- Email: [email protected]
Key Factors Associated with Major Depression in a National Sample of Stroke Survivors
ARTICLE IN PRESS
Key Factors Associated with Major Depression in a National Sample of Stroke Survivors Sarah Hirata,
MS,*
Bruce Ovbiagele, MD, MSc, MAS,† Daniela Markovic, Amytis Towfighi, MD§‖
MS,‡
and
Background: Depression, one of the most common complications encountered after stroke, is associated with poorer outcomes. The aim of this study was to determine the factors independently associated with and predictive of poststroke depression (PSD). Methods: We assessed the prevalence of depression (Patient Health Questionnaire [PHQ-8] score >10) among a national sample of adults (≥20 years) with stroke who participated in the National Health and Nutrition Examination Surveys from 2005 to 2010. Logistic regression and random forest models were used to determine the factors associated with and predictive of PSD, after adjusting for sociodemographic and clinical factors. Results: Of the 17,132 individuals surveyed, 546 stroke survivors were screened for depression, and 17% had depression, corresponding to 872,237 stroke survivors with depression in the United States. In the logistic regression model, after adjustment for sociodemographic variables, poverty (poverty index <200% versus ≥200%, odds ratio [OR] 2.61, 95% confidence interval [CI] 1.23-5.53) and 3 or more medical comorbidities (OR 1.59, 95% CI 1.01-2.49) were associated with higher odds of PSD; increasing age was associated with lower odds of PSD (per year OR .95, 95% CI .94-.97). In the random forest model, the 10 most important factors predictive of PSD were younger age, lower education level, higher body mass index, black race, poverty, smoking, female sex, single marital status, lack of cancer history, and previous myocardial infarction (specificity = 70%, sensitivity = 64%). Conclusion: Although numerous factors were predictive of developing PSD, younger age, poverty, and multiple comorbidities were strong and independent factors. More aggressive screening for depression in these individuals may be warranted. Key Words: Depression—stroke— NHANES—poststroke depression—factors. © 2015 National Stroke Association. Published by Elsevier Inc. All rights reserved.
Introduction From the *University of Hawaii John A. Burns School of Medicine, Honolulu, Hawaii; †Department of Neurology, Medical University of South Carolina, Charleston, South Carolina; ‡Department of Biomathematics, University of California, Los Angeles, California; §Department of Neurology, University of Southern California, Los Angeles, California; and ‖Department of Neurology, Rancho Los Amigos National Rehabilitation Center, Downey, California. Received June 10, 2015; revision received November 28, 2015; accepted December 30, 2015. Grant support: American Heart Association National Scientist Development Award (11SDG7590160). Address correspondence to Amytis Towfighi, MD, 7601 E. Imperial Highway, Downey, CA 90242. E-mail: [email protected] 1052-3057/$ - see front matter © 2015 National Stroke Association. Published by Elsevier Inc. All rights reserved. http://dx.doi.org/10.1016/j.jstrokecerebrovasdis.2015.12.042
Poststroke depression (PSD) is one of the most common consequences of stroke, affecting approximately one third of stroke survivors.1-5 Recent studies have indicated an association between PSD and poststroke physical morbidity and mortality; those with PSD have a higher risk for suboptimal recovery, recurrent vascular events, and an overall poorer quality of life.1-4,6 Although the adverse effects of PSD have been well studied, the nature of PSD itself remains unclear. Factors most consistently associated with PSD include living alone, poststroke isolation or distress, physical disability, and a history of depression, stroke, or other psychiatric conditions.1 The roles of factors such as sex, age, and socioeconomic status remain less clear.1,7,8 Identification of
Journal of Stroke and Cerebrovascular Diseases, Vol. ■■, No. ■■ (■■), 2016: pp ■■–■■
1
ARTICLE IN PRESS S. HIRATA ET AL.
2
factors associated with PSD may permit earlier recognition of PSD and prompt initiation of interventions to prevent or treat PSD. The purpose of this study was to determine which factors were independently associated with and predictive of depression among individuals with a history of stroke in a national sample of the U.S. population.
other sociodemographic factors (education level, poverty index, marital status). The second model also adjusted for major medical comorbidities as individual factors or by total number of comorbidities (0, 1, 2, and ≥3). The discrimination ability of the logistic model was determined by computing the specificity, sensitivity, and accuracy using the receiver operating characteristic curve analysis under the final model.
Methods Patient Population The National Health and Nutrition Examination Surveys (NHANES) are cross-sectional samples of the U.S. civilian noninstitutionalized population conducted by the National Center for Health Statistics (NCHS), a branch of the Centers for Disease Control and Prevention. We assessed data from NHANES from 2005 to 2010. The NCHS institutional review board approved the protocols for conduct of the NHANES and informed consent was obtained from all participants. The sampling plan followed a complex, stratified, multistage probability cluster design, with oversampling of non-Hispanic blacks, Mexican Americans, and the elderly, to enhance the precision of prevalence estimates in those groups.9 Interviews were conducted in sampled households, and all subjects were invited to participate in medical examinations that were conducted at nearby mobile examination centers. The interviews collected demographic, socioeconomic, dietary, and health-related information. Mobile examinations consisted of medical and dental examinations, physiological measurements, and laboratory tests. Detailed descriptions of the plan and operation of each survey have been previously published.9 NHANES weights were calculated according to the analytical guidelines provided in the NHANES online documentation. All estimates were weighted in order to obtain nationally representative estimates.
