Sage Journals: Discover world-class research

Abstract

Objective: A large number of people with obstructive sleep apnea (OSA) also suffer from major depressive disorder (MDD), leading to underdiagnosis due to overlapping symptoms. Polysomnography has been considered to identify MDD. However, limited access to sleep clinics makes this challenging. In this study, we propose a model to detect MDD in people with OSA using an electrocardiogram (ECG) during sleep. Methods: The single-lead ECG data of 32 people with OSA (OSAD-) and 23 with OSA and MDD (OSAD+) were investigated. The first 60 min of their recordings after sleep were segmented into 30-s segments and 13 parameters were extracted: PR, QT, ST, QRS, PP, and RR; mean heart rate; two time-domain HRV parameters: SDNN, RMSSD; and four frequency heart rate variability parameters: LF_power, HF_power, total power, and the ratio of LF_power/HF_power. The mean and standard deviation of these parameters were the input to a support vector machine which was trained to separate OSAD- and OSAD+. Results: The proposed model distinguished between OSAD+ and OSAD- groups with an accuracy of 78.18%, a sensitivity of 73.91%, a specificity of 81.25%, and a precision of 73.91%. Conclusion: This study shows the potential of using only ECG for detecting depression in OSA patients.

Keywords

Electrocardiogram major depressive disorder obstructive sleep apnea screening wearable devices

Introduction

Obstructive sleep apnea (OSA) has partial or complete blockages of the upper airway during sleep which causes pauses in the air supply to the person during sleep.¹ Apnea-hypopnea index (AHI) is a measure of the number of hypopnea and apnea events and is used to determine the severity of OSA. A number of comorbidities such as cardiovascular, respiratory, endocrine, metabolic, and psychiatric symptoms are associated with OSA.² OSA association with depression symptoms has been established,^2–4 with 5%–63% OSA showing such symptoms and thus the coexistence of these two conditions.² The very big range indicates possible misdiagnosis.⁵

OSA and depression present a complex interplay of symptoms that can mask each other, making diagnosis and treatment challenging.^1,6 Patients with OSA due to disturbed sleep and oxygen during sleep have a higher tendency to sleepiness, poor concentration and alertness during the day which can lead to a number of difficulties such as accidents and quality of life.⁷

Reduced production of neurotransmitters in the brain is associated with major depressive disorder (MDD).⁸ OSA and MDD present common symptoms such as anxiety, restlessness, fatigue, and poor concentration. Identifying MDD is essential prior to treating OSA⁹ and patients exhibiting symptoms of both OSA and depression necessitate treatment for both conditions.¹⁰

Polysomnography (PSG) is the gold standard for diagnosis of OSA.¹¹ It requires overnight monitoring of the person and recording of electroencephalogram (EEG), electrooculogram (EOG), electromyogram (EMG), and electrocardiogram (ECG). However, this requires the person to sleep in the clinic while connected to a large number of electrical leads, making the assessment uncomfortable for the patient. Further, with only a few specialized sleep centers which are all located in major urban centers, patients often have long waiting periods before assessment. An alternate option is for home-based assessment, and this is feasible if there are a small number of sensors that can be easily placed on the user. For this purpose, only ECG and PPG have been proposed¹² which has the potential for home-based assessment.^13,14

Clinical interviews are the “gold standard” for validating self-rated questionnaires that are used for detecting MDD.¹⁵ However, both of these are subjective, time-consuming, and thus expensive. Mental health sleep clinics are related to Psychiatry and have emerged as the gold standard for detecting and treating MDD and OSA.^16,17 However, these are few and located in large urban centers. Thus, the patients who report OSA symptoms are investigated for sleep disorders while the MDD condition is often under-diagnosed.^18,19 To overcome this, machine learning models were proposed which use EEG, ECG, and breathing recordings as input to detect OSA patients with MDD.²⁰ However, recordings such as EEG require multiple leads and masks to be connected to the patient, making this unsuitable for home-based assessment.

ECG has been found to distinguish between healthy people and people with suicidal thoughts by Khandoker et al.²¹ The authors reported the difference in heart rate variability of people with MDD. This work has identified the changes to the cardiac activity that are associated with MDD and are useful for detecting MDD. However, a number of MDD symptoms are similar to OSA, and thus there is a need to detect the differences between MDD and non-MDD patients with OSA.

Previous studies^22–25 have demonstrated significant differences in ECG signals between individuals with OSA and healthy controls, identifying specific ECG markers such as variations in heart rate variability (HRV), changes in R-R intervals, and altered autonomic function indicative of OSA. Various studies^26–28 have successfully utilized ECG data for detecting OSA, achieving high accuracy. For instance, Varon et al.²⁹ and Zarei et al.²⁷ used single-lead ECG recordings for sleep apnea detection, reporting accuracies up to 97.21%. Similarly, ECG has been effectively employed to identify mental disorders.^27,30–34 Kobayashi et al.³² and Khandoker et al.³⁴ achieved high sensitivity and specificity in detecting psychiatric conditions using ECG.

The association between OSA and MDD is well-documented, with overlapping symptoms such as fatigue, daytime sleepiness, poor concentration, and anxiety. Individuals with both conditions also exhibit a heightened risk of cardiovascular disease.³⁵ Despite these findings, no study has yet attempted to use only ECG to detect the comorbidity of MDD within the OSA cohort. To address this gap, we investigated sleep ECG parameters among OSA patients to screen for coexistence of MDD, expanding on previous research that focused on using ECG to differentiate OSA patients from healthy individuals.

