Serum Cytokines as Biomarkers for Comorbid Anxiety in Postpartum Depression: A Machine Learning Approach

Abstract

Background:

This study aimed to investigate the serum levels of interleukin 2, interleukin 6 (IL-6), interleukin 10, and tumor necrosis factor-alpha in patients with postpartum depression (PPD) and to explore their potential as biomarkers for PPD and comorbid anxiety using machine learning techniques.

Methods:

Serum samples were collected from 53 patients diagnosed with PPD and 35 healthy controls. Cytokine levels were measured using a flow cytometer analyzer. Machine learning models, including Multinomial Logistic Regression, Decision Trees, Random Forest, and Support Vector Machines (SVMs), were developed to predict PPD and comorbid anxiety based on cytokine levels.

Results:

Patients with PPD exhibited significantly elevated serum levels of IL-6 compared to the control group. A positive correlation was found between psychological anxiety scores and IL-6 levels (r = 0.483, P < .001). Machine learning models, particularly the Random Forest and SVMs, demonstrated high accuracy in predicting PPD and comorbid anxiety, with IL-6 being identified as a key predictor.

Conclusion:

The activation of serum cytokines is evident in PPD patients, with IL-6 potentially serving as an auxiliary biomarker for the diagnosis of PPD and comorbid anxiety. The incorporation of machine learning techniques has enhanced the understanding of the complex relationships between cytokines and PPD, with IL-6 levels showing a correlation to the severity of clinical symptoms.

Main Points

• Elevated interleukin 6 (IL-6) levels in postpartum depression (PPD) patients, with a more pronounced increase in those with comorbid anxiety.

• Positive correlations between IL-6 and Hamilton Anxiety Rating Scale scores, suggesting a link between IL-6 and anxiety in PPD.

• The development of predictive models using machine learning techniques, highlighting the potential of IL-6 as a biomarker for PPD comorbid with anxiety.

Introduction

Postpartum depression (PPD) is a severe form of depression that can manifest within 6 weeks after childbirth,¹ affecting 10%-20% of women.² It is one of the most common complications in the postnatal period and is a significant cause of maternal mortality.^3,4 The suffering of PPD also impairs family harmony and imposes a considerable socioeconomic burden.⁵ The symptoms of PPD include mood lability, anxiety, irritability, feeling overwhelmed, and sleep disturbance.^6,7 Fallah-Hassani et al⁸ reported prevalence rates of comorbid anxiety symptoms and depressive symptoms were among 3.5%-9.2% in the first 24 weeks postpartum.^9,10 A study in Egypt showed that 21.2% of 500 postpartum females suffered from comorbid anxiety and depression, and the rate of patients suffering from anxiety alone is higher than that with PPD alone.¹¹ In addition, comorbid symptoms of depression and anxiety are frequent and seem to be a sign of severity.¹² Patients suffering from depression accompanied by anxiety showed significantly lower treatment response rates than their counterparts who do not have anxiety.¹³ Thus, the ability to predict anxiety levels in patients with PPD serves as a vital component in the management of the illness and the prediction of its outcomes.

Emerging evidence suggests that dysregulations in immune-inflammatory pathways play a pivotal role in the pathophysiology of mood disorders.¹⁴ There are several mechanisms characterized to elucidate the bidirectional interplay between mood disorders and immune dysfunction.¹⁵ Emotional symptoms frequently accompany neuroinflammatory diseases, and the activation of the immune system is known to induce sickness behaviors that manifest during depressive episodes.¹⁶ Studies have shown that cytokine fluctuations and interactions in biomarkers of function might be involved in the development of neuropsychiatric disorders or the exacerbation of existing disorders during pregnancy and postpartum.¹⁷ Although there have been numerous studies examining the relationships between PPD and various biomarkers, as well as their mutual interactions about clinical presentation, observed associations have been inconsistent across different studies.¹⁸ They are highlighting the imperative for further investigation.

To further elucidate the impact of cytokines on the pathophysiology of PPD, this study aims to elucidate the role of cytokines in the pathophysiology of PPD by examining alterations in serum inflammatory markers. Concurrently, inflammatory markers in the serum of patients with bipolar disorder were assessed to evaluate the potential heterogeneity in correlations between these markers and the disorder. Additionally, the implications of these alterations on the spectrum of emotional symptoms exhibited by PPD patients were observed.