Statistical Analyses Stroke was defined by self-report. Presence of depression was defined as Patient Health Questionnaire (PHQ8) scores of 10 or more. Using the NHANES 2005-2010, the predictors of depression among stroke survivors aged 20 years or more in the United States were assessed. The following potential predictors were assessed: age, sex, race/ ethnicity, household income as measured by the income to poverty index ratio, education level (less than high school, high school, more than high school), marital status, body mass index, smoking, and medical history including congestive heart failure, coronary artery disease, angina pectoris, myocardial infarction (MI), diabetes, emphysema, chronic bronchitis, liver disease, and cancer (except skin cancer). We performed two logistic regression models. The first model adjusted for age, sex, race/ethnicity, and
Random Forest Random forest model was used to identify the factors that were associated with and predictive of PSD, as another multivariate strategy to the above logistic model. The random forest model consists of an ensemble of classification trees calculated on bootstrap samples of the original data (training samples). For each tree, a random subset of the input variables is used to find the best split at each node of the tree. Observations that are not used in building the current tree constitute the so-called “out of bag” (OOB) sample (testing samples). For each observation, the final class prediction is made by a majority vote across all trees.10 We used N = 2000 as the total number of trees in the forest and n = 4 as the number of input variables that are randomly chosen at each split.
Variable Importance and Final Model Selection We sorted the 17 variables in the order of importance using the mean decrease in Gini index. To select the final model, we used backward elimination of variables using out of box error.11 We chose the model with the smallest number of variables and whose out of box error was within 1 standard error of the best fitting model.
Prototypes Using the 10 variables chosen, we computed a proximity or ten-dimensional distance measure among all n observations. For each index observation, we could determine how many of K = 86 nearest neighbors belonged to the same group (depression or not depression) as the index observation. We used K = 86 because the smaller group had 87. We defined the “best/purest” observation as that observation with the greatest number of its 86 nearest neighbors who belonged to the same group, and called this group of up to 86 neighbors the “prototype group.” We then reported summary statistics (“medoid”) on the prototype depression group compared to the prototype nondepression group. Prototypes for each class were computed using the function “ClassCenter” (“RandomForest” package in R version 2.15.1). All analyses were performed in SAS version 9.4 (Copyright © 2002-2012 by SAS Institute Inc., Cary, NC) and in R version 2.15.1 (RandomForest package).
ARTICLE IN PRESS FACTORS ASSOCIATED WITH DEPRESSION AFTER STROKE
Results Of the 17,132 individuals surveyed from 2005 to 2010, a total of 678 stroke survivors were identified, corresponding to 6,235,528 individuals in the United States. Of those stroke survivors, 546 (81%) completed the PHQ-8 survey; 17% (n = 87) had depression, corresponding to 872,237 stroke survivors with depression in the United States. The 132 individuals who did not complete the PHQ-8 were approximately 4.5 years older on average than those who completed the PHQ-8. There were no other significant differences in the distribution of covariates in the group who completed the PHQ-8 versus those who did not. In the bivariate analysis, compared to stroke survivors without depression, depressed stroke survivors were younger (55.0 versus 65.8 years) and more likely to have an income to poverty ratio less than 200% (70% versus 44%; Table 1). Stroke survivors with depression were more likely to smoke than those without depression (50% versus 19%). PSD was seen less frequently among individuals with a history of cancer and more frequently seen in overweight individuals and those with history of an MI; however, the trends only approached statistical significance. The initial random forest model was performed using all 17 variables. The mean decrease in Gini index was used to sort the 17 variables in order of importance (Table 2). After backward elimination using OOB error rate, the final model solution had the following 10 variables in the order of decreasing importance: age, education level, body mass index, race/ethnicity, poverty index, smoking, female sex, marital status, history of cancer, and history of MI (Table 2). Once these 10 factors were known, the addition of the other 7 factors to the model did not further improve the predictions (Fig 1). The final random forest model based on these 10 variables had an overall accuracy of 69%, correctly predicting 322 out of 459 stroke survivors without depression and 56 out of 87 stroke survivors with depression (specificity: 70% and sensitivity: 64%). In the final model based on 10 variables, the depression prototype group was characterized by individuals with a median age of 47 years, less than high school education (55.6%), predominantly obese (50%), white (61%), black (31%), poverty to income index less than 200% (86%), current smokers (82%), female (44%), not married (67%), did not have history of cancer (91%), and had a history of heart attack (52%) (Table 3). On the other hand, the prototype for individuals without PSD was characterized by persons who had a median age of 80 years, had more than high school education (72%), were predominantly nonobese (70%), were white (96%), had poverty to income index greater than 200% (67%), did not smoke (100%), were female (13%), were married (59%), had a history of cancer (39%), and did not have a history of heart attack (89%) (Table 3).