Methods

Participants and data collection

Data were collected from participants with OSA at the American Center for Psychiatry and Neurology (ACPN) in Abu Dhabi after receiving ethics approval from the ACPN Institutional Review Board (IRB): approval 0019/2017. Written and verbal consent was obtained from all the participants prior to data collection. Patients having comorbid cardiovascular diseases were excluded from this study. The information of the participants is shown in Table 1. The prevalence of MDD in the OSA population is 23%.³⁶ Using a 95% confidence level and a margin error of 5% for the sample size calculation, the required number of participants for each group is 273. Our pilot study included 23 OSAD+ and 32 OSAD- participants. In similar research, sample sizes for detecting OSA using ECG recordings have ranged from 70 to 95 participants,^29,37–41 and for detecting MDD using ECG recordings, sample sizes have ranged from 23 to 78 participants.^32–34 Our sample size falls within these ranges, demonstrating its adequacy.

Table 1.

Demographic and clinical data of OSA patients with and without MDD.

	OSAD- (OSA without MDD	OSAD+ (OSA with MDD)	p-values
Participants	32	23	-
Gender(M/F)	22/10	13/10	-
Age	43.37 ± 14.31	42.48 ± 10.91	0.92
BMI (kg/m²)	34.81 ± 9.75	32.01 ± 6.97	0.29
(AHI/hr)	27.62 ± 17.01	25.36 ± 14.06	0.82

A consultant psychiatrist assessed OSA patients with mental health concerns using the Patient Health Questionnaire-9 (PHQ-9).^42,43 This questionnaire, consisting of nine items, is utilized for assessing the severity of depressive symptoms through self-reporting. Participants select one or more response options, ranging from 0 (no presence of the symptom) to 3 (almost every day). The total score is calculated by summing the highest response for each item, falling within the range of 0 to 27. Scores of 5, 10, 15, and 20 on the PHQ-9 correspond to mild, moderate, moderately severe, and severe depression, respectively.⁴⁴ People with a positive score greater than 9 were considered to have MDD.

ECG signals were recorded at the sampling frequency of 100 Hz using the MiniScreen Plus (Löwenstein, Germany). The data has previously been reported in the work reported by Moussa et al.²⁰ and the deidentified recordings are available on request.

Data analysis

Segmentation

The first one-hour ECG data immediately after the patient fell asleep was taken for analysis. This was done to streamline the machine-based analysis without requiring the sleep labels. All the stages of NREM cover within 1 hour of the first sleep cycle,⁴⁵ which has been reported as suitable for cardiac analysis.⁴⁶ A window of 30 s duration was selected for the parameter’s extraction from the ECG signal. The 30-s ECG data is considered for the calculation of parameters because it provides a long enough time window to capture important information about the heart’s electrical activity, while still being practical for clinical use. Analyzing a short-term ECG signal, typically less than 60 s, is effective in finding the optimal length and number of segments for analysis.^47,48

After segmentation, baseline wandering, and power-line noise were suppressed using a 0.5-45 Hz bandpass filter. Noisy segments having movement artifacts or loss of signal recording were removed from the analysis using manual inspection, as shown in Figure 1. Figure 1(b) – (d) shows examples of signals that were found to be unsuitable because of the presence of noise and artifacts. In (b), although the R peaks are visible, the signal is highly noisy, making it challenging to accurately identify and analyze other critical components such as the P, Q, S, and T peaks. This noise can interfere with precise annotations and potentially lead to unreliable parameter extraction. Figure 1(c) exhibits pronounced movement artifacts, resulting in abrupt, large amplitude deflections unrelated to cardiac activity, which significantly distort the ECG signal and make it difficult to accurately identify the morphology of the P, Q, R, S, and T peaks. Additionally, there are intervals of signal loss in Figure 1(c), further complicating the identification process and disrupting the continuity of the ECG recording. In (d), the ECG signal shows irregularities and high levels of noise throughout the recording, making it difficult to consistently identify and annotate the P, Q, R, S, and T peaks.

Figure 1.

Accepted and rejected frames: (a) accepted segment (b-d) rejected segments.

Feature extraction

Pan–Tompkins algorithm⁴⁹ was used to identify the QRS complex while ECGdeli⁵⁰ was used to identify the P and T. The six different intervals (PR, QT, ST, QRS, PP, and RR) were calculated further. The ECG wave with the segments is shown in Figure 2.

Figure 2.

Morphological features of an example ECG signal.

From the RR intervals, we calculated heart rate (HR), HRV time (SDNN, RMSDD), and frequency parameters (Lf_power, Hf_power, total power, and ratio of LF_power/Hf_power).⁵¹ Table 2 provides the equations used to calculate these 13 parameters from ECG segments.

Table 2.

Equations and definitions of parameters calculated from ECG segments.

