Background
Schizophrenia [
1] is a mental disorder characterized by abnormal social behavior and failure to understand what is real. Common symptoms include false beliefs, unclear or confused thinking, hearing voices that others do not hear, reduced social engagement and emotional expression, and a lack of motivation. It not only produce great pain to the patients but also bring a heavy burden to their family.
fNIRS is a haemodynamic-based technique for the assessment of functional activity in the human brain [
2]. Based on the tight coupling of neural activity and oxygen delivery [
3], changes in the concentration of oxygenated and deoxygenated haemoglobin are noninvasively measured by fNIRS and taken as indicators for cortical activation. The typical fNIRS signal observed after neural activation is a decrease of deoxygenated accompanied by an increase of oxygenated comparable in time course to the blood oxygenation level dependent signal of fMRI [
4]. fNIRS provides comprehensive information about haemodynamics consisting of oxygenated, deoxygenated and changes in total haemoglobin. It is characterised by its straightforward application which resembles in the outward appearance more an electroencephalogram. Thus, the data collection is comfortable for the subjects because of the less constrictive measurement circumstances which probably lead to more ecologically valid conditions than in other neuroimaging methods [
5]. These inherent advantages accompanied by the rapid developments in technology and methodology enabled fNIRS to easily enter psychological, psychiatric and basic research on children, adults and elderly subjects.
Because the symptom of schizophrenia is similar with other diseases, such as depression and anxiety. Doctors can only use the information of genetic predisposition, substance abuse, living conditions and prenatal stressors to predict the schizophrenia is triggered or not. Usually it is not generate immediately but takes years for the disease to surface. So design a computer aided identification method can help improving the doctor’s diagnostic result. With that being the case, many patients can have access to a proper medication, as such, the wellbeing of the patients and the medical quality of hospitals will increase.
During the past several years, many studies have applied the fNIRS technique to investigate the brain activation patterns in patients with schizophrenia. Converging evidence suggests schizophrenia patients are often associated with reduced activities and inappropriate activity timing around the bilateral prefrontal cortex during a verbal fluency task or other cognitive tasks [
6]. Based on these findings, some studies have attempted to apply the fNIRS signal as a diagnostic tool with different pattern recognition methods. In [
7], authors measured the changes of the oxy-Hb signal during multiple cognitive tasks from two fNIRS channels located in the bilateral prefrontal areas and then applied stepwise linear discriminant analysis to distinguish patients with schizophrenia from healthy subjects. They separated the total sample into two groups, and each group consisted of 60 subjects (including 30 patients with schizophrenia and 30 age-and gender-matched healthy controls). The experimental results demonstrated that there was an accuracy rate of 88.3% for classification in the first group, and the discrimination function derived from the first group correctly differentiated 75% of the subjects in the second group. To integrate spatial and temporal information in multichannel fNIRS, [
8] employed a novel probabilistic pattern recognition method called Gaussian process classifier for the diagnostic classification of schizophrenia. Using the temporal patterns of fNIRS data measured during a working memory task, an overall accuracy of 76% was achieved in a group of 80 samples. And [
9] applied a 52 channel fNIRS system to identify the significantly different regions in the prefrontal cortex during a verbal fluency test and then utilized a k-means clustering method for discriminant analysis between schizophrenia patients and healthy subjects. The results indicated 68.69% and 71.72% of the participants were correctly classified as schizophrenic or healthy subjects with all 52 channels and six significantly different channels, respectively. And [
10] proposed a method using principal component analysis and SVM to discriminate patients with schizophrenia from health controls using a large sample of 52 schizophrenia patients and 38 healthy controls. They achieved a satisfactory classification with the accuracy of 93.33%, 100% for schizophrenia samples and 84.62% for healthy controls.
Human brain network is one of the complex networks [
11,
12]. Researchers have used the complex network theory [
13,
14] to construct the brain network, then analyze the constructed brain network using complex network theory and calculate index of the brain network for further study. The brain network can be divided into structural and functional brain network [
15,
16]. The nodes and edges are two key elements in the brain network. Diffusion tensor imaging and diffusion spectrum imaging [
17,
18] are two imaging techniques used in structural brain network. Since the two techniques can track the direction of the mental fiber electrical signal, the structural brain network is considered as directional. The definition of the node of functional brain network is changing with different imaging techniques. Generally we define the connections between nodes by calculating pearson correlation or partial correlation which is used to describe the statistical significance of functional brain signals over a period of time. The functional network is non-directional since the correlation between nodes only reflects their statistical significance, no causal relationship.
In this paper, we designed a cognitive task and recruited a group of subjects to perform this task. The group included 42 schizophrenia patients and 34 healthy controls and a 52 multichannel fNIRS system was used to examine the hemodynamic signals in the bilateral prefrontal and superior temporal cortices during the cognitive task. Then we used CBNA to extract the effective features between schizophrenia patients and healthy controls. Finally we trained the SVM classifier and evaluated it with leave-one-out cross validation. The results show that the proposed approach has the high potential to be a promising clinical tool in the objective diagnosis and treatment of psychiatric disorders.