3
Table 1. Bivariate analysis of risk factors for depression among stroke survivors aged 20 years or older, NHANES 2005-2010
Factor Age, mean (SE) Female, % (SE) Race, % (SE) Black NH Hispanic/other White NH Education level, % (SE) Less than high school High school More than high school Income to poverty ratio, % (SE) <125% 125%-200% 200%-400% >400% Marital status: married, % (SE) Major medical comorbidity, % (SE) Congestive heart failure Coronary artery disease Angina History of myocardial infarction Diabetes Emphysema Chronic bronchitis Liver disease Cancer (excluding skin) BMI (kg/m2), % (SE) <25 25-29 ≥30 Smoking status: current smoker, % (SE)
Without depression N = 459
With depression N = 87
65.8 (1.0) 54.9% (3.1)
55.0 (1.3) 63.1% (7.0)
14.6% (1.7) 10.5% (2.0) 74.9% (2.7)
19.0% (3.9) 11.3% (3.6) 69.6% (5.6)
P value <.001 .31 .43
.80 29.2% (2.6)
31.9% (4.7)
26.8% (2.5) 44.0% (3.2)
29.1% (7.4) 39.0% (7.9) .03
25.7% (2.8) 18.2% (2.2) 35.0% (2.9) 21.1% (2.8) 53.7% (3.0)
38.5% (8.2) 31.0% (5.6) 21.2% (7.9) 9.2% (4.0) 47.2% (4.9)
.27
15.5% (2.0)
22.3% (4.7)
.12
16.2% (1.8)
21.9% (5.7)
.30
12.2% (1.8) 19.9% (2.2)
10.9% (3.5) 28.8% (5.3)
.75 .07
30.8% (2.4) 7.0% (1.5) 14.8% (2.6)
28.5% (5.4) 10.5% (3.4) 23.1% (5.6)
.68 .31 .12
5.8% (1.6) 17.6% (2.2)
11.0% (4.2) 9.5% (3.3)
.19 .07 .08
23.9% (1.9) 31.9% (2.6) 44.2% (3.0) 19.2% (2.0)
13.1% (3.9) 41.0% (7.1) 45.8% (5.2) 49.8% (5.3)
<.001
Abbreviations: NH, non-Hispanic; NHANES, National Health and Nutrition Examination Survey; SE, standard error.
ARTICLE IN PRESS S. HIRATA ET AL.
4
Table 2. Variable importance using mean decreases in Gini index Predictor
Mean decrease in Gini index*
Age Education level Body mass index Race Poverty index Smoker Female Married Cancer Myocardial infarction Diabetes Congestive heart failure Coronary artery disease Chronic bronchitis Angina Emphysema Liver disease
17.44 4.56 4.49 4.12 2.91 2.73 2.49 2.34 2.21 2.03 1.71 1.65 1.42 1.34 1.28 1.01 .98
*The final model includes the first 10 predictors after variable selection.
In the first multivariate logistic model, after adjusting for age, sex, race/ethnicity, education level, poverty index, and marital status, increasing age was associated with lower odds of PSD (per year odds ratio [OR] .95, 95% confidence interval [CI] .94-.97) and poverty was associated with nearly threefold higher odds of PSD (poverty index (PI) <200% versus ≥200%, OR 2.61, 95% CI 1.235.53; Table 4). These associations remained after further adjusting for medical comorbidities including congestive heart failure, coronary artery disease, angina pectoris, MI, diabetes, emphysema or chronic bronchitis, liver
Table 3. Prototypes of stroke survivors with and without depression Class prototypes*
Median age, years Education: less than high school Education: high school Education: more than high school BMI: normal (<25 kg/m2) BMI: overweight (25-29 kg/m2) BMI: obese (≥30 kg/m2) Race White Black Other Poverty index <200% Current smoker Female Married Cancer Heart Attack
Without depression n = 86
With depression n = 35
80 10.9%
47 55.6%
17.4% 71.7%
22.2% 22.2%
19.6% 50.0%
5.6% 44.4%
30.4%
50.0%
96.2% 1.4% 2.4% 33.4% .0% 13.0% 58.9% 39.2% 11.2%
61.4% 30.9% 7.7% 85.9% 81.5% 44.4% 33.2% 9.3% 52.1%
Abbreviation: BMI, body mass index. *Prototype—representative observations of each class based on proximities (similarity matrix).
disease, and cancer (per year OR .95, 95% CI .93-.97 and PI <200% versus ≥200%, OR 2.70, 95% CI 1.27-5.73; Table 4). In addition, congestive heart failure was associated with higher odds of PSD (OR 1.64, 95% CI .98-2.74) and individuals with at least 3 medical comorbidities were more likely to have PSD (OR 1.59, 95% CI 1.01-2.49).
Discussion
Figure 1.
Variable importance plot—mean decrease in Gini index.
In this national cross-sectional study of communitydwelling stroke survivors, 17% had PSD. Using a random forest model, we identified 10 factors associated with PSD: age, education level, body mass index, race/ethnicity, poverty index, smoking, female sex, marital status, history of cancer, and history of MI. Using a multivariate model controlling for numerous sociodemographic and medical factors, younger age, poverty, congestive heart failure, and higher number of comorbidities were independently associated with PSD. The prevalence of PSD noted in this study is slightly lower than previous prevalence estimates among noninstitutionalized stroke survivors. Ayerbe et al’s systematic review and meta-analysis of 43 cohorts published prior to August 2011 (n = 20,293) revealed a pooled prevalence of PSD of 25%, CI 19-32 beyond 1 year.12 NHANES’
ARTICLE IN PRESS FACTORS ASSOCIATED WITH DEPRESSION AFTER STROKE
5
Table 4. Multivariate logistic models for evaluating factors associated with presence of depression among stroke survivors, NHANES 2005-2010 Model I* Predictor Age (per year) Poverty index (<200% versus ≥200%) Congestive heart failure Coronary artery disease
Model II†
OR
P value
OR
P value
.95 (.94-.97) 2.60 (1.23-5.53) – –
<.001 .01 – –
<.001 .01 .06 .13
–
–
.95 (.93-.97) 2.70 (1.27-5.73) 1.64 (.98-2.74) 1.92 (.83-4.45) Alternative model 1.59 (1.01-2.49)
Three or more major comorbidities (versus <3)
.04
Abbreviations: NHANES, National Health and Nutrition Examination Survey; OR, odds ratio. *Model I adjusts for age, sex, race, and sociodemographic factors including education level, poverty index, and marital status. †Model II also adjusts for major medical comorbidities as individual factors (congestive heart failure, coronary artery disease, angina pectoris, myocardial infarction, diabetes, emphysema or chronic bronchitis, liver disease, and cancer) or by total number of major comorbidities (0, 1, 2, or ≥3).