Parameters	Formulas	Definition/remarks
$m e a n (PR)$	$\frac{1}{n} \sum_{i = 1}^{n} (\frac{\| R_{s a m p l e s} [i] - P_{s a m p l e s} [i] \|}{F_{s}} \times 1000)$	$R_{s a m p l e s} [i]$ $P_{s a m p l e s} [i]$ , $T_{s a m p l e s} [i]$ , $S_{s a m p l e s} [i]$ and $Q_{s a m p l e s} [i]$ are the samples corresponding to the R, P, T, S and Q peaks for the ith heart cycle, respectively. Here, n is the total number of heart cycles present in the segment.
$m e a n (Q T)$	$\frac{1}{n} \sum_{i = 1}^{n} (\frac{\| T_{s a m p l e s} [i] - Q_{s a m p l e s} [i] \|}{F_{s}} \times 1000)$
$m e a n (S T)$	$\frac{1}{n} \sum_{i = 1}^{n} (\frac{\| T_{s a m p l e s} [i] - S_{s a m p l e s} [i] \|}{F_{s}} \times 1000)$
$m e a n (Q R S)$	$\frac{1}{n} \sum_{i = 1}^{n} (\frac{\| S_{s a m p l e s} [i] - Q_{s a m p l e s} [i] \|}{F_{s}} \times 1000)$
$m e a n (R R)$	$\frac{1}{n - 1} \sum_{i = 1}^{n - 1} (\frac{\| R_{s a m p l e s} [i + 1] - R_{s a m p l e s} [i] \|}{F_{s}} \times 1000)$
$m e a n (P P)$	$\frac{1}{n - 1} \sum_{i = 1}^{n - 1} (\frac{\| P_{s a m p l e s} [i + 1] - P_{s a m p l e s} [i] \|}{F_{s}} \times 1000)$
HR	$\frac{60}{m e a n (R R)}$
SDNN	$\sqrt{\frac{1}{n - 2} \sum_{i = 1}^{n - 1} {((\frac{R_{s a m p l e s} [i + 1] - R_{s a m p l e s} [i]}{F_{s}}) - m e a n (R R))}^{2}}$
RMSSD	$\sqrt{\frac{1}{n - 2} \sum_{i = 1}^{n - 2} {(\frac{R_{s a m p l e s} [i + 2] - R_{s a m p l e s} [i + 1]}{F_{s}} - \frac{R_{s a m p l e s} [i + 1] - R_{s a m p l e s} [i]}{F_{s}})}^{2}}$
LF power	$b a n d p o w e r (p, f, [0.04, 0.15],' p s d')$	Power in the low-frequency range (0.04-0.15 Hz) using the lomb-scargle periodogram. Here, p are the PSD values at specific frequencies, and f are the corresponding frequency values.
HF power	$b a n d p o w e r (p, f, [0.15, 0.4],' p s d')$
Total power	$L F p o w e r + H F p o w e r$
LF/HF Ratio	$\frac{L F p o w e r}{H F p o w e r}$

The mean ( $μ_{j}$ ) and standard deviation ( $σ_{j}$ ) of each parameter were calculated across all selected frames.

μ_{j} = \frac{1}{N} \sum_{i = 1}^{N} p_{i j}

σ_{j} = \sqrt{\frac{1}{N} \sum_{i = 1}^{N} {(p_{i j} - μ_{j})}^{2}}

where N is the total number of selected frames and

[μ_{j}, σ_{j}]

of each feature. As shown in Figure 3, a total of 26 features were obtained.

Figure 3.

Feature extraction of 1-h ECG data.

Feature selection

A ReliefF algorithm⁵² was employed to rank the features within the dataset. To obtain the most appropriate number of features, the highest-ranked feature was the input to a support-vector machine (SVM) and its performance was obtained. This was repeated by incrementally adding the next ranked feature and repeated for all 26 features. The feature set with the highest accuracy was selected.

Classification

The ECG recordings of OSA patients were classified into two classes: with MDD (OSAD+) and with no MDD (OSAD-). An SVM classifier with an RBF kernel was employed for classification. Hyperparameters C and gamma were optimized using a grid search approach, exploring values from 2⁻¹ to 2⁷ for C and 2⁻³ to 2⁴ for gamma, as these parameters have a strong influence on the performance of the classifiers.^37,53 A leave-one-out (LOO) cross-validation method was used to evaluate the model’s performance. An incremental feature inclusion strategy was applied, progressively including top-ranked features and performing a grid search for each subset. This process identified the optimal feature set and hyperparameters, yielding the best-performing SVM model for the classification task. The model was evaluated based on different parameters such as accuracy, sensitivity, specificity, precision, NPV, F1, positive Likelihood Ratio (LR+) and negative Likelihood Ratio (LR-) calculated based on True Positive (TP), True Negative (TN), False Positive (FP) and Falise Negative (FN).

Once the optimum feature set was obtained, five other classifiers, namely: (i) XGBoost, (ii) KNN, (iii) Logistic Regression, (iv) Random Forest, and (v) Decision Tree, were also tested to determine which classifier could provide the best result.

Accuracy = (TP + TN) / (TP + TN + FP + FN)

Sensitivity = TP / (TP + FN)

Specificity = TN / (TN + FP)

Precision = TP / (TP + FP)

Negative Predictive Value (NPV) = TN / (TN + FN)

F 1 = 2 * (precision * sensitivity) / (precision + sensitivity)

LR + = sensitivity / 1 - specificity

LR - = 1 - sensitivity / specificity

Results

In the analysis of 1 hour of ECG data for each subject, divided into 30-s segments, we obtained a total of 120 segments per subject. Table 3 shows the total number of segments, the number of selected segments, and the number of rejected segments for both the OSAD+ and OSAD- groups.

