Predicting the Treatment Response of Patients With Major Depressive Disorder to Selective Serotonin Reuptake Inhibitors Using Machine Learning Techniques and EEG Functional Connectivity Features
Abstract
Background: Escitalopram and sertraline are first-line medications for treating depression. They belong to selective serotonin reuptake inhibitors (SSRIs) and are widely used due to their effectiveness and fewer side effects. However, despite the significant efficacy of escitalopram and sertraline, there is a large variation among individuals. Therefore, predicting symptom improvement based on the baseline period is crucial.
Methods: In this study, we conducted functional connectivity (FC) analysis of electroencephalogram (EEG) data during resting-state with eyes closed, resting-state with eyes open, watching neutral videos, negative videos, and comedy videos for 30 untreated depression patients over 2 weeks. Each modality yielded 18 EEG FC features. Based on the treatment response at 8 weeks, patients were divided into treatment-effective and treatment-ineffective groups. The dataset was randomly split into a 75% training set and a 25% independent test set. Feature selection was performed on these FC features in the training set, and the selected features were used to classify the effective and ineffective groups using the support vector machine (SVM) machine learning algorithm. Fivefold cross-validation was conducted on the training set to obtain validation results, followed by testing on the test set. The Spearman’s correlation method was used to analyze the correlation between each EEG feature value and the reduction rate of the Hamilton Depression Rating Scale for Depression (HAMD-17) scores from baseline to 8 weeks, with Bonferroni correction applied.
Results: The study found that out of all modalities, 33 features achieved classification accuracies of over 95% on the validation set, and two features achieved classification accuracies of over 85% on the independent test set. A total of 58 feature values were found to be correlated with the reduction rate of HAMD-17 scores from baseline to 8 weeks.
Conclusions: The findings from this research suggest that EEG FC features at baseline can be used to differentiate between effective and ineffective groups with high accuracy using machine learning models. Multiple feature values and HAMD-17 scores were found to be correlated with the reduction rate of HAMD-17 scores from baseline to 8 weeks, and these correlated feature values can be used to predict treatment efficacy.
Trial Registration: ClinicalTrials.gov identifier: NCT05775809
1. Introduction
Major depressive disorder (MDD) is a chronic mental illness characterized by remission and relapse. Despite the variety of available treatments, up to 40%–50% of patients do not respond [1]. Additionally, the success of antidepressant treatment can only be evaluated after 6 weeks. Therefore, it often takes months or even years for depression patients to achieve successful treatment. This can heighten their feelings of hopelessness and despair, which are already prominent in their condition [2]. Sertraline and escitalopram are frequently regarded as first-line medications for treating depression [3]. Previous analyses using biomarkers from genetic and serum sources have not captured the significant differences in treatment responses, nor do they directly reflect brain signals [4]. In contrast, noninvasive and secure electroencephalography is more capable of capturing directly measured, instantaneous brain responses. Using EEG biomarkers, we might elucidate the biological roots of clinical presentations and personalize medical treatment [5].
It has been observed that serotonergic stimulation can reduce intrinsic functional connectivity (FC) in brain regions associated with mood regulation [6], such as the ventral anterior cingulate cortex and limbic structures like the amygdala [7]. However, the comprehensive effects of serotonergic and antidepressant treatments on human brain FC have not been extensively explored. To address this gap, I not only analyzed resting-state EEG but also examined EEG responses while participants watched videos eliciting different emotional states. This approach allows for a more nuanced understanding of how serotonergic and antidepressant treatments influence brain connectivity under varying emotional conditions, providing deeper insights into their effects on mood regulation mechanisms.
The latest systematic review proposes the application of FC analysis to uncover alterations in patterns during various activities experienced by individuals with depression [8]. FC analysis denotes the linkages among detected brain areas that are interconnected [9]. Research using complex network analysis originates from the quantitative analysis of network structures called graph theory [10]. A graph provides a conceptual model of a network, composed of vertices, and the pathways between them (edges) [10]. Datasets on brain connectivity encompass networks where brain areas are linked through physical structures or operational relationships [11].
A study utilizing resting-state EEG data discovered heightened FC between the prefrontal cortex and the posterior cingulate cortex in patients with MDD who are in remission, indicating that EEG FC could serve as a neurological indicator of depression [12]. A clinical research discovered that FC was negatively correlated with depression severity [13]. A previous MRI study showed that FC changed after taking antidepressants, with increased connectivity reported between frontal and limbic brain regions [14]. Peng et al. [15] accurately distinguished individuals who responded to medication from those who did not with a 96.67% accuracy by using baseline resting-state spatial network topology as distinguishing features. Recent research indicates that electroencephalographic connectivity can differentiate various subtypes of MDD [16] and forecast individual responses to antidepressant treatments [17]. In conclusion, previous studies have demonstrated that EEG FC offers valuable information for precision medicine in MDD.
