1. Chinese university students with depressive symptoms exhibited hemodynamic changes in bilateral medial prefrontal cortices (MPFC), left dorsolateral prefrontal cortex, and left temporal lobe.

2. Among the five machine learning algorithms, the K-Nearest Neighbors (KNN) model demonstrated optimal performance in differentiating depressed and non-depressed university students.

3. The left MPFC contributed most to the KNN model’s classification accuracy, suggesting its crucial role in depression.

The prevalence of depression is increasing and constitutes a significant mental health issue on Chinese university campuses. Factors such as separation from parents, inadequate adaptation to new environments, academic stress, and career planning challenges contribute to this phenomenon, with approximately 15% to 35% of Chinese university students report experiencing depression [1]. Early identification is essential to mitigate the adverse effects of depressive symptoms on students’ academic and occupational performance [2]. Currently, the detection of depressive symptoms predominantly relies on self-reported questionnaires, including the Patient Health Questionnaire-9 (PHQ-9), the Hamilton Depression Rating Scale (HAMD), and Beck’s Depression Inventory (BDI). These tools may introduce subjective bias into the assessment results, and their flexibility and accuracy are relatively limited [3]. Consequently, there is an urgent need for objective, quantitative measures that can accurately identify depressive symptoms in Chinese university students and elucidate their underlying neural mechanisms.

Researchers are increasingly developing technological tools to address the challenges described. Functional near-infrared spectroscopy (fNIRS), a non-invasive brain imaging technique based on neurovascular coupling, is widely used to understand the neurobiology of depression [4]. Compared with electroencephalography and functional magnetic resonance imaging, fNIRS has many advantages, such as relatively low cost, safety, portability, high temporal resolution, and insensitivity to motion artifacts [5]. fNIRS can utilize the tight coupling between oxygenation levels associated with neural activity and localized cerebral blood flow to monitor hemodynamic changes [6]. Therefore, it has the potential to provide mechanistic insights into the neurobiological alterations associated with depression [4]. The verbal fluency task (VFT) is often used with fNIRS to evaluate an individual’s ability to convey information verbally within a defined timeframe or category, serving as an indicator of linguistic competence and cognitive flexibility [2, 7]. Hence, the fNIRS-VFT paradigm offers a promising approach to explore the neural mechanisms of depression [8]. Previous study has reported significant hemodynamic changes in depression utilizing the fNIRS-VFT paradigm, further supporting the link between fNIRS signals and depression [9]. Given the distinct linguistic characteristics and retrieval strategies inherent in English and Chinese [2, 10, 11], the findings derived from the fNIRS-English VFT paradigm are not directly transferable to the Chinese population. Therefore, we employed the fNIRS-Chinese VFT paradigm as an objective measure to investigate the neural activity of brain regions in Chinese university students.

Rapid and accurate methodologies are required for large-scale depression screening among the Chinese university student population. With the development of artificial intelligence, machine learning algorithms offer a powerful approach to automatically learn from existing data, establish functional relationships, identify latent patterns, and make predictions [12]. Previous studies have utilized fNIRS to develop machine learning models that discriminate depression by employing various feature combinations during the VFT [13, 14]. A recent study employed fNIRS and support vector machine (SVM) to classify mild and severe depression in a cohort of 140 subjects [14]. Yi et al. [13] used fNIRS signals from 25 depressed individuals and 30 healthy controls during a resting state to develop an SVM model for the automatic classification of depression. However, fewer studies have utilized fNIRS in the context of the Chinese VFT to construct machine learning models for identifying depressive symptoms in Chinese university students.

This study aimed to develop an fNIRS-based system for the identification of depressive symptoms among Chinese university students by employing five distinct machine learning algorithms. The primary objective was to identify specific hemodynamic alterations in Chinese university students exhibiting depressive symptoms. Furthermore, the study aimed to investigate the correlation between regional hemodynamic changes and the severity of depressive symptoms. Lastly, we aimed to evaluate the efficacy of machine learning models in differentiating between individuals with and without depressive symptoms based on fNIRS data.

This study developed an fNIRS-based system to identify depressive symptoms among Chinese university students by utilizing five machine learning algorithms. There were three main findings of our study. First, individuals with depression exhibit significant hemodynamic changes in the bilateral MPFC, left DLPFC, and TL compared with controls, suggesting that these regions are involved in cognitive processing during a VFT. Second, fNIRS signals in these brain regions were also significantly associated with depression, highlighting their critical role in the neural mechanisms underlying depression. Third, these distinct fNIRS alterations were utilized to develop classification models for identifying depressive symptoms among university students. Among the five machine learning methods, the KNN algorithm exhibited the highest performance, with a mean AUC of 66.51% and an accuracy of 65.63%. Notably, the left MPFC contributed the most to the KNN model’s classification accuracy, highlighting its crucial role in depression. These findings suggest that the integration of fNIRS with machine learning classifiers could serve as an objective and precise tool to complement traditional diagnostic methods.

