RECOGNITION AND CLASSIFICATION OF DEPRESSION UNDER DEEP NEURAL NETWORK AND REHABILITATION EFFECT OF MUSIC THERAPY

1. Introduction

Depression is a common mental illness clinically. It mainly manifests as low mood, negative pessimism, slow thinking, easy self-blame, poor sleep quality, headache, loss of interest in everything, obsessive-compulsive behaviors, and a sense of worthlessness; in severe cases, suicidal thoughts and behaviors even occur.^1,2,3 Severe depression will also affect other social functions due to the loss of the will to get up. For example, students may be inattentive, and their grades will decline; workers may arrive late and leave early, change jobs frequently, and have conflicts with colleagues.^4,5 There are many reasons for depression such as excessive stress, emotional stimulation, abuse of drugs, hormone disorders in vivo, and genetic factors.^6,7,8 Depression has now become one of the major diseases that has a serious impact on human beings worldwide, and the suffering caused to patients and their families is far more serious than other diseases.⁹ Because people lack the correct understanding of depression, depressed patients are often reluctant to go to the hospital for prejudice, and their families and relatives cannot give the required understanding and emotional support. Relevant data show that only 2% of depressed patients have received treatment, and many patients will delay it until the deterioration and finally the serious consequence of suicide.¹⁰

Currently, the clinical treatment of depression mainly includes transmitter-modifying drugs, implantation surgery of vagus nerve pacemakers, and auxiliary psychological counseling treatment.³ However, all related drugs are maintenance drugs in nature which cannot cure depression completely. Once the drug is stopped, depression will recur with more serious condition and greater harm to patients.^11,12 Direct electrical stimulation of neurons by implanting a vagus nerve pacemaker into the hypothalamus can allow patients to return to normal life in a very long time, but the surgery has the disadvantages of high risk, high cost, and great limitations.^13,14,15 Music therapy utilizes the tone, connotation, nature, performance, and method of music. Through various forms of music experience, including appreciation, performance, and creation, it acts on subhealthy, diseased, and normal people to achieve the purpose of looking into the soul, stimulating memories, relieving emotions, enhancing self-cultivation, balancing psychological status, and finally treating diseases.⁸ Although music therapy is an emerging treatment method, it has been applied in many fields such as intellectual disability, depression, visual impairment, physical impairment, and hearing impairment.

With the rapid development of medical technology and imaging technology, an increasing amount of medical image data need to be stored and managed, and it is very important to find an appropriate method for medical image analysis.^16,17 In recent years, many studies have proven that the occurrence of depression is closely related to changes in cerebral function. Magnetic resonance imaging (MRI) is a common method for cerebral examination of patients, as it can display the cerebral cortex and internal tissue structure clearly, having the advantages of high resolution and multiple parameters. Deep neural networks are a technology in machine learning. It has multiple hidden layers in the input layer and output layer, which can simulate complex nonlinear relationships; thus, it has good application prospects in image processing.^18,19,20

In summary, the evaluation of the condition of patients with depression based on deep learning technology is a trend in clinical studies. However, there are a few studies on relevant recognition algorithms. On the grounds of a graph convolutional neural network (GCNN), a depression recognition algorithm of an optimized GCNN (O-GCNN) was proposed in this work. A total of 120 patients with different degrees of depression were included as the research subjects; 60 cases were divided into the test group, and the other 60 cases were in the control group depending on different treatment plans. Those in the control group were given Western medicine treatment alone, while those in the test group were given music therapy in combination with Western medicine treatment. The general data, the content of urinary free cortisol (UFC), bilateral hippocampal regional homogeneity (ReHo) value, fractional amplitude of low frequency fluctuation (fALFF) value, Self-rating Depression Scale (SDS) score and Hamilton depression scale (HAMD) score of patients were all compared between the two groups. Thereby, an in-depth analysis was made on the application of the O-GCNN in cerebral MRI image analysis of patients with depression and the effect of Western medicine combined with music therapy in the treatment of depression, which provided help for clinical diagnosis and treatment of patients with depression.

2. Materials and Methods

2.1. Research objects

A total of 120 patients with different degrees of depression who were treated in the outpatient department of the Hospital from 1 December 2019 to 1 December 2021, were selected as the research subjects. They were composed of 37 males and 83 females aged 28–58 years old. This research was approved by the ethics committee of the hospital. All the patients participated voluntarily and signed an informed consent form before the implementation of this project.