Predictions for individual patients can be made using multivariable machine learning analysis, which directly evaluates generalization to new patients through cross-validation. In other words, machine learning methods can predict new data in the test set based on pattern recognition in the training data [8]. The fact that machine learning can be used for binary response prediction has led to the search for a reliable tool that could inform clinical practice [9]. Most previous studies have used a single modality of brain network features to predict drug efficacy. Therefore, this study analyzes EEG FC features under various conditions: the eyes-open and eyes-closed resting states, as well as while subjects are watching the neutral, negative, and comedy videos. It compares the different types of FC features under different modalities based on machine learning models to distinguish between depression patients who respond to sertraline or escitalopram treatment and those who do not. Treatment efficacy is defined as a reduction of no less than 50% in the HAMD-17 total score from baseline after 8 weeks of adequate sertraline or escitalopram use. Additionally, the study further explores the correlation between these FC features and the percentage decrease of the initial HAMD-17 total score after 8 weeks. In summary, the innovation of this study lies in its integration of multivariate machine learning, multimodal brain network feature analysis, and the exploration of the relationship between these features and responses to antidepressant medications, providing new perspectives and methods for precision medicine in depression.
2. Methods and Materials
2.1. Assessment Instruments
The Hamilton Depression Rating Scale for Depression (HAMD-17) is employed to evaluate the intensity of depressive symptoms; the Hamilton Anxiety Scale (HAMA) is employed to evaluate the intensity of anxiety symptoms.
2.2. Pharmacotherapy
This study included patients who had not taken any medication for more than 2 weeks. The starting dose on the first day was 5 mg of escitalopram or 50 mg of sertraline per day, with half a tablet added every 3 days until reaching the target dose. One week later, all patients regularly took sufficient doses of sertraline or escitalopram for the following 7 weeks. Thirty patients were followed up after 8 weeks of medication.
2.3. Experimental Equipment
This study collected EEG data using a 64-channel electrode cap from NeuroScan, reducing the resistance to below 5 kΩ, with a sampling rate of 1000 Hz. The subject machine resolution was 1024 × 768 pixels.
2.4. Experimental Task
2.4.1. Closed-Eye Resting State
Collected 5 min of EEG data from subjects with their eyes closed but remaining awake.
2.4.2. Open-Eye Resting State
Recorded 5 min of EEG data from subjects staring at a cross in the middle of the display.
2.4.3. Video Stimuli
Negative videos were from the Chinese Emotional Video System: a 257-s clip from “Fist of Fury” and a 117-s clip from “Conman In Tokyo.” Neutral and positive videos were from the Emotional EEG Dataset of Shanghai Jiao Tong University, with 240-s and 226-s videos clips from the documentary “China’s World Cultural Heritage”; a 241-s clip from “Lost in Thailand” and a 266-s clip from “Flirting Scholar.” The videos resolution was 1024 × 576, at 25 frames per second.
2.4.4. Positive Affect and Negative Affect Scale
After each video, subjects completed the Positive Affect and Negative Affect Scale based on their feelings from watching the video.
2.5. EEG Data Preprocessing
First, some electrodes (“F11,” “F12,” “FT11,” “FT12,” “CB1,” “CB2,” “VEOG,” “HEOG,” “TRIGGER”) and nonscalp electrodes M1 and M2 were removed, leaving 56 electrodes. Then, band-pass filtering (1–50 Hz) and a 48–52 Hz Parks–McClellan notch filter were applied. Noisy segments were manually removed, and eye drift, blinking, and other independent component analysis components were eliminated. The data were segmented into 2-s intervals, removing segments exceeding ±70 μV. Finally, the data were re-referenced to the average.
2.6. EEG Feature Extraction
Calculate the features of various functional connections between every two electrodes among the 56 electrodes, with each type of feature being 1540 dimensions. Spatial Laplacian transformation was performed to obtain the current source density. To study the differences in FC across different frequency bands, EEG data were also band-pass filtered to extract the following frequency ranges: all bandwidth (1–45 Hz), delta waves (1–4 Hz), theta waves (4–8 Hz), alpha waves (8–13 Hz), beta waves (13–30 Hz), and gamma waves (30–45 Hz). The following FC indices were calculated to obtain the connectivity matrix.
2.6.1. Coherence (COH)
COH measures the degree of linear relationship between two EEG channels at a specific frequency. One indicates maximum linear correlation, and 0 indicates no linear correlation.
