1.IntroductionDepression is a prevalent psychological disorder characterized by persistent feelings of sadness and loss of interest in activities, significantly impacting the lives and work of individuals during the onset of the illness. Recognized as a pressing concern by the WHO’s Mental Health Gap Action Program (mhGAP),1 of the global population suffers from this condition. The severity of depression, as assessed by the Hamilton Depression Rating Scale (HAMD),2 can be roughly categorized as mild, moderate, and severe. However, the diagnosis process often involves subjective evaluations, such as clinical interviews and professional medical assessments, which may impede the timely and accurate diagnosis of severe depression. Moreover, due to the shortage of mental health workers, only 9.2% of individuals with depression receive appropriate treatment, whereas some individuals with anxiety disorders or mild depression are mistakenly prescribed antidepressant medications.3 Research4 has indicated that antidepressants are ineffective in treating mild depression. Thus, there is an urgent need to develop an objective and efficient approach to quickly differentiate between mild and major depression for early diagnosis and effective treatment evaluation. Several studies5,6 have found evidence of structural and functional changes in various brain regions among individuals with severe depression. These changes include abnormal activity in prefrontal, limbic, thalamic, and cortical areas. In prefrontal regions, increased depression severity is associated with decreased neural activity in the low-frequency range and increased activity in the high-frequency range.7 In addition, severe depression is characterized by extremely low mood, which may be accompanied by severe sleep problems, loss of appetite, and loss of interest in daily activities. Even for mild depression, the changes in the brain and physiology may be relatively mild but can still lead to problems with mood, concentration, and cognition.8,9 Depression severity has been found to be significantly associated with cognitive performance in situational memory, executive functioning, and processing speed, and as depression severity increases, these cognitive abilities decline.10 Impaired cognitive function in depressed patients results in reduced regional cerebral blood flow (rCBF) in the frontal, temporal, and anterior cingulate gyrus during cognitive tasks.11 Functional near-infrared spectroscopy (fNIRS) serves as a non-invasive, portable tool for monitoring hemodynamic responses in the cerebral cortex, commonly utilized to observe hemodynamic changes in psychiatric patients undertaking cognitive tasks due to its simplicity, safety, and resistance to interference. When performing a cognitive task, the brain experiences an oversupply of rCBF, leading to the hemodynamic response that can be measured through fNIRS.12 Researchers13–15 found that values of oxyhemoglobin concentration in the prefrontal lobe measured by the fNIRS device were negatively correlated with HAMD scores, which were associated with depression severity, suggesting that changes in oxygen-hemoglobin are a potential biomarker for recognizing depression severity. In recent years, several researchers16,17 utilized machine learning techniques to construct prediction models that differentiate between severe depression and healthy individuals based on fNIRS data, obtaining favorable classification performance. Although these studies13,18–20 suggest that machine learning prediction models developed based on fNIRS data serve as effective analytical tools for identifying patients with severe depression, few studies have focused on discriminating depression severity. Ramasubbu et al.21 developed a predictive model for distinguishing between patients with mild depression and severe depression but did not achieve high accuracy. Richter et al.22 proposed a diagnostic system (25 clinical anxiety or depression patients, 76 healthy control participants) and achieved an accuracy of 81.69% in distinguishing the mixed anxiety/depression group from the control group, but it only achieved a success rate of 50.66% in differentiating the anxiety group from the depression group. Inspired by previous researches,14,15,23 we found that fNIRS measurement can be instrumental in distinguishing the severity of depression. In this study, we proposed a predictive model based on fNIRS data to differentiate between patients with mild and severe depression. This model provides mental health professionals with an objective and effective diagnosis method for determining the severity of depression. 2.Materials2.1.ParticipantsIn this study, 140 patients with drug-naïve, first-episode depression were recruited with their permission and informed consent. All subjects had no other neurological disorders (stroke, brain tumor, severe concussion, migraine, etc.), cardiovascular (myocardial infarction, arrhythmia, etc.) diseases, and no obvious impairment in vision and hearing. The experiment was conducted under the guidance of a professional psychiatrist. No data identifying the participants were recorded during the entire research process. A senior professional psychiatrist conducted interviews and observations, utilizing the HAMD2 scale to evaluate depressive symptom severity. According to the diagnosis of the psychiatrist, among these 140 first-episode depression patients, 58 individuals were assessed as mildly depressed, whereas the remaining 82 subjects were classified as having severe depression. More details of the participants are shown in Table 1. Table 1Subject information of our data.
