The early diagnosis of blast-induced traumatic brain injury (bTBI) is of great clinical significance for prognostication and treatment. Here, we report a new strategy for early bTBI diagnosis through serum and cerebrospinal fluid (CSF) based on terahertz time-domain spectroscopy (THz-TDS). The spectral differences of serum and CSF for different degrees of experimental bTBI in rats have been demonstrated in the early period. In addition, the THz spectra of total protein in the hypothalamus and hippocampus were investigated at different time points after blast exposure, which both showed clear differences with time increasing compared with that in the normal brain. This might help to explain the neurological symptoms caused by bTBI. Moreover, based on the THz absorption spectra of serum and CSF, the principal component analysis and machine learning algorithms were performed to automatically identify the degree of bTBI. The highest diagnostic accuracy was up to 95.5%. It is suggested that this method has potential as an alternative method for high-sensitive, rapid, label-free, economical and early diagnosis of bTBI.
© 2020 Optical Society of America under the terms of the OSA Open Access Publishing Agreement
Traumatic brain injury (TBI), as a worldwide health problem, has a global annual incidence of over 10 million new cases . The blast-induced TBI (bTBI) is a kind of TBI resulting from direct or indirect exposure to an explosion. The mechanisms of bTBI are complex and distinctive, which can be categorized into four phases according to the different physical aspects of the blast phenomenon [2,3]. With the development of high-energy weapons and the increasement of explosions caused by industrial accidents, bTBI has become the main type of injury in both military conflicts and civilian explosions [4,5]. Unlike other types of TBI (caused by direct impact on head), which has been conducted a number of studies, the scientific understanding of bTBI is still not well developed [6–8]. In particular, the exact recognitions of mild and moderate bTBIs are the bottleneck for current label-free diagnosis technologies in neurosurgery. According to the different detection principles, the existing methods for diagnosing bTBI can be divided into three categories. The first is neurological impairment assessment, such as Glasgow Coma Scale (GCS), which is generally based on self-reported or clinical observed neurological symptoms. It has high sensitivity at the cost of specificity and objectivity [9,10]. The second is widely used neuroimaging in clinic, such as CT, PET and MRI. But it is less effective at detecting mild bTBI, and there are still some technical issues concerning scanning protocols, analysis techniques and different contrast agents [11–13]. The third is a biomarker method, such as enzyme linked immunosorbent assay (ELISA) and western blotting (WB), which can measure the content of characteristic substances associated with bTBI to inference the injury degree. However, these methods have the disadvantages of being time-consuming and expensive [14,15]. Considering bTBI has long-term effects, there is an urgent need for a rapid, sensitive and label-free method that can be used for early diagnosis of post-traumatic phase of bTBI.
Compressed-air driven shock wave generators, such as bio-shock tube, have attracted much attention because of their safety and high repeatability . This type of instruments can simulate varying degrees of blast injuries in an indoor environment by adjusting air pressure. It is a powerful tool for animal models of bTBI. It has been showed that blast brain injury is associated with cerebral vascular injury similar to that resulting from non-blast TBI . Clinically, serum and cerebrospinal fluid (CSF) are often used as important indicators to reflect some diseases . Considering several potential serum biomarkers for bTBI have been identified (such as serum vascular endothelial growth factor, the glial protein S100β, neuron-specific enolase, and glial fibrillary acidic protein), it would be clinically meaningful for diagnosis if their specificity could be proven in the future [14,15,19]. CSF can communicate freely with the brain interstitial fluid that bathes the neurons, so it could be used as a detection matrix to reflect the central nervous system (CNS) injury in theory. Nevertheless, no neurochemical evidence has yet been found in CSF samples according to a study of human bTBI . In addition, neuropsychiatric abnormalities in the post-traumatic period are common symptoms of bTBI. These manifestations suggest that the regions of the brain associated with memory or emotion are affected. The level-dependent reductions in relative blood flow in the hippocampus has been proved in a blast injury model of rats . According to a biomarker study in a mouse model of bTBI, the concentration of ganglioside GM2 significantly increased in the hypothalamus and hippocampus after a single blast exposure . Therefore, it is of great significance to study the performance of hippocampus and hypothalamus after blast exposure.
