Classifying depression using blood biomarkers: A large population study

Abstract

Background:

Depression is a common mood disorder characterized by persistent low mood or lack of interest in activities. People with other chronic medical conditions such as obesity and diabetes are at greater risk of depression. Diagnosing depression can be a challenge for primary care providers and others who lack specialized training for these disorders and have insufficient time for in-depth clinical evaluation. We aimed to create a more objective low-cost diagnostic tool based on patients’ characteristics and blood biomarkers.

Methods:

Blood biomarker results were obtained from the National Health and Nutrition Examination Survey (NHANES, 2007–2016). A prediction model utilizing random forest (RF) in NHANES (2007–2014) to identify depression was derived and validated internally using out-of-bag technique. Afterwards, the model was validated externally using a validation dataset (NHANES, 2015–2016). We performed four subgroup comparisons (full dataset, overweight and obesity dataset (BMI≥25), diabetes dataset, and metabolic syndrome dataset) then selected features using backward feature selection from RF.

Results:

Family income, Gamma-glutamyl transferase (GGT), glucose, Triglyceride, red cell distribution width (RDW), creatinine, Basophils count or percent, Eosinophils count or percent, and Bilirubin were the most important features from four models. In the training set, AUC from full, overweight and obesity, diabetes, and metabolic syndrome datasets were 0.83, 0.80, 0.82, and 0.82, respectively. In the validation set, AUC were 0.69, 0.63, 0.66, and 0.64, respectively.

Conclusion:

Results of routine blood laboratory tests had good predictive value for distinguishing depression cases from control groups not only in the general population, but also individuals with metabolism-related chronic diseases.

1. Introduction

Depression is a common mental illness that adversely impacts how we think, feel, and act. Depression can cause various emotional and physical problems, as well as potentially reduce a person’s ability to work. According to a report by the World Health Organization (WHO), there are a total of 332 million people living with depression globally (World Health Organization 2017, 2017, 2017). Further, Hasin et al. (2018) reported that the lifetime prevalence of depression was 20.6% based on findings from the National Epidemiologic Survey on Alcohol and Related Conditions III (NESARC-III) (Hasin et al., 2018). More seriously, people with other chronic medical conditions such as obesity and diabetes are at greater risk of depression (Anderson et al., 2001).

Obesity is a growing health concern globally. Based on a report from Centers for Disease Control and Prevention (CDC), the prevalence of obesity increased from 30.5% in 1999–2000 to 42.4% in 2017–2018 in the United State (US) (Hales et al., 2020). Additionally, prospective studies have documented that depression was strongly associated with elevated likelihood of becoming obese (Simon et al., 2008; Tashakori et al., 2016). This is potentially due to people who are overweight were more likely to diet and experience worse physical health, increasing vulnerability to mental disorders such as depression (Roberts et al., 2000). A meta-analysis involving fifteen studies encompassing 58,745 individuals reported that obesity was associated with increased odds of depression (Luppino et al., 2010).

Similarly, depressive symptoms adversely affect up to one-third of people with diabetes, and depression was significantly associated with variety of diabetes-related complications (De Groot et al., 2001; Holt et al., 2014). One meta-analysis including 13 studies with 6916 cases observed that depression was associated with baseline diabetes (Mezuk et al., 2008).

In fact, obesity and diabetes are all related to metabolic disorders. Additionally, metabolic syndrome, which is a cluster of metabolic disorders, is associated with elevated risk of obesity and diabetes (Eckel et al., 2005). Previous studies documented that people with metabolic syndrome had higher rates of depression than people without metabolic syndrome (Dunbar et al., 2008; Moazzami et al., 2019; Rhee et al., 2014; Viinamäki et al., 2009). In a meta-regression study that include 17 studies, totaling 31,880 individuals, reported that depression was associated with metabolic syndrome (Gheshlagh et al., 2016).

Literature has documented that individuals with chronic diseases (e.g. diabetes, metabolic syndrome) experience greater mental disorders and emotional stress which can contribute to elevated risk of other chronic illnesses, due to lack of physical exercise, poor diet, and non-adherence to prescribed medication (Sartorius, 2018). Therefore, early recognition of depression in high-risk groups such as obese, diabetics, and those with metabolic syndrome can help alleviate patients from both adverse physical and psychological health.