Material and Methods

Participant Population

This was a single-center, retrospective case-control study of patients with PPD between July 2020 and December 2023. All patients were diagnosed with PPD according to the Diagnostic and Statistical Manual of Mental Disorders, 5th edition (DSM-V) by qualified psychiatrists. The Edinburgh Postnatal Depression Scale (EPDS) was used at 2 weeks postpartum to assess PPD with a cut-off score over or equal to 13. All procedures were reviewed and approved by the ethical committee of the Fourth Affiliated Hospital, Zhejiang University School of Medicine (Approval No: K2024051 Date: 2024.03.06). Written informed consent was obtained after a complete description of the study. The inclusion criteria were: 1) patients of 20-40 years of age; 2) EPDS ≥13. The exclusion criteria were 1) participants with records of bipolar disorders, schizophrenia, or other mental illnesses; 2) participants with major chronic diseases; 3) participants with obstetric and pregnancy complications, including placenta previa, placental abruption, preeclampsia/eclampsia, primary postpartum infection, stillbirth, significant congenital anomalies, or neonates with a birth weight of less than 1500 g; and 4) Patients who had previously taken antipsychotics.

Sample Collection and Measurements

Venipuncture collected fasting peripheral venous blood samples (5 mL) without anticoagulants between 07:00 am and 09:00 am. The serum was obtained by centrifugation at 3000 rpm for 15 minutes. The serum was separated, divided into aliquots, and stored at −80°C in a refrigerator before laboratory assays. Serum levels of IL-2, IL-6, IL-10, and TNF-α were measured with an ACEA NovoCyte flow cytometer analyzer. The assays were performed by the same technician, who was blinded to the samples’ ID and clinical information.

Analysis

All statistical analyses were conducted using the R programming environment (version 4.3.1) within the RStudio integrated development environment (version 2023.12.1 Build 402). The analytical process was divided into several stages to ensure a comprehensive examination of the data. Demographic and obstetric characteristics were described and compared between women with PPD and control subjects. Categorical variables were assessed using the chi-square test, while the independent samples t-test was employed for continuous variables to determine any significant differences in demographic and obstetric profiles. Subsequently, the differences in serum levels of inflammatory cytokines between the PPD group and the control group were examined. To compare the levels of cytokines between the 2 groups (HC and PPD, PPD and PPDA), the normality of the data was first assessed using the Shapiro–Wilk test. The Shapiro–Wilk test was conducted separately for each group to determine if the data followed a normal distribution. Subsequently, the homogeneity of variances between the 2 groups was evaluated using Levene’s test. Based on the results of these tests, the appropriate statistical method for comparison was selected. If both groups followed a normal distribution and had homoscedastic variances (P > .05 for both normality and homoscedasticity tests), the independent t-test was used to compare the means of the 2 groups. If both groups followed a normal distribution but had heteroscedastic variances (P > .05 for normality tests and P < .05 for the homoscedasticity test), Welch’s t-test was used. If at least 1 group did not follow a normal distribution (P < .05 for the normality test), the Mann–Whitney U test was used to compare the medians of the 2 groups. The summary statistics, including mean ± standard error or median with interquartile range, were calculated for each group. All statistical analyses were performed using R software with the necessary packages.

To explore the relationship between PPD severity and inflammatory markers, the Pearson correlation coefficient was utilized. This method was applied to assess the correlations between the Hamilton Depression Rating Scale (HAMD) and Hamilton Anxiety Rating Scale (HAMA) scores in PPD women and the levels of inflammatory cytokines. Additionally, the Pearson correlation coefficient was used to investigate the interrelationships among serum inflammatory cytokine levels. A P-value of less than .05 was considered to indicate statistical significance for all tests performed. For the development of predictive models and associated predictions, several R packages known for their robust statistical capabilities were leveraged. These included the randomForest package (version 4.7.1.1) for constructing classification and regression models, the caret package (version 6.0.94) for machine learning, the e1071 package (version 1.7.14) providing functions for robust statistics, and the nnet package (version 7.3.19) for multinomial logistic regression (MLR). The optimal number of trees in the random forest (RF) model was determined based on minimizing the out-of-bag error, ensuring the model’s predictive accuracy and generalizability.

To assess the performance of the MLR model and RF model in classifying different groups (HC, PPD, PPDA), receiver operating characteristic (ROC) curves were generated for both the training and test datasets. The dataset was split into training and test sets. An MLR model or RF model was fit to the training data. For both training and test datasets, class probabilities were predicted using the trained model. For each class in the training and test sets, the ROC curve was computed using the predicted probabilities and actual class labels. The area under the ROC curve (AUC) was calculated to quantify the model’s performance for each class. The ROC curves for each class were plotted on the same graph. Each class was represented in a different color for clear distinction. The AUC values for each class were displayed on the plot to provide a numerical measure of the model’s performance.