2.6.2. Weighted Phase Lag Index (wPLI)
The wPLI calculates the probability of phase synchronization. wPLI values range from [0, 1], with 1 indicating maximum phase synchronization, and 0 indicating minimum phase synchronization.
2.6.3. Phase Locking Value (PLV)
PLV assesses FC between electrode channel signals through phase synchronization. PLV values range from [0, 1], with 1 indicating maximum phase synchronization, and 0 indicating minimum phase synchronization.
2.7. Feature Selection and Machine Learning Modeling
Recursive feature elimination is a powerful technique for feature selection that iteratively removes the least important features based on the model’s weights. Using support vector machine (SVM) as the base model ensures that the feature selection process is robust and effective, given SVM’s strength in handling high-dimensional data and its ability to find a hyperplane that best separates the classes. A recursive feature elimination method based on SVM was used to evaluate the importance of features, and the best feature subset was determined based on SVM. In total, 75% of the data was randomly selected as the training set, and 25% of the data as the independent test set. Feature selection was performed only on the training set, and it was conducted on each of the 1540-dimensional features. Models were built using the selected features based on the SVM algorithm to distinguish between patients with effective and ineffective treatment, with fivefold cross-validation on the training set and testing on the independent test set. Effective treatment is defined as a reduction in HAMD-17 score by more than 50% following an 8-week therapeutic regimen in contrast with baseline. By using GridSearchCV, the optimal hyperparameters for the SVM classifier were determined in a robust and systematic manner. This approach ensures that the selected hyperparameters generalize well to new, unseen data, thereby enhancing the model’s predictive performance.
2.8. Spearman Correlation
The Spearman’s rank correlation coefficient assesses the relationship between two variables with a nonparametric statistical method, especially their monotonic relationship. The correlation between selected feature subsets in each feature type and the proportion of reduction in HAMD-17 score following an 8-week treatment period in contrast with baseline was analyzed, and the Bonferroni correction method was applied for adjustment. Bonferroni correction was applied within each paradigm and frequency band.
3. Results
3.1. Demographic and Clinical Information
This study included 30 patients with depression who had not taken any medication for at least 2 weeks and were visiting the psychiatric outpatient department of Sir Run Run Shaw Hospital affiliated with Zhejiang University. There were no statistically significant differences in age, gender, and education level between patients who were responsive and nonresponsive to the 8-week medication treatment. All depression patients had a baseline Hamilton Depression Rating Scale (HAMD-17) score of 17 or higher. All patients are over 18 years old, have been interviewed using the Chinese version of the Mini International Neuropsychiatric Interview, meet the DSM-5 diagnostic criteria for MDD, plan to use monotherapy with a single type of antidepressant medication, have an education level above elementary school, and can understand the study content. All patients currently or in the past have not been diagnosed with any other major mental disorders that meet the DSM-5 criteria in addition to MDD, do not have physical illnesses, and have not received physical therapy within the 3 months prior to screening. Demographic characteristics and clinical information are detailed in Table 1. All subjects voluntarily participated in this study, cooperated with EEG examinations, and signed informed consent forms. The Ethics Committee of Sir Run Run Shaw Hospital affiliated with Zhejiang University approved this study. Approval no: Sir Run Run Shaw Hospital Ethics Review 2022 Research No. 0146.
| Demographic and medical details | Effective group | Ineffective group | p-Value | U-statistic/chi-square statistic |
|---|---|---|---|---|
| Age (years) | 24.437 (6.387) | 28.857 (8.411) | 0.095 | 71.5 |
| Gender (male/female) | 6/10 | 4/10 | 0.897 | 0.016 |
| Education (grade) | 4.812 (0.403) | 4.500 (0.759) | 0.247 | 134.0 |
| HAMD (baseline) | 22.625 (3.593) | 23.642 (2.951) | 0.346 | 89.0 |
| HAMD (baseline 8 weeks) | 17.000 (3.829) | 7.642 (2.373) | <0.05 | 223.5 |
| HAMD ([baseline 8 weeks]/baseline) | 0.752 (0.124) | 0.317 (0.074) | <0.05 | 224.0 |
| HAMA (baseline) | 20.437 (2.965) | 22.214 (3.826) | 0.188 | 80.0 |
| PANAS (neutral videos) | 1.750 (2.938) | 0.357 (2.741) | 0.196 | 143.5 |
| PANAS (negative videos) | 0.656 (2.83) | −0.535 (1.780) | 0.268 | 139.0 |
| PANAS (positive videos) | 2.312 (3.187) | 2.392 (2.435) | 0.786 | 119.0 |
- Note: Continuous variables are presented as mean (standard deviation). For education, four represents a high school diploma, and five represents a college degree. Drug effectiveness refers to a reduction in the baseline ratio of HAMD-17 by more than 50% after 8 weeks of adequate treatment with sertraline or escitalopram.