Diagnosing depression can be a challenge for primary care providers and other clinicians that lack specialized training or have insufficient time for in-depth clinical evaluation. Individuals that had a traumatic childhood, overly paranoid, or introvert often underestimate symptoms, which adds to the challenge. Additionally, a systematic review by Gulliver et al. reported that young people with mental illness would not like to seek help because of they feel humiliated and embarrassed about poor mental health literacy (Gulliver et al., 2010). Therefore, it is necessary to build an easy-to-use diagnostic tool to detect depression. Recently, some discriminative classification methods have been used to distinguish patients with different mental illnesses through individual brain images and DNA methylation (Bartlett et al., 2019; Schnyer et al., 2017). However, the cost to collect brain image data or DNA methylation data is very expensive. In our study, we used blood biomarkers and demographic information that can be obtained in most clinical situations to build four random forest classifiers: (1) participants with depression from participants without depression; (2) high BMI participants with depression from high BMI participants without depression; (3) diabetic participants with depression from diabetic participants without depression; and (4) metabolic syndrome participants with depression from metabolic syndrome participants without depression. The first RF classifier will be used for the general population, and the other RF classifiers in a specified population. Blood biomarkers are usually available in routine physical examinations, which may reflect the severity of chronic diseases such as diabetes, and may also be involved in the mechanism of depression (Lyons and Basu, 2012). Therefore, this approach will contribute towards building an objective and easy-to-use diagnostic tool to detect depression in the general population and among people with chronic diseases, facilitating early prevention and intervention for depression.

2. Methods

2.1. Data sources

The National Health and Nutrition Examination Survey (NHANES) is a nationwide survey of adults and children in the US. NHANES aims to monitor the health and nutritional status of people within the U.S. (National Center for Health Statistics, 2008). Complex multi-stage probability sampling design method was used to select participants in NHANES. Since 1999, NHANES has conducted data collection in a two-year cycle. In each two-year cycle, approximately thirty counties were selected in the U.S. In selected areas of these counties, around 5000 participants were chosen and provided written informed consent prior to inclusion in the study. The overall interview response rate is approximately 80%.

The present study used data from NHANES 2007–2016 that included completed blood biomarker tests. We assigned NHANES 2007–2014 as training dataset and remaining as the validation dataset. In the training dataset, there were 7702 participants >20 years of age, where 522 participants with depression and 7180 participants without depression defined as healthy control (HC). In the validation dataset, there were 1752 participants >20 years of age, where 117 participants with depression and 1635 HC participants.

2.2. Outcome measure

An interview-based survey was conducted by trained interviewer. During the interview, participants were given the Patient Health Questionnaire (PHQ-9) to measure depression symptoms. The PHQ-9 is a 9-item depression screening (e.g. little interest, feeling down, trouble falling sleep, feeling tired, overeating, feeling bad about yourself, trouble concentrating on things, moving or speaking so slowly, and thoughts of hurting yourself) that ask about the frequency of depression symptoms in the past 2 weeks (Kroenke et al., 2001). Each question of PHQ-9 is scored from 0 to 3, and the PHQ-9 total score ranges from 0 to 27. Since the PHQ-9 score may not be easily comprehended and utilized by primary care providers and other clinical staff that lack training on this instrument, we defined depression as PHQ-9 ≥ 10, which was reported to have a sensitivity of 88% and specificity of 88% for major depression recommended by the National Quality Forum (Kroenke et al., 2001).

2.3. Predictors

There were a total of 48 features including demographic, biometric, and laboratory data which can be easily obtained in most clinical settings. These features included gender, age, ethnicity, family income to poverty, body mass index (BMI), waist, systolic blood pressure, diastolic blood pressure, albumin, alanine aminotransferase (ALT), aspartate aminotransferase (AST), alkaline phosphatase (ALP), blood urea nitrogen, calcium, cholesterol, bicarbonate, creatinine, gamma glutamyl transferase (GGT), glucose, iron, lactate dehydrogenase, phosphorus, total bilirubin, total protein, triglycerides, uric acid, sodium, potassium, chloride, osmolality, globulin, LDL, white blood cell count (WBC), lymphocyte percent, lymphocyte count, monocyte percent, monocyte count, neutrophils percent, neutrophils count, eosinophils percent, eosinophils count, basophils percent, basophils count, red blood cell count (RBC), hemoglobin, red cell distribution width (RDW), platelet count, and glycohemoglobin.