2.2.Activation Task (Verbal Fluency Task)The verbal fluency task (VFT) is a cognitive task commonly used in fNIRS research, which primarily reflects executive function and has been correlated with a number of basic neurocognitive activities (e.g., working memory, motivation, and attention).24 In China, the Chinese version of the VFT has been widely used in the diagnosis of psychiatric disorders as a sensitive indicator for assessing deficits in domains of cognition and executive function that depend on activation in prefrontal regions.25 Before the VFT task, the participants were instructed to stay in a separate, quiet room and sit on a chair, ensuring that their eyes were directly in front of the monitor at a distance of 60 cm. Participants were asked to look at the “+” symbol on the monitor and keep their heads still. At the beginning of each experiment, during a 30-s rest period before the task, a guiding voice prompt was played: “Check begins, please repeat the reading 12345.” Concurrently, the fNIRS data were recorded. Next, during a 60-s task period, the participants were required to complete three tasks involving the formation of words using the three characters “zhong,” “ri,” and “lan,” respectively. Each word formation task had a time limit of 20 s. Finally, during a 30-s resting period after the task completion, the instructional audio prompt stating, “Please repeat reading 12345 until the end of the check” was played and the fNIRS data acquisition was terminated. The entire VFT process is shown in Fig. 1(b). 2.3.fNIRS Data AcquisitionIn the experiments, an fNIRS instrument (ETG-4100 Optical Topography System) manufactured by Hitachi Medical is utilized to obtain data. This device utilizes near-infrared light at two specific wavelengths ( and ) to penetrate the scalp, skull, and cerebrospinal fluid, enabling irradiation of the cerebral cortex. The instrument consists of 17 transmitters and 16 detectors, forming a total of 52 channels, including 18 in the prefrontal lobes and 17 channels each in the right and left frontal lobes. The distance between each pair of transmitters and detectors was 3.0 cm, with a sampling frequency of the instrument was 10 Hz. During the VFT, neural activity leads to changes in CBF, causing changes in brain oxygen-hemoglobin () and deoxygen-hemoglobin ().6 According to Lambert-Beer,26 changes in and can be deduced from the NIRS signals measured during the task. These changes provide a more accurate reflection of the activation state of the cerebral cortex in relation to cognitive tasks. The entire process of data collection in the experiment is shown in Fig. 1. 3.MethodsTo predict the severity of depression, we developed a prediction model for mild and severe depression. The model involves data preprocessing, feature extraction, feature selection, feature fusion, and classification prediction, as shown in Fig. 2. 3.1.Data Pre-ProcessingWe utilized the HOMER2 toolbox in MATLAB 2018b to process NIRS data and obtained the change in and using the Beer-Lambert law.27 These hemodynamic parameters are recognized as sensitive indicators for investigating psychiatric disorders.28 However, the raw hemodynamic signals often contain various types of noise that can affect the feature extraction process. Therefore, preprocessing of the fNIRS data is essential to remove noise before further analysis.29 We utilized a polynomial regression model to estimate linear or nonlinear trends and then subtracted this trend from the original hemoglobin concentration signal to obtain detrended data. The temporal derivative distribution repair method30 was employed for motion artifact correction. In addition, a third-order Chebyshev type II filter31 with a cutoff frequency of 0.01 Hz and a stopband frequency of 0.2 Hz was used to remove most physiological noise [e.g., respiration (0.2 to 0.4 Hz) and heartbeat (0.5 to 2.0 Hz)]. However, various artifacts, such as ambient light, electrical interference from the instrumentation, and positional movements, still contaminated the effective frequency band. In previous studies,32,33 empirical mode decomposition (EMD) and ensemble empirical mode decomposition (EEMD) methods have been used to eliminate motion artifacts and baseline drifts in physiological signals. In this study, to enhance the signal-to-noise ratio (SNR) and facilitate subsequent feature extraction, we applied a complete EEMD with adaptive noise-wavelet threshold (CEEMDAN-WPT) denoising method to eliminate noise contaminating the effective frequency band of fNIRS data. The flowchart of the CEEMDAN-WPT algorithm is shown in Fig. 3, and the specific implementation steps are detailed below.