Table 3.

Summary of ECG segments analyzed for OSAD+ and OSAD- groups.

Group	Total segments	Selected	Rejected
OSAD+	2760	2375	385
OSAD-	3840	3133	707

For the OSAD + group, out of 2760 total segments (from 23 subjects), 2375 segments were selected and 385 segments were rejected. For the OSAD- group, out of 3840 total segments (from 32 subjects), 3133 segments were selected, and 707 segments were rejected. These exclusions indicate that 13.94% of the segments in the OSAD + group and 18.41% of the segments in the OSAD- group were rejected.

The performance of the SVM model for screening MDD patients within the OSA cohort, i.e., OSAD+, is shown in Figure 4. When top-ranked features were introduced cumulatively to the model. The bar plot illustrates the classification accuracy achieved with cumulative top-ranked features ranging from 1 to 26. Each bar represents the accuracy of the SVM model using an increasing number of features, starting from the top-ranked feature and progressively adding more.

Figure 4.

Classification accuracy of the SVM model with cumulative top-ranked features.

The highest classification accuracy of 0.78 is marked by a green bar, which corresponds to the inclusion of the top 5 features. For the top-ranked 5 features, the best parameters identified were C = 1 and gamma = 0.25. Including more features beyond this point does not lead to a significant improvement in accuracy; in fact, the accuracy tends to stabilize or even decrease slightly with additional features. This trend suggests that while the top 5 features provide the most relevant information for classification, adding more features may introduce noise or redundancy. This observation emphasizes that a smaller, well-chosen subset of features can lead to better classification outcomes compared to using a larger set of features indiscriminately.

The result obtained from the ReliefF feature ranking algorithm followed by iteratively determining the number of features using SVM showed that the top five features provide the best classification result for distinguishing between OSA patients with and without MDD. The features were SD-HR, SD-PP, SD-RR, Mean-TP, and Mean-LF-Power. The descriptive statistics (mean and standard deviation) of these five features are shown in Table 4.

Table 4.

Descriptive statistics of top five features.

Parameters	OSAD-mean (SD)	OSAD + mean (SD)
SD_PP	35.30 (13.76)	32.68 (16.45)
SD_HR	2.93 (1.03)	3.13 (1.41)
SD_RR	35.30 (13.77)	32.66 (16.44)
Mean_TP	0.36 (0.05)	0.37 (0.04)
Mean_lf_power	0.20 (0.06)	0.21 (0.08)

After obtaining the optimal feature set using the SVM model, we evaluated the performance of five other classifiers: KNN, Logistic Regression, Random Forest, Decision Tree, and XGBoost. The performance metrics for each classifier, including accuracy, specificity, sensitivity, precision, and F1 score, are summarized in Table 5.

Table 5.

Performance metrics of various classifiers using the optimal feature set.

Model	Accuracy	Specificity	Sensitivity	Precision	F1
SVM	0.7818	0.8125	0.7391	0.7391	0.7391
KNN	0.7455	0.8438	0.6087	0.7368	0.6667
Logistic regression	0.5818	0.5938	0.5652	0.5000	0.5306
Decision tree	0.6364	0.7188	0.5217	0.5714	0.5455
XGBoost	0.6545	0.7500	0.5217	0.6000	0.5581

From Table 5, it is evident that the SVM model outperformed the other classifiers across all performance metrics. The best SVM model from grid-search resulted in 78.18% accuracy, 73.91% sensitivity, 81.25% specificity, and 73.91% precision. Figure 5 shows the corresponding confusion matrix for differentiating between the two groups.

Figure 5.

Confusion matrix for differentiating between OSAD- (OSA without MDD) and OSAD+ (OSA with MDD).

Discussion

The prevalence of MDD in the OSA population is estimated to be between 20-30%, with shared symptoms such as fatigue, sleep disturbances, and cognitive impairments.^36,54,55 These commonalities often lead to the underdiagnosis of MDD in OSA patients, resulting in inadequate treatment and poorer health outcomes. Thus, there is a pressing need to screen for MDD in patients with OSA due to the high prevalence and overlapping symptoms of both conditions. Early screening for MDD in OSA patients is crucial,⁵⁶ as addressing both conditions simultaneously is essential for comprehensive care and improved patient outcomes.⁵⁷

In this study, we applied single-lead ECG to screen for MDD within an OSA patient group, expanding on previous research that focused on using ECG to differentiate OSA patients from healthy individuals. In addition, while earlier studies have examined OSA and MDD separately, this is the first to explore their coexistence using single-lead ECG, offering a more accessible and efficient screening tool.

Heart Rate Variability (HRV) analysis indicates the changes in autonomic function and is an indicator of OSA. OSA patients typically exhibit increased sympathetic activity and reduced parasympathetic modulation during apneic events, which can be captured through HRV metrics such as SDNN, LF/HF ratio, RMSSD, and entropy-based features like Approximate Entropy (ApEn) and Sample Entropy (SampEn). HRV analysis is also a well-established method for identifying autonomic dysfunction in MDD patients, with key features such as SDNN, RMSSD, LF/HF ratio, and HF power indicating reduced parasympathetic activity and increased sympathetic dominance. Leveraging this knowledge, we explored both HRV and interval-based features to develop a model capable of classifying ECG recordings of OSA patients into two categories: those with MDD (OSAD+) and those without MDD (OSAD-). This approach incorporates the known autonomic disturbances in both OSA and MDD, providing a comprehensive foundation for our classification model.