2.4. Assessment of overweight and obesity

We defined overweight and obese as BMI ≥25. There were 5396 overweight and obese participants in NHANES 2007–2014, where 398 participants classified as having depression and 4998 HC participants. In NHANES 2015–2016, there were 1278 overweight and obese participants, with 90 participants classified as having depression and 1188 HC participants.

2.5. Assessment of diabetes

Diabetes was determined based on participants self-report from interview and blood draw from an examination. According to the definition, diabetes is defined if at least one of the following criteria were met: (1) a fasting glucose concentration ≥ 126 mg/dL; (2) Glycosylated hemoglobin (HbA1c) ≥ 6.5%; (3) participant answered yes to the question “have you ever been told by a doctor that you have diabetes”; (4) participant answered yes to the question “are you now taking diabetic pills to lower your blood sugar”; and (5) participant answered yes to the question “are you taking insulin now”. There were 1336 diabetes participants in NHANES 2007–2014, where 131 participants were classified as having depression and 1205 HC participants. In NHANES 2015–2016, there were 380 diabetes participants, where 35 had depression and 345 were HC.

2.6. Assessment of metabolic syndrome

The National Cholesterol Education Program (NCEP) Adult Treatment Group III developed the definition of metabolic syndrome in 2001. According to the definition, metabolic syndrome is defined if at least three of the following criteria were met: (1) diastolic blood pressure ≥ 85 mmHg or Systolic blood pressure ≥ 130 mmHg; (2) triglyceride ≥ 150 mg/dL; (3) fasting high-density lipoprotein (HDL) cholesterol < 40 mg/dL for male or HDL cholesterol < 50 mg/dL for female; (4) fasting blood sugar ≥ 100 mg/dL; and (5) waist circumference > 102 cm for male or waist circumference > 88 cm for female (Huang, 2009). There were 2905 metabolic syndrome participants in NHANES 2007–2014, where 272 had depression and 2633 were HC. In NHANES 2015–2016, there were 725 metabolic syndrome participants, where 52 had depression and 673 were HC.

2.7. Synthetic Minority Over-sampling technique

Due to severely skewed class distribution, most machine learning algorithms will have poor performance and must be modified to avoid only predicting majority class in all cases. Synthetic Minority Oversampling Technique (SMOTE) is a popular machine learning method to solve imbalance problems, and was used in our study (Chawla et al., 2002). SMOTE performs by connecting the points of minority groups with line segments, and then placing artificial points on these lines. There are several advantages for using SMOTE to rebalance the dataset, including 1) generating synthetic sample instead of copying sample to avoid overfitting problems; 2) easy to implement and explanation; and 3) no information loss (Chawla et al., 2002).

2.8. Feature selection and prediction

Random Forest (RF) is a classifier method that can be used to predict class membership or group means. It aggregates many smaller tree models grown using the algorithm Classification and Regression Trees (CART) (Breiman, 2001). CART grows a decision tree whose first decision node attempts to split all individuals into two groups based on a cut-point of a predictor. The threshold and predictor are chosen by exhaustive search to find the pair that maximizes the decrease of the Gini Impurity Index (Menze et al., 2009). To decrease the variance of the predictions and the inherent overfitting of one CART model, numerous trees are grown over many bootstrap samples of the original dataset. A bootstrap sample is a sample of the same size as the original data set but drawn with replacement. The excluded data, roughly 36% of the original sample, is called “out-of-bag” (OOB) observations. These observations are used to appraise predictions of class membership on data not used to grow the tree. The OOB process is very similar to cross-validation. RF has another important characteristic, where adding a noise term to each feature to exam changes in Gini index or overall error to measure the importance of each feature. The greater the Gini index change or error rate caused by this feature, the more important this feature is for the prediction model.