Terahertz (THz) wave usually refers to the electromagnetic wave between microwave and infrared band with frequency from 0.1 to 10 THz. Since the rotational and vibrational energy levels of most biomolecules, such as DNA and proteins, are located in this frequency range, THz wave is sensitive for investigating the unique spectral features of biological samples [23,24]. In particular, our previous work has proved that impact-induced TBI can be well distinguished with THz continuous-wave (CW) transmission imaging, though this method needs slicing samples due to the high absorption of water . THz time-domain spectroscopy (THz-TDS) is a spectral technology that can obtain phase and intensity information at the same time. Generally speaking, the transmission and reflection modes are widely used in THz-TDS. However, transmission of THz wave through polar substances suffers from significant absorption, while the reflection mode has to avoid diffuse reflection and combine several complex reference calibrations [25,26]. To solve this problem, the attenuated total reflection (ATR) mode, which was widely used in mid-infrared spectroscopy, has been shown to be practical in the THz band . In the ATR THz-TDS system, highly absorbing samples in solid and liquid states on the top of an ATR prism can be rapidly detected with high sensitivity. Thus, it is an ideal tool for analyzing the THz spectral properties of biological samples without pre-processing or labeling.
In this paper, we propose a new strategy for early diagnosis of bTBI through serum and CSF based on ATR THz-TDS. The mild and moderate bTBI models of rats were established using bio-shock tube, the reliability of which was verified using a neurological assessment method. The spectral differences of serum and CSF for different degrees of experimental bTBI in rats were studied in the early period after blast exposure. In addition, the THz spectra of total protein in the hypothalamus and hippocampus were demonstrated at different time points after blast exposure for investigating the neurological pathogenesis. Moreover, the principal component analysis (PCA) and machine learning algorithms were used respectively in THz absorption spectra of serum and CSF to automatically identify the degree of bTBI. The results in this study provide a high-sensitive, rapid and label-free method for early diagnosis of bTBI.
2. Samples and methods
2.1 Blast model and biological methods
Nine-week-old male Sprague-Dawley rats ranging in weight from 240 to 260g were purchased from the Experimental Center of Medical Animal of the Daping Hospital/Research Institute of Surgery, Army Medical University (Chongqing, China). All animal procedures used in this study were performed in accordance with the China Animal Welfare Legislation and were approved by the Army Medical University Committee on Ethics for the Care and Use of Laboratory Animals.
A compressed air-driven bio-shock tube (BST-I) apparatus (Daping hospital, Army Medical University) was used as a shock wave generator, which was described in Refs. [16,28]. Unanesthetized rats were placed into individual cages to simulate the human experience and to eliminate confounds of anesthesia and surgery inherent in other models . The cages were fixed at the same vertical plane to ensure equal pressure exposure and to prevent subsequent secondary and tertiary blast injuries. A pressure gauge fixed on a metal cage was used to measure pressure changes near the rat’s body . The experimental rats were randomly divided into mild and moderate bTBI groups (number = 20 rats for each group), corresponding to the overpressure peaks of 4 MPa and 5 MPa in the driving section of shock tube respectively. Accordingly, a sham group without shock wave exposure was set as the blank control (number = 12 rats).
Considering biomarkers in serum undergo rapid changes over a period of time after trauma, particularly within 6 hours after the trauma, the related analyses were implemented at 3 h after blast exposure . The neurological deficits of rats were evaluated by modified neurological severity score (mNSS) ; brain water content (BWC) was measured using the dry-wet weight method ; brain tissues were sectioned coronally at a thickness of 2 mm and then the slices were stained with 2, 3, 5-triphenyltetrazolium chloride (TTC) at 37 °C for 25 min . These experiments were all performed to identify the different severities of bTBI models (number = 4 rats for each group). And the procedures were also performed on the sham group (number = 4 rats).
2.2 Sample preparation
Serum and CSF samples were prepared at 3 h after blast exposure (number = 8 rats for each group). Pentobarbital sodium powder was dissolved into aseptic saline to produce a 1% (w/v) solution, which was then intraperitoneally injected into rats at a dose of 30 mg/kg body weight. CSF was collected by muscle puncture in the medulla oblongata pool. First, the head of the rat was mounted on a stereotaxic frame and the fur was shaved. A longitudinal incision (1.5-2 cm) was made in the scalp to expose the occipital bone. Then, the CSF was extracted slowly from the medulla oblongata pool (with a depth of about 0.2 mm) using a microinjector. When about 100 µl of CSF was extracted, the microinjector was quickly withdrawn and the collected CSF was centrifuged within 20 minutes (4°C, 4500 r/ min, 10 minutes).