3.2.Methodology for the Effective Feature and Identified ModelIn our study, we employed a methodology to extract key features for effective classification of patients with mild and severe depression. This involved identifying critical features from the data obtained from channels. Subsequently, we established a recognition model using these key features to classify patients with severe depression and those with mild depression. Figure 4 shows a detailed description of feature selection and classification model construction process in this study. To ensure an objective assessment of the classification performance, the analysis was conducted exclusively on the training set within the blue-bordered box, whereas the final selected features were validated on the test set within the orange-bordered box. For comprehensive model evaluation, we utilized nested cross-validation (black-bordered box) to obtain model metrics. The principle of nested cross-validation, as shown in Fig. 4, involves two loops: an outer loop and an inner loop. The inner loop consists of cross-validation and search for the best hyperparameters of the model, such as grid search, which determines the optimal hyperparameters for the outer loop. Meanwhile, the outer loop provides training data to the inner loop while retaining some data for testing the models within the inner loop. The entire process was repeated for all possible fold combinations within the outer loop, providing a robust estimation of the model performance. This approach prevents information leakage and overfitting. 3.2.1.Feature extractionTemporal featureConsidering that the actual task time for VFT is only 60 s, we extracted the temporal features (TF) within this time period. These features encompass time domain characteristics, frequency domain attributes, and information entropy. The time domain features consist of maximum, minimum, mean, rectification average, skewness, and peak factors. Mean squared frequency was selected as frequency domain feature. Detailed information about the selected time and frequency domain features can be found in Table S1 in the Supplementary Material. Power spectral entropy and singular spectral entropy serve as the information entropy features, calculated as follows: In the equation, represents the power spectral entropy, and denotes the energy distribution of the signal in the frequency domain: where is the singular spectral entropy and represents the singular value spectrum obtained through the singular value decomposition of the signal.Correlation featureIn our study, the correlation among channels was extracted as the correlation features (CF). These were obtained by calculating the Pearson correlation coefficient19 of the change in and between 18 frontal channels of each participant. The equation for calculating the CF is as follows: where and represent the concentration change of hemoglobin in different channels of the participant, respectively.Therefore, a total of TF and CF can be extracted from the change curves of and in the 18 frontal channels of each participant. 3.2.2.Feature selectionIn the classification task, when the sample size is small and the number of features is large, it can lead to the problem of dimensionality curse and reduce the accuracy. To address this issue, recursive feature elimination with cross-validation (RFECV) method34 can be utilized to select the optimum features from TF and CF. The RFECV method utilizes RFE35 to obtain the importance of each feature and find the optimal number of features based on cross-validation accuracy. Compared to dimension reduction methods, such as PCA,36 the RFECV method can eliminate redundant feature information and determine the most impactful features for optimal classification performance. In our prediction model, we applied REFCV method twice. First, the optimal 47 TF and 54 CF were screened separately. Then, after fusing the optimal CF and TF selected in the first step, we further identified the best 35 features for distinguishing between mild and severe depression. 3.2.3.Classification modelsGiven the limited number of samples in depression data, we employed machine learning techniques to construct the classification prediction. Four classical algorithms were utilized, including two supervised learning methods [logistic regression (LR)37 and support vector machines (SVM)38], an ensemble learning algorithm [random forest (RF)],39 and an artificial neural networks [multi-layer perceptron (MLP)].40 According to the participants, the dataset was split into two parts: a training set comprising 60% (49 severe depression and 35 mild depression) and a testing set comprising 40% (33 severe depression and 23 mild depression). To identify the optimal prediction model for each pattern, a grid search41 based on hyperparameter turning42 was employed to determine the model parameter with the highest accuracy. 3.3.Model MetricsThe denoising performance of the CEEMDAN-WPT algorithm is evaluated using two metrics: SNR and root mean square error (RMSE). SNR quantifies the relationship between signal and noise intensity, where a high SNR indicates stronger useful signals compared to noise, resulting in clearer and more reliable signals. RMSE measures the average deviation between predicted values and actual values. The equations for calculating SNR and RMSE are as follows: where represents the original signal, is the the signal length, and denotes the denoised signal.For the classification models, we calculated the area under the ROC curve (AUC) for different models to compare their performance differences. In addition, several other metrics, including specificity, sensitivity, accuracy, and score, were calculated. Supplementary Material provide detailed calculation methods for these metrics. 4.Results and Analysis4.1.Result on DenoisingAs shown in Fig. 5, the original signal is decomposed into five high-frequency components and five low-frequency components. The high-frequency components denoised by wavelet packet threshold and low-frequency components are reconstructed into a new signal, which is visibly smoother. Figure 6 shown the spectrum of the original signal and the signal processed using the CEEMDAN-WPT method. In Fig. 6(a), it can be observed that even after filtering, a small amount of noise remains mixed within the effective frequency range. However, our proposed method effectively removes this noise while retaining the information relevant to neural activity. In Table 2, compared to EMD,43 the CEEMDAN-WPT method significantly improves SNR and reduces the RMSE. Compared to wavelet filtering,44 although there is a slight difference in RMSE, the SNR is improved by 4.1 dB. Table 2Performance comparison of three algorithms.