Stepwise backward variable selection was used in the RF to maximize Area Under the Curve (AUC) (Lin et al., 2021). The stepwise backward variable selection algorithm utilized in this study is as follows. Suppose we have N features:

1. Start with a full features model with all predictors (N features) to get the AUC.

2. Least important feature is removed from current model, and new AUC is calculated from RF.

3. Repeat step 2 until all features are removed from the model.

4. Select a final model from N models which has the best AUC. This final model’s features are used in analyses.

RF was used to distinguish between HC and participants with depression based on the best features identified from the backward variable selection processes. Our results report accuracy, sensitivity, specificity, and AUC. The rankings of the importance of the variables were reported as well, which were measured by the mean decrease in the Gini impurity index.

2.9. Validation

For internal validation, RF itself has an OOB method which uses bootstrap sampling to build trees, and OOB samples to test errors.

For external validation, NHANES 2015–2016 was used as validation to validate the model we built from NHANES 2007–2014. The external validation AUC, accuracy, sensitivity, and specificity were reported.

3. Results

Table 1 summarizes the demographic characteristics for NHANES 2007–2016 participants that completed blood biomarker tests. The average age for depression in NHANES 2007–2014 and NHANES 2015–2016 were 49.11 (SD 15.00) and 48.94 (SD 16.53), respectively. There were no significant differences between NHANES 2007–2014 depression group and NHANES 2015–2016 depression group by age, gender, race/ethnicity, education, BMI, and family income to poverty.

Table 1.

Demographic of NHANES 2007–2016 who had completed blood biomarker test.

Demographic
	NHANES 2007–2014		NHANES 2015–2016		p-value compare between depressiona
Variables	HC n = 7180	Depression n = 522	HC n = 1635	Depression n = 117
Age	49.36 (17.79)	49.11 (15.00)	49.84 (17.26)	48.94 (16.53)	0.913
Gender
Male	3642 (50.72%)	192 (36.78%)	809 (49.48%)	48 (41.03%)	0.453
Female	3538 (49.28%)	330 (63.22%)	826 (50.52%)	69 (58.97%)
Ethnicity
Hispanic	1729 (24.08%)	153 (29.31%)	495 (30.28%)	40 (34.19%)	0.253
Non-Hispanic Asian	426 (5.93%)	7 (1.34%)	181 (11.07%)	4 (3.42%)
Non-Hispanic Black	1357 (18.90%)	101 (19.35%)	308 (18.84%)	25 (21.37%)
Non-Hispanic White	3435 (47.84%)	246 (47.13%)	594 (36.33%)	44 (37.61%)
Non-Hispanic Other	233 (3.25%)	15 (2.87%)	57 (3.49%)	4 (3.42%)
Education
Less than 12th grade	1675 (23.33%)	199 (38.12%)	329 (20.12%)	42 (35.90%)	0.652
HS diploma or GED	1612 (22.45%)	127 (24.33%)	366 (22.39%)	33 (28.21%)
2 years college, AA degree	2080 (28.97%)	140 (26.82%)	491 (30.03%)	33 (28.21%)
4 years college or above	1807 (25.17%)	56 (10.73%)	449 (27.46%)	9 (7.69%)
Body mass index	28.70 (6.46)	30.70 (7.89)	29.40 (6.99)	30.93 (7.80)	0.775
Family income to poverty	2.59 (1.64)	1.63 (1.33)	2.53 (1.60)	1.48 (1.25)	0.291

^aSignificant difference between groups determined by Chi-square test (categorical variables) or t-test (continuous variables).

Fig. 1 presents the correlation by each biometric and laboratory data. We observed LDL was strongly positively correlated with cholesterol (0.921), and lymphocyte percent was negatively correlated with neutrophil percent (−0.933).

Fig. 1.

Plots of correlation of biometric and laboratory data.

3.1. HC vs depression – completed blood biomarker tests

For the dataset that included all participants, we found ratio of income to poverty (RIP), gamma-glutamyl transferase (GGT), glucose, Triglyceride, and red blood cell distribution width (RDW) as the five most important features (Fig. 2a). The AUC, accuracy, sensitivity, and specificity from random forest were 0.830, 0.758, 0.821, and 0.697, respectively (Fig. 3a). When we consecutively removed the top 3 features - RIP, GGT, and glucose, the AUC dropped to 0.797, 0.765, and 0.752, respectively (Fig S1). With bootstrap method, the AUC without top 3 important features was significantly different compared to initial AUC (p-value = <0.0001).