Inclusion criteria included the following requirements. (1) Patients could offer complete clinical data. (2) They were over 18 years old. (3) They stopped taking related antidepressant drugs one week before the research. (4) Their course of disease was longer than 14 days. (5) They had no contraindications to MRI scanning.

Exclusion criteria were formulated as follows. (1) Patients with immune dysfunction. (2) Patients who have complications of severe heart, liver, or kidney disease. (3) Their depression was caused by organic mental disorders. (4) Patients dropped out of the research due to personal reasons. (5) Patients who had hematopoietic disorders. (6) Those who were alcoholic and drug dependence. (7) They were pregnant women.

2.2. Intervention methods

According to the different treatment plans, the 120 patients were divided into the test group and the control group, with 60 cases in each. In the control group, only Western medicine treatment was given, while in the test group, they were treated with music therapy on the basis of Western medicine treatment.

For Western medicine treatment, the antidepressant fluoxetine was given to the patients. The dose could be increased according to the personal condition of patients, generally 15mg per day, and the maximum dose did not exceed 35mg.

Music therapy was conducted as follows. (1) Before treatment, the patients were gathered for propagandization and education of related knowledge. The principle, purpose, and process of music therapy were explained, and patients were taught for relaxation training. (2) During treatment, a quiet and comfortable environment was first prepared with soft tones, and then the patients were instructed to sit in a comfortable chair and relax the body and mind. (3) According to the specific situation of each patient, appropriate music was chosen. For excited and depressed cases, music in D was selected, such as Flowing Water, Mind between the Clouds and Water, and Holiday Beach. For nervous and anxious patients, music in G was chosen, such as A Moonlit Night on the Spring River, lullaby, and Laputa: Castle in the Sky. For irritable cases, music in E was played, such as The Moon Over a Fountain, The Spring of Northland, and The Wandering Songstress. For the somber and dispirited persons, music in C was played, including Rhapsody in Blue, Along the Songhua River, and Fern-leaf Hedge Bamboo under Moonlight. (4) When playing music to patients, the volume should be controlled at approximately 65 dB. After 5mins, the prepared narration was read, and adjustments were made according to the patients’ response to arouse the emotional resonance of patients. The treatment lasted for 40min per time, twice a day for 6 weeks.

2.3. MRI examination methods

Resting-state MRI scanning was performed on all patients, and the scanning sequence was echo-planar imaging. The scanning parameters included time of repetition (TR) 1500ms, time of echo (TE) 25ms, field of view $215\times 215\times 150$mm, matrix $156\times 156$, slice thickness 4.5mm, and slice spacing 0.5mm. The collected cranial MRI images were sent to the workstation for processing, and the ReHo value and fALFF value of the patient’s bilateral hippocampus were extracted.

2.4. Depression recognition algorithm under the GCNN

In this work, the GCNN shown in Fig. 1 was used for feature extraction and pattern discrimination in cerebral images. The convolutional layer and pooling layer performed feature extraction, and the fully connected layer performed pattern discrimination.

The filter was denoted as $L$, and the adjacency matrix was $B$. Then, the filter was defined as a polynomial of the adjacency matrix, from which the following equation could be determined :

where $B_1$ represents the neighbor whose distance is 1 from a specific vertex, $B_n$ represents the neighbor whose distance is $n$ from a specific vertex, and $E$ represents the vertex being processed. $l_0,l_1,\dots ,l_{n-1},l_n$ are the influence parameters of the neighbors on the vertices. Through the convolution of vertices and filters, the following equation can be obtained : where $W^{″}$ denotes a linear filter, and $w$ denotes a vertex. However, the above convolution was based on the spatial distance among vertices, and each area of the brain was functionally connected; thus, some adjustments are needed to be made to the meaning of the above equations. The adjacency matrix $B$ was taken as the adjacency tensor $B^{\ast }$, and the number of edge features was set as $m$. Then, a linear filter was defined to perform the convex combination of each adjacency matrix, which can be the following equation :

Thus, the dimensional information of each sample adjacency tensor can be expressed as the following equation :

where $u$ represents the number of nodes. If the sample size was $m$, then the set of adjacency matrix tensors was shown in the following equation :

Similarly, for multiple edge features of each node, a set of filter coefficients can be established as the following equation :

where $d$ represents the number of features of the node, and Eq. (3) can be updated as the following equation :

Thus, the edge feature filter corresponding to each node was expressed as the following equation :

where $L^{(d)}$ represents a slice of the filter, $l_m^{(d)}$ is the weighting coefficient, and $B_m$ represents the adjacency matrix corresponding to the $m$th edge feature.