Data are expressed as mean ± S.E.M. if they have a normal distribution. If the data do not show a normal distribution, the median (Q1-Q3) values are given. In all cases, P < .05 was considered statistically significant.

Results

Clinical Characteristics and Chemokine Concentrations in Postpartum Depression Patients and Healthy Controls

A cohort of 53 PPD patients and 35 healthy controls (HCs) was enrolled, comprising 34 first-time, 17 second-time, and 2 third-time mothers. Ages were comparable between PPD patients and HCs, with no significant difference observed (Table 1, P > .05). Postpartum depression patients demonstrated significantly elevated HAMD (HAMD-17) scores (25.96 ± 0.73 vs. 8.06 ± 0.63 for HCs, P < .001), with 37 patients reporting suicidal ideation.

Table 1.

Concentration of Some Serum Chemokines in PPDs

	PPDs (n = 53)	HCs (n = 35)	P
Age (years)	28 [26-31]	29 [26-31]	.611
HAMD	25.96 ± 0.73	8.06 ± 0.63	<.001
IL-2 (pg/mL)	0.980 [0.740-1.030]	0.860 [0.635-1.405]	.742#
TNF-α (pg/mL)	3.650 [1.780-9.020]	3.420 [3.330-9.065]	.911#
IL-10 (pg/mL)	1.810 [1.510-2.160]	1.670 [1.340-1.920]	.181#
IL-6 (pg/mL)	2.670 [1.950-3.850]	1.820 [1.450-2.190]	.001#
Delivers (times)	1.396 ± 0.078

^#Mann–Whitney U test.

Serum concentrations of tumor necrosis factor-alpha (TNF-α), interleukin-2 (IL-2), and interleukin-10 (IL-10) were not significantly different between PPD and HC groups. However, PPD patients exhibited higher concentrations of interleukin-6 (IL-6), with IL-6 levels at 2.670 [1.950-3.850] pg vs. 1.820 [1.450-2.190] (P < .001) for PPD and HC groups, respectively.

Cytokine Concentrations in Postpartum Depression Patients with Comorbid Anxiety

Among the 53 PPD patients, 28 with elevated HAMA scores (HAMA ≥ 21) were identified as having comorbid depression and anxiety (PPDA). The PPDA group showed higher HAMD and HAMA scores and increased IL-6 concentrations compared to the PPD group, while TNF-α, IL-2, and IL-10 levels were not significantly different, suggesting a link between IL-6 concentrations and anxiety in PPD patients (Table 2).

Table 2.

Concentration of Some Serum Chemokines in PPDA and PPD

	PPDA (n =28)	PPD (n = 25)	P
Age (years)	28.00 [26.00-31.00]	29.00 [20.00-30.00]	.844#
HAMA	23.00 [22.00-24.25]	17.00 [14.00-18.00]	<.001#
HAMD	27.821 ± 0.976	23.880 ± 0.961	.006
IL-2 (pg/mL)	0.980 [0.873-1.073]	0.900 [0.660-0.990]	.051#
TNF-α (pg/mL)	2.105 [1.780-9.795]	3.720 [1.800-6.290]	.864#
IL-10 (pg/mL)	1.833 ± 0.101	1.882 ± 0.127	.760
IL-6 (pg/mL)	3.150 [2.115-5.200]	2.270 [1.400-3.130]	.011#
Delivers (times)	1.00 [1.00-1.25]	1.00 [1.00-2.00]	.099#

^#Mann–Whitney U test.

Associations of Cytokine Concentrations with Clinical Characteristics

The analysis commenced with simple linear regression to investigate the relationships between cytokine concentrations and psychological scores. Positive correlations were observed between IL-6 concentrations and both HAMA scores (r = 0.483, P < .001, Figure 1A) and HAMD scores (r = 0.382, P < .001, Figure 1B). While, IL-2 concentrations did not exhibit significant correlations with HAMA (r = −0.138, P = .201, Figure 1C) and HAMD (r = −0.172, P = .109, Figure 1D). the IL-10 concentrations also were not correlated with HAMA scores (r = −0.349, P = .747, Figure 1E) and HAMD scores (r = −0.077, P = .474, Figure 1F). Additionally, a robust correlation was identified between HAMA and HAMD scores (r = 0.903, P < .001, Figure 1G), underscoring a strong association between anxiety and depressive symptoms.