use of PHQ-8 as a screening tool for depression likely underestimated true PSD prevalence. A study involving 72 individuals more than 3 weeks poststroke revealed that a PHQ-9 threshold of 9 had a sensitivity of .69; 95% CI .39-.91 and specificity of .78; 95% CI .65-.88).13 The stroke literature includes an abundance of studies evaluating potential predictors of PSD; however, findings have been variable and inconsistent. Three recent systematic reviews attempted to elucidate the most common and consistent variables associated with PSD. A systematic review of studies assessing predictors of PSD published prior to August 2011 (10 studies; n = 16,045) revealed that 2 of 5 studies found a relationship between baseline disability and PSD; another 2 of 5 found that disability was associated with PSD at follow-up; and 4 of 5 studies found premorbid depression or treatment for depression as a predictor for PSD.12 Age and sex did not predict depression in 6 out of the 7 studies that investigated the associations.12 A second systematic review of studies published prior to October 2012 (24 studies = 14,642) showed that sex was evaluated as a predictor of PSD in 21 studies; 13 studies found no association between PSD and sex, but 7 studies identified female sex as a risk factor for PSD; and 1 study reported male sex as a risk factor.14 Age was not associated with PSD in 16 studies, but 4 reported a link (1 study showed that age <68 years was an independent predictor of PSD within the first year after stroke and 3 other studies revealed that older age was a significant risk factor for PSD). Older patients were more often depressed in the “acute” poststroke phase. In our study of community-dwelling stroke survivors, most participants were likely in the chronic phase. In addition, they found that medical history, predisposing illness, or comorbidity was associated with PSD in 5 of 10 studies,14 consistent with our finding of a higher number of comorbidities associated with PSD. A third systematic review of studies published prior to May 2013 (23 studies;
n = 18,374) revealed that only 10 variables were assessed for their relationship with depression after stroke in 5 or more of the 23 included studies.15 A personal history of depression before stroke was associated with later depression in 4 of 7 studies. Cognitive impairment was associated with depression in 2 of 4 studies. The most consistent associations were severe stroke (in 4 of 6 studies), early physical disability (in 4 of 5 studies), and later disability (mild-to-moderate in 5 of 5 studies, and major in 7 of 8 studies). Older age was associated with PSD in 3 of 16 studies and female sex in 8 of 18 studies.15 The significant variability in predictors of PSD identified in these studies reflects variations in study setting (acute inpatient, rehabilitation, chronic care facility, or outpatient), study populations, method used to diagnose stroke, time since stroke onset, inclusion/exclusion criteria, PSD screening method, threshold for PSD diagnosis, potential PSD predictors evaluated, and statistical methods. Of note, women are more likely to screen positively for depressive symptoms; therefore, higher prevalence of depression is generally seen among women in comparison to men.16 The higher frequency of PSD seen among women in this study may be partially attributed to the higher prevalence of depression among women, limiting the strength of sex as a predictor of PSD.1 This study also revealed a higher prevalence of smoking among individuals with PSD, a factor not highlighted in previous studies.1,8 Whereas PSD was more frequently observed in smokers, as well as those with a history of an MI, both smoking and heart disease are also wellestablished risk factors for stroke. The higher frequency of PSD may in part be due to the higher prevalence of stroke among these specific populations, limiting the predictive significance of these factors.17 This study has several limitations. First, NHANES relies on self-reported demographic, socioeconomic, and healthrelated information including history of stroke. Although
ARTICLE IN PRESS S. HIRATA ET AL.
6
NHANES has not validated self-reporting of stroke, selfreported stroke has a positive predictive value of 79%, a sensitivity of 80%, and a specificity of 99%, suggesting that self-report can reliably be used in epidemiological analysis.18 Second, NHANES does not collect information regarding stroke type (ischemic versus hemorrhagic), location, severity, symptoms, poststroke disability, or date of stroke onset—factors that may have an impact on—or have been shown to have an effect on—the development of PSD. Although one meta-analysis of individuals within 2 months of stroke onset showed that PSD was most often seen in left anterior, compared with left posterior lesions, and in left anterior, compared with right anterior lesions,19 a meta-analysis of 35 cohorts published prior to August 199920 and a subsequent systematic review and meta-analysis of 43 cohorts published prior to January 2014 (n = 5507)21 did not show an association between laterality or location and stroke. Third, the NHANES interview did not collect information regarding premorbid conditions prior to the stroke (such as a history of depression). Fourth, uncertainty remains regarding the ideal screening tool for diagnosing depression in individuals with prior stroke. The NHANES used PHQ8, a method with reasonable sensitivity and specificity.13 Finally, NHANES evaluates noninstitutionalized individuals; therefore, patients with severe strokes who reside in skilled nursing facilities were not captured in the interviews. The strengths of this study include the rigorous, validated, standardized, and consistent methods for collecting data in NHANES each year, the nationwide scope, with oversampling of minority populations, and the number of factors evaluated. Identifying sociodemographic and medical factors that are consistently associated with PSD may assist us in targeting PSD prevention efforts and determining which specific population subsets may benefit from systematic screening.