- Abbreviations: HAMA, Hamilton Anxiety Rating Scale; HAMD, Hamilton Depression Rating Scale; PANAS, Positive and Negative Affect Schedule.
3.2. The Results of Fivefold Cross-Validation on the Training Set
Table 2 presents the fivefold cross-validation classification results of the SVM classifier on the validation set, using three types of EEG FC metrics under different types of stimuli and frequency bands. Among them, the average accuracy of 33 types of feature types is greater than 0.950. You can see the complete Table 2 in Table S1 of the Supporting Information.
| Features | Mean accuracy | Mean precision | Mean recall | Mean F1 score |
|---|---|---|---|---|
| Negative videos_Alpha band_wPLI | 1 | 1 | 1 | 1 |
| Negative videos_Beta band_wPLI | 1 | 1 | 1 | 1 |
| Negative videos_Delta band_COH | 1 | 1 | 1 | 1 |
| Negative videos_Delta band_PLV | 1 | 1 | 1 | 1 |
| Negative videos_Delta band_wPLI | 1 | 1 | 1 | 1 |
| Negative videos_Gamma band_PLV | 1 | 1 | 1 | 1 |
| Negative videos_Gamma band_wPLI | 1 | 1 | 1 | 1 |
| Negative videos_Theta band_wPLI | 1 | 1 | 1 | 1 |
| Neutral videos_Alpha band_wPLI | 1 | 1 | 1 | 1 |
| Neutral videos_Delta band_wPLI | 1 | 1 | 1 | 1 |
| Neutral videos_Gamma band_wPLI | 1 | 1 | 1 | 1 |
| Positive videos_Alpha band_COH | 1 | 1 | 1 | 1 |
| Positive videos_Alpha band_PLV | 0.96 | 0.967 | 0.967 | 0.96 |
| Positive videos_Alpha band_wPLI | 1 | 1 | 1 | 1 |
| Positive videos_Beta band_COH | 1 | 1 | 1 | 1 |
| Positive videos_Beta band_PLV | 1 | 1 | 1 | 1 |
| Positive videos_Beta band_wPLI | 1 | 1 | 1 | 1 |
| Positive videos_Gamma band_COH | 1 | 1 | 1 | 1 |
| Positive videos_Gamma band_wPLI | 1 | 1 | 1 | 1 |
| Positive videos_Theta band_PLV | 1 | 1 | 1 | 1 |
| Positive videos_Theta band_wPLI | 1 | 1 | 1 | 1 |
| Resting state (eyes-closed)_All band_COH | 1 | 1 | 1 | 1 |
| Resting state (eyes-closed)_Beta band_PLV | 1 | 1 | 1 | 1 |
| Resting state (eyes-closed)_Delta band_PLV | 1 | 1 | 1 | 1 |
| Resting state (eyes-closed)_Gamma band_wPLI | 1 | 1 | 1 | 1 |
| Resting state (eyes-closed)_Theta band_PLV | 1 | 1 | 1 | 1 |
| Resting state (eyes-closed)_Theta band_wPLI | 1 | 1 | 1 | 1 |
| Resting state (eyes-open)_Alpha band_COH | 0.96 | 0.967 | 0.967 | 0.96 |
| Resting state (eyes-open)_Beta band_COH | 1 | 1 | 1 | 1 |
| Resting state (eyes-open)_Beta band_wPLI | 1 | 1 | 1 | 1 |
| Resting state (eyes-open)_Delta band_wPLI | 1 | 1 | 1 | 1 |
| Resting state (eyes-open)_Theta band_PLV | 1 | 1 | 1 | 1 |
| Resting state (eyes-open)_Theta band_wPLI | 1 | 1 | 1 | 1 |
- Note: The first part represents paradigms: resting state (eyes closed) refers to the eyes-closed resting state paradigm. Resting state (eyes open) refers to the eyes-open resting state paradigm. Neutral videos represent the neutral videos stimulus paradigm. Angry videos represent the negative videos stimulus paradigm. Happy videos represent the positive videos stimulus paradigm. The second part represents frequency band ranges: all band represents 1–45 Hz. Delta band represents 1–4 Hz. Theta band represents 4–8 Hz. Alpha band represents 8–13 Hz. Beta band represents 13–30 Hz. Gamma band represents 30–45 Hz. The third part represents FC metrics: COH represents coherence. wPLI represents weighted phase lag index. PLV represents phase locking value. Explanation of the features: Resting state (eyes closed)_All band_COH represents the coherence features of EEG FC obtained using the eyes-closed resting state paradigm within the 1–45 Hz frequency band range. The others can be inferred in the same manner.
