Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Non-linear functional connectivity of resting-state fmri data from patients with brain insults: a time-varying analysis

Simos Nikolaos-Ioannis

Full record

---

Summary

Mild traumatic brain injuries (mTBI) are a fairly common type of brain injury in which 90% of brain injury cases are classified as mild. However, existing diagnostic criteria for mTBI often fail to identify positive cases due to the inefficiency of diagnostic methods. For this reason, it has become necessary to develop techniques of greater diagnostic accuracy. In this paper, we thus propose the use of resting-state functional magnetic resonance imaging (rs-fMRI), a promising technique with positive research and diagnostic samples in various clinical groups. In the current utilization of rs-fMRI, an innovative combination being developed with the use of static and dynamic functional connectivity networks (SFC and DFC, respectively). The proposed methodology is applied to a comparable number of subjects/patients to accurately identify and separate chronic mTBI from normal brain activity, which to our knowledge has not been explored previously. SFC and DFC are calculated using bivariate linear and non-linear correlation indices. An approach based on temporal connectivity states was employed to characterize the networks of each time frame as either integrated or segregated. A reduction of the generated brain networks using Orthogonal Minimum Spanning Trees was applied to produce networks that maximize efficient information flow. Graph metrics were used to quantify various functional and topological features in SFC and DFC networks. Different types of features representing regional connectivity estimated by SFC/DFC are combined in the final stage to produce a reliable and efficient model. A novel machine learning model combination technique was compared with existing methods. The proposed diagnostic methodology resulted in a classification accuracy of 80% achieved by XGBoost models combined with logistic regression, with nested cross-validation of consensus type and built-in feature selection technique. Most of the regions derived from the machine learning model are in agreement with previous research in the field with the addition of some new findings with interesting interpretations. This experimental combination of approaches seems to offer promising results towards neurodiagnostic imaging of mTBI with high-precision tools, opening new windows of exploration.

Available Files

Services