Fig. 2.

Discriminating Depression from Healthy Controls - Feature Importance. a. all NHANES 2007–2014 participants who had completed blood biomarker test; b. High BMI (BMI≥25) NHANES 2007–2014 participants who had completed blood biomarker test; c. Diabetes NHANES 2007–2014 participants who had completed blood biomarker test; and d. Metabolic Syndrome NHANES 2007–2014 participants who had completed blood biomarker test (Red represents more important and green less important). (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.)

Fig. 3.

Discriminating Depression from Healthy Controls- AUC, Accuracy, Sensitivity, and Specificity. a. all NHANES 2007–2014 participants who had completed blood biomarker test; b. High BMI (BMI≥25) NHANES 2007–2014 participants who had completed blood biomarker test; c. Diabetes NHANES 2007–2014 participants who had completed blood biomarker test; and d. Metabolic Syndrome NHANES 2007–2014 participants who had completed blood biomarker test (SE: Sensitivity. SP: Specificity. Acc: Accuracy. AUC: Area Under the Curve).

NHANES 2015–2016 with all participants was applied to validate the performance of our model, AUC, accuracy, sensitivity, and specificity from the validation dataset, where results were 0.691, 0.613, 0.726, and 0.605, respectively (Fig. 3a).

Since the frequency of abnormal blood test may increase with age, we grouped age as: 20–45, >45–64, and >64. We observed that this age grouping was not as important (Fig S5). This is potentially due to loss of information when categorizing continuous variables into several categories (Irwin and McClelland, 2003).

3.2. HC vs depression – with overweight and obesity (BMI≥25)

For the dataset with participants whose BMI were ≥25, we found GGT, RIP, Creatinine, RDW, and glucose as the five most important features (Fig. 2b). The AUC, accuracy, sensitivity, and specificity from random forest were 0.799, 0.730, 0.787, and 0.671, respectively (Fig. 3b). When we consecutively removed the top 3 features - GGT, RIP, and Creatinine, the AUC dropped to 0.778, 0.763, and 0.751, respectively (Fig S2). With bootstrap method, the AUC without top 3 important features was significantly different compared to initial AUC (p-value = <0.0001).

We applied NHANES 2015–2016 with participants whose BMI were ≥25 to validate the performance of our model, the AUC, accuracy, sensitivity, and specificity from the validation dataset, where the results were 0.634, 0.509, 0.744, and 0.492, respectively (Fig. 3b).

3.3. HC vs depression – with diabetes

For the dataset with participants who had diabetes, we found GGT, Eosinophils count, RIP, Basophils percent, and Eosinophils percent as the five most important features (Fig. 2c). The AUC, accuracy, sensitivity, and specificity from random forest were 0.818, 0.740, 0.674, and 0.805, respectively (Fig. 3c). When we consecutively removed the top 3 features - GGT, Eosinophils count, and RIP, the AUC dropped to 0.803, 0.787, and 0.766, respectively (Fig S3). With bootstrap method, the AUC without top 3 important features were significantly different compared to initial AUC (p-value < 0.0001).

We applied NHANES 2015–2016 with participants who had diabetes to validate the performance of our model, the AUC, accuracy, sensitivity, and specificity from validation dataset, where the results were 0.658, 0.674, 0.571, and 0.684, respectively (Fig. 3c).

3.4. HC vs depression – with metabolic syndrome

For the dataset with participants who were defined as having metabolic syndrome, we found RIP, GGT, Eosinophils count, Bilirubin, and Basophils count as the five most important features (Fig. 2d). The AUC, accuracy, sensitivity, and specificity from random forest were 0.816, 0.744, 0.707, and 0.782, respectively (Fig. 3d). When we consecutively removed the top 3 features - RIP, GGT, and Eosinophils, the AUC dropped to 0.797, 0.773, and 0.753, respectively (Fig S4). With bootstrap method, the AUC without top 3 important features was significantly different compared to initial AUC (p-value = <0.0001).