Current state of art for ECG analysis for detecting OSA and MDD

The use of ECG for detecting OSA has been proposed to reduce the burden of investigating patients in sleep clinics. Varon et al.²⁹ employed 80 single-lead ECG recordings to detect sleep apnea, with 70 sourced from the publicly available Apnea-ECG database and 10 obtained from the Sleep Laboratory of the University Hospital Leuven in Belgium. Zarei et al.³⁷ used 95 ECG recordings for obstructive sleep apnea detection, comprising 70 from the Apnea-ECG database and 25 from the UCD database. Atri et al., Li et al., and Mashrur et al.^38–40 each used 70 ECG recordings from the Apnea-ECG database for OSA detection. Urtnasan et al.⁴¹ utilized PSG data from 82 participants for the detection of OSA events and normal events.

Song et al.⁵⁸ used a discriminative hidden Markov model (HMM) to detect OSA and achieved an accuracy of 97.1%. Yücelbaş et al.²⁶ considered 16 ECG morphological features to identify sleep parameters and their accuracy was 78%. Yang et al.²⁸ implemented a one-dimensional squeeze-and-excitation (SE) residual group network, achieving an accuracy of 90.3% for per-segment detection in the Apnea-ECG database. Varon et al.²⁹ utilized 4 features from ECG recordings and employed a least-squares SVM (RBF kernel), achieving an accuracy of about 85% for per-segment OSA detection. Zarei et al.²⁷ utilized 108 entropy-based and non-linear features from wavelet transform coefficients, and OSA detection accuracies were 94.63% and 95.71% using SVM (RBF kernel) for per-minute and participant-based classifications, respectively. Atri et al.³⁸ extracted spectrum and bi-spectrum features from HRV and EDR signals, utilizing an SVM classifier to achieve an accuracy of 95.57%. Li et al.³⁹ employed a deep neural network and HMM model for OSA detection, attaining an accuracy of 85%. Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) Network were used by Zarei et al.²⁷ The authors achieved an accuracy of 97.21% for the Apnea-ECG database and 95.82% for the UCD database. These studies have shown that there are significant differences between ECG of people with and without OSA and have also developed the baseline to facilitate the ECG analysis as an alternative to traditional sleep studies. Previous studies have demonstrated significant differences in ECG signals between individuals with OSA and healthy controls, identifying specific ECG markers such as variations in heart rate variability (HRV), changes in R-R intervals, and altered autonomic function indicative of OSA.

Kobayashi et al.³² employed 23 participants (16 healthy and 7 psychiatric patients) to develop a Mental Disorder Screening System using an SVM model from single-lead electrocardiography, achieving a sensitivity of 71.4% and specificity of 93.8%. Khandoker et al.²¹ used 61 participants (32 patients with MDD; 16 of whom had suicidal ideation) to predict depressed patients with suicidal ideation from ECG recordings, obtaining 94.83% accuracy in predicting MDD subjects with suicidal ideation. Byun et al.³³ employed a sample size of 78 individuals, consisting of 37 patients diagnosed with MDD and 41 healthy controls, to detect MDD using linear and nonlinear heart rate variability analysis, achieving 74.4% accuracy, 73% sensitivity, and 75.6% specificity with two optimal features selected by SVM-RFE. Tasci et al.³¹ used 198 participants (62 bipolar, 17 depressions, 119 schizophrenia) for automatic psychiatric disorders classification using ECG signals. The lead-by-lead accuracy range of the proposed model using the ANN classifier was from 73.67% to 89.19%. Khare et al.³⁰ used 233 subjects taken from two publicly available datasets with the ECGPsychNet model for the automatic detection of psychiatric disorders (schizophrenia, bipolar disorder, and depression) using ECG signals. The developed ECGPsychNet obtained the highest classification accuracy of 98.15%. These studies have demonstrated the potential of employing ECG for the identification of mental disorders.

The association between OSA and MDD is known,^2–4 and a number of self-presented symptoms by the patients of both categories such as fatigue, sleepiness during the day, poor concentration, and anxiety are often similar. It has also been found that individuals concurrently afflicted with OSA and MDD exhibit heightened susceptibility to cardiovascular disease.³⁵ However, no study has yet been reported that has attempted to use only ECG to detect comorbidity of MDD from the OSA cohort.

Our work in relation to state-of-the-art

In this study, we used single-lead ECG data of 55 participants (23 OSAS+ and 32 OSAS-) to create an SVM-based model to detect depression in OSA. After identifying the 5 highest ranked ECG features were: (i) SD-HR, (ii) SD-PP, (iii) SD-RR, (iv) Mean-TP, and (v) Mean-LF-Power. With an accuracy of 78.18%, sensitivity of 73.91%, and specificity of 81.25%, these findings indicate the potential for utilizing this method to screen OSA patients for the detection of MDD. Single-lead ECG can be easily recorded using standard Holter monitors and thus can be conducted in the homes of the patients without requiring access to sleep clinics. The current method for detecting MDD in OSA individuals requires the use of sleep clinics and PSG and no method uses only ECG for this purpose. Ours is the first study that reports the use of only ECG for this aim and the performance is moderate.