Figure 1.

Correlations between cytokine concentrations and psychological scores. A: Correlation between IL-6 concentrations and HAMA scores. A positive correlation is observed (r = 0.483, P < .001). B: Correlation between IL-6 concentrations and HAMD scores. A significant positive correlation is noted (r = 0.382, P < .001). C: Correlation between IL-2 concentrations and HAMA scores. No significant association is found (r = −0.138, P = .201). D: Correlation between IL-2 concentrations and HAMD scores. No significant association is found (r = −0.172, P = .109). E: Correlation between IL-10 concentrations and HAMA scores. No significant association is detected (r = −0.349, P = .747). F: Lack of correlation between IL-10 concentrations and HAMD scores. No significant association is found (r = −0.077, P = .474). G: Strong correlation between HAMA and HAMD scores. A very strong link is indicated between anxiety and depressive symptoms (r = 0.903, P < .001).

Further analysis using multiple linear regression confirmed significant correlations between the concentrations of IL-2 and IL-6 and the scores on the HAMA and HAMD scales (Table 3). An MLR model incorporating HAMA scores and cytokine concentrations as predictors was then constructed. The model, which included IL-6, IL-2, and HAMD as variables, demonstrated the lowest Akaike Information Criterion, signifying the optimal model fit. Subsequent analysis revealed that incorporating IL-10 into the model containing IL-6 and HAMD did not significantly enhance predictive accuracy. This suggests that the combination of IL-6, IL-2, and HAMD is adequate for the model’s predictive capabilities.

Table 3.

Summary of MLR Analysis Results

		IL-2	TNF-α	IL-10	IL-6	Intercept	F-statistic
HAMD	β	−0.725	−0.043	−0.119	2.347	14.073	F ( 4 ,83) = 4.222, P = .004, Adjusted R2 = 0.129
	t	−1.089	−0.346	−0.092	3.706	5.174
	P	.279	.730	.927	<.001*	<.001*
HAMA	β	−0.419	0.011	−0.038	2.303	9.191	F ( 4 ,83) = 6.741, P < .001, Adjusted R2 = 0.209
	t	−0.855	0.125	−0.039	4.941	4.591
	P	.395	.901	.969	<.001*	<.001*

Highlight the t-values of the MLR (Multiple Linear Regression) analysis results with P values smaller than 0.001

Model Development and Prediction Analysis

A predictive diagnostic model was endeavored to be constructed, leveraging the established correlations between IL-6 and HAMD scores, as well as HAMA ratings. The approach involved a random division of the dataset, allocating 70% for training the models and reserving the remaining 30% for validation purposes. The MLR, Decision Trees (DT), RF, and Support Vector Machines (SVMs) with linear kernels were employed for this task. Notably, the MLR, RF, and SVM models achieved accuracies over 80%, while the DT model only reached 64%, and the confusion matrix of each model was shown in Figure 2A-D. The feature importance for the RF model and DT models was shown in Figure 2E and F, indicating that HAMD and IL-6 concentration were important for these models.

Figure 2.

Model development and prediction analysis based on the concentrations of serum cytokines. A-D): Heatmaps showing the confusion matrices of MLR (A), DT (B), RF (C), and SVM (D) models, respectively. E): Importance of IL-6 in the Random Forest model. F): Importance of IL-6 in the Decision Tree (DT) model. G): Training ROC curves of the MLR model for PPDA and PPD based on the concentrations of IL-6, IL-2, and HAMD scores. H): Test ROC curves of the MLR model for PPDA and PPD based on the concentrations of IL-6, IL-2, and HAMD scores. I): Training ROC curves of the Random Forest (RF) model for PPDA and PPD based on the concentrations of IL-6, IL-2, and HAMD scores. J): Test ROC curves of the Random Forest (RF) model for PPDA and PPD based on the concentrations of IL-6, IL-2, and HAMD scores. K): Training ROC curves of the MLR model for PPDA and PPD based on the concentrations of IL-6, IL-2, and HAMD scores. L): Test ROC curves of the MLR model for PPDA and PPD based on the concentrations of IL-6, IL-2, and HAMD scores.