The activation functions used in this work were the sigmoid function²¹ and Softmax function.²² The sigmoid function was used in convolutional layers, pooling layers, and fully connected layers to enhance features. It can be described as the following equation :

where $\beta (x)$ is the sigmoid function. The softmax function was in the output layer to normalize the output exponentially, from which the following equation was computed :

The value range of $\alpha (x_i)$ was [0,1]. In this work, the Adam algorithm²³ was adopted as the optimizer of the model. The first-order matrix estimation was set as $p_t$, and the second-order matrix estimation was $q_t$. Then, the two could be calculated as the following equations :

where ${\lambda }_1$ and ${\lambda }_2$ are parameters, and $G_t$ represents the partial derivative result of the loss function, which can be expressed as the following equation :

After further processing of the first-order matrix estimation and the second-order matrix estimation, the following equations were acquired :

Finally, the weight matrix was updated, and the following equation (16) was obtained :

where $V_{t+1}$ is the updated weight matrix, and $\gamma$ is a constant taking a value of 10${}^{-7}$. The optimized algorithm based on the deep GCNN was set as the O-GCNN.

2.5. Algorithm evaluation indicators

In this work, the accuracy (Acc), sensitivity (Sen), and precision (Pre) were taken as the evaluation indicators of the algorithm recognition. The CNN model,²⁴ the backpropagation (BP) algorithm,²⁵ and the forward propagation (FP) algorithm²⁶ were introduced for comparison with the O-GCNN algorithm.

TP, TN, FP, and FN represent the true positives, true negatives, false positives, and false negatives, respectively.

2.6. Observation indicators

General data of patients were recorded, including sex, mean age, mean body mass index (BMI), conditions (mild, moderate, or severe), and course of disease. The content of UFC was detected through chemiluminescence immunoassay and urine extraction techniques before, during, and after treatment. The HMAD score and SDS score were adopted for evaluating the psychological and emotional status of patients before, during, and after treatment. The Self-Rating Symptom Checklist 90 (SCL-90) was also used for the evaluation of the mental health status of patients before, during, and after treatment, including somatization, obsessive-compulsive symptoms, interpersonal sensitivity, depression, anxiety, hostility, phobic anxiety, paranoid ideation, and psychoticism. Patients’ curative status was recorded in the follow-up, which was denoted as effective, markedly effective, and ineffective.

2.7. Statistical methods

SPSS 19.0 was utilized for data processing. Measurement data are expressed as the mean$\pm$standard deviation ($\bar{x}\pm s$), and enumeration data are expressed as a percentage (%). One-way analysis of variance was adopted for pairwise comparisons between the test group and the control group. The difference was believed to be statistically significant at $P<0.05$.

3. Results

3.1. Comparison of classification results of different algorithms

In Fig. 2, the Acc for classification was 89.45% of the O-GCNN algorithm, the Sen was 90.31%, and the Pre was 78.33%. The Acc, Sen, and Pre of the CNN model for classification were 83.26%, 81.44%, and 67.34%, respectively. Those of the BP model were 81.05%, 76.58%, and 64.28%, respectively, while those of the FP model were 76.83%, 69.58%, and 68.03%, respectively. The Acc, Sen, and Pre for classification were remarkably higher for the O-GCNN algorithm than for the CNN model, the BP mode, and the FP model; the differences were statistically significant ($P<0.05$).

3.2. Comparison of general data of patients between the two groups

Figure 3 shows the general data of patients in the two groups. There was no statistical significance between the groups in terms of sex, mean age, mean BMI, condition (mild, moderate, or severe), or course of disease ($P<0.05$).

