Detection of Mild Traumatic Brain Injury by Machine Learning Classification Using Resting State Functional Network Connectivity and Fractional Anisotropy

Victor M. Vergara, Andrew R. Mayer, Eswar Damaraju, Kent A. Kiehl, Vince D. CalhounJournal of Neurotrauma 2016 · DOI

TL;DR

This study explores machine learning to detect mild traumatic brain injury (mTBI) using brain imaging data. They found that resting-state functional network connectivity (rsFNC) alone achieved higher classification accuracy than diffusion MRI (dMRI) or a combination of both, suggesting rsFNC may be a more promising biomarker for mTBI.

What Problem Does This Paper Address?

Diagnosing mild traumatic brain injury (mTBI) is notoriously difficult, yet it can lead to significant long-term neurological and psychological issues. Current diagnostic methods lack definitive biomarkers, prompting researchers to explore advanced neuroimaging techniques. This paper addresses the challenge of identifying reliable, objective biomarkers for mTBI by comparing the effectiveness of two distinct neuroimaging modalities: diffusion MRI (dMRI), which assesses white matter integrity, and resting-state functional MRI (rsFMRI), which measures brain network connectivity during rest. The goal is to determine which modality, or combination thereof, is superior for classifying individuals with mTBI.

Methods

The researchers employed a machine learning approach using a linear Support Vector Machine (SVM) for classification. To ensure robust evaluation, they utilized a leave-one-out cross-validation (LOOCV) strategy.

For dMRI data, they extracted features using two methods:

  • Enhanced Z-score Microstructural Assessment for Pathology (EZ-MAP): This method identifies abnormal fractional anisotropy (FA) values compared to a normative dataset.
  • Distribution Corrected Z-score (DisCo-Z): An alternative to EZ-MAP, aiming to capture subject-specific abnormalities and correct for estimation biases.
  • All Voxels: Features were also extracted from all white matter voxels.

For rsFMRI data, they used:

  • Group Independent Component Analysis (gICA): To identify independent resting-state networks (RSNs).
  • Functional Network Connectivity (rsFNC): Calculated as the correlation between the time courses of these RSNs.

The study also incorporated a novel pre-processing step for rsFNC data, which involved accounting for head motion variance before the gICA analysis, aiming to minimize confounds and improve accuracy.

Datasets

The study utilized data from a cohort of 50 patients with mTBI and 50 age-sex matched healthy controls (HC). The data was collected at the Mind Research Network and Lovelace Biomedical and Environmental Research Institute. The paper does not introduce a new benchmark dataset but rather re-analyzes existing data with refined pre-processing and analysis techniques.

Key Insights and Ideas

  • rsFNC Outperforms dMRI: The most striking finding is that rsFNC alone achieved significantly higher classification accuracy than dMRI features. This suggests that disruptions in functional brain communication during rest might be more sensitive or readily detectable in mTBI than structural white matter changes.
  • Cerebellum as a Key Region: Connectivity involving the cerebellum emerged as particularly important. Significant increases in rsFNC were observed between the cerebellum and sensorimotor networks, as well as between the left angular gyrus and precuneus in mTBI patients. This highlights the cerebellum's potential role as a critical region for mTBI detection.
  • Pre-processing Matters: The paper emphasizes the impact of pre-processing choices, particularly the order of head motion correction in rsFNC analysis. Their refined approach, addressing motion variance before gICA, likely contributed to improved classification performance.
  • Multimodal Fusion Not Always Better: Interestingly, combining dMRI and rsFNC features did not yield better classification results than using rsFNC alone. This could be due to the sheer number of dMRI features overwhelming the rsFNC signal or limitations in the current fusion techniques.

Key Results

  • Classification Accuracy:
    • rsFNC: 84.1%
    • dMRI (FA voxels): 75.5%
    • Combined dMRI and rsFNC: 74.5%
  • Significant Group Differences in rsFNC: Five specific rsFNC pairs showed significant differences between mTBI patients and healthy controls after false discovery rate correction. These included connections involving the cerebellum, sensorimotor area, and the left angular gyrus/precuneus.

Limitations and Future Work

  • Data Dimensionality and Sample Size: The authors acknowledge the common limitations of high data dimensionality and relatively small sample sizes in neuroimaging studies, which can impact classification robustness.
  • Feature Extraction Methods: While their refined pre-processing for rsFNC was beneficial, they suggest that more advanced feature extraction methods for dMRI might be needed to improve its performance.
  • Multimodal Fusion: The lack of improvement when combining modalities suggests that current fusion approaches may not be optimal. Future work could explore advanced data fusion techniques to better integrate information from different imaging types.
  • Timing of Scans: All mTBI patients were scanned within 21 days post-injury, meaning the findings reflect the acute phase. Future research could investigate how these connectivity patterns evolve over time.
  • Simple Feature Selection: The study used a relatively simple feature selection method. More sophisticated approaches might yield better results, especially for multimodal data.

Summary updated

Abstract

Traumatic brain injury (TBI) may adversely affect a person's thinking, memory, personality, and behavior. While mild TBI (mTBI) diagnosis is challenging, there is a risk for long-term psychiatric, neurologic, and psychosocial problems in some patients that motivates the search for new and better biomarkers. Recently, diffusion magnetic resonance imaging (dMRI) has shown promise in detecting mTBI, but its validity is still being investigated. Resting state functional network connectivity (rsFNC) is another approach that is emerging as a promising option for the diagnosis of mTBI. The present work investigated the use of rsFNC for mTBI detection compared with dMRI results on the same cohort. Fifty patients with mTBI (25 males) and age-sex matched healthy controls were recruited. Features from dMRI were obtained using all voxels, the enhanced Z-score microstructural assessment for pathology, and the distribution corrected Z-score. Features based on rsFNC were obtained through group independent component analysis and correlation between pairs of resting state networks. A linear support vector machine was used for classification and validated using leave-one-out cross validation. Classification achieved a maximum accuracy of 84.1% for rsFNC and 75.5% for dMRI and 74.5% for both combined. A t test analysis revealed significant increase in rsFNC between cerebellum versus sensorimotor networks and between left angular gyrus versus precuneus in subjects with mTBI. These outcomes suggest that inclusion of both common and unique information is important for classification of mTBI. Results also suggest that rsFNC can yield viable biomarkers that might outperform dMRI and points to connectivity to the cerebellum as an important region for the detection of mTBI.
BibTeX
@article{vergara2016detection,
  author = {Victor M. Vergara and Andrew R. Mayer and Eswar Damaraju and Kent A. Kiehl and Vince D. Calhoun},
  title = {{Detection of Mild Traumatic Brain Injury by Machine Learning Classification Using Resting State Functional Network Connectivity and Fractional Anisotropy}},
  journal = {Journal of Neurotrauma},
  year = {2016},
  doi = {10.1089/neu.2016.4526},
}
Figure