In individuals with depression, significant hemodynamic changes were observed in the bilateral MPFC, left DLPFC, and left TL, as represented by channels 4, 16, 21, 26, 32, 43, 44, and 47, which were also correlated with PHQ-9 scores. These findings regarding hemodynamic alterations are consistent with previous research. Lim and Park [26] investigated the hemodynamic changes in individuals with subclinical depression, identifying significant differences in the prefrontal cortex. Sun et al. [27] observed decreased activation in the DLPFC among individuals with depressive symptoms. Similarly, Fan et al. [9] reported diminished activation in the prefrontal cortex, particularly within the DLPFC in depressed individuals. Yang et al. [28] also demonstrated reduced activity in the left TL and DLPFC among individuals with depression. A recent study involving 72 depressed university students identified a significant correlation between the left DLPFC and depression [29]. The MPFC, DLPFC, and TL have been implicated in cognitive, emotional, and behavioral regulation [30, 31, 32]. The VFT task activated the prefrontal cortex, including the MPFC and DLPFC, which are primarily responsible for emotional regulation and cognitive control, language comprehension, and memory retrieval [33, 34]. Depressed individuals may be unable to effectively mobilize the full functionality of the MPFC, DLPFC, and TL, leading to correspondingly flattened brain responses and low variability during the VFT task. Chinese university students experiencing depression demonstrated significant alterations in the MPFC, DLPFC, and TL, potentially linked to impaired emotional regulation and cognitive control [30, 31, 32]. Consequently, we speculate that dysfunction in the MPFC, DLPFC, and TL contributed to the observed symptoms in individuals with depression, such as flat affect, low mood, delayed decision-making, weakened executive functions, and challenges in language generation and semantic retrieval. Frontotemporal hemodynamic changes detected by fNIRS during VFT may serve as an objective measure for assessing depression among Chinese university students. Previous studies have reported similar findings. In one study involving 72 depressed university students and 67 healthy controls using fNIRS during a VFT, the depression group exhibited lower oxyhemoglobin levels in the right DLPFC and Broca’s area, along with reduced frontotemporal connectivity. Importantly, these alterations were correlated with the severity of depression [35]. Park et al. [36] identified decreased activation in the right frontopolar cortex and MPFC in the depression group. In addition, a recent study conducted by Kang et al. [37], involving 204 older adults with depression revealed diminished prefrontal hemodynamic responses during the Stroop test. However, no significant differences were observed in performance on the digit span backward task or the VFT. Collectively, these findings suggested that frontal hemodynamic changes detected by fNIRS during a VFT have the potential to be an objective measure for assessing depressive symptoms among Chinese university students.

We employed five distinct machine learning algorithms to capitalize on their unique strengths and facilitate a comparative analysis using a consistent dataset. Each algorithm exhibits specific characteristics that may yield different performance. This diverse selection of classifiers ensures a balanced evaluation encompassing linear, nonlinear, probabilistic, and instance-based learning methodologies, thereby enabling a comprehensive assessment of classification performance and model generalizability in the context of depression-related data analysis. The KNN model demonstrated the highest classification accuracy. Owing to its low computational demands, straightforward implementation, and robust classification performance, KNN is a prevalent and effective classification technique in the fields of data mining and statistics [38]. However, the classification accuracy achieved by the KNN model did not attain an ideal level. Several factors may have contributed to this outcome. First, the feature extraction and selection processes from the fNIRS data may not have fully captured all critical neural activity patterns associated with depression. Despite utilizing a grid search strategy to optimize model parameters, parameter selection remains inadequate for achieving optimal results. Second, the hemodynamic changes reflected in the fNIRS data may have been influenced by biological variables, such as individual differences, environmental factors, or experimental conditions, all of which can impact the model’s accuracy. Third, this study primarily concentrated on binary classification (depression versus control) without stratifying participants according to depression severity scores. Indeed, we observed fewer differences between university students with depressive symptoms and controls than between patients with clinical depression and controls. This may be due to the mild depressive symptoms exhibited by participants, as defined by a PHQ-9 score of $\ge$5 for Chinese university students. Future research will incorporate a grading system for depression severity, potentially utilizing the PHQ-9 or other standardized clinical assessments, to improve classification accuracy and explore whether varying severity levels manifest distinct neurobiological or behavioral characteristics. Moreover, deep learning models, particularly Convolutional Neural Networks (CNNs), are capable of effectively capturing highly non-linear relationships in data through their multi-layer neural structures [39]. These models are adept at identifying and extracting significant features from data automatically [40], thereby reducing the dependency on manual feature engineering when processing complex data types such as images, audio, and time-series data [41]. To improve the predictive performance of models, we will integrate CNNs with attention mechanisms to more effectively capture spatial and temporal patterns in fNIRS signals.