The major limitation of this study is that it is based on the data collected in a single clinic. The dataset is relatively small which was obtained by a cross-sectional study, and hence factors such as the effect of ethnicity and demographics could not be investigated. Further, the dataset does not have information such as diet, alcohol consumption, and exercise, which could help improve the accuracy of such a model. Another limitation of this work is that the detection of the noisy segments was done manually. It is essential to automate the detection of noisy segments for this approach to be suitable for computerized analysis. One of the factors that may influence the ECG is the medications taken by the individuals such as selective serotonin reuptake inhibitors (SSRIs) which are known to lead to QT prolongation potentially.⁵⁹ However, the database does not provide this information, which is a limitation of this study. Yet another limitation of this work is that the assessment of the patients was not done by cardiologists, thus the pathophysiology underpinning the differences between the two groups cannot be explored.

Conclusion

In this study, we have developed and tested an ECG-based system that can identify MDD patients in a cohort of people with OSA. We have identified the five most relevant ECG parameters, and our system’s performance includes an accuracy of 78.18%, sensitivity of 73.91%, specificity of 81.25%, and precision of 73.91%. This model has the potential to be used for screening MDD patients among OSA patients based solely on using their single-lead ECG which makes it suitable for being used in the homes of the patients.

Footnotes

Author contributions

VS- data analysis and first draft; QN- result analysis and editing; NP- data analysis; GO- data analysis; AK- data management and editing; PM- editing; DP- editing; DK- project planning and editing.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: The work has been supported in part by PhD scholarship received by Vikash Shaw and Guilherme Oliveira, a grant from Khalifa University (8474000376 (KAU-KU-2021-02 awarded to Ahsan Khandoker)), and a grant from ICMR Delhi, India (5/3/8/34/ITR-F/2020-ITR 1).

Ethical statement

ORCID iD

Quoc Cuong Ngo

Data availability statement

The de-identified data will be made available on request to the corresponding author.*

References

Schroder

O'Hara

. Depression and obstructive sleep apnea (OSA). Ann Gen Psychiatr 2005; 4: 13. DOI: 10.1186/1744-859X-4-13.

Ejaz

Khawaja

Bhatia

, et al. Obstructive sleep apnea and depression: a review. Innov Clin Neurosci 2011; 8: 17–25.

Sivertsen

Overland

Glozier

, et al. The effect of OSAS on sick leave and work disability. Eur Respir J 2008; 32: 1497–1503. DOI: 10.1183/09031936.00044908.

Rey de Castro

Rosales-Mayor

. Depressive symptoms in patients with obstructive sleep apnea/hypopnea syndrome. Sleep Breath 2013; 17: 615–620. DOI: 10.1007/s11325-012-0731-1.

Pan

Tsao

Hsu

, et al. Bidirectional association between obstructive sleep apnea and depression: a population-based longitudinal study. Medicine 2016; 95: e4833. DOI: 10.1097/MD.0000000000004833.

Jehan

Auguste

Pandi-Perumal

, et al. Depression, obstructive sleep apnea and psychosocial health. Sleep Med Disord 2017; 1: 20171027.

Sforza

de Saint Hilaire

Pelissolo

, et al. Personality, anxiety and mood traits in patients with sleep-related breathing disorders: effect of reduced daytime alertness. Sleep Med 2002; 3: 139–145. DOI: 10.1016/s1389-9457(01)00128-9.

Hasler

. Pathophysiology of depression: do we have any solid evidence of interest to clinicians? World Psychiatr 2010; 9: 155–161. DOI: 10.1002/j.2051-5545.2010.tb00298.x.

Peppard

Szklo-Coxe

Hla

, et al. Longitudinal association of sleep-related breathing disorder and depression. Arch Intern Med 2006; 166: 1709–1715. DOI: 10.1001/archinte.166.16.1709.

10.

Means

Lichstein

Edinger

, et al. Changes in depressive symptoms after continuous positive airway pressure treatment for obstructive sleep apnea. Sleep Breath 2003; 7: 31–42. DOI: 10.1007/s11325-003-0031-x.

11.

Berry

Brooks

Gamaldo

, et al. AASM scoring manual updates for 2017 (Version 2.4). J Clin Sleep Med 2017; 13: 665–666. DOI: 10.5664/jcsm.6576.

12.

Motin

Karmakar

Palaniswami

, et al. Multi-stage sleep classification using photoplethysmographic sensor. R Soc Open Sci 2023; 10: 221517. DOI: 10.1098/rsos.221517.

13.

Golpe

Jiménez

Carpizo

. Home sleep studies in the assessment of sleep apnea/hypopnea syndrome. Chest 2002; 122: 1156–1161. DOI: 10.1378/chest.122.4.1156.

14.

Douglas

. Home diagnosis of the obstructive sleep apnoea/hypopnoea syndrome. Sleep Med Rev 2003; 7: 53–59. DOI: 10.1053/smrv.2001.0205.

15.

Davison

McCabe

Mellor

. An examination of the “gold standard” diagnosis of major depression in aged-care settings. Am J Geriatr Psychiatr 2009; 17: 359–367. DOI: 10.1097/JGP.0b013e318190b901.

16.

Stubbs

Vancampfort

Veronese

, et al. The prevalence and predictors of obstructive sleep apnea in major depressive disorder, bipolar disorder and schizophrenia: a systematic review and meta-analysis. J Affect Disord 2016; 197: 259–267. DOI: 10.1016/j.jad.2016.02.060.