Results and discussion
The testing result of schizophrenics and healthy controls is shown in Table
1, where 39 of the 42 schizophrenia and 26 of 34 health controls were discriminated successfully on Oxy-Hb/Deoxy-Hb signal. The method based on CBNA and SVM successfully discriminated 65 (39 schizophrenia and 26 healthy persons) signals with an overall accuracy of 85.5% for fNIRS classification on testing set. And on total signal, where 39 of the 42 schizophrenia and 22 of 34 health controls were discriminated successfully. The method successfully discriminated 61 (39 schizophrenia and 22 healthy persons) signals with an overall accuracy of 80.3%.
Table 1
The finally testing result
Oxy-Hb | 85.5% | 76.5% | 92.8% |
Deoxy-Hb | 85.5% | 76.5% | 92.8% |
Total | 80.3% | 64.7% | 92.8% |
Our study was a binary classification, and we first defined the class of schizophrenia patient as positive and the class of healthy control as negative. Then, TP is the number of schizophrenia patients correctly predicted; TN is the number of healthy controls correctly predicted; FP is the number of healthy controls classified as schizophrenia patients; and the FN is the number of schizophrenia patients classified as healthy controls. Finally, the performance of a classifier can be quantified by using the ACC, SS and specificity or TNR; These measures are defined as follows:
$$\begin{array}{@{}rcl@{}} ACC&=&\frac{TP+TN}{TP+FP+TN+FN} \end{array} $$
(3)
$$\begin{array}{@{}rcl@{}} SS&=&\frac{TP}{TP+FN} \end{array} $$
(4)
$$\begin{array}{@{}rcl@{}} TNR&=&\frac{TN}{TN+FP} \end{array} $$
(5)
ACC represents the ration between correctly classified samples and total samples. SS represents the ratio between correctly classified schizophrenic patients and total schizophrenic patients. TNR represents the ratio between correctly classified health controls and total health controls. Therefore, a good fNIRS-aided diagnostic classifier is assumed to have larger ACC and TNR values.
After 76 rounds of cross validation, there are totally 11 misclassified cases on Oxy-Hb signal. Including 8 cases of normal people and 3 cases of schizophrenic patients. The classification accuracy is 85.5%, specificity is 76.5%, sensitivity is 92.8%. Table
2 shows more details about the testing results on Oxy-Hb signal. For the Deoxy-Hb signal, there are also 11 misclassified cases. Including 8 cases of normal people and 3 cases of schizophrenic patients. The classification accuracy is 85.5%, specificity is 76.5%, sensitivity is 92.8%, the same as Oxy-Hb signal. For the total signal is shown in Table
3. There are 15 misclassified cases. Including 12 cases of normal people and 3 cases of schizophrenic patients. The classification accuracy is 80.3%, specificity is 64.7% and sensitivity is 92.8%. Here we choose Oxy-Hb signal to discriminate schizophrenic. This accuracy is especially satisfactory for the discrimination.
Table 2
Testing result of schizophrenic and healthy on Oxy-Hb signal
CBNA+SVM | | | Classified results |
| | | 1(Schizophrenia) | -1(healthy) |
| | | 47 | 29 |
Data set | 1(Schizophrenia) | 42 | 39 | 3 |
| -1(healthy) | 34 | 8 | 26 |
Accuracy of schizophrenia (SS) | 39/42=92.8% |
Accuracy of healthy((TNR)) | 26/34=76.5% |
Classification accuracy(ACC) | 65/76=85.5% |
Table 3
Testing result of schizophrenic and healthy on total signal
CBNA+SVM | | | Classified results |
| | | 1(Schizophrenia) | -1(healthy) |
| | | 61 | 25 |
Data set | 1(Schizophrenia) | 42 | 39 | 3 |
| -1(healthy) | 34 | 12 | 22 |
Accuracy of schizophrenia (SS) | 39/42=92.8% |
Accuracy of healthy((TNR)) | 22/34=64.7% |
Classification accuracy(ACC) | 61/76=80.3% |
Conclusion
Our study demonstrated that the designed task is an effective experimental paradigm. Compared with healthy controls, the multichannel fNIRS results on the sample confirmed that schizophrenia patients in the Chinese population had significant lower brain activation over the prefrontal cortex and superior temporal cortex. Finnally, we achieved a considerable overall classification accuracy of 85.5% (65/76) using the SVM classifier and CBNA based feature selection on the oxy-Hb signal. Thus, SVM had the good classification performance especially after performing the CBNA based feature selection. Our results illustrated that, by using the appropriate classification method, fNIRS represents a promising diagnostic tool to differentiate schizophrenia patients from healthy controls.
Acknowledgements
The authors would like to thank the staff of Peking University Sixth Hospital for welcoming and assisting the research team and the interview participants for their time.