The left MPFC contributed the most to the optimal KNN model’s classification accuracy, highlighting the crucial role of the left MPFC regions in the neural mechanisms of depression [42]. The MPFC is involved not only in emotional regulation but also in cognitive control processes including decision-making and impulse control [30]. Its functional role is closely linked to the brain’s functional lateralization, with the left hemisphere typically associated with positive emotion processing. Whereas the right hemisphere is more engaged in the regulation of negative emotions [43, 44]. Depressed individuals exhibit an exaggerated response to negative emotional stimuli and weakened cognitive control functions, which may be attributed to abnormal activity in the left MPFC [45]. Given the critical role of the left MPFC’s pivotal role in regulating positive emotions and inhibiting negative ones [46], we speculate that dysfunction in the left MPFC may contribute to the persistence of negative emotions, deficits in positive emotional experiences, and impaired cognitive control associated with depressive symptoms. When faced with complex tasks, as evidenced by fNIRS data obtained from the VFT, individuals with depressive symptoms may manifest flat affect, difficulties in language generation, and semantic retrieval. Consequently, alterations in fNIRS signals in the left MPFC could serve as a potential biomarker and predictor of depressive symptoms among Chinese university students.

However, this study has several limitations. First, the design of the VFT task requires rapid syllable changes every 20 seconds during the task period. This methodology was employed to reduce variability in silent intervals, thereby potentially diminishing activation effects associated with vocalization. Future research should consider incorporating more comprehensive cognitive paradigms to further elucidate the neural mechanisms underlying depressive symptoms. Additionally, fNIRS signals exhibit considerable inter-subject variability, attributable to differences in head size, head shape, and the spatial distribution of brain functional regions. Subsequent research will be conducted at the individual subject level, taking into account differences in head size and array placement. Second, multiple comparison correction methods were not utilized in the statistical analyses due to the small sample size. Feature selection aims to enhance the performance of prediction models and provide a deeper understanding of the data-generating process [47]. Previous research has shown that features can be selected as model inputs without correcting for multiple comparisons [48]. For instance, Fourdain et al. [49] presented the fNIRS results without applying Bonferroni or False Discovery Rate (FDR) correction. When the number of features exceeds the number of participants, this high feature-to-sample ratio may increase the risk of overfitting and reduce the generalizability of the classification results [50]. Future studies with larger sample sizes will facilitate the application of more rigorous multiple comparison correction methods. Third, the optimal KNN model did not achieve the desired accuracy. The PHQ-9 alone is not sufficiently rigorous for diagnosing depression. To fully realize its potential applicability, the classification model should be capable of accurately distinguishing the onset of depression and differentiating between mild, moderate, and severe forms of depression. Future research will explore the integration of deep learning models, multi-modal features, or the HAMD to potentially enhance classification accuracy. Lastly, our analysis did not include an external dataset for model validation and the absence of multicenter data from diverse regions may limit the representativeness of our findings. Without external validation, the model’s generalizability is potentially compromised. Moreover, training the model exclusively on participants from a single university introduces population homogeneity, further reducing representativeness. To establish the generalizability of the classification model among university students, future studies should attempt to validate the model using independent datasets.

Our fNIRS results demonstrate that Chinese university students with depressive symptoms exhibited significant alterations in frontotemporal activity. Among the five machine learning algorithms evaluated, the KNN model exhibited superior classification performance. The left MPFC contributed most to the KNN model’s classification accuracy, highlighting its central role in the emotional processing and cognitive flexibility impairments associated with depressive symptoms.

The integration of fNIRS with these machine learning classifiers holds potential for advancing the development of an objective, automated, and user-friendly assessment tool and intelligent system for detecting specific cerebral hemodynamic changes, thereby enhancing the early identification of depressive symptoms in Chinese university students. Future research should focus on validating these findings and refining classification models to improve their applicability on a large scale.

The fNIRS data in this study are available upon request. Further information and requests should be directed to and will be fulfilled by Dr. Yange Wei.

YW, YC, YM and XX were involved in the study’s conceptualization, design, and manuscript writing. PL, JY and LY conducted data collection. Statistical analysis was conducted by NW, HZ, and RL The initial manuscript draft was prepared by YW. GJ, WZ and ZZ conducted the analyses. All authors contributed to editorial changes in the manuscript. All authors read and approved the final manuscript. All authors have participated sufficiently in the work and agreed to be accountable for all aspects of the work.

Eligibility screening was conducted for individuals who provided verbal consent, followed by the collection of written informed consent. The study was approved by the institutional review boards of the Second Affiliated Hospital of Xinxiang Medical University (XYEFYLL-2023-35-4), in accordance with the Declaration of Helsinki’s Ethical Principles of Medical Research Involving Human Subjects.

The authors thank Guang Yang, Shisen Qin, and Jilong Chen of the Mental Health and Artificial Intelligence Research Center for their assistance with this study, and the Brain and Intelligence Group of the National Clinical Research Center for Mental Disorders. We are grateful to all subjects for their participation.

This research was supported by the Postgraduate Education Reform Project of Henan Province (No. 2023SJGLX063Y to YW), Medical Science and Technique Foundation of Henan Province (No. SBGJ202403043 to YW), and Xinxiang Medical University Graduate Innovation Research Project (grant number YJSCX202411Z to YC).

The authors declare no conflict of interest. Wei Zheng is serving as one of the Editors-in-Chief and a Guest Editor of this journal. We declare that Wei Zheng had no involvement in the peer review of this article and has no access to information regarding its peer review. Full responsibility for the editorial process for this article was delegated to Francesco Bartoli.