References 1. Teasell R, Foley N, Salter K, et al. Evidence-based review of stroke rehabilitation: executive summary. Top Stroke Rehabil 2009;16:463-488. 2. Altieri M, Maestrini I, Mercurio A, et al. Depression after minor stroke: prevalence and predictors. Eur J Neurol 2012;19:517-521. 3. Ellis C, Zhao Y, Egede LE. Depression and increased risk of death in adults with stroke. J Psychosom Res 2010;68:545-551. 4. Morris R, Jones J, Wilcox J, et al. Depression and anxiety screening after stroke: adherence to guidelines and future directions. Disabil Rehabil 2012;34:733-739.
5. Hackett ML, Yapa C, Parag V, et al. Frequency of depression after stroke a systematic review of observational studies. Stroke 2005;36:1330-1340. 6. Bushnell C. Depression and the risk of stroke in women: an identification and treatment paradox. Stroke 2011;42:2718-2719. 7. Bos MJ, Linden T, Koudstaal PJ, et al. Depressive symptoms and risk of stroke: the Rotterdam study. J Neurol Neurosurg Psychiatry 2008;79:997-1001. 8. Ouimet MA, Primeau F, Cole MG. Psychosocial risk factors in poststroke depression: a systematic review. Can J Psychiatry 2001;46:819-828. 9. Centers for Disease Control and Prevention. National Center for Health Statistics (NCHS): National Health and Nutrition Examination Survey questionnaire, examination protocol, and laboratory protocols. Hyattsville: Department of Health and Human Services., 2005-2006. 10. Breiman L. Random forests. Mach Learn 2001;24:123140. 11. Diaz-Uriarte R, Alvarez de Andres S. Variable section from random forests using both backwards. Am Stat Assoc 2005;92:548-560. 12. Ayerbe L, Ayis S, Wolfe CD, et al. Natural history, predictors and outcomes of depression after stroke: systematic review and meta-analysis. Br J Psychiatry 2013;202:14-21. 13. Turner A, Hambridge J, White J, et al. Depression screening in stroke: a comparison of alternative measures with the structured diagnostic interview for the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (major depressive episode) as criterion standard. Stroke 2012;43:1000-1005. 14. De Ryck A, Brouns R, Geurden M, et al. Risk factors for poststroke depression: identification of inconsistencies based on a systematic review. J Geriatr Psychiatry Neurol 2014;27:147-158. 15. Kutlubaev MA, Hackett ML. Part II: predictors of depression after stroke and impact of depression on stroke outcome: an updated systematic review of observational studies. Int J Stroke. 2014;9:1026-1036. 16. Salokangas RK, Vaahtera K, Pacriev S, et al. Gender differences in depressive symptoms. An artefact caused by measurement instruments? J Affect Disord 2002;68:215220. 17. Sacco RL, Benjamin EJ, Broderick JP, et al. American Heart Association Prevention Conference. IV. Prevention and Rehabilitation of Stroke. Risk factors. Stroke 1997;28:15071517. 18. Engstad T, Bonaa KH, Viitanen M. Validity of self-reported stroke: the Tromso study. Stroke 2000;31:1602-1607. 19. Robinson RG. The clinical neuropsychiatry of stroke. Cambridge; New York: Cambridge University Press, 2006. 20. Carson AJ, MacHale S, Allen K, et al. Depression after stroke and lesion location: a systematic review. Lancet 2000;356:122-126. 21. Wei N, Yong W, Li X, et al. Post-stroke depression and lesion location: a systematic review. J Neurol 2015;262:8190.
Key Factors Associated with Major Depression in a National Sample of Stroke Survivors Sarah Hirata,
MS,*
Bruce Ovbiagele, MD, MSc, MAS,† Daniela Markovic, Amytis Towfighi, MD§‖
MS,‡
and
Background: Depression, one of the most common complications encountered after stroke, is associated with poorer outcomes. The aim of this study was to determine the factors independently associated with and predictive of poststroke depression (PSD). Methods: We assessed the prevalence of depression (Patient Health Questionnaire [PHQ-8] score >10) among a national sample of adults (≥20 years) with stroke who participated in the National Health and Nutrition Examination Surveys from 2005 to 2010. Logistic regression and random forest models were used to determine the factors associated with and predictive of PSD, after adjusting for sociodemographic and clinical factors. Results: Of the 17,132 individuals surveyed, 546 stroke survivors were screened for depression, and 17% had depression, corresponding to 872,237 stroke survivors with depression in the United States. In the logistic regression model, after adjustment for sociodemographic variables, poverty (poverty index <200% versus ≥200%, odds ratio [OR] 2.61, 95% confidence interval [CI] 1.23-5.53) and 3 or more medical comorbidities (OR 1.59, 95% CI 1.01-2.49) were associated with higher odds of PSD; increasing age was associated with lower odds of PSD (per year OR .95, 95% CI .94-.97). In the random forest model, the 10 most important factors predictive of PSD were younger age, lower education level, higher body mass index, black race, poverty, smoking, female sex, single marital status, lack of cancer history, and previous myocardial infarction (specificity = 70%, sensitivity = 64%). Conclusion: Although numerous factors were predictive of developing PSD, younger age, poverty, and multiple comorbidities were strong and independent factors. More aggressive screening for depression in these individuals may be warranted. Key Words: Depression—stroke— NHANES—poststroke depression—factors. © 2015 National Stroke Association. Published by Elsevier Inc. All rights reserved.