4.2.Result on Feature SelectionThrough the RFECV algorithm, we identified 35 optimal features that are related to the severity of depression. As shown in Fig. 7, we plotted some violin figures regarding the features. It can be seen from the figure that there are significant differences between mild depression and severe depression subjects in terms of peak factor, skewness, singular spectral entropy, and correlation in certain channels. Conversely, the differences in maximum, minimum, mean, and mean square frequency are relatively small. The SHAP diagram in Fig. 7 demonstrated that peak factor and skewness in certain channels have a significant impact on the prediction results, whereas the influence of maximum, minimum, average, and mean square frequency is relatively minor. 4.3.Activation AnalysisIndividual-level statistical analysis was performed on preprocessed fNIRS data using the general linear model to calculate beta values representing brain activation intensity during the VFT task. The activation levels of individual channels were evaluated using a one-sample -test, and the heat map of channel activation -values is presented in Fig. 8(a). Channels, such as ch1, ch3, ch9, ch10, ch11, ch12, ch17, ch19, ch33, and ch40, did not show significant activation during the VFT task, whereas the remaining channels exhibited significant activation. In Fig. 8(a), 52 channels were mapped onto the brain, with non-activated channels mainly located in the left and right temporal lobes. To compare brain activation intensity between the severe depression and mild depression groups, a paired -test was conducted. The results indicated that, during the VFT task, the activation levels in the mild depression group were higher than those in the severe depression group. Notably, there was a significant difference in channel activation intensity in the frontal lobe region, as depicted in Figs. 8(c) and 8(d). The Supplementary Material contain detailed results of both the one-sample and paired sample -tests. 4.4.Result on Classification ModelWe evaluated the TF, the CF, and the fusion of both as inputs for the four classification models and validated the results on the test set, as shown in Table 3. Across all models, the fusion of temporal and CFs consistently outperformed individual temporal or CFs. Furthermore, in LR, MLP, and SVM models, the AUC of the fusion features was greater than 0.9, indicating its ability to better distinguish between mild and severe depression. Notably, the fusion features of temporal and correlation fusion features showed outstanding performance in the SVM classification model, with an accuracy of 92.8%, a sensitivity of 91.6%, a specificity of 93.7%, and an -score of 92.7%. Table 3The results of classification models.
Abbreviation: RF, random forest; LR, logistic regression; MLP, multi-layer perceptron; SVM, support vector machine; TF, temporal features; CF, correlation features; TCF, temporal and correlation fusion features. 5.DiscussionCerebral blood flow (CBF) is associated with brain activity and metabolism. During cognitive tasks, rCBF increases12 while significant changes in CBF occur in patients with depression, particularly in the prefrontal region. Specifically, rCBF is lower in patients with mild depression compared to the normal control group, whereas it is higher in patients with moderate to severe depression compared to the rCBF in most cortical regions of the normal control group.45,46 The changes in rCBF can lead to hemodynamic responses, which can be measured using an fNIRS device, making it an important tool for studying depression. However, the fNIRS device is sensitive to motion artifacts and noise, posing a challenge in filter out noise while retaining depression-related information. Some researchers47 attempted to use wavelet filtering to remove the motion artifacts from fNIRS data and compared it with other methods for correcting motion artifacts, such as principal component analysis, spline interpolation, and Kalman filtering. The results indicated that wavelet filtering had better performance. However, in this paper, the CEEMADN-WPT method was utilized to remove noise from fNIRS data, which is a more powerful technique. Compared to signals filtered by wavelet filtering, the signal filtered by CEEMADN-WPT had higher SNR (as shown in Table 2). With this denoising method, we effectively reduced the noise mixed in the valid frequency band in fNIRS data, thus preserving information related to depression. Shallow tissues, such as the scalp, skull, and meninges, are abundant in capillaries. The concentration of hemoglobin within these capillaries fluctuates due to respiration, heartbeat, and task-related autonomous neural activity. When near-infrared light penetrates these shallow tissues, changes in hemoglobin concentration lead to variations in fNIRS light attenuation, referred to as shallow physiological noise. The “short-separation” method (channels where the emitter-detector distance is below 1.0 cm) is recognized as an effective approach of removing shallow noise. This method involves using additional short-separation fNIRS channels to record shallow physiological noise and then subtract it from the signal. The absence of physiological noise data during our data collection process has posed a certain limitation on our preprocessing of fNIRS data. In future studies, we plan to collect both fNIRS data and physiological noise data to better capture data related to brain neural activity. In this study, temporal and channel CFs extracted from fNIRS data of the prefrontal lobe of patients with severe and mild depression were compared. Our findings suggest that using TFs as inputs leads to higher accuracy compared to using only CFs. Furthermore, the fusion of temporal