra lay on the Riemann space rather than the Euclidean.The Riemannian vectorization operator transforms the cross-spectra to Gaussian distribution with the space projection from the Riemannian manifold to the Euclidean vector space.The transformation retains the bi-linearity of the cross spectra and provides a solid theoretical multivariate q EEG description parameter.(2)This dissertation developed an efficient and privacy-preserving data-sharing protocol solving the dilemma of multi-center data cooperation.This study collected 1792 participants from 9 countries,11 sites,and 14 studies.By sending the unified running code to the collaborators,instead of the raw data,here collected the cross-spectral tensor and anonymized metadata directly.It accomplishes a)Avoided complex audits of raw data sharing and privacy protection;b)Reduced the collection and processing burden of the center by cutting down the storage and database management;c)Accelerated the research process.The cross-spectra calculation can be processed parallelly at each site,and the site dataset transferred.(3)Designed a multi-level data quality control strategy which effectively guarantees the accuracy and reliability of multinational q EEG norm analysis.Due to the lack of unified standards for multinational data collection and preprocessing,this study formulated a three-level quality control strategy:(a)quality control within the data sharing site;(b)EEG viewing check with the experts based on the topographies and shapes of log-spectra within the central site;(c)Automatic outlier detection visually at the central site,which avoids batch data preprocessing deviations and outlier samples providing a robust data guarantee for the construction of q EEG norms.(4)Based on the generalized linear mixed model,this dissertation built a Har MNq EEG model solving the multinational data fusion problem of q EEG norm analysis for generating the human EEG brain development surface(norm)for the first time.This dissertation established a more generalized hierarchical standardized harmonization model based on the limitations of the existing harmonization methods that cannot deal with the nonlinear,high-dimensional big data analysis and model redundancy to detect and remove the "batch effects" induced by data collection equipment,experimental protocols,countries,and regions and find the optimal harmonization function with the developed fast kernel regression method and model information criteria.With the support of the developed model and algorithm,this dissertation draws the standard trajectory of brain development across the entire human life cycle,providing a benchmark for human brain development research and brain function disease diagnosis.(5)The multinational harmonized q EEG norms established in this dissertation obtain a higher diagnostic accuracy than univariate variable and local area q EEG norm analysis,which helps promote the global development trend of q EEG norm analysis.z-scores based on the Har MNq EEG model achieved higher diagnostic accuracy in predicting brain dysfunction caused by malnutrition in the first year of life at school age and in identifying functional brain impairment induced by COVID-19.Among the discrimination,the multivariate variable description parameters provide connectivity information for diagnosing brain diseases and new explanatory cognitive disorders and psychiatric diseases.(6)The established Har MNq EEG model allows the brain developmental bias calculation and multinational data harmonization distributed.This dissertation shared the normative norms of human brain development and provides the corresponding opensource platform.The proposed Har MNq EEG model realized local data harmonization processing and z-score calculation by matching the existing site norms in the database without sharing,which is convenient for researchers in various countries and regions for z-score online calculation and real-time disease diagnosis and monitoring.In addition,the Har MNq EEG model and data sharing strategy meet the federated learning strategy of parameter transmission,solving the problem of data island,and providing theoretical and technical support for the integration and computing of other data types.