After CSF collection, serum was extracted by heart exsanguination. First, the anesthetized rats were fixed in a supine position. The xiphoid process was depilated and disinfected. Then, the syringe tip was obliquely inserted into the chest cavity of rats below the xiphoid process. As the tip passes through the diaphragm into the ventricles, the plasma was drawn into the syringe. Finally, the plasma was centrifuged twice (2000-2500 r/ min, 2 minutes) within 15 minutes to obtain the serum sample (about 1 mL). Here, it should be mentioned that, after the serum and CSF collections, the brain tissues of these rats were used for BWC or TTC staining.
Total protein samples were prepared at several time points (e.g. 3 hours, 6 hours, and 24 hours) after blast exposure (number = 4 rats for each group). The total protein samples at 3 hours were from those used for mNSS testing. First, the intact rat brain tissues were extracted respectively and immediately preserved in a sample cassette kept at low temperature (4°C). Then, the hippocampus and hypothalamus samples were separated in the cold state. After the tissue has been cut into tiny pieces, it was lysed using cell lysis buffer [20 mM Tris pH 7.5, 150 mM NaCl, 1% Triton X-100, 2.5 mM sodium pyrophosphate, 1 mM EDTA, 1% Na3VO4, 0.5 µg/ml leupeptin, 1 mM phenylmethylsulfonyl fluoride (PMSF)] at the ration of 150 microliters of lysate per 20 mg of tissue. A glass homogenizer was used to homogenize the slurry until the tissues was fully cracked and then the lysates were collected by centrifugation (4°C, 10,000 g, 5 minutes,). The concentrations of total protein from different tissues were determined using the bicinchoninic acid (BCA) Protein Assay kit (Pierce, Rockford, Illinois, USA) according to the instructions provided with the kit and then adjusted to the same level (4 mg/ml) before the THz-TDS measurement.
2.3 THz-TDS measurement
To acquire the complex refractive indices of liquid biological samples, the commercially available THz time-domain spectrometer (Advantest Corp., TAS7500SP) worked in the ATR geometry was used. The spectrum was measured from 0.1 to 4.0 THz, with a frequency resolution setting at 7.6 GHz and dynamic range of 70 dB. Figure 1 shows the experiment protocol of ATR THz-TDS measurement. A Dove prism made of Si crystal (n = 3.42 in THz range), was fixed at the focal position of the incident THz wave (p polarized). The sample was put on the top surface of the prism. The THz wave was totally reflected at the top surface of prism at the incident angle of 57° along with generating an evanescent wave interacting with the sample. The minimum thickness of the sample during the measurement is required to be greater than the penetration depth of the evanescent field to record the spectrum that corresponds to the bulk sample. Considering the penetration depth of the evanescent wave in distilled water at 1 THz is 24 µm, multiple tests of different samples were carried out to make sure the signals from repeated measurements can be overlapped in a certain time. Thus, the sample volume used for each measurement was chosen to be approximately 60-80 µl. After being recorded by a detector, the THz time-domain waveforms of the reference and sample measurements are Fourier transformed to acquire the frequency-dependent parameters. The spectral measurement time is about 10 seconds. Each measurement was repeated for 5 times. All the measurements were carried out at 23 ± 1°C with a nitrogen purge to avoid excess absorption of water vapor. The sample evaporation can be ignored during the whole process.
2.4 Spectral analysis
The spectral analysis method can be divided into three stages: pre-processing, feature extraction and classifier model. The appropriate spectral pre-processing is required to remove any irrelevant information such as noise and scattering effect and avoid data distortion at the same time. PCA is a widely used statistical technique which attempts to explain the covariance structure of data by using a small number of components. These principal components (PCs) are linear combinations of the original variables, and often allow for an interpretation and a better understanding of the different sources of variation. PCA is concerned with data reduction and commonly used as a preliminary step of high-dimensional data analysis, followed by further multivariate statistical methods.
k-Nearest neighbor (kNN) and support vector machine (SVM) are commonly used methods in data classification. kNN is classified by measuring the distance between different characteristic values. Specifically, if most of the k nearest neighbors of a sample in the feature space belong to a certain category, the sample also belongs to this category (k<20). It means that the selected neighbors are all objects that have been correctly classified. Therefore, this method only determines the category of samples according to the category of the nearest samples. SVM is a machine learning algorithm based on structural risk minimization (SRM) that reduces the risk of data over-fitting. Thus, SVM-based classifiers worked well in dealing with a small or limited size of dataset. At first, SVM was only designed to study the problem with binary classification. In the clinical setting, it is necessary to make a useful and robust extension of the conventional SVM model to adapt to multi-class classification problems. Moreover, the linearly inseparable problems can be solved by introducing different kernel functions in SVM.