Introduction From the *University of Hawaii John A. Burns School of Medicine, Honolulu, Hawaii; †Department of Neurology, Medical University of South Carolina, Charleston, South Carolina; ‡Department of Biomathematics, University of California, Los Angeles, California; §Department of Neurology, University of Southern California, Los Angeles, California; and ‖Department of Neurology, Rancho Los Amigos National Rehabilitation Center, Downey, California. Received June 10, 2015; revision received November 28, 2015; accepted December 30, 2015. Grant support: American Heart Association National Scientist Development Award (11SDG7590160). Address correspondence to Amytis Towfighi, MD, 7601 E. Imperial Highway, Downey, CA 90242. E-mail: [email protected] 1052-3057/$ - see front matter © 2015 National Stroke Association. Published by Elsevier Inc. All rights reserved. http://dx.doi.org/10.1016/j.jstrokecerebrovasdis.2015.12.042
Poststroke depression (PSD) is one of the most common consequences of stroke, affecting approximately one third of stroke survivors.1-5 Recent studies have indicated an association between PSD and poststroke physical morbidity and mortality; those with PSD have a higher risk for suboptimal recovery, recurrent vascular events, and an overall poorer quality of life.1-4,6 Although the adverse effects of PSD have been well studied, the nature of PSD itself remains unclear. Factors most consistently associated with PSD include living alone, poststroke isolation or distress, physical disability, and a history of depression, stroke, or other psychiatric conditions.1 The roles of factors such as sex, age, and socioeconomic status remain less clear.1,7,8 Identification of
Journal of Stroke and Cerebrovascular Diseases, Vol. ■■, No. ■■ (■■), 2016: pp ■■–■■
1
ARTICLE IN PRESS S. HIRATA ET AL.
2
factors associated with PSD may permit earlier recognition of PSD and prompt initiation of interventions to prevent or treat PSD. The purpose of this study was to determine which factors were independently associated with and predictive of depression among individuals with a history of stroke in a national sample of the U.S. population.
other sociodemographic factors (education level, poverty index, marital status). The second model also adjusted for major medical comorbidities as individual factors or by total number of comorbidities (0, 1, 2, and ≥3). The discrimination ability of the logistic model was determined by computing the specificity, sensitivity, and accuracy using the receiver operating characteristic curve analysis under the final model.
Methods Patient Population The National Health and Nutrition Examination Surveys (NHANES) are cross-sectional samples of the U.S. civilian noninstitutionalized population conducted by the National Center for Health Statistics (NCHS), a branch of the Centers for Disease Control and Prevention. We assessed data from NHANES from 2005 to 2010. The NCHS institutional review board approved the protocols for conduct of the NHANES and informed consent was obtained from all participants. The sampling plan followed a complex, stratified, multistage probability cluster design, with oversampling of non-Hispanic blacks, Mexican Americans, and the elderly, to enhance the precision of prevalence estimates in those groups.9 Interviews were conducted in sampled households, and all subjects were invited to participate in medical examinations that were conducted at nearby mobile examination centers. The interviews collected demographic, socioeconomic, dietary, and health-related information. Mobile examinations consisted of medical and dental examinations, physiological measurements, and laboratory tests. Detailed descriptions of the plan and operation of each survey have been previously published.9 NHANES weights were calculated according to the analytical guidelines provided in the NHANES online documentation. All estimates were weighted in order to obtain nationally representative estimates.
Statistical Analyses Stroke was defined by self-report. Presence of depression was defined as Patient Health Questionnaire (PHQ8) scores of 10 or more. Using the NHANES 2005-2010, the predictors of depression among stroke survivors aged 20 years or more in the United States were assessed. The following potential predictors were assessed: age, sex, race/ ethnicity, household income as measured by the income to poverty index ratio, education level (less than high school, high school, more than high school), marital status, body mass index, smoking, and medical history including congestive heart failure, coronary artery disease, angina pectoris, myocardial infarction (MI), diabetes, emphysema, chronic bronchitis, liver disease, and cancer (except skin cancer). We performed two logistic regression models. The first model adjusted for age, sex, race/ethnicity, and
Random Forest Random forest model was used to identify the factors that were associated with and predictive of PSD, as another multivariate strategy to the above logistic model. The random forest model consists of an ensemble of classification trees calculated on bootstrap samples of the original data (training samples). For each tree, a random subset of the input variables is used to find the best split at each node of the tree. Observations that are not used in building the current tree constitute the so-called “out of bag” (OOB) sample (testing samples). For each observation, the final class prediction is made by a majority vote across all trees.10 We used N = 2000 as the total number of trees in the forest and n = 4 as the number of input variables that are randomly chosen at each split.
Variable Importance and Final Model Selection We sorted the 17 variables in the order of importance using the mean decrease in Gini index. To select the final model, we used backward elimination of variables using out of box error.11 We chose the model with the smallest number of variables and whose out of box error was within 1 standard error of the best fitting model.