- Abbreviations: COH, coherence; PLV, phase locking value; SVM, support vector machine; wPLI, weighted phase lag index.
3.3. The Results of the Test Set
Table 3 shows the classification results of the SVM classifier on the test set using three types of EEG FC metrics under different types of stimuli and frequency bands. The accuracy of two types of feature types exceeded 0.850. The accuracy of 14 types of feature types was not less than 0.750. Ten types of FC features derived from electroencephalographic data induced by video stimuli achieved an accuracy of no less than 75% on an independent test set, and four types of FC features derived from resting-state EEG data achieved an accuracy of no less than 75% on an independent test set. You can see the complete Table 3 in Table S2 of the Supporting Information.
| Features | Mean accuracy | Mean precision | Mean recall | Mean F1 score |
|---|---|---|---|---|
| Resting state (eyes-closed)_All band_PLV | 1 | 1 | 1 | 1 |
| Neutral videos_Alpha band_wPLI | 0.875 | 0.917 | 0.833 | 0.855 |
| Resting state (eyes-closed)_Theta band_COH | 0.75 | 0.8 | 0.8 | 0.75 |
| Resting state (eyes-open)_Beta band_wPLI | 0.75 | 0.8 | 0.8 | 0.75 |
| Resting state (eyes-open)_Gamma band_PLV | 0.75 | 0.733 | 0.733 | 0.733 |
| Neutral videos_All band_COH | 0.75 | 0.733 | 0.733 | 0.733 |
| Neutral videos_Gamma band_wPLI | 0.75 | 0.733 | 0.733 | 0.733 |
| Positive videos_All band_COH | 0.75 | 0.857 | 0.667 | 0.667 |
| Positive videos_All band_PLV | 0.75 | 0.733 | 0.733 | 0.733 |
| Positive videos_Beta band_COH | 0.75 | 0.857 | 0.667 | 0.667 |
| Positive videos_Delta band_PLV | 0.75 | 0.8 | 0.8 | 0.75 |
| Positive videos_Gamma band_COH | 0.75 | 0.857 | 0.667 | 0.667 |
| Positive videos_Theta band_COH | 0.75 | 0.8 | 0.8 | 0.75 |
| Positive videos_Theta band_wPLI | 0.75 | 0.8 | 0.8 | 0.75 |
- Note: The first part represents paradigms: resting state (eyes closed) refers to the eyes-closed resting state paradigm. Resting state (eyes open) refers to the eyes-open resting state paradigm. Neutral videos represent the neutral videos stimulus paradigm. Angry videos represent the negative videos stimulus paradigm. Happy videos represent the positive videos stimulus paradigm. The second part represents frequency band ranges: all band represents 1–45 Hz. Delta band represents 1–4 Hz. Theta band represents 4–8 Hz. Alpha band represents 8–13 Hz. Beta band represents 13–30 Hz. Gamma band represents 30–45 Hz. The third part represents functional connectivity metrics: COH represents coherence. wPLI represents weighted phase lag index. PLV represents phase locking value. Explanation of the features: Resting state (eyes closed)_All band_COH represents the coherence features of EEG functional connectivity obtained using the eyes-closed resting state paradigm within the 1–45 Hz frequency band range. The others can be inferred in the same manner.
- Abbreviations: COH, coherence; PLV, phase locking value; SVM, support vector machine; wPLI, weighted phase lag index.
3.4. Spearman Correlation
According to Table 4, the characteristics of EEG FC between eight electrode pairs during closed-eye resting state and the ratio of reduction in HAMD-17 scores following an 8-week therapeutic regimen in contrast with baseline are related. The characteristics of EEG FC between 13 electrode pairs during open-eye resting state and the ratio of reduction in HAMD-17 scores following an 8-week therapeutic regimen in contrast with baseline are related. The characteristics of EEG FC between 12 electrode pairs during neutral videos watching and the ratio of reduction in HAMD-17 scores following an 8-week therapeutic regimen in contrast with baseline are related. The characteristics of EEG FC between 13 electrode pairs during negative videos watching and the ratio of reduction in HAMD-17 scores following an 8-week therapeutic regimen in contrast with baseline are related. The characteristics of EEG FC between 12 electrode pairs during positive videos watching and the ratio of reduction in HAMD-17 scores following an 8-week therapeutic regimen in contrast with baseline are related.