Here in our study, Savitzky-Golay smoothing and boxplot methods were used to reduce noise and reject outliers respectively. In the feature extraction phase, PCA was employed to reduce dimension of THz spectral data and improve the efficiency of the classifier. Then, scores of retained PCs were used as input data for SVM. For model calibration and validation, 20% of the data in each category was selected as the testing set and a 5-fold cross validation was utilized in the remained data (training set) due to the relatively small sample size. Finally, the trained kNN and SVM classifiers were used to predict the test set and their performance was evaluated respectively.
3. Results and discussion
3.1 Neurology and pathology
To verify whether the bTBI model caused different degrees of injury to the rats, mNSS was used to evaluate the degree of nerve function defect at 3 h after blast exposure. Considering cerebral edema is an important symptom for many brain diseases, we measured BWC in different groups at 3 h after blast exposure by dry-wet weighting method. The results of mNSS and BWC are both shown in Fig. 2(a). Obviously, the mNSS scores of bTBI groups were higher than that in the control group (*p < 0.05 versus sham; #p < 0.05 versus mild injury). The severity of nerve function impairment was positively correlated with the intensity of blast wave produced by the shock tube. In other words, the rats of 4 MPa group had mild bTBI, while the rats of 5 MPa group had moderate. Nevertheless, it is found that there was no difference in BWC between the bTBI and sham rats (p = 0.05). The results of TTC staining of brain slices are shown in Fig. 2(b). The locations of the hypothalamus and hippocampus were labelled with blue and yellow dashed boxes, respectively. It can be seen that white areas were absent in both the trauma and sham groups. This indicated that there was no ischemic infarction in the brains of the rats with bTBI, which was consistent with the test of BWC.
First, we measured the serum of bTBI and normal rats by a THz-TDS system in ATR mode, as described above. In order to ensure the reliability of the data, frequency range from 0.2 to 2.0 THz was chosen from the original spectra with high signal-to-noise ratio. The 13 points Savitzky-Golay smoothing was used for attenuating noise interference in spectral data. Since the instability of the system, it was necessary to measure the same sample repeatedly. Boxplot method was used to reject outliers in spectra data obtained by reduplicative measurements. Figure 3(a) shows an example of the phase shift spectra of a serum sample from sham group obtained by repeated measurements. It can be seen that the spectrum (red) measured at the second time was significantly different from the others. Moreover, the phase shift value at 1.2 THz was extracted as a reference to draw the boxplot, as shown in Fig. 3(b). The median, the mean, and the quartile were calculated. The interquartile range (IQR) was used to observe the dispersion of data. Outliers were defined as data that exceeds 1.5IQR in the boxplot and clearly marked on the boxplot. It can be seen that the boxplot method can effectively eliminate the wrong data caused by the instability of test means and equipment, so that the processed spectral data can reflect the characteristics of biological samples more veritably and effectively.
Then, the THz absorption and the refractive index spectra were calculated by averaging over the obtained data, as shown in Fig. 4. The standard deviation was shown as error bars. Steady increasing of the absorption coefficients for serum from all groups were observed when the frequency increased from 0.2 to 2.0 THz, as shown in Fig. 4(a). However, absorption coefficient of serum of bTBI rats was significantly higher than that of the control group in the spectral range greater than 0.5 THz. Furthermore, the absorption coefficient was positively correlated with the degree of damage in the high frequency region. In other words, for THz absorption spectra of serum, the impact of moderate bTBI was greater than that of mild bTBI. Consistently, the refractive index of the sham group decreased slowly from 2.31 to 2.06 over a wide THz range from 0.5 to 2.0 THz, while those of the mild and moderate groups displayed a gradual decrease from 2.32 to 1.98 and from 2.30 to 1.87 respectively, as shown in Fig. 4(b). It is clearly that the refractive index of the serum samples from the trauma groups was significantly lower than that of the sham group in the high frequency region. The above analysis indicated that the serum composition of the rats changed at 3 h after the shock wave injury experiment. To investigate specific damage to the brain caused by shock waves, the THz absorption coefficient and refractive index spectra of CSF were acquired using the same method, as shown in Figs. 4(c)-(d). Similar to the spectra of the serum, significant differences between the trauma groups and the sham group were observed in the high frequency range of both the THz absorption coefficient and refractive index spectra of CSF samples. The THz spectral differences also showed a positive correlation with the degree of bTBI. This was consistent with the spectral data of serum. It is noteworthy that there were differences in the spectral characteristics of serum and CSF samples. At 1.6 THz, for example, the difference between the mean absorption coefficient of the sham group and the mild group was -29.20 cm-1 in serum samples, while the value was -15.32 cm-1 in CSF. Similarly, at the same frequency, the difference between the mean refractive index of the sham group and the mild group was 0.065 in serum samples, while it was 0.037 for CSF. Namely, compared to serum, the THz spectral differences between the sham group and the mild group were smaller in CSF. The above analysis indicates that CSF and serum can be used as early diagnostic indicators for mild and moderate bTBI, and the serum is more sensitive to shock waves at lower doses than the CSF. This will provide a method for the early diagnosis of bTBI, especially for the mild bTBI.