and CFs obtained the highest accuracy with an AUC of 96.3% on SVM models. These results indicate that studying the functional connectivity of the brain alone is insufficient in distinguishing between patients with mild and severe depression. By combining the temporal and CFs of fNIRS data, the classification accuracy can be improved. Through feature visualization, significant differences can be observed between patients with mild depression and those with severe depression in certain TFs, such as peak factor, skewness, and singular spectral entropy. In addition, the peak factor and skewness of and curves play a crucial role in the output of the model. The research results indicate that temporal and channel CFs extracted from and are important for distinguishing the severity of depression, and these features have the potential to serve as biomarkers for distinguishing between mild and severe depression, which is consistent with previous research results.13 The findings of our research have several potential applications in clinical services. The diagnostic support system based on fNIRS that we propose provides a non-invasive, objective, and quantitative method for assessing the severity of depression, which can serve as a supplement to traditional diagnostic methods, such as clinical interviews or self-report questionnaires. By combining fNIRS technology and machine learning algorithms, our model can extract valuable information from brain activity patterns that are difficult to detect using traditional methods. These results can serve as an auxiliary tool for psychiatrists, helping them accurately assess the severity of depression in patients and develop more personalized treatment plans. In addition, our research outcomes can be utilized for monitoring treatment progress. By regularly conducting fNIRS tests, clinicians can objectively understand changes in patient’s conditions and make timely adjustments to treatment plans for better therapeutic efficacy. It is important to note that this study has several limitations. First, the small sample size of 140 participants, including 82 cases of severe depression and 58 cases of mild depression, may have limited the precision of our model. A larger dataset is needed to improve the effectiveness of the model. The participants in the study were drug-naive, first-episode depression patients receiving outpatient treatment at the Mental Health Center in Chongqing City. Therefore, our findings are not influenced by medication, or previous treatments, as these data were not included in the analyzed dataset. However, our study did not consider potential differences among the participants, such as long-term smoking, alcohol consumption, the number of past depressive episodes, or a family history of mental illness, all of which could have had an impact on the severity of depression. In addition, the diagnostic data were obtained from a single psychiatric research center, and the generalizability of our model to other hospitals has not been validated. Furthermore, due to the small sample size, patients with moderate depression were not included in the study, necessitating further research in this area. Future work will involve collecting a larger sample size to investigate the severity of depression and analyzing clinical information among participants. We also plan to develop a deep learning-based predictive model for classifying mild, moderate, and severe depression and embed the model into the software used by hospitals. By connecting to the fNIRS database and integrating clinical information from patients, we aim to obtain more accurate and personalized diagnostic results through big data analysis, providing psychiatrics with a more reliable and efficient tool to support their decision-making in depression treatment. 6.ConclusionsThis paper proposes a depression severity prediction model based on the temporal and CFs extracted from fNIRS data, achieving high classification accuracy for mild and severe depression. As an objective auxiliary tool, this model can improve the diagnostic efficiency of depression severity and assist doctors in clinical diagnosis, which is of great significance for the treatment of depressed patients. Code and Data AvailabilityThe data and materials that support the findings of this study are available upon request. Due to privacy and confidentiality concerns, the raw data cannot be publicly shared. However, interested researchers can contact Zhiyong Huang at zyhuang@cqu.edu.cn to request access to the data used in this study. The code used to generate the results and figures is available in a Github repository: https://github.com/liumanliu/fNIRS_depression. Ethical ApprovalThe experiment was conducted under the supervision of a professional psychiatrist. No identifiable data of the participants were recorded during the whole study. The experiment proposal was approved by the Ethics Committee of Chongqing Mental Health Center (2020-037). Authors’ ContributionsMan Liu and Zhiyong Huang wrote the main manuscript text, whereas the experiment was conducted under the supervision of professional psychiatrist Hui Yang. Yunlan Zhao and Xiao Han reviewed the manuscript. Huan Chen and Yaju Feng contributed equally to this work. They jointly took the management and coordination responsibility for the research activity planning and execution. AcknowledgmentsThe work was supported by grants from the Natural Science Foundation of Chongqing (Grant No. cstc2020jcyj-msxmX0641), the Fundamental Research Funds for the Central Universities (Grant No. 2020CDJ-LHZZ-025), and the Medical Research Project of Chongqing Health Commission (Grant No. 2020FYYX031). References, “Depression,”
https://www.who.int/news-room/fact-sheets/detail/depression
(2023).