| Features | Spearman’s correlation coefficient | Bonferroni adjusted p-value |
|---|---|---|
| Resting state (eyes-closed)_All band_PLV_“FPZ”_“AF4” | −0.692 | 0.002 |
| Resting state (eyes-closed)_Beta band_PLV_“FZ”_“CZ” | −0.598 | 0.028 |
| Resting state (eyes-closed)_Beta band_PLV_“FC5”_“PZ” | −0.662 | 0.004 |
| Resting state (eyes-closed)_Beta band_wPLI_“P1”_“P2” | −0.674 | 0.002 |
| Resting state (eyes-closed)_Delta band_COH_“F2”_“PO3” | −0.606 | 0.043 |
| Resting state (eyes-closed)_Gamma band_wPLI_“T8”_“CPZ” | 0.633 | 0.047 |
| Resting state (eyes-closed)_Gamma band_wPLI_“CP4”_“P5” | −0.638 | 0.041 |
| Resting state (eyes-closed)_Theta band_wPLI_“F7”_“CP5” | −0.672 | 0.003 |
| Resting state (eyes-open)_All band_PLV_“CP1”_“PO4” | −0.634 | 0.009 |
| Resting state (eyes-open)_All band_PLV_“CP6”_“P1” | −0.596 | 0.026 |
| Resting state (eyes-open)_All band_wPLI_“PO4”_“OZ” | −0.634 | 0.010 |
| Resting state (eyes-open)_Beta band_wPLI_“FPZ”_“C6” | −0.620 | 0.026 |
| Resting state (eyes-open)_Beta band_wPLI_“F1”_“T8” | −0.613 | 0.032 |
| Resting state (eyes-open)_Beta band_wPLI_“FC5”_“FCZ” | −0.675 | 0.004 |
| Resting state (eyes-open)_Beta band_wPLI_“CPZ”_“PO3” | −0.621 | 0.024 |
| Resting state (eyes-open)_Beta band_wPLI_“P4”_“PO7” | −0.643 | 0.012 |
| Resting state (eyes-open)_Delta band_PLV_“CZ”_“P1” | −0.637 | 0.024 |
| Resting state (eyes-open)_Delta band_PLV_“PO7”_“PO4” | −0.620 | 0.041 |
| Resting state (eyes-open)_Delta band_wPLI_“P4”_“OZ” | −0.598 | 0.025 |
| Resting state (eyes-open)_Theta band_COH_“AF4”_“C6” | −0.641 | 0.019 |
| Resting state (eyes-open)_Theta band_PLV_“C6”_“TP8” | −0.659 | 0.007 |
| Neutral videos_All band_PLV_“CZ”_“PO3” | −0.625 | 0.011 |
| Neutral videos_All band_PLV_“C4”_“O2” | −0.650 | 0.005 |
| Neutral videos_Alpha band_wPLI_“C5”_“TP7” | −0.574 | 0.049 |
| Neutral videos_Beta band_COH_“TP7”_“P1” | −0.752 | <0.001 |
| Neutral videos_Beta band_wPLI_“F7”_“FC6” | −0.622 | 0.012 |
| Neutral videos_Beta band_wPLI_“FC1”_“P4” | −0.572 | 0.048 |
| Neutral videos_Delta band_COH_“F5”_“CZ” | −0.647 | 0.006 |
| Neutral videos_Delta band_COH_“C5”_“CZ” | −0.610 | 0.017 |
| Neutral videos_Delta band_PLV_“FC4”_“P6” | −0.626 | 0.046 |
| Neutral videos_Delta band_wPLI_“CP5”_“PO4” | 0.628 | 0.011 |
| Neutral videos_Gamma band_wPLI_“T8”_“TP7” | 0.624 | 0.018 |
| Neutral videos_Theta band_COH_“F3”_“P2” | −0.599 | 0.028 |
| Negative videos_All band_COH_“CP4”_“P7” | −0.605 | 0.046 |
| Negative videos_All band_PLV_“CZ”_“PO3” | −0.616 | 0.044 |
| Negative videos_Beta band_wPLI_“FP2”_“TP7” | −0.651 | 0.012 |
| Negative videos_Beta band_wPLI_“F1”_“CP5” | −0.628 | 0.025 |
| Negative videos_Delta band_PLV_“F4”_“O1” | 0.607 | 0.030 |
| Negative videos_Delta band_PLV_“CP6”_“P1” | −0.675 | 0.004 |
| Negative videos_Delta band_wPLI_“FCZ”_“C5” | 0.629 | 0.026 |
| Negative videos_Gamma band_COH_“F4”_“FC1” | 0.588 | 0.034 |
| Negative videos_Gamma band_wPLI_“FC5”_“PZ” | 0.623 | 0.027 |
| Negative videos_Theta band_COH_“F8”_“C5” | −0.577 | 0.042 |
| Negative videos_Theta band_COH_“FC6”_“CP6” | −0.588 | 0.031 |
| Negative videos_Theta band_COH_“C5”_“CZ” | −0.594 | 0.027 |
| Negative videos_Theta band_wPLI_“FC6”_“OZ” | 0.664 | 0.013 |
| Positive videos_All band_COH_“AF4”_“CZ” | −0.620 | 0.028 |
| Positive videos_All band_COH_“P5”_“PO4” | −0.630 | 0.020 |
| Positive videos_Alpha band_COH_“AF4_“CZ” | −0.722 | 0.001 |
| Positive videos_Beta band_COH_“P5”_“PO4” | −0.594 | 0.034 |
| Positive videos_Delta band_COH_“FPZ”_“AF4” | −0.604 | 0.047 |
| Positive videos_Delta band_COH_“F4”_“CP1” | −0.607 | 0.043 |
| Positive videos_Delta band_COH_“CZ”_“PO3” | −0.663 | 0.008 |
| Positive videos_Delta band_wPLI_“CP4”_“P3” | −0.606 | 0.023 |
| Positive videos_Theta band_COH_“AF4”_“CZ” | −0.639 | 0.049 |
| Positive videos_Theta band_COH_“C1”_“PO8” | −0.687 | 0.010 |
| Positive videos_Theta band_COH_“P5”_“PO8” | −0.653 | 0.031 |
| Positive videos_Theta band_wPLI_“CP4”_“PZ” | −0.600 | 0.031 |