3.3. Cerebral MRI image manifestations in patients with depression

To compare the characteristics of cerebral MRI images of patients with depression (Figs. 4(b)–4(f)), the brain MRI image of normal humans (Fig. 4(a)) was introduced. Compared with normal people, the brains of patients with depression showed subtle abnormalities (red and blue dots in the images). The amygdala in the deep center of the brain was more active in patients with depression, the size of the hippocampus was smaller, and the activity of the insula and the dorsal prefrontal cortex was also decreased.

3.4. Comparison of SDS and HAMD scores between the two groups before and after treatment

As presented in Fig. 5, no significant difference was found before treatment in the SDS and HAMD scores between the test group and the control group ($P>0.05$). The SDS score and HAMD score of both groups were significantly lower after treatment than before treatment ($P<0.05$). The SDS and HAMD scores of the test group during and after treatment were remarkably lower than those of the control group, and the differences were statistically significant ($P<0.05$).

3.5. Comparison of SCL-90 scores between the two groups before and after treatment

In Fig. 6, there was no statistically significant difference before treatment in the SCL-90 scores of somatization, obsessive-compulsive symptoms, interpersonal sensitivity, depression, and anxiety between the test group and the control group ($P>0.05$). These scores after treatment were dramatically lower than those before treatment in both groups, with statistically significant differences ($P<0.05$). The scores of somatization, interpersonal sensitivity, depression, and anxiety were much lower in the test group than in the control group, suggesting statistically significant differences ($P<0.05$). Compared with the control group, the scores of obsessive-compulsive symptoms of patients in the test group during and after treatment were not significantly different ($P>0.05$).

In Fig. 7, no significant difference was found in the scores of hostility, phobic-anxiety, paranoid ideation, and psychoticism between the two groups before treatment ($P>0.05$). The scores of the four items during and after treatment were significantly lower than those before treatment ($P<0.05$). The scores of psychoticism in the test group were remarkably lower than those in the control group both during and after treatment, and the difference was statistically significant ($P<0.05$). There was no significant difference in the scores of hostility, phobic-anxiety, and paranoid ideation between the two groups during and after treatment ($P>0.05$).

3.6. Comparison of the 24-h UFC content of patients in the two groups

In Fig. 8, the 24-h UFC content was not significantly different between the test group and the control group before treatment ($P>0.05$). The 24-h UFC content of both groups after treatment was notably lower than that before treatment; the differences were also statistically significant ($P<0.05$). The 24-h UFC content of patients in the test group was significantly lower than that in the control group during and after treatment, suggesting statistically significant differences ($P<0.05$).

3.7. Comparison of MRI parameters between the two groups before and after treatment

As presented in Fig. 9, no significant difference was shown in the ReHo value of the right hippocampus of patients between the two groups before, during, and after treatment ($P>0.05$). The ReHo values of the left hippocampus in both groups after treatment were dramatically lower than those before treatment, which showed statistically significant differences ($P<0.05$). In the test group, the left hippocampal ReHo values of patients during and after treatment were much lower than those in the control group, as the difference was statistically significant ($P<0.05$).

As presented in Fig. 10, the fALFF values of the right hippocampus of patients before, during, and after treatment were not significantly different between the groups ($P>0.05$). The fALFF values of the left hippocampus of both groups after treatment were much lower than those before treatment, as the differences were calculated to be statistically significant ($P<0.05$). The left hippocampal fALFF values in the test group during and after treatment were significantly lower than those of the control group, and the differences were also statistically significant ($P<0.05$).

3.8. Comparison of curative statuses for patients in the two groups

In Fig. 11, in the test group, it was effective in 30 cases, markedly effective in 26 cases, and ineffective in 4 cases. In the control group, it was effective, markedly effective, and ineffective in 24, 25, and 11 cases, respectively. The effective rate of the test group (93.33%) was extraordinarily higher than that of the control group (81.67%), with a statistically significant difference ($P<0.05$).

