From Covariance Regressions to Nonparametric Dynamic Causal Modelling (CoreDCM)

点击翻译按钮获取中文摘要

The global incidence of brain injuries is increasing, with at least 80% being classified as mild. These mild injuries are often not visible on routine clinical imaging such as magnetic resonance imaging (MRI) and computed tomography (CT). Magnetoencephalography (MEG) imaging is currently the only technique that has a high rate of success in detecting mild brain injuries both in the acute and chronic phase. However, this rate depends on the accuracy of MEG data analysis. The current statistical analysis of MEG data suffers from the following major limitations: a) The sensor data are assumed stationary but this assumption is not true in practice. b) The underlying locations of active sources are assumed time-independent, which is invalid according to physical principles. c) The effective connectivity analysis, based on a raw approximation such as canonical microcircuits to nonlinear neural differential equations, may miss some important factors. For example, fitting both canonical microcircuits and the proposed nonparametric model to some resting state MEG data, a few significant terms were found to appear in the nonparametric model but not in the microcircuit model. d) Group-level analyses prevalently used in the field may not be applicable for diagnosis of individuals due to the ecological fallacy, failure in inference about an individual based on aggregate data for a group. Therefore, these limitations underscore the need of an enhanced statistical analysis of MEG data on single-case basis to further improve its success rate of diagnosis of brain diseases and to differential different brain disorders. In this project, we propose to tackle these limitations by advancing MEG data modelling in the context of dynamic source localisation and nonparametric causal modelling. The proposed research will produce a flexible and accurate statistical inference tool for the diagnosis of brain conditions with a direct impact on mental healthcare practices. In conclusion, the proposed new method for MEG data analysis is able:- To cope with non stationary MEG data- To exploit time-varying behaviour of source locations- To move toward a fully nonparametric dynamic causal modelling (DCM) to avoid the missing of important factors as occurred in the conventional DCMs- To have supports from a rigorous statistical theoryCollectively, these unique features represent a step change beyond the methods available today and help pave a way for a large scale clinical applications. Therefore, the proposed research with exciting promise could save human life and have both economic and social impacts. It is timely because a clinical role for MEG in brain injury could become evident with further investigation identified, for which a strong interdisciplinary team has been assembled, the research can be directly implemented by the industrial partner Innovision IP, and a PDRA will be trained in this high-demand research area. The possibility of exposure to relevant end users in healthcare scenarios can be maximised through pathways to impact activities and working with the proposed research partners.The proposed research is closely aligned to one of the current EPSRC grand challenges of healthcare technology in optimising disease prediction, diagnosis and intervention and particularly in addressing both physical and mental health with techniques that optimise patient-specific illness prediction, accurate diagnosis and effective intervention.