We applied NHANES 2015–2016 with participants who were defined as having metabolic syndrome to validate the performance of our model, AUC, accuracy, sensitivity, and specificity from validation dataset, where the results were 0.644, 0.497, 0.788, and 0.474, respectively (Fig. 3d).

4. Discussion

In the present study, the prevalence of depression was 6.8% (639/9454) which is similar to the national average of 7.5% for the 12-month prevalence of major depressive disorder (MDD) reported by National Comorbidity Survey (NCS) (Avenevoli et al., 2015). Additionally, based on the validation AUC from the four RF classifiers, we observed that the RF classifier for either general population or specified population have similar predictive for depression. The objective diagnostic criteria for mental health, especially depression was clinically significant. Jihoon et al. (2019) identified several biological biomarkers using deep learning methods to reliably predict depression, a proof of concept research to prove that objective standards are feasible (Oh et al., 2019). The present study identified demographic and blood biomarker features from routine hospital examinations that can be utilized to predict depression from control participants with good accuracy. Random forest is an accurate and effective classification method to combine multiple features to create a classifier. This classifier will allow for predicting depression in patients not previously diagnosed using selected features. These patients tend to avoid treatment due to stigma and avoidance of symptoms, therefore, benefiting from appropriate diagnosis and treatment. Additionally, we not only built classifiers to separate depression between all participants, but also built classifiers to separate patients with depression from patients without depression in obese patients, patients with diabetes and patients with metabolic syndrome. We identified nine important features from the above models relevant to depression, which included ratio of income to poverty, GGT, glucose, Triglyceride, RDW, creatinine, Basophils number or percent, Eosinophils number or percent, and Bilirubin.

4.1. Ratio of income to poverty

Ratio of income to poverty is one of the most important features among the models. Numerous studies have reported that having a low income is associated with greater likelihood of developing depression compared to households of higher income (Burns et al., 2017; Patel et al., 2018; Pickett and Wilkinson, 2015; Ribeiro et al., 2017; Wilkinson and Pickett, 2017). In a study among 1,743,948 individuals in Taiwan observed that low income women aged 18 to 44 and men aged 45–64 had the highest incidence of depression (Lee et al., 2016). This was potentially due to residing in a low-income household is associated with increased chronic stress as a result of greater barriers to accessing needed resources (e.g. medical care, grocery) leading to elevated risk of mental illnesses (Ahern and Galea, 2006). Those of lower income are more likely to reside in an area with greater exposure to violent crime, inadequate housing, and other social environmental stressors linked to elevated risk of depressive symptoms (Williams, 2018). An epidemiological study reported that the combined increased in income and reduction in stressful experiences may reduce clinically significant depressive symptoms (Barrett et al., 2021). Moreover, a U.S. study by Swift et al. (2020) reported that financial shocks during an economic recession were associated with greater depressive symptoms, further strengthening the importance of inclusion of the ratio of income to poverty feature in our model (Swift et al., 2020).

4.2. Gamma-glutamyl transferase (GGT)

GGT was among the five most important features in all models. GGT is primarily observed in liver, kidney, and pancreatic cells. GGT is used to diagnose and monitor hepatobiliary disease and is currently the most sensitive enzyme indicator for diseases of the liver (Burtis and Ashwood, 1994). Hideki et al. (2020) documented that over a 5-year period during annual medical check-ups, the overall incidence in fat changes were greater among individuals with elevated GGT levels compared to individuals with normal GGT (Fujii et al., 2020). The study also reported that frequent elevated GGT levels is potentially a good predictor of fatty liver changes (Fujii et al., 2020). However, several studies reported that patients with non-alcoholic fatty liver disease (NAFLD) had a higher prevalence of depression (Castro et al., 1996; Liebman et al., 1983; Weinstein et al., 2011; Youssef et al., 2013). In a U.S. study among 10, 484 individuals observed that the prevalence of depression and functional impairment due to depression was higher in people with NAFLD than those without (Kim et al., 2019). This is potentially the result of patients with chronic illnesses may also experience pain, disability, or social isolation leading to risk of depression (Simon, 2001). Additionally, Huang et al. (2017) reported that GGT was higher in NAFLD patients with depression than those without depression, further highlighting the importance of the GGT feature (Huang et al., 2017).