17.

Plante

. The evolving nexus of sleep and depression. Am J Psychiatr 2021; 178: 896–902. DOI: 10.1176/appi.ajp.2021.21080821.

18.

Edwards

Mukherjee

Simpson

, et al. Depressive symptoms before and after treatment of obstructive sleep apnea in men and women. J Clin Sleep Med 2015; 11: 1029–1038. DOI: 10.5664/jcsm.5020.

19.

Knechtle

Economou

Nikolaidis

, et al. Clinical characteristics of obstructive sleep apnea in psychiatric disease. J Clin Med 2019; 8: 534. DOI: 10.3390/jcm8040534.

20.

Moussa

Alzaabi

Khandoker

. Explainable computer-aided detection of obstructive sleep apnea and depression. IEEE Access 2022; 10: 110916–110933. DOI: 10.1109/Access.2022.3215632.

21.

Khandoker

Luthra

Abouallaban

, et al. Predicting depressed patients with suicidal ideation from ECG recordings. Med Biol Eng Comput 2017; 55: 793–805. DOI: 10.1007/s11517-016-1557-y.

22.

Dissanayake

Bin

Ucak

, et al. Association between autonomic function and obstructive sleep apnea: a systematic review. Sleep Med Rev 2021; 57: 101470. DOI: 10.1016/j.smrv.2021.101470.

23.

Qin

Fietze

Mazzott

, et al. Obstructive sleep apnea heterogeneity and autonomic function: a role for heart rate variability in therapy selection and efficacy monitoring. J Sleep Res 2024; 33: e14020. DOI: 10.1111/jsr.14020.

24.

Qin

Steenbergen

Glos

, et al. The different facets of heart rate variability in obstructive sleep apnea. Front Psychiatr 2021; 12: 642333. DOI: 10.3389/fpsyt.2021.642333.

25.

Qin

Keenan

Mazzotti

, et al. Heart rate variability during wakefulness as a marker of obstructive sleep apnea severity. Sleep 2021; 44: zsab018. DOI: 10.1093/sleep/zsab018.

26.

Yücelbas

Tezel

, et al. Identification of full-night sleep parameters using morphological features of ECG signals: a practical alternative to EEG and EOG signals. Biomed Signal Proces 2024; 88: 105633. DOI: 10.1016/j.bspc.2023.105633.

27.

Zarei

Beheshti

Asl

. Detection of sleep apnea using deep neural networks and single-lead ECG signals. Biomed Signal Proces 2022; 71: 103125. DOI: 10.1016/j.bspc.2021.103125.

28.

Yang

Zou

Wei

, et al. Obstructive sleep apnea detection from single-lead electrocardiogram signals using one-dimensional squeeze-and-excitation residual group network. Comput Biol Med 2022; 140: 105124. DOI: 10.1016/j.compbiomed.2021.105124.

29.

Varon

Caicedo

Testelmans

, et al. A novel algorithm for the automatic detection of sleep apnea from single-lead ECG. Ieee T Bio-Med Eng 2015; 62: 2269–2278. DOI: 10.1109/Tbme.2015.2422378.

30.

Khare

Gadre

Acharya

. ECGPsychNet: an optimized hybrid ensemble model for automatic detection of psychiatric disorders using ECG signals. Physiol Meas. 2023; 44(11): 115004. doi:10.1088/1361-6579/ad00ff.

31.

Tasci

Dogan

, et al. A novel ternary pattern-based automatic psychiatric disorders classification using ECG signals. Cogn Neurodyn 2022; 18: 95–108. DOI: 10.1007/s11571-022-09918-8.

32.

Kobayashi

Sun

Shinba

, et al. Development of a mental disorder screening system using support vector machine for classification of heart rate variability measured from single-lead electrocardiography. 2019 IEEE sensors applications symposium (Sas): IEEE, 2019, pp. 1–6.

33.

Byun

Kim

Jang

, et al. Detection of major depressive disorder from linear and nonlinear heart rate variability features during mental task protocol. Comput Biol Med 2019; 112: 103381. DOI: 10.1016/j.compbiomed.2019.103381.

34.

Khandoker

Luthra

Abouallaban

, et al. Predicting depressed patients with suicidal ideation from ECG recordings. Med Biol Eng Comput 2017; 55: 793–805. DOI: 10.1007/s11517-016-1557-y.

35.

Liu

Peng

Zhou

, et al. Depression with obstructive sleep apnea lead to high cardiovascular disease morbidity/all-cause mortality: findings from the SHHS cohort. J Sleep Res 2023; 32: e13828. DOI: 10.1111/jsr.13828.

36.

Jackson

Tolson

Bartlett

, et al. Clinical depression in untreated obstructive sleep apnea: examining predictors and a meta-analysis of prevalence rates. Sleep Med 2019; 62: 22–28. DOI: 10.1016/j.sleep.2019.03.011.

37.

Zarei

Asl

. Automatic detection of obstructive sleep apnea using wavelet transform and entropy-based features from single-lead ECG signal. Ieee J Biomed Health 2019; 23: 1011–1021. DOI: 10.1109/Jbhi.2018.2842919.

38.

Atri

Mohebbi

. Obstructive sleep apnea detection using spectrum and bispectrum analysis of single-lead ECG signal. Physiol Meas 2015; 36: 1963–1980. DOI: 10.1088/0967-3334/36/9/1963.