Moreover, it is necessary to study the changes in encephalic region of rats after blast exposure, so as to further probe the pathogenesis of bTBI, especially those related to neurology (such as memory loss and PTSD). Hypothalamus and hippocampus were isolated separately from rat brain tissue at 3 hours, 6 hours and 24 hours after blast exposure. The total protein samples were prepared by lyase and adjusted to a certain concentration before the THz measurement, according to the results of the BCA kit. The absorption and refractive index spectra (0.2-2 THz) of total protein in the hypothalamus was shown in Figs. 5(a)-(b). The insets showed the enlarged view in the range of 1-1.6 THz. Overall, the absorption coefficient and refractive index of total protein in the hypothalamus raised and decreased monotonically with frequency increasing respectively. In order to observe the temporal changes more clearly, the normalized absorption coefficient values and refractive index values at 1.6 THz were selected, as shown in Fig. 5(c). The x-coordinate of 0 h represented the sham group. At each time node, the normalized values were obtained by dividing the absorption coefficient and refractive index by the corresponding value of the sham group. At 3 h after blast exposure, the absorption coefficients of the mild and moderate groups were significantly higher than that of the sham group. With the passage of time (6 h and 24 h), these values began to decline gradually. A gradual tendency to return to normal level can be observed, but the differences between the trauma groups and the sham group still remained at 24 h after blast exposure. However, at the same time point, there was no significant difference in the THz absorption spectra of total protein in hypothalamus from rats with different degree of bTBI. Accordingly, the trend of refractive index in THz range was opposite to that of absorption coefficient. It reached the minimum at 3 h and then began to increase slowly. Similarly, there was no recovery to normal level at 24 h after blast exposure. Additionally, the spectra of refractive index for the trauma groups with different degree showed no difference at the same time point.
Using the same method, the THz spectra of total protein in hippocampal and its temporal changes were demonstrated in Figs. 6(a)-(c). It can be seen that both the THz absorption and refractive index spectra showed no significant change at 3 h after blast exposure, which were obviously different from the phenomenon in hypothalamus. However, at 6 h after blast exposure, the THz absorption coefficient of total protein in hippocampus increased accompanied by a decrease in THz refractive index. In addition, those changes were more obvious with the increase of trauma degree. At 24 h after blast exposure, the absorption coefficient still increased slightly, compared to that of 6 h. The refractive index decreased correspondingly, during this period of time. According to the above analysis, it is suggested that the bTBI caused some changes in the total protein of the hypothalamus and hippocampus of the rats, which could be attributed to a decrease or increase in the content of certain biomolecules. Such changes may affect the normal function of these brain regions, leading to the symptoms of neurological impairment.
The THz absorption coefficient spectra of serum and CSF were classified and identified using a method combined PCA and machine learning classifiers (kNN and SVM). First of all, the absorption spectral data after preprocessing were summarized to respectively establish the THz spectral databases of serum and CSF. The total numbers of THz spectra of serum and CSF were 108 and 93, respectively. The detailed number of each group is shown in Table 1. To improving the efficiency of classification, PCA method was used to reduce the dimension of data (237 frequency points). Upon PCA calculation, PC scores and loadings were generated. Considering the first 5 PCs, Figs. 7(a)-(b) respectively illustrate the eigenvalues of each principal component and their cumulative contribution to the total variance of spectra data from serum and CSF samples. It can be observed that the eigenvalues dropped off rapidly with increasing PC numbers and the first few PCs retained the maximum variance of the data being considered. For instance, the first PC accounts for the largest variance within the spectral data sets, which were 93.7% and 92.0% of the total variance for serum and CSF respectively. The first two PCs collectively account for 98.7% (serum) and 96.6% (CSF) of the total variance. The succeeding PCs account for a small proportion of the total variance. It indicated that the characteristic data set composed of first two PCs can accurately describe the original spectral data. Moreover, the scores on the first two PCs based on the spectra data of serum and CSF were shown in the Figs. 7(c)-(d), respectively. Each scatter represented a sample, and their classes were colored differently. On the score graph, confidence ellipses were plotted based on 95% confidence interval. It can be seen that the three groups of samples had good clustering on each two-dimensional score scatter plot with only a few samples falling outside the confidence ellipses.