Google Scholar
M. Zimmerman et al.,
“Severity classification on the Hamilton depression rating scale,”
J. Affect. Disord., 150
(2), 384
–388 https://doi.org/10.1016/j.jad.2013.04.028 JADID7
(2013).
Google Scholar
H. Baumeister,
“Inappropriate prescriptions of antidepressant drugs in patients with subthreshold to mild depression: time for the evidence to become practice,”
J. Affect. Disord., 139
(3), 240
–243 https://doi.org/10.1016/j.jad.2011.05.025 JADID7
(2012).
Google Scholar
P. A. Vöhringer and S. N. Ghaemi,
“Solving the antidepressant efficacy question: effect sizes in major depressive disorder,”
Clin. Ther., 33
(12), B49
–B61 https://doi.org/10.1016/j.clinthera.2011.11.019
(2011).
Google Scholar
S. Wang et al.,
“Aberrant patterns of brain cerebral blood flow in Chinese Han first-episode drug-naive depressive patients with and without a family history of depression,”
Oncotarget, 8
(45), 79906 https://doi.org/10.18632/oncotarget.20306
(2017).
Google Scholar
H. S. Mayberg,
“Modulating dysfunctional limbic-cortical circuits in depression: towards development of brain-based algorithms for diagnosis and optimised treatment,”
Br. Med. Bull., 65
(1), 193
–207 https://doi.org/10.1093/bmb/65.1.193 BMBUAQ 0007-1420
(2003).
Google Scholar
J. Xiao et al.,
“Decoding depression severity from intracranial neural activity,”
Biol. Psychiatry, 94
(6), 445
–453 https://doi.org/10.1016/j.biopsych.2023.01.020 BIPCBF 0006-3223
(2023).
Google Scholar
Y. Lu et al.,
“A regressional analysis of maladaptive rumination, illness perception and negative emotional outcomes in Asian patients suffering from depressive disorder,”
Asian J. Psychiatry, 12 69
–76 https://doi.org/10.1016/j.ajp.2014.06.014
(2014).
Google Scholar
G. Y. Lim et al.,
“Prevalence of depression in the community from 30 countries between 1994 and 2014,”
Sci. Rep., 8
(1), 2861 https://doi.org/10.1038/s41598-018-21243-x SRCEC3 2045-2322
(2018).
Google Scholar
L. M. McDermott and K. P. Ebmeier,
“A meta-analysis of depression severity and cognitive function,”
J. Affect. Disord., 119
(1–3), 1
–8 https://doi.org/10.1016/j.jad.2009.04.022 JADID7
(2009).
Google Scholar
K. Oda et al.,
“Regional cerebral blood flow in depressed patients with white matter magnetic resonance hyperintensity,”
Biol. Psychiatry, 53
(2), 150
–156 https://doi.org/10.1016/S0006-3223(02)01548-2 BIPCBF 0006-3223
(2003).
Google Scholar
P. Pinti et al.,
“The present and future use of functional near-infrared spectroscopy (fNIRS) for cognitive neuroscience,”
Ann. N. Y. Acad. Sci., 1464
(1), 5
–29 https://doi.org/10.1111/nyas.13948 ANYAA9 0077-8923
(2020).
Google Scholar
C. S. H. Ho et al.,
“Diagnostic and predictive applications of functional near-infrared spectroscopy for major depressive disorder: a systematic review,”
Front. Psychiatry, 11 378 https://doi.org/10.3389/fpsyt.2020.00378
(2020).
Google Scholar
M. Kawano et al.,
“Correlation between frontal lobe oxy-hemoglobin and severity of depression assessed using near-infrared spectroscopy,”
J. Affect. Disord., 205 154
–158 https://doi.org/10.1016/j.jad.2016.07.013 JADID7
(2016).
Google Scholar
T. Noda et al.,
“Frontal and right temporal activations correlate negatively with depression severity during verbal fluency task: a multi-channel near-infrared spectroscopy study,”
J. Psychiatric Res., 46
(7), 905
–912 https://doi.org/10.1016/j.jpsychires.2012.04.001
(2012).
Google Scholar
H. Song et al.,
“Automatic depression discrimination on FNIRS by using general linear model and SVM,”
in 7th Int. Conf. Biomed. Eng. and Inf.,
278
–282
(2014). https://doi.org/10.1109/BMEI.2014.7002785 Google Scholar
J. Chao et al.,
“fNIRS evidence for distinguishing patients with major depression and healthy controls,”
IEEE Trans. Neural Syst. Rehabilit. Eng., 29 2211
–2221 https://doi.org/10.1109/TNSRE.2021.3115266
(2021).