Prototypes Using the 10 variables chosen, we computed a proximity or ten-dimensional distance measure among all n observations. For each index observation, we could determine how many of K = 86 nearest neighbors belonged to the same group (depression or not depression) as the index observation. We used K = 86 because the smaller group had 87. We defined the “best/purest” observation as that observation with the greatest number of its 86 nearest neighbors who belonged to the same group, and called this group of up to 86 neighbors the “prototype group.” We then reported summary statistics (“medoid”) on the prototype depression group compared to the prototype nondepression group. Prototypes for each class were computed using the function “ClassCenter” (“RandomForest” package in R version 2.15.1). All analyses were performed in SAS version 9.4 (Copyright © 2002-2012 by SAS Institute Inc., Cary, NC) and in R version 2.15.1 (RandomForest package).
ARTICLE IN PRESS FACTORS ASSOCIATED WITH DEPRESSION AFTER STROKE
Results Of the 17,132 individuals surveyed from 2005 to 2010, a total of 678 stroke survivors were identified, corresponding to 6,235,528 individuals in the United States. Of those stroke survivors, 546 (81%) completed the PHQ-8 survey; 17% (n = 87) had depression, corresponding to 872,237 stroke survivors with depression in the United States. The 132 individuals who did not complete the PHQ-8 were approximately 4.5 years older on average than those who completed the PHQ-8. There were no other significant differences in the distribution of covariates in the group who completed the PHQ-8 versus those who did not. In the bivariate analysis, compared to stroke survivors without depression, depressed stroke survivors were younger (55.0 versus 65.8 years) and more likely to have an income to poverty ratio less than 200% (70% versus 44%; Table 1). Stroke survivors with depression were more likely to smoke than those without depression (50% versus 19%). PSD was seen less frequently among individuals with a history of cancer and more frequently seen in overweight individuals and those with history of an MI; however, the trends only approached statistical significance. The initial random forest model was performed using all 17 variables. The mean decrease in Gini index was used to sort the 17 variables in order of importance (Table 2). After backward elimination using OOB error rate, the final model solution had the following 10 variables in the order of decreasing importance: age, education level, body mass index, race/ethnicity, poverty index, smoking, female sex, marital status, history of cancer, and history of MI (Table 2). Once these 10 factors were known, the addition of the other 7 factors to the model did not further improve the predictions (Fig 1). The final random forest model based on these 10 variables had an overall accuracy of 69%, correctly predicting 322 out of 459 stroke survivors without depression and 56 out of 87 stroke survivors with depression (specificity: 70% and sensitivity: 64%). In the final model based on 10 variables, the depression prototype group was characterized by individuals with a median age of 47 years, less than high school education (55.6%), predominantly obese (50%), white (61%), black (31%), poverty to income index less than 200% (86%), current smokers (82%), female (44%), not married (67%), did not have history of cancer (91%), and had a history of heart attack (52%) (Table 3). On the other hand, the prototype for individuals without PSD was characterized by persons who had a median age of 80 years, had more than high school education (72%), were predominantly nonobese (70%), were white (96%), had poverty to income index greater than 200% (67%), did not smoke (100%), were female (13%), were married (59%), had a history of cancer (39%), and did not have a history of heart attack (89%) (Table 3).
3
Table 1. Bivariate analysis of risk factors for depression among stroke survivors aged 20 years or older, NHANES 2005-2010
Factor Age, mean (SE) Female, % (SE) Race, % (SE) Black NH Hispanic/other White NH Education level, % (SE) Less than high school High school More than high school Income to poverty ratio, % (SE) <125% 125%-200% 200%-400% >400% Marital status: married, % (SE) Major medical comorbidity, % (SE) Congestive heart failure Coronary artery disease Angina History of myocardial infarction Diabetes Emphysema Chronic bronchitis Liver disease Cancer (excluding skin) BMI (kg/m2), % (SE) <25 25-29 ≥30 Smoking status: current smoker, % (SE)
Without depression N = 459
With depression N = 87
65.8 (1.0) 54.9% (3.1)
55.0 (1.3) 63.1% (7.0)
14.6% (1.7) 10.5% (2.0) 74.9% (2.7)
19.0% (3.9) 11.3% (3.6) 69.6% (5.6)
P value <.001 .31 .43
.80 29.2% (2.6)
31.9% (4.7)
26.8% (2.5) 44.0% (3.2)
29.1% (7.4) 39.0% (7.9) .03
25.7% (2.8) 18.2% (2.2) 35.0% (2.9) 21.1% (2.8) 53.7% (3.0)
38.5% (8.2) 31.0% (5.6) 21.2% (7.9) 9.2% (4.0) 47.2% (4.9)
.27
15.5% (2.0)
22.3% (4.7)
.12
16.2% (1.8)
21.9% (5.7)
.30
12.2% (1.8) 19.9% (2.2)
10.9% (3.5) 28.8% (5.3)
.75 .07
30.8% (2.4) 7.0% (1.5) 14.8% (2.6)
28.5% (5.4) 10.5% (3.4) 23.1% (5.6)
.68 .31 .12
5.8% (1.6) 17.6% (2.2)
11.0% (4.2) 9.5% (3.3)
.19 .07 .08
23.9% (1.9) 31.9% (2.6) 44.2% (3.0) 19.2% (2.0)
13.1% (3.9) 41.0% (7.1) 45.8% (5.2) 49.8% (5.3)
<.001
Abbreviations: NH, non-Hispanic; NHANES, National Health and Nutrition Examination Survey; SE, standard error.
ARTICLE IN PRESS S. HIRATA ET AL.
