Identification Of Mild Traumatic Brain Injury With FMRI Based On Support Vector Machine

Objective: The recursive feature elimination(RFE)method was used to select voxels with strong discrimination ability,and linear support vector machine(SVM)method was applied to classify for seeking characteristic imaging parameters to distinguish the mild traumatic brain injury(mTBI)from the normal controls,and to identify the most critical brain regions or networks that are most sensitive to classification.Methods: Twenty-four mTBI patients(15 males and 9 females;mean age: 38.88±13.33 years;mean education: 8.88±3.58 years)and 24 healthy matched state volunteers(13 males and 11 females;mean age : 40.46 ± 11.4 years;mean education: 8.54 ± 3.41 years)each underwent resting-state functional magnetic resonance imaging exam and behavioral questionnaire assessment.Seven imaging indicators randomly combined into 126 combinations,including amplitude of low-frequency fluctuation(ALFF),fractional amplitude of low-frequency fluctuation(fALFF),regional homogeneity(ReHo),degree centrality(DC),voxel-mirrored homotopic connectivity(VMHC),long-range The functional connectivity density(FCD)and short-range FCD.The SVM-RFE machine learning classification method was used to evaluate the ability of the 126 combinations to distinguish the normal control group from the mTBI group.The classification accuracy,sensitivity,specificity,and area under the curve(AUC)were observed.After reducing the dimensionality of these features by the SVM-RFE method,the important features with the best classification performance are obtained,their degree of contributions are normalized,and they are mapped to the surface of the three-dimensional brain region.Results: The ability for any single imaging index to distinguish the normal control group from the mTBI group is not high.The discrimination ability of ALFF is highest in the seven imaging index,including ALFF,fALFF,ReHo,DC,VMHC,long-range FCD and short-range FCD.However,the AUC value,accuracy rate,optimal sensitivity and optimal specificity of the ALFF index for distinguishing the mild traumatic brain injury(mTBI)from the normal controls were only 0.692,70.89%,67% and 75%,respectively.Subsequently,we merged multiple indicator features to distinguish the two groups separately.The combination that integrated the characteristics of the five indicators of the ALFF,fALFF,DC,VMHC,and short-range FCD showed the best classification performance for distinguishing the normal control group from the mTBI group in the 126 combinations using the linear SVM-RFE machine learning classification method.The AUC value,accuracy rate,optimal sensitivity,and optimal specificity of the combination were increased to 0.778,81.11%,88%,and 75%,respectively.The brain regions of this combination that have a strong ability to discriminate the mTBI group from the normal control group mainly include the bilateral cerebellum,the left orbitofrontal cortex,the left cuneus,the left temporal pole,the right inferior occipital cortex,the bilateral parietal lobe and the left supplementary motor area.Conclusions: Multi-indicator feature combination can improve the classification performance of linear SVM-RFE machine learning.Understanding the evolution of brain function changes in mTBI patients will provide a reliable basis for early clinical diagnosis,timely treatment,and prognostic evaluation,and provide important theoretical basis for revealing the neurobiological mechanism of mTBI.