Google Scholar
Z. Li et al.,
“Identifying neuroimaging biomarkers of major depressive disorder from cortical hemodynamic responses using machine learning approaches,”
EBioMedicine, 79 104027 https://doi.org/10.1016/j.ebiom.2022.104027
(2022).
Google Scholar
Q. Yu et al.,
“GNN-based depression recognition using spatio-temporal information: a fNIRS study,”
IEEE J. Biomed. Health Inf., 26
(10), 4925
–4935 https://doi.org/10.1109/JBHI.2022.3195066
(2022).
Google Scholar
R. Wang et al.,
“Depression analysis and recognition based on functional near-infrared spectroscopy,”
IEEE J. Biomed. Health Inf., 25
(12), 4289
–4299 https://doi.org/10.1109/JBHI.2021.3076762
(2021).
Google Scholar
R. Ramasubbu et al.,
“Accuracy of automated classification of major depressive disorder as a function of symptom severity,”
NeuroImage Clin., 12 320
–331 https://doi.org/10.1016/j.nicl.2016.07.012
(2016).
Google Scholar
T. Richter et al.,
“Machine learning-based diagnosis support system for differentiating between clinical anxiety and depression disorders,”
J. Psychiatric Res., 141 199
–205 https://doi.org/10.1016/j.jpsychires.2021.06.044 JPYRA3 0022-3956
(2021).
Google Scholar
S. Y. Dong et al.,
“Prefrontal functional connectivity during the verbal fluency task in patients with major depressive disorder: a functional near-infrared spectroscopy study,”
Front. Psychiatry, 12 659814 https://doi.org/10.3389/fpsyt.2021.659814
(2021).
Google Scholar
J. Huang et al.,
“Increased prefrontal activation during verbal fluency task after repetitive transcranial magnetic stimulation treatment in depression: a functional near-infrared spectroscopy study,”
Front. Psychiatry, 13 876136 https://doi.org/10.3389/fpsyt.2022.876136
(2022).
Google Scholar
T. Yang et al.,
“The fNIRS evaluation of frontal and temporal lobe cortical activation in Chinese first-episode medication-naïve and recurrent depression during a verbal fluency task,”
Front. Psychiatry, 14 1132666 https://doi.org/10.3389/fpsyt.2023.1132666
(2023).
Google Scholar
F. Scholkmann et al.,
“A review on continuous wave functional near-infrared spectroscopy and imaging instrumentation and methodology,”
NeuroImage, 85 6
–27 https://doi.org/10.1016/j.neuroimage.2013.05.004 NEIMEF 1053-8119
(2014).
Google Scholar
S. F. Husain et al.,
“Cortical haemodynamic response measured by functional near infrared spectroscopy during a verbal fluency task in patients with major depression and borderline personality disorder,”
EBioMedicine, 51 102586 https://doi.org/10.1016/j.ebiom.2019.11.047
(2020).
Google Scholar
F. Herold et al.,
“Applications of functional near-infrared spectroscopy (fNIRS) neuroimaging in exercise-cognition science: a systematic, methodology-focused review,”
J. Clin. Med., 7
(12), 466 https://doi.org/10.3390/jcm7120466
(2018).
Google Scholar
M. A. Yücel et al.,
“Best practices for fNIRS publications,”
Neurophotonics, 8
(1), 012101 https://doi.org/10.1117/1.NPh.8.1.012101
(2021).
Google Scholar
F. A. Fishburn et al.,
“Temporal derivative distribution repair (TDDR): a motion correction method for fNIRS,”
NeuroImage, 184 171
–179 https://doi.org/10.1016/j.neuroimage.2018.09.025 NEIMEF 1053-8119
(2019).
Google Scholar
A. Faress and T. Chau,
“Towards a multimodal brain-computer interface: combining fNIRS and fTCD measurements to enable higher classification accuracy,”
NeuroImage, 77 186
–194 https://doi.org/10.1016/j.neuroimage.2013.03.028 NEIMEF 1053-8119
(2013).
Google Scholar
G. Li et al.,
“Baseline wander removal for ECG signals based on improved EMD,”
in 15th IEEE Int. Conf. Signal Process. (ICSP),
484
–487
(2020). https://doi.org/10.1109/ICSP48669.2020.9320927 Google Scholar
J. Jenitta and A. Rajeswari,
“Denoising of ECG signal based on improved adaptive filter with EMD and EEMD,”
in IEEE Conf. Inf. Commun. Technol.,
957
–962
(2013). https://doi.org/10.1109/CICT.2013.6558234 Google Scholar
C. Wang, Z. Xiao and J. Wu,
“Functional connectivity-based classification of autism and control using SVM-RFECV on RS-fMRI data,”
Phys. Med.-Eur. J. Med. Phys., 65 99
–105 https://doi.org/10.1016/j.ejmp.2019.08.010
(2019).