Two machine learning classification methods (kNN and SVM) were applied to offer an improved model performance in predicting the class of unknown subject. In kNN algorithm, k value was set as 3 and Euclidean distance was calculated. In SVM algorithm, linear function was introduced as kernel function, and ‘one-against-one’ strategy (OAO) was adopted to realize multi-class classification. In order to optimize the classifier parameters and prevent overfitting, 5-fold cross validation was used in the training set randomly selected from preconditioned data (80% of each group). Four classifiers (serum-based kNN, serum-based SVM, CSF-based kNN and CSF-based SVM) were trained by the above method of partitioning dataset. After the process of models training had been completed, the performance of each classifier was evaluated with the testing set. Figure 8 shows the predicted value of the trained classifier for each sample in the test set. Three classification performance parameters (accuracy, sensitivity, and specificity) were calculated based on the predicted results, as shown in Table 2. The accuracy rate represents the proportion of all predicted samples to the total test set; sensitivity (true positive rate) represents the proportion of the samples that are actually traumatic and judged to be ‘mild’ or ‘moderate’, ignoring the prediction error between the two groups; specificity (true negative rate) refers to the proportion of the samples that are actually normal and judged to be ‘sham’. In clinical practice, the above sensitivity can be used as a reference for the omission diagnosis rate (its value equals to 1-sensitivity), and specificity is used for the misdiagnosis rate (its value equals to 1-specificity). It can be seen from Table 2 that the diagnostic accuracy of all classifiers has reached more than 88.9%, and the SVM classifier based on serum samples has reached the diagnostic accuracy up to 95.5%. Notably, the diagnostic sensitivity of most of the classifiers reached 100%, except for the serum-based kNN classifier (93.3%). Similarly, the diagnostic specificity of most of the classifiers reached 100%, except for the serum-based SVM classifier (85.7%). Particularly, the serum-based SVM classifier has the best predicting performance, achieving a diagnostic accuracy of up to 95.5%. Based on the above analysis, it can be concluded that the spectral classification algorithm used in this paper can automatically identify the degree of bTBI of rats based on the serum and CSF samples using the ATR THz-TDS system.
In summary, an early diagnosis method for bTBI in rats has been proposed based on ATR THz-TDS system. The bTBI models of rats were established using bio-shock tube, the reliability of which was verified by mNSS. The diagnostic matrixes (serum and CSF) and their spectral differences were analyzed. In addition, for investigating the neurological pathogenesis of bTBI, the THz spectra of total protein in hypothalamus and hippocampus at different time points after blast exposure have been demonstrated. The results indicate that the temporal changes of THz spectra affected by bTBI showed different trends in the hypothalamus and hippocampus. It might help to explain the neurological symptoms caused by bTBI. Moreover, the THz absorption spectra of serum and CSF were analyzed by combining PCA and machine learning algorithms (kNN and SVM) to automatically identify the degree of bTBI. The serum-based SVM classifier performed best on the test set with the highest diagnostic accuracy of 95.5%. The results indicate that this method has great potential as an alternative method for high-sensitive, rapid, label-free, economical and early diagnosis of bTBI.
The National Basic Research Program of China (973) (2015CB755403); National Natural Science Foundation of China (61771332, 61775160, 62011540006, U1837202).
The authors declare that there are no conflicts of interest related to this article.
