Dynamic functional network connectivity discriminates mild traumatic brain injury through machine learning

Victor M. Vergara, Andrew R. Mayer, Kent A. Kiehl, Vince D. CalhounNeuroImage 2018 · DOI

TL;DR

This paper introduces dynamic functional network connectivity (dFNC) as a more sensitive biomarker for mild traumatic brain injury (mTBI) than traditional static measures. By analyzing distinct brain states, they achieved a 92% classification accuracy in distinguishing mTBI patients from healthy controls, highlighting the cerebellum and sensorimotor networks as key areas of dysfunction.

What Problem Does This Paper Address?

Mild traumatic brain injury (mTBI) is a common but often difficult-to-diagnose condition. Current diagnostic methods rely heavily on subjective self-reports and have limited accuracy, leading to misdiagnosis and persistent deficits for many patients. While neuroimaging techniques like resting-state functional network connectivity (FNC) show promise, static FNC (sFNC) – which averages connectivity over time – may overlook crucial temporal dynamics of brain activity. This study aims to leverage the temporal variability of brain networks through dynamic FNC (dFNC) to develop a more accurate and objective diagnostic tool for mTBI.

Methods

The researchers employed a machine learning approach using resting-state fMRI data from 48 mTBI patients and 48 age- and gender-matched healthy controls.

  • Dynamic Functional Network Connectivity (dFNC): Instead of analyzing connectivity over the entire scan, they used a sliding window approach to capture temporal fluctuations in brain network interactions.
  • State Identification: K-means clustering was used to identify distinct "states" of functional connectivity, representing recurring patterns of brain network co-activation.
  • Classification: A linear support vector machine (LSVM) was trained to classify subjects into mTBI or healthy control groups based on the dFNC patterns within each identified state.
  • Feature Selection: To improve classification accuracy and identify the most informative features, they incorporated a nested leave-one-out cross-validation (LOOCV) loop. This involved selecting features (specific dFNC connections) based on their statistical significance (t-values from group comparisons) within each state.
  • Validation: The classification performance was evaluated using the area under the curve (AUC) metric, with LOOCV used to ensure robust and generalizable results.

Datasets

The study utilized resting-state fMRI data from a cohort previously described in prior work.

  • Participants: 48 mTBI patients and 48 healthy controls (total N=96).
  • Data Source: Resting-state fMRI scans collected at the Mind Research Network and the University of New Mexico.

Key Insights and Ideas

  • Temporal Dynamics Matter: The study strongly suggests that the dynamic, fluctuating nature of brain connectivity is more informative for mTBI detection than static averages. Different "states" of connectivity reveal distinct patterns of brain activity.
  • State-Specific Biomarkers: Not all dFNC states are equally useful for classification. The researchers identified that a specific state (State 2) was significantly more effective in discriminating between mTBI patients and controls.
  • Cerebellum-Sensorimotor Link: The most significant dFNC differences were observed between the cerebellum and sensorimotor networks, specifically showing increased connectivity in mTBI patients within State 2. This highlights these regions as critical areas affected by mTBI.
  • Machine Learning for Feature Selection: The integration of statistical tests (t-tests) for feature selection within the machine learning framework allowed for the identification of the most discriminative dFNC connections, leading to higher classification accuracy.

Key Results

  • High Classification Accuracy: The dFNC approach achieved a classification accuracy of 92% AUC for State 2, significantly outperforming static FNC (82% AUC) from a previous analysis of the same cohort.
  • Optimal State Identification: State 2 was identified as the most discriminative state, showing significant group differences in connectivity and being consistently selected for classification.
  • Significant dFNC Increases: In State 2, mTBI patients exhibited significantly increased dFNC between the cerebellum and sensorimotor networks, a finding that survived FDR correction.
  • Feature Selection Effectiveness: The nested LOOCV with feature selection demonstrated that using a subset of the most informative dFNC connections (selected based on t-values) improved classification performance.

Limitations and Future Work

  • State Occupancy Variability: A key limitation is that not all subjects consistently exhibited all identified dFNC states during the scan. The absence of a critical state like State 2 in a subject's scan could hinder accurate classification.
  • Temporary Disturbance: It remains unclear whether the observed dFNC abnormalities in State 2 are a temporary disturbance or a persistent marker of mTBI.
  • Scan Duration: The 5-minute resting-state scan might have been too short to capture State 2 in all participants. Future work could explore longer scan durations or methods to ensure State 2 detection.
  • Longitudinal Studies: Longitudinal studies are needed to track the temporal evolution of dFNC abnormalities post-injury and to determine their predictive value for long-term outcomes.
  • Real-time Applications: Developing real-time techniques to identify and measure specific dFNC states could pave the way for more immediate clinical applications.
  • Mental Training: Exploring techniques like mental training to encourage dwelling in specific dFNC states could enhance their utility in diagnostic or therapeutic contexts.

Summary updated

Abstract

Mild traumatic brain injury (mTBI) can result in symptoms that affect a person's cognitive and social abilities. Improvements in diagnostic methodologies are necessary given that current clinical techniques have limited accuracy and are solely based on self-reports. Recently, resting state functional network connectivity (FNC) has shown potential as an important imaging modality for the development of mTBI biomarkers. The present work explores the use of dynamic functional network connectivity (dFNC) for mTBI detection. Forty eight mTBI patients (24 males) and age-gender matched healthy controls were recruited. We identified a set of dFNC states and looked at the possibility of using each state to classify subjects in mTBI patients and healthy controls. A linear support vector machine was used for classification and validated using leave-one-out cross validation. One of the dFNC states achieved a high classification performance of 92% using the area under the curve method. A series of <i>t</i>-test analysis revealed significant dFNC increases between cerebellum and sensorimotor networks. This significant increase was detected in the same dFNC state useful for classification. Results suggest that dFNC can be used to identify optimal dFNC states for classification excluding those that does not contain useful features.
BibTeX
@inproceedings{vergara2018dynamic,
  author = {Victor M. Vergara and Andrew R. Mayer and Kent A. Kiehl and Vince D. Calhoun},
  title = {{Dynamic functional network connectivity discriminates mild traumatic brain injury through machine learning}},
  booktitle = {NeuroImage},
  year = {2018},
  doi = {10.1016/j.nicl.2018.03.017},
}
Figure