0025
Thalamic Connectome Based Machine Learning for Predicting Individual Symptoms after Mild Traumatic Brain Injury1Department of Biomedical Imaging and Radiological Sciences, National Yang-Ming University, Taipei, Taiwan, 2Neuroscience Research Center, Taipei Medical University, Taipei, Taiwan, 3Translational Imaging Research Center, Taipei Medical University Hospital, Taipei, Taiwan, 4Department of Radiology, School of Medicine, College of Medicine, Taipei Medical University, Taipei, Taiwan, 5Department of Medical Imaging, Taipei Medical University Hospital, Taipei, Taiwan
Synopsis
Mild traumatic brain injury (mTBI) can cause persistent post-concussion symptoms in 15-20% patients, however the presented type and severity of symptoms differ largely between patients. This study recruited 53 mTBI patients and 44 healthy controls to demonstrate the feasibility in using the imaging features of thalamic connectome combined with machine-learning regression models for the individualized prediction of clinical symptoms.
Background and Purpose
Mild traumatic brain injury (mTBI), commonly referred to as concussion, typically occurs with absence of structural lesions on clinical MRI examinations. Accordingly, it is difficult to correlate the heterogeneous post-concussion symptoms with relevant neuroanatomy and provide neuroimaging evidence to support diagnosis or therapeutic evaluation. In our previous work, we have demonstrated that the mTBI-induced shear injury can damage the thalamo-cortical tracts leading to the alterations of tract-based fractional anisotropy and functional connectivity of thalamo-cortical networks1. In this study, we further hypothesize that the specific pattern of altered thalamic connectome, including structure and functional perspectives, may elucidate the presented clinical symptoms in individuals. Considering the diversity of post-concussion symptoms between patients, we proposed a thalamic connectome-based machine learning approach that can predict individual symptoms according to the subject’s imaging features to fulfill the demands of personalized medicine.Materials and Methods
This study was approved by the local Institutional Review Board and the written informed consent was provided by each participant. Fifty-three patients with mTBI and 44 age- and gender-matched healthy controls (HC) were recruited (Table 1). Inclusion criteria for patients were witnessed closed-head trauma, no focal neurologic deficit, and initial Glasgow Coma Scale higher than 132. Clinical assessments were performed to evaluate post-concussion symptoms. MRI data, including a 3D T1-MPRAGE (TR/TE: 2300/3.26 ms; voxel size: 1.0x1.0x1.0 mm3), diffusion tensor imaging (DTI) with 64 gradient directions and 10 sets of b0 (TR/TE: 7500/59 ms; voxel size: 0.86x0.86x3.0 mm3), and BOLD resting-state fMRI (TR/TE: 2000/20 ms; voxel size: 2.2x2.2x3.5 mm3, 190 volumes) were acquired on a 3T MR scanner (Siemens MAGNETOM Prisma). Patients received MR scans within 4 weeks after mTBI.
Thalamic connectome analysis between 8 thalamic nuclei (defined by the Thalairach atlas) and 31 cortical regions (defined by the Brodmann area) in each cerebral hemisphere was carried out based on structural connectivity assessed on DTI probabilistic tractography and functional connectivity on fMRI (Table 2). Specifically, two hundred and eight features describing the mean, minimal, and maximal fractional anisotropy (FA) along each thalamo-cortical tract were computed3. The preprocessed fMRI data using SPM12 were used to extract regional BOLD signals followed by the calculation of Pearson’s correlation coefficient between regions (with bandpass-filtered between 0.01 and 0.10 Hz and Fisher’s r-to-z transform) to assess 664 thalamic-related functional connectivities. Finally, the small-world topological parameters were calculated based on functional networks to yield another 80 connectome features quantifying the network properties4.
Overall 972 connectome features were pooled and further ranked according to their correlation coefficients with each estimated clinical assessment, including Beck Depression Inventory (BDI), Beck Anxiety Inventory (BAI), post-concussion symptoms questionnaire (PCSQ), and Dizziness Handicap Inventory (DHI). The first 100 top-rank features were subsequently selected to train a support vector machine (SVM) regression model with the 5-fold cross validation for each prediction of clinical symptom5 (Figure 1).
Results and Discussion
The mTBI patients had elevated scores with larger standard deviations of BDI, BAI, PCSQ, and DHI compared to healthy controls (Table 1). The diversity of these clinical assessments in mTBI patients can be also observed in the right column of Figure 3 (blue data points of Participant #45 to #97 within blue areas, namely the patients with mTBI). The connectome features that selected for the model training are mainly the categories of functional connectivity and tract-based FA (Figure 2). None of the small-world topological parameter was selected for the model training. During the iterative optimization of hyperparameters of each SVM regression model, the continuous reduction of loss function was presented to reach the minimum value and confirm the model convergence (plots in the left column of Figure 3). With these trained regression models based on the selected thalamic connectome features, the BDI, BAI, PCSQ, and DHI scores for each individual participant can therefore be predicted (red data points in the plots of right column in Figure 3). The performance of each prediction was evaluated by a root mean square error (RMSE), a common method for quantifying the difference between observed (true) and predicted values6. Satisfactory predictions were reached in the prediction models for BDI (RMSE=7.80), BAI (RMSE=6.73), and PCSQ (RMSE=9.21), and suboptimal performance for DHI prediction (RMSE=16.39). As shown in Figure 3, the predicted scores of BAI and PCSQ (blue dotted lines in the right column) match well with the profiles of the observed scores (red dotted lines in the right column).Conclusions
This study showed the feasibility of individualized prediction of post-concussion symptoms based on the thalamic connectome in mTBI. Future studies with a larger study cohort and refined machine-learning models may further facilitate the clinical applications of connectome-based symptom prediction.Acknowledgements
This study was funded in part by the Ministry of Science and Technology (MOST106-2314-B-010-058-MY2, MOST104-2923-B-038-003-MY3), Taipei, Taiwan.References
1. Lu CF, Kao YC, Hsieh LC, et al. Association Between Concussion-induced Dizziness and Damage of Thalamo-cortical Connectivity after mild Traumatic Brain Injury. 2017 ISMRM, Abstract #0511.
2. Head J. Definition of mild traumatic brain injury. J Head Trauma Rehabil. 1993;8(3):86-7.
3. Chen YJ, Lo YC, Hsu YC, et al. Automatic whole brain tractâbased analysis using predefined tracts in a diffusion spectrum imaging template and an accurate registration strategy. Human brain mapping 2015;36(9):3441-58.
4. Rubinov M, Sporns O. Complex network measures of brain connectivity: Uses and interpretations. Neuroimage. 2010;52:1059–1069.
5. Smola AJ, Schölkopf B. A tutorial on support vector regression. Statistics and computing. 2004;14(3):199-222.
6. Marsland S. Machine Learning: An Algorithmic Perspective, Second Edition. Chapman and Hall/CRC, October 8, 2014.
Figures