- Note: The first part represents paradigms: resting state (eyes closed) refers to the eyes-closed resting state paradigm. Resting state (eyes open) refers to the eyes-open resting state paradigm. Neutral videos represent the neutral videos stimulus paradigm. Angry videos represent the negative videos stimulus paradigm. Happy videos represent the positive videos stimulus paradigm. The second part represents frequency band ranges: all band represents 1–45 Hz. Delta band represents 1–4 Hz. Theta band represents 4–8 Hz. Alpha band represents 8–13 Hz. Beta band represents 13–30 Hz. Gamma band represents 30–45 Hz. The third part represents functional connectivity metrics: COH represents coherence. wPLI represents weighted phase lag index. PLV represents phase locking value. The last two parts represent electrode pairs. Resting state (eyes closed)_All band_COH represents the coherence features of EEG FC obtained using the eyes-closed resting state paradigm within the 1–45 Hz frequency band range. The others can be inferred in the same manner.
- Abbreviations: COH, coherence; PLV, phase locking value; wPLI, weighted phase lag index.
3.5. Top 20 Weighted Features
Figures 1 and 2 and Supporting Scale 1–12 show top 20 weighted electrode connection pairs in the machine learning model for predicting the 8-week drug response from baseline data, based on 14 types of EEG FC features that achieved an accuracy of no less than 0.75 on the test set. You can find Supporting Scale 1–12 figures in the Supporting information.
4. Discussion
A meta-analysis [18] that synthesized five studies [19–23] with a total of 325 patients indicated an accuracy of 81.41%, a median sensitivity of 77.78%, and a median specificity of 82.06% for predicting antidepressant treatment response using EEG. Notably, most of these predictive models utilized features derived from resting-state EEG data, with only two studies [22, 24] incorporating features extracted from task-based EEG data. Therefore, this study aims to investigate the feasibility of utilizing task-based EEG data for predicting treatment outcomes, particularly employing a diverse emotional videos stimulation paradigm to assess depression-related sensitivity to various emotional cues. Additionally, there is a lack of research directly comparing the predictive performance of resting-state and task-based EEG data for the same antidepressant treatment. We found that the EEG FC under neutral and positive video stimuli performed no worse than that of resting-state EEG FC on an independent test set. However, the EEG FC under negative video stimuli performed relatively poorly on the independent test set. Exploring the efficacy of different modalities of EEG data in predicting treatment response may provide insights into potential mechanisms of action and enrich the biomarkers available for predicting treatment outcomes. Furthermore, there is a need for greater emphasis on evaluating model performance using independent test sets. The aforementioned systematic review reported that 60% of the included studies did not assess the accuracy on independent data, relying solely on internal cross-validation, which may lead to overly optimistic results [18].
Previous research [25] suggests that the connectivity between the subcortical network and the ventral attention network could be a potential target of escitalopram treatment and is associated with changes in depressive states. Baseline connectivity strength within this network may serve as a predictor of treatment response to selective serotonin reuptake inhibitors (SSRIs). FC metrics may be valuable predictors of treatment response. This study utilized three types of EEG FC metrics obtained from five types of stimuli and six types of frequency bands. The classification results obtained through fivefold cross-validation on the validation set were promising. Among 90 types of EEG FC metrics, the average accuracy of 33 feature types exceeded 0.950. Given the limited sample size of this study and the risk of overfitting, a separate test set consisting of 25% of the data was randomly allocated to evaluate the generalization ability of the models. On the test set, the accuracy of 14 feature types using the SVM classifier was no less than 0.750. Notably, using Resting state (eyes-closed)_All band_PLV and Neutral videos_Alpha band_wPLI features yielded accuracies exceeding 85%. These findings suggest that FC features computed from baseline data can effectively differentiate between responders and nonresponders to 8 weeks of medication treatment.