39.

Pan

, et al. A method to detect sleep apnea based on deep neural network and hidden Markov model using single-lead ECG signal. Neurocomputing 2018; 294: 94–101. DOI: 10.1016/j.neucom.2018.03.011.

40.

Mashrur

Islam

Saha

, et al. SCNN: scalogram-based convolutional neural network to detect obstructive sleep apnea using single-lead electrocardiogram signals. Comput Biol Med 2021; 134: 104532. DOI: 10.1016/j.compbiomed.2021.104532.

41.

Urtnasan

Park

Joo

, et al. Automated detection of obstructive sleep apnea events from a single-lead electrocardiogram using a convolutional neural network. J Med Syst 2018; 42: 104. DOI: 10.1007/s10916-018-0963-0.

42.

Salkind

. Beck depression inventory in general practice. J Roy Coll Gen Pract 1969; 18: 267–271.

43.

Gilbody

Richards

Brealey

, et al. Screening for depression in medical settings with the patient health questionnaire (PHQ): a diagnostic meta-analysis. J Gen Intern Med 2007; 22: 1596–1602. DOI: 10.1007/s11606-007-0333-y.

44.

Kroenke

Spitzer

Williams

. The PHQ-9: validity of a brief depression severity measure. J Gen Intern Med 2001; 16: 606–613. DOI: 10.1046/j.1525-1497.2001.016009606.x.

45.

Patel

Reddy

Shumway

, et al. Physiology, sleep stages. In: StatPearls. Treasure Island (FL): StatPearls Publishing LLC, 2023.

46.

Malik

. Sleep-wake classification via quantifying heart rate variability by convolutional neural network. Physiol Meas 2018; 39: 085004. DOI: 10.1088/1361-6579/aad5a9.

47.

Steinberg

O'Connell

, et al. Thirty-second gold standard definition of atrial fibrillation and its relationship with subsequent arrhythmia patterns: analysis of a large prospective device database. Circ Arrhythm Electrophysiol 2018; 11: e006274. DOI: 10.1161/CIRCEP.118.006274.

48.

Erdenebayar

Kim

Park

, et al. Automatic prediction of atrial fibrillation based on convolutional neural network using a short-term normal electrocardiogram signal. J Kor Med Sci 2019; 34: e64. DOI: 10.3346/jkms.2019.34.e64.

49.

Sedghamiz

. Matlab implementation of Pan Tompkins ECG QRS detector. Code Available at the File Exchange Site of MathWorks: MathWorks, 2014.

50.

Pilia

Nagel

Lenis

, et al. ECGdeli-An open source ECG delineation toolbox for MATLAB. SoftwareX 2021; 13: 100639. DOI: 10.1016/j.softx.2020.100639.

51.

Sammito

Böckelmann

. Reference values for time- and frequency-domain heart rate variability measures. Heart Rhythm 2016; 13: 1309–1316. DOI: 10.1016/j.hrthm.2016.02.006.

52.

Robnik-Sikonja

Kononenko

. Theoretical and empirical analysis of ReliefF and RReliefF. Mach Learn 2003; 53: 23–69. DOI: 10.1023/A:1025667309714.

53.

Syarif

Prugel-Bennett

Wills

. SVM parameter optimization using grid search and genetic algorithm to improve classification performance. TELKOMNIKA (Telecommunication Computing Electronics and Control) 2016; 14: 1502–1509.

54.

Zhao

Zhang

Sun

, et al. Association between obstructive sleep apnea severity and depression risk: a systematic review and dose-response meta-analysis. Sleep Breath 2024; 28: 2175–2185. DOI: 10.1007/s11325-024-03083-4.

55.

Edwards

Almeida

Ford

. Obstructive sleep apnea and depression: a systematic review and meta-analysis. Maturitas 2020; 142: 45–54. DOI: 10.1016/j.maturitas.2020.06.002.

56.

Peppard

Szklo-Coxe

Hla

, et al. Longitudinal association of sleep-related breathing disorder and depression. Arch Intern Med 2006; 166: 1709–1715. DOI: 10.1001/archinte.166.16.1709.

57.

Chinvararak

Garcia-Borreguero

. Comorbid major depressive disorder and obstructive sleep apnea. Case Rep Psychiatry 2022; 2022: 2943059. DOI: 10.1155/2022/2943059.

58.

Song

Liu

Zhang

, et al. An obstructive sleep apnea detection approach using a discriminative hidden Markov model from ECG signals. Ieee T Bio-Med Eng 2016; 63: 1532–1542. DOI: 10.1109/Tbme.2015.2498199.

59.

Kahl

Stapel

Correll

. Psychological and psychopharmacological interventions in psychocardiology. Front Psychiatr 2022; 13: 831359. DOI: 10.3389/fpsyt.2022.831359.

Screening major depressive disorder in patients with obstructive sleep apnea using single-lead ECG recording during sleep

Abstract

Keywords

Introduction

Methods

Participants and data collection

Data analysis

Segmentation

Feature extraction

Feature selection

Classification

Results

Discussion

Current state of art for ECG analysis for detecting OSA and MDD

Our work in relation to state-of-the-art

Conclusion

Footnotes

Author contributions

Declaration of conflicting interests

Funding

Ethical statement

ORCID iD

Data availability statement

References