4.3. Glucose

Glucose is another important feature in models that included all participants and participants with high BMI. Glucose is the primary sugar found in blood which derives from food consumed and a main energy source (Schulman et al., 1997). Among people with depression, a common characteristic of the disorder is loss of appetite, which can adversely impact blood sugar levels. Recent studies reported that patients with depression had higher fasting glucose, and these glucose levels were significantly correlated with depression score (De la Roca-Chiapas et al., 2013; Kahn et al., 2011). In a prediction study by Song et al. (2018), observed that glucose was the second important feature for separating healthy controls and depression groups, consistent with our findings (Song et al., 2018). Among patients with depression, this is potentially due to brain regions including the amygdala, prefrontal cortex, and hippocampus have defects in glucose metabolism that are involved in the emotional process (Drevets et al., 2002b; Song et al., 2018). In addition, Drevets et al. (2002) reported that long-term use of antidepressant treatment may reduce glucose metabolism in the amygdala and ventral anterior cingulate cortex, indicating a continuous positive treatment response (Drevets et al., 2002a).

4.4. Triglycerides

Triglycerides are a type of fat in the blood stored in fat cells. In humans, calories are converted that are not needed into triglycerides, and the hormone releases triglycerides between meals to provide energy. In the present study, triglycerides were one of the top five important features in models that included all participants, and participants with high BMI. Previous studies have reported that triglycerides were highly correlated with depression, where successful management of triglyceride levels may contribute to reducing depression (Almeida et al., 2007; Sutin et al., 2010; Toker et al., 2008). In a study by Cepeda et al. (2020), reported that high level of triglycerides were associated with increased risk of depression (Cepeda et al., 2020). This may be due to interleukin-2 lowers cholesterol, increases triglycerides, and inhibits melatonin secretion, thereby decreasing in brain serotonin and causing depression and suicidal tendencies (Sheikh et al., 2004). Moreover, a UK study with 367,703 unrelated middle-aged participants observed that interleukin-6, C-reactive protein, and triglycerides might be causally linked with depression, which could be used for treatment and prevention of depression (Khandaker et al., 2020).

4.5. Red blood cell distribution width (RDW)

Red blood cells carry oxygen from your lungs to other parts of your body. Any phenomenon in which the width or volume of red blood cells is outside the normal range indicates that there may be a problem with the body function, which may affect the flow of oxygen into various parts of the body. In this study, RDW was among the top five important features for the models including all participants and participants with high BMI. Several studies have documented that higher hematological inflammatory markers including white blood cell (WBC) and RDW are associated with elevated depression and anxiety (Gundogmus et al., 2019; May et al., 2013; Shafiee et al., 2017). Demircan et al. (2015) reported that the RDW level in patients with depression was significantly higher than the control group. However, after treatment with selective serotonin reuptake inhibitor (SSRI), the significant differences between RDW level in patients with depression and control group was eliminated (Demircan et al., 2016). However, another study comparing individuals with depression to those with no mental illnesses reported no significant difference for RDW level between groups (Maes et al., 1996). The differences between this study and ours could be the result of small sample size in their study, resulting in low statistical power.

4.6. Creatinine

Creatinine is a waste product that comes from normal wear and tear of human muscles. A high serum creatinine level indicates potential problem with kidney function. In our models, we observed that creatinine was an important feature for participants with high BMI. A study by Zheng et al. (2012) identified key metabolites to separate depression and healthy controls including amino acid, lipid/protein complexes, lipid metabolism-related molecules, and energy metabolism-related molecules (creatine, creatinine) (Zheng et al., 2012). Further, Song et al. (2018) reported that creatinine was a predictive feature to separate depression and no depression groups, consistent with the present study (Song et al., 2018). This is potentially the result of energy deficiencies of creatinine association with depression. Kious et al. (2019) observed that disruption in brain energy production, storage, and utilization are related to the development and maintenance of depression, and creatine which was converted to creatinine by non-enzymatic dehydration has the potential to improve these disruption in some patients (Kious et al., 2019).