4. Discussion

Depression, as a common mental illness worldwide, is characterized by continuous and long-term low mood and is the most important type of modern mental illness. Depression affects all aspects of people, including psychological, social, and physical functions; it has become one of the major factors endangering human health.²⁷ Music therapy is a method of using music tones and rhythms to treat patients with physical or mental diseases, as music repertoires are played to promote health and assist in eliminating psychosomatic disorders of patients.²⁸ Therefore, 120 patients with different degrees of depression were included as the research subjects, and the test group and the control group of 60 cases with different treatment plans were divided. Western medicine treatment and music therapy were carried out in the test group, while Western medicine treatment was only conducted in the control group. The patients in both groups received MRI scanning. To further improve the quality of MRI images, the depression recognition algorithm O-GCNN was put forward under deep GCNN. The results showed that the Acc, Sen, and Pre for classification of the O-GCNN algorithm were all significantly higher than those of the CNN model, BP model, and FP model, which were significantly different ($P<0.05$). Thus, this indicated that the O-GCNN algorithm was more effective for MRI image processing and auxiliary diagnosis of depression than traditional algorithms, suggesting application feasibility. In addition, the general data of patients were compared between the two groups. The sex, mean age, mean BMI, severity (mild, moderate, and severe), and course of disease were not significantly different between the two groups ($P<0.05$), which provided reliability for later trials.

For clinical effects, the SCL-90 scores were first compared. The scores of somatization, interpersonal sensitivity, depression, psychoticism, and anxiety in the test group were remarkably lower than those in the control group during and after treatment, which showed statistically significant differences ($P<0.05$). The SCL-90 is a practical, simple, and valuable scale for mental health status identification and group mental health censuses. It can be used for assessing mental health status within a specific period of time. The results demonstrated that music therapy combined with Western medicine treatment could be more effective in improving anxiety and negative psychological state in patients with depression.²⁹ The SDS and HAMD scores of the test group were also much lower than those in the control group during and after treatment; the differences were statistically significant as well ($P<0.05$). This is similar to the research results of Sanders et al.³⁰ It was indicated that music therapy combined with Western medicine treatment could improve the negative emotions of patients and relieve depression and anxiety psychologically. The left hippocampal ReHo and fALFF of patients in the test group were notably lower than those in the control group during and after treatment, with significant differences statistically ($P<0.05$). A large number of studies have proven that the cerebral MRI parameters ReHo and fALFF of patients with depression are lower than those of normal people. This result suggested that music therapy in combination with Western medicine could improve the bilateral hippocampal ReHo and fALFF of patients with depression and then improve the condition of patients with depression. UFC is the direct excretion of free cortisol in the blood, and it is one of the important indicators for the evaluation of depression.³¹ The 24-h UFC content of patients in the test group was dramatically lower than that in the control group during and after treatment in this research, with significant differences ($P<0.05$). In comparison with Western medicine treatment alone, music therapy combined with Western medicine could reduce the UFC content more effectively, which even tended to be normal, thereby improving depression. In this research, the physiological indicators of patients were not collected. The effect of music therapy on the expression of serum factors among patients with depression is still unclear. Hence, the duration of follow-up should be prolonged in follow-up research. In addition, all indicators of patients before and after treatment were recollected.

5. Conclusion

A total of 120 patients with depression of different degrees were included as the research subjects. They were divided into the test group of 60 cases (western medicine treatment+music therapy) and the control group of the other 60 cases (western medicine treatment alone). All patients underwent MRI scanning. To further improve the quality of MRI images, the improved depression recognition algorithm O-GCNN was constructed under deep GCNN. The effect of the O-GCNN on MRI image processing and auxiliary diagnosis of depression was proven to be better than that of traditional algorithms, so there was an application feasibility of this research. Music therapy combined with Western medicine treatment would improve the anxiety and negative psychological state more effectively in patients with depression, increase the values of ReHo and fALFF of the bilateral hippocampus, and reduce the content of UFC. Therefore, the condition of patients with depression could be improved. However, the patient samples included were all from the same hospital, and the generalizability of the results has not been verified. Without follow-up after treatment, there was also a lack of data on the long-term prognosis of patients. Therefore, in future research, a wider range of patient data should be reincluded for analyzing the effect of music therapy on the prognosis of patients with depression. In conclusion, the results of this work provide a reference for the formulation of drug-assisted music therapy for patients with depression.

Acknowledgment

Xueting Li and Canrui Chen contributed equally to this work. This work was supported by Innovation Outstanding Young Talents Program of Department of Education of Guangdong Province (Grant no. 2021WQNCX001).