1. J. Ghajar, “Traumatic brain injury,” Lancet 356(9233), 923–929 (2000). [CrossRef]
2. A. Nakagawa, G. T. Manley, A. D. Gean, K. Ohtani, R. Armonda, A. Tsukamoto, H. Yamamoto, K. Takayama, and T. Tominaga, “Mechanisms of Primary Blast-Induced Traumatic Brain Injury: Insights from Shock-Wave Research,” J. Neurotrauma 28(6), 1101–1119 (2011). [CrossRef]
3. R. G. DePalma, D. G. Burris, H. R. Champion, and M. J. Hodgson, “Current concepts: Blast injuries,” New Engl. J. Med. 352(13), 1335–1342 (2005). [CrossRef]
4. Y. Bhattacharjee, “Neuroscience - Shell shock revisited: Solving the puzzle of blast trauma,” Science 319(5862), 406–408 (2008). [CrossRef]
5. C. W. Hoge, D. McGurk, J. L. Thomas, A. L. Cox, C. C. Engel, and C. A. Castro, “Mild traumatic brain injury in US Soldiers returning from Iraq,” New Engl. J. Med. 358(5), 453–463 (2008). [CrossRef]
6. J. P. Coles, T. D. Fryer, P. Smielewski, K. Rice, J. C. Clark, J. D. Pickard, and D. K. Menon, “Defining ischemic burden after traumatic brain injury using O-15 PET imaging if cerebral physiology,” J. Cereb. Blood Flow Metab. 24(2), 191–201 (2004). [CrossRef]
7. J. Shi, Y. Wang, T. Chen, D. Xu, H. Zhao, L. Chen, C. Yan, L. Tang, Y. He, H. Feng, and J. Yao, “Automatic evaluation of traumatic brain injury based on terahertz imaging with machine learning,” Opt. Express 26(5), 6371–6381 (2018). [CrossRef]
8. Y. Xiong, A. Mahmood, and M. Chopp, “Animal models of traumatic brain injury,” Nat. Rev. Neurosci. 14(2), 128–142 (2013). [CrossRef]
9. D. W. Marion, K. C. Curley, K. Schwab, R. R. Hicks, and T. B. I. D. W. m, “Proceedings of the Military mTBI Diagnostics Workshop, St. Pete Beach, August 2010,” J. Neurotrauma 28(4), 517–526 (2011). [CrossRef]
10. M. S. Xydakis, G. S. F. Ling, L. P. Mulligan, C. H. Olsen, and W. C. Dorlac, “Epidemiologic aspects of traumatic brain injury in acute combat casualties at a major military medical center: A cohort study,” Ann. Neurol. 72(5), 673–681 (2012). [CrossRef]
11. H. S. Levin, E. Wilde, M. Troyanskaya, N. J. Petersen, R. Scheibel, M. Newsome, M. Radaideh, T. Wu, R. Yallampalli, Z. Chu, and X. Li, “Diffusion Tensor Imaging of Mild to Moderate Blast-Related Traumatic Brain Injury and Its Sequelae,” J. Neurotrauma 27(4), 683–694 (2010). [CrossRef]
12. M. F. Mendez, E. M. Owens, G. R. Berenji, D. C. Peppers, L. J. Liang, and E. A. Licht, “Mild traumatic brain injury from primary blast vs. blunt forces: Post-concussion consequences and functional neuroimaging,” NeuroRehabilitation 32(2), 397–407 (2013). [CrossRef]
13. C. L. Mac Donald, A. M. Johnson, D. Cooper, E. C. Nelson, N. J. Werner, J. S. Shimony, A. Z. Snyder, M. E. Raichle, J. R. Witherow, R. Fang, S. F. Flaherty, and D. L. Brody, “Detection of Blast-Related Traumatic Brain Injury in U.S. Military Personnel,” New Engl. J. Med. 364(22), 2091–2100 (2011). [CrossRef]
14. S. Mondello, U. Muller, A. Jeromin, J. Streeter, R. L. Hayes, and K. K. W. Wang, “Blood-based diagnostics of traumatic brain injuries,” Expert Rev. Mol. Diagn. 11(1), 65–78 (2011). [CrossRef]
15. A. Gyorgy, G. Ling, D. Wingo, J. Walker, L. Tong, S. Parks, A. Januszkiewicz, R. Baumann, and D. V. Agoston, “Time-Dependent Changes in Serum Biomarker Levels after Blast Traumatic Brain Injury,” J. Neurotrauma 28(6), 1121–1126 (2011). [CrossRef]
16. Y.-L. Ning and Y.-G. Zhou, “Shock tubes and blast injury modeling,” Chin. J. Traumatol. 18(4), 187–193 (2015). [CrossRef]
17. D. S. DeWitt and D. S. Prough, “Blast-Induced Brain Injury and Posttraumatic Hypotension and Hypoxemia,” J. Neurotrauma 26(6), 877–887 (2009). [CrossRef]