Except for the negative videos stimuli, machine learning models constructed using EEG features from eyes-closed resting state, eyes-open resting state, neutral videos, and positive videos stimuli all demonstrated good generalization performance. The promising results of this study may be attributed to the choice of the SVM algorithm. Previous research [18] has shown that SVMs outperform other algorithms in this context.
Regarding the comparison at baseline, previous studies have found that compared to responders, nonresponders show enhanced network connectivity, mainly in the form of long-distance connections between the temporal and occipital lobes [15]. The relevant network characteristics provide quantitative evidence of more intense brain activity among nonresponders compared to responders. As verified in previous studies [26], nonresponders experienced greater abnormal activation and excessive communication compared to responders, particularly in the alpha band, which may have led to an increase in FC within the default mode network in nonresponders. These findings align with our study, which observed predominantly negative correlations between 8-week HAMD-17 score reduction from baseline and EEG FC in both resting-state and task-based conditions. Moreover, we found that the electrode distribution of FC features associated with 8-week score reduction from baseline, as well as top 20 weighted features ranked by weight in the medication response classification model, spanned various brain regions. Similarly, the frequency bands of these features were distributed across multiple frequency ranges. These observations suggest that FC across multiple brain regions and frequency bands, in both resting-state and task-based paradigms, influences the efficacy of medication.
4.1. Limitations
The lack of a large sample size limits the application of our method in clinical practice. This study was conducted at a single center, which may limit the generalizability of the results. Multicenter studies would help to ensure that the findings are applicable across different populations and settings. The study focused on specific SSRI drugs, which may not represent the effects of other types of antidepressants or even other SSRIs. Future research should include a broader range of antidepressant medications to determine if the observed effects are consistent across different treatments.
5. Conclusion
This study explored the use of baseline EEG data from five modalities (eyes-closed resting state, eyes-open resting state, neutral videos stimulation, negative videos stimulation, and positive videos stimulation) to predict the response to escitalopram or sertraline treatment in patients with depression. Apart from the baseline EEG data from negative videos stimulation, the baseline EEG data from eyes-closed resting state, eyes-open resting state, neutral videos stimulation, and positive videos stimulation all achieved an accuracy rate of no less than 75% on an independent validation set. Notably, features from Resting state (eyes-closed)_All band_PLV and Neutral videos_Alpha band_wPLI achieved an accuracy rate of over 85%. Moreover, the study found that there are FC values at some electrodes in each modality that correlate with the reduction rate from baseline to week 8, and most of these correlations are negative. This indicates that the lower the FC values between electrodes in these modalities, the better the therapeutic effect of the medication. Collectively, these findings consistently suggest that the baseline EEG FC features from the five modalities (eyes-closed resting state, eyes-open resting state, neutral videos stimulation, negative videos stimulation, and positive videos stimulation) serve as sensitive biomarkers capable of predicting the efficacy of medication over 8 weeks, providing a feasible approach for clinical treatment strategies.
Ethics Statement
The Ethics Committee of Sir Run Run Shaw Hospital affiliated with Zhejiang University approved this study.
Consent
All subjects provided written informed consent.
Conflicts of Interest
The authors declare no conflicts of interest.
Author Contributions
Jinsong Tang guided this research. Fanglan Wang collected data, analyzed data, and wrote the manuscript. Zifan You, Tingkai Zhang, Kai Xu, Liangliang Wang, and Jingqi He helped to collect electroencephalogram (EEG) data and assisted in reviewing the manuscript. All authors made significant scientific contributions to the research in the manuscript, approved its claims, and agreed to be authors. Fanglan Wang is the first author.
Funding
This research was supported by the joint funds of the Zhejiang Provincial Natural Science Foundation of China under grant no. LBD23H090001.
Acknowledgments
This research was supported by the joint funds of the Zhejiang Provincial Natural Science Foundation of China under grant no. LBD23H090001. Fanglan Wang used Kimi to polish the language, mainly to refine some complex and difficult sentences in the article. After using this tool/service, the author reviewed and edited the content as needed and took full responsibility for the content of the publication.
Supporting Information
Additional supporting information can be found online in the Supporting Information section.
Open Research
Data Availability Statement
The data that were used in this study would be available if the readers have reasonable aims to use these data and acquire an agreement from the corresponding author.