4.7. Basophils count or percent

Basophils are a type of white blood cells. High levels of basophils or low levels of basophils would both cause diseases such as myeloproliferative disorders, allergic reactions, inflammatory disorders, and depression (Baek et al., 2016; Chapuy et al., 2014; Hoyle et al., 1989; Siracusa et al., 2013). In this study, Basophils count and basophils percent were two important features in models for patients with diabetes, and patients with metabolic syndrome. A prior study among 709 participants observed that basophils was negatively associated with depression (Baek et al., 2016). This is potentially due to less histamine being released from fewer basophils, where the histamine system has an important role in regulating alertness, anxiety, memory, and mood (Dere et al., 2010). As a result of limited publications to date, it is difficult to conclude cause and effect relationship between basophils and depression.

4.8. Eosinophils count or percent

Similar to basophils, eosinophils are another type of white blood cell. In our study, eosinophils count and eosinophils percent were two important features in models among patients with diabetes, and patients with metabolic syndrome. A previous study by Steel et al. (2010) observed that high levels of pain, fatigue, and depression were associated with elevated eosinophil percentages (Steel et al., 2010). Further, Taft et al. (2019) observed that depressive symptoms would contribute greatly to eosinophilic esophagitis patients failing to comply with medicine, and associated with persistent upper abdominal pain and greater sleep disturbance (Taft et al., 2019). Additionally, Singh et al. (2020) reported that eosinophilia of the rectal sigmoid colon is associated with a higher anxiety score, while eosinophil density is associated with a depression score for irritable bowel syndrome patients (Singh et al., 2020).

4.9. Bilirubin

Bilirubin is formed after red blood cells are broken down and passes through the liver, gallbladder, and digestive tract before excretion. Bilirubin is an effective antioxidant and the final product of heme cleavage pathway, which can be catalyzed through heme oxygenase and biliverdin reductase (Stocker et al., 1987). In the present study, bilirubin was ranked as the fourth important feature in the model of patients with metabolic syndrome. In a Hong Kong study among 635 patients reported that high bilirubin level is associated with post-stroke depression. This study also observed that the level of bilirubin was an important biological indicator for risk of depression in the patients with ischemic stroke (Tang et al., 2013). This is potentially due to psychological disorders might cause bilirubin oxidation in vivo, and further increased concentration of bilirubin oxidation metabolites (Yamaguchi et al., 2002). Miyaoka et al. (2004) documented that the concentration of bilirubin oxidation metabolites in patients with depression was significantly higher than control groups (Miyaoka et al., 2005). Additionally, a study with 9 depression and 7 with no depression observed that nocturnal bilirubin levels were lower in depression than non-depressed groups (Oren et al., 2002). However, this inconsistent result may be due to the small sample size.

4.10. Limitation

We must note several limitations. First, since NHANES data is collected from thirty U.S. counties every two year cycle, our findings may not be generalizable. However, the RF model can be used to estimate the prevalence of depression in areas where individual mental health surveys have not been conducted. Second, the number of participants with depression is much smaller than participants with no depression, which could cause the model to predict the majority class in all situations. To address this limitation, rebalance method called SMOTE was applied in our analysis. Third, the health status at the time of blood collection may affect the results of biomarkers, therefore, it cannot be ruled out that certain disease information not obtained in the investigation affects the biomarkers. Fourth, by using the PHQ-9 defined depression, we may fail to detect the severity of depression. However, we compared the relationship between major biomarkers and PHQ-9 score or PHQ-9 defined depression. We observed that the associated direction between PHQ-9 score or PHQ-9 defined depression and biomarker were the same (table S1). Finally, data on substance-related disorders were not accounted for in this study.

5. Conclusion

The present study observed that results from blood biomarkers can be used to separate patients with depression from those with no depression, and separate patients with depression and chronic diseases (obesity, and metabolic syndrome) from those with no depression, having good accuracy using RF. Our findings revealed an association between depression and blood biomarker test, where common laboratory markers can facilitate clinical diagnosis of depression. Additionally, blood biomarkers not used in this study can be obtained from laboratory tests to potentially enhance prediction accuracy. Further research is needed to verify the use of these biomarkers in the diagnosis of depression.

Appendix A. Supplementary data

Supplementary data to this article can be found online at https://doi.org/10.1016/j.jpsychires.2021.05.070.