The subsequent exploration with MLR, RF, and SVMs models, using serum cytokine concentrations and HAMD scores to predict the health condition (HC), PPD, or PPDA (Figures 2G-L) was conducted. The MLR model, when enriched with these variables, yielded the most impressive accuracy rate of over 80%. The MLR model was then refined with these predictors (Table 4), which resulted in precise patient group predictions. Additionally, for the prediction, the MLR model showed good performance, with the AUC reached to 0.8 and 0.9 for PPD and PPDA, respectively (Figures 2G and H). The RF model reached the highest accuracy rate, and the AUC reached 0.87 and 0.94 for PPD and PPDA (Figures 2I and J). While the SVMs model failed to predict the PPD, it showed higher accuracy rate to the PPDA (Figures 2K and L), therefore, the RF model demonstrates robust predictive performance, and the partial dependence of each feature was illustrated in Figures 3A-I.

Table 4.

Coefficients of the Multinomial Logistic Regression model using IL-6, IL-2, and HAMD as Predictors

	Class		IL-2	IL-6	HAMD	Intercept
Group ~ IL6+IL2+HAMD	PPD	Coefficients	−57.842	−4.155	12.545	−100.619
		Standard errors	30.331	8.729	2.553	33.050
	PPDA	Coefficients	−58.830	−3.405	12.837	−109.365
		Standard errors	30.344	8.736	2.550	33.155

Figure 3.

Variable importance in the Random Forest model. A-C): One-way partial dependence plots of HAMD for the classification of HC, PPD and PPDA. D-F): One-way partial dependence plots of IL-6 for the classification of HC, PPD and PPDA. G-I): One-way partial dependence plots of IL-2 for the classification of HC, PPD and PPDA.

Discussion

The present study evaluated serum chemokine concentrations in patients with PPD and HCs, revealing elevated IL-6 levels in PPD patients. Notably, patients with PPD who also exhibited comorbid anxiety (as indicated by HAMA ≥ 21) displayed increased IL-6 levels compared to those with depression alone. Linear regression analyses underscored a positive correlation between IL-6 and HAMA scores, suggesting a link between IL-6 levels and the presence of anxiety in PPD. Furthermore, an RF model and an MLR model were built, both of which have excellent performance in predicting PPDA, indicating that serum IL-6 may be a potential biomarker for PPD comorbid anxiety.

The findings are in dialogue with existing research highlighting the delicate immunological balance in females during pregnancy, which involves an adaptive shift in immune function from an anti-inflammatory to a pro-inflammatory state post-delivery.¹⁹ The psychosocial stressors of pregnancy can precipitate changes that may disrupt psychological equilibrium, with well-documented associations between stress and inflammation.²⁰ Peripheral immunological signals can influence the central nervous system, potentially contributing to the pathophysiology of psychiatric disorders.²¹

The results imply a correlation between IL-6 levels and emotional disorders in PPD, particularly anxiety. Interleukin 6 may affect brain connectivity in regions associated with anxiety, such as the anterior cingulate cortex and amygdala.²² The immune system’s role in emotional fluctuations in PPD requires further investigation, considering the unique physiological and social stressors associated with childbirth.

In line with previous studies, a significant correlation was found between serum IL-6 and depression scores or other scale scores.²³ The lack of association between TNF-α and depressive symptoms in this study²⁴ further emphasizes the complexity of immune system alterations in PPD.

Clinically, the findings suggest that IL-6 could serve as a potential predictor for the severity of anxiety and prognosis in PPD patients. The identification of distinct inflammatory patterns in patients with and without anxiety symptoms may provide valuable clues for the development of biomarkers for PPD. This could aid clinicians in distinguishing between different patient characteristics and guiding more precise treatments.

This study has several limitations, including the small sample size and cross-sectional design, which limit the generalizability of the findings and the ability to establish causality. Future research should employ larger cohorts and longitudinal designs to clarify the temporal relationships between cytokine levels and PPD symptoms. Additionally, the exploration of other immune markers, genetic factors, and environmental influences could provide a more comprehensive understanding of the immunological aspects of PPD.

In conclusion, this study contributes to the understanding of the immunological underpinnings of PPD and its comorbid anxiety. The potential of IL-6 as a biomarker and therapeutic target warrants further investigation. By elucidating the role of cytokines in PPD, progress is made toward developing targeted interventions and improving the lives of those affected by this debilitating condition.

Funding Statement

The authors declared that this study has received no financial support.

Data Availability Statement:

The data that support the findings of this study are available from the corresponding author upon reasonable request.