18. A. Petzold, G. Keir, J. Warren, N. Fox, and M. N. Rossor, “A systematic review and meta-analysis of CSF neurofilament protein levels as biomarkers in dementia,” Neurodegener. Dis. 4(2-3), 185–194 (2007). [CrossRef]
19. D. V. Agoston and M. Elsayed, “Serum-based protein biomarkers in blast-induced traumatic brain injury spectrum disorder,” Front. Neurol. 3, 107 (2012). [CrossRef]
20. K. Blennow, M. Jonsson, N. Andreasen, L. Rosengren, A. Wallin, P. A. Hellstrom, and H. Zetterberg, “No neurochemical evidence of brain injury after blast overpressure by repeated explosions or firing heavy weapons,” Acta Neurol. Scand. 123(4), 245–251 (2011). [CrossRef]
21. C. Bir, P. VandeVord, Y. Shen, W. Raza, and E. M. Haacke, “Effects of variable blast pressures on blood flow and oxygen saturation in rat brain as evidenced using MRI,” Magn. Reson. Imaging 30(4), 527–534 (2012). [CrossRef]
22. A. S. Woods, B. Colsch, S. N. Jackson, J. Post, K. Baldwin, A. Roux, B. Hoffer, B. M. Cox, M. Hoffer, V. Rubovitch, C. G. Pick, J. A. Schultz, and C. Balaban, “Gangliosides and Ceramides Change in a Mouse Model of Blast Induced Traumatic Brain Injury,” ACS Chem. Neurosci. 4(4), 594–600 (2013). [CrossRef]
23. H. B. Liu, Y. Q. Chen, G. J. Bastiaans, and X. C. Zhang, “Detection and identification of explosive RDX by THz diffuse reflection spectroscopy,” Opt. Express 14(1), 415–423 (2006). [CrossRef]
24. M. Bernier, F. Garet, and J.-L. Coutaz, “Precise Determination of the Refractive Index of Samples Showing Low Transmission Bands by THz Time-Domain Spectroscopy,” IEEE Trans. Terahertz Sci. Technol. 3(3), 295–301 (2013). [CrossRef]
25. P. C. Ashworth, E. Pickwell-MacPherson, E. Provenzano, S. E. Pinder, A. D. Purushotham, M. Pepper, and V. P. Wallace, “Terahertz pulsed spectroscopy of freshly excised human breast cancer,” Opt. Express 17(15), 12444–12454 (2009). [CrossRef]
26. S. Fan, E. P. J. Parrott, B. S. Y. Ung, and E. Pickwell-MacPherson, “Calibration method to improve the accuracy of THz imaging and spectroscopy in reflection geometry,” Photonics Res. 4(3), A29–A35 (2016). [CrossRef]
27. H. Hirori, K. Yamashita, M. Nagai, and K. Tanaka, “Attenuated total reflection spectroscopy in time domain using terahertz coherent pulses,” Jpn. J. Appl. Phys. 43(No. 10A), L1287–L1289 (2004). [CrossRef]
28. Z. G. Wang, L. Y. Sun, Z. H. Yang, H. G. Leng, J. X. Jiang, H. G. Yu, J. H. Gu, and Z. F. Li, “Development of serial bio-shock tubes and their application,” Chin. Med. J. 111, 109–113 (1998).
29. O. Uddin, P. E. Studlack, S. Parihar, K. Keledjian, A. Cruz, T. Farooq, N. Shin, V. Gerzanich, J. M. Simard, and A. Keller, ““Chronic pain after blast-induced traumatic brain injury in awake rats,” Neurobiol. Pain 6, 100030 (2019). [CrossRef]
30. H.-Y. Shen, Y. Zhao, X.-Y. Chen, R.-P. Xiong, J.-L. Lu, J.-F. Chen, L.-Y. Chen, and Y.-G. Zhou, “Differential Alteration of Heat Shock Protein 90 in Mice Modifies Glucocorticoid Receptor Function and Susceptibility to Trauma,” J. Neurotrauma 27(2), 373–381 (2010). [CrossRef]
31. M. Lu, J. L. Chen, D. Y. Lu, L. Yi, A. Mahmood, and M. Chopp, “Global test statistics for treatment effect of stroke and traumatic brain injury in rats with administration of bone marrow stromal cells,” J. Neurosci. Methods 128(1-2), 183–190 (2003). [CrossRef]
32. C. Zhang, J. Chen, and H. Lu, “Expression of aquaporin-4 and pathological characteristics of brain injury in a rat model of traumatic brain injury,” Mol. Med. Rep. 12(5), 7351–7357 (2015). [CrossRef]
33. S. Yu, T. Hayashi, Y. Kaneko, E. Bae, C. E. Stahl, Y. Wang, and C. V. Borlongan, “Severity of Controlled Cortical Impact Traumatic Brain Injury in Rats and Mice Dictates Degree of Behavioral Deficits,” Brain Res. 18, 241–242 (2009). [CrossRef]