Google Scholar
I. Guyon et al.,
“Gene selection for cancer classification using support vector machines,”
Mach. Learn., 46 389
–422 https://doi.org/10.1023/A:1012487302797 MALEEZ 0885-6125
(2002).
Google Scholar
I. T. Jolliffe and J. Cadima,
“Principal component analysis: a review and recent developments,”
Philos. Trans. R. Soc. A Math. Phys. Eng. Sci., 374
(2065), 20150202 https://doi.org/10.1098/rsta.2015.0202
(2016).
Google Scholar
S. Sperandei,
“Understanding logistic regression analysis,”
Biochem. Med., 24
(1), 12
–18 https://doi.org/10.11613/bm.2014.003
(2014).
Google Scholar
J. Cervantes et al.,
“A comprehensive survey on support vector machine classification: applications, challenges and trends,”
Neurocomputing, 408 189
–215 https://doi.org/10.1016/j.neucom.2019.10.118 NRCGEO 0925-2312
(2020).
Google Scholar
P. Probst, M. N. Wright and A. L. Boulesteix,
“Hyperparameters and tuning strategies for random forest,”
Wiley Interdiscip. Rev. Data Mining Knowl. Discov., 9
(3), e1301 https://doi.org/10.1002/widm.1301
(2019).
Google Scholar
F. Rossi and B. Conan-Guez,
“Functional multi-layer perceptron: a non-linear tool for functional data analysis,”
Neural Netw., 18
(1), 45
–60 https://doi.org/10.1016/j.neunet.2004.07.001 NNETEB 0893-6080
(2005).
Google Scholar
S. M. LaValle, M. S. Branicky and S. R. Lindemann,
“On the relationship between classical grid search and probabilistic roadmaps,”
Int. J. Rob. Res., 23
(7–8), 673
–692 https://doi.org/10.1177/0278364904045481 IJRREL 0278-3649
(2004).
Google Scholar
L. Yang and A. Shami,
“On hyperparameter optimization of machine learning algorithms: Theory and practice,”
Neurocomputing, 415 295
–316 https://doi.org/10.1016/j.neucom.2020.07.061
(2020).
Google Scholar
Y. Gu et al.,
“Empirical mode decomposition-based motion artefact correction method for functional near-infrared spectroscopy,”
J. Biomed. Opt., 21
(1), 015002 https://doi.org/10.1117/1.JBO.21.1.015002 JBOPFO 1083-3668
(2016).
Google Scholar
B. Molavi and G. A. Dumont,
“Wavelet-based motion artefact removal for functional near-infrared spectroscopy,”
Physiol. Meas., 33
(2), 259 https://doi.org/10.1088/0967-3334/33/2/259 PMEAE3 0967-3334
(2012).
Google Scholar
T. Chithiramohan et al.,
“Investigating the association between depression and cerebral haemodynamics: a systematic review and meta-analysis,”
J. Affect. Disord., 299 144
–158 https://doi.org/10.1016/j.jad.2021.11.037 JADID7
(2022).
Google Scholar
M. T. Abou-Saleh et al.,
“Single photon emission tomography with 99m Tc-HMPAO in Arab patients with depression,”
J. Affect. Disord., 55
(2–3), 115
–123 https://doi.org/10.1016/S0165-0327(98)00209-2 JADID7
(1999).
Google Scholar
S. Brigadoi et al.,
“Motion artefacts in functional near-infrared spectroscopy: a comparison of motion correction techniques applied to real cognitive data,”
NeuroImage, 85 181
–191 https://doi.org/10.1016/j.neuroimage.2013.04.082 NEIMEF 1053-8119
(2014).
Google Scholar
BiographyZhiyong Huang is a professor and doctoral supervisor at the School of Microelectronics and Communication Engineering, Chongqing University, China. He serves as the principal investigator for the “Active Health and Aging Technology Response” project under the Chinese National Support Program. He has also led several projects, including the National Key R&D Program, Key Projects of Technological Innovation and Application Development in Chongqing Municipality, Chongqing Natural Science Foundation, and Key Collaborative Projects in Sports Science Research in Chongqing Municipality. |
Data modeling
Near infrared spectroscopy
Signal to noise ratio
Machine learning
Spectroscopy
Denoising
Neurophotonics