4
Table 2. Variable importance using mean decreases in Gini index Predictor
Mean decrease in Gini index*
Age Education level Body mass index Race Poverty index Smoker Female Married Cancer Myocardial infarction Diabetes Congestive heart failure Coronary artery disease Chronic bronchitis Angina Emphysema Liver disease
17.44 4.56 4.49 4.12 2.91 2.73 2.49 2.34 2.21 2.03 1.71 1.65 1.42 1.34 1.28 1.01 .98
*The final model includes the first 10 predictors after variable selection.
In the first multivariate logistic model, after adjusting for age, sex, race/ethnicity, education level, poverty index, and marital status, increasing age was associated with lower odds of PSD (per year odds ratio [OR] .95, 95% confidence interval [CI] .94-.97) and poverty was associated with nearly threefold higher odds of PSD (poverty index (PI) <200% versus ≥200%, OR 2.61, 95% CI 1.235.53; Table 4). These associations remained after further adjusting for medical comorbidities including congestive heart failure, coronary artery disease, angina pectoris, MI, diabetes, emphysema or chronic bronchitis, liver
Table 3. Prototypes of stroke survivors with and without depression Class prototypes*
Median age, years Education: less than high school Education: high school Education: more than high school BMI: normal (<25 kg/m2) BMI: overweight (25-29 kg/m2) BMI: obese (≥30 kg/m2) Race White Black Other Poverty index <200% Current smoker Female Married Cancer Heart Attack
Without depression n = 86
With depression n = 35
80 10.9%
47 55.6%
17.4% 71.7%
22.2% 22.2%
19.6% 50.0%
5.6% 44.4%
30.4%
50.0%
96.2% 1.4% 2.4% 33.4% .0% 13.0% 58.9% 39.2% 11.2%
61.4% 30.9% 7.7% 85.9% 81.5% 44.4% 33.2% 9.3% 52.1%
Abbreviation: BMI, body mass index. *Prototype—representative observations of each class based on proximities (similarity matrix).
disease, and cancer (per year OR .95, 95% CI .93-.97 and PI <200% versus ≥200%, OR 2.70, 95% CI 1.27-5.73; Table 4). In addition, congestive heart failure was associated with higher odds of PSD (OR 1.64, 95% CI .98-2.74) and individuals with at least 3 medical comorbidities were more likely to have PSD (OR 1.59, 95% CI 1.01-2.49).
Discussion
Figure 1.
Variable importance plot—mean decrease in Gini index.
In this national cross-sectional study of communitydwelling stroke survivors, 17% had PSD. Using a random forest model, we identified 10 factors associated with PSD: age, education level, body mass index, race/ethnicity, poverty index, smoking, female sex, marital status, history of cancer, and history of MI. Using a multivariate model controlling for numerous sociodemographic and medical factors, younger age, poverty, congestive heart failure, and higher number of comorbidities were independently associated with PSD. The prevalence of PSD noted in this study is slightly lower than previous prevalence estimates among noninstitutionalized stroke survivors. Ayerbe et al’s systematic review and meta-analysis of 43 cohorts published prior to August 2011 (n = 20,293) revealed a pooled prevalence of PSD of 25%, CI 19-32 beyond 1 year.12 NHANES’
ARTICLE IN PRESS FACTORS ASSOCIATED WITH DEPRESSION AFTER STROKE
5
Table 4. Multivariate logistic models for evaluating factors associated with presence of depression among stroke survivors, NHANES 2005-2010 Model I* Predictor Age (per year) Poverty index (<200% versus ≥200%) Congestive heart failure Coronary artery disease
Model II†
OR
P value
OR
P value
.95 (.94-.97) 2.60 (1.23-5.53) – –
<.001 .01 – –
<.001 .01 .06 .13
–
–
.95 (.93-.97) 2.70 (1.27-5.73) 1.64 (.98-2.74) 1.92 (.83-4.45) Alternative model 1.59 (1.01-2.49)
Three or more major comorbidities (versus <3)
.04
Abbreviations: NHANES, National Health and Nutrition Examination Survey; OR, odds ratio. *Model I adjusts for age, sex, race, and sociodemographic factors including education level, poverty index, and marital status. †Model II also adjusts for major medical comorbidities as individual factors (congestive heart failure, coronary artery disease, angina pectoris, myocardial infarction, diabetes, emphysema or chronic bronchitis, liver disease, and cancer) or by total number of major comorbidities (0, 1, 2, or ≥3).
use of PHQ-8 as a screening tool for depression likely underestimated true PSD prevalence. A study involving 72 individuals more than 3 weeks poststroke revealed that a PHQ-9 threshold of 9 had a sensitivity of .69; 95% CI .39-.91 and specificity of .78; 95% CI .65-.88).13 The stroke literature includes an abundance of studies evaluating potential predictors of PSD; however, findings have been variable and inconsistent. Three recent systematic reviews attempted to elucidate the most common and consistent variables associated with PSD. A systematic review of studies assessing predictors of PSD published prior to August 2011 (10 studies; n = 16,045) revealed that 2 of 5 studies found a relationship between baseline disability and PSD; another 2 of 5 found that disability was associated with PSD at follow-up; and 4 of 5 studies found premorbid depression or treatment for depression as a predictor for PSD.12 Age and sex did not predict depression in 6 out of the 7 studies that investigated the associations.12 A second systematic review of studies published prior to October 2012 (24 studies = 14,642) showed that sex was evaluated as a predictor of PSD in 21 studies; 13 studies found no association between PSD and sex, but 7 studies identified female sex as a risk factor for PSD; and 1 study reported male sex as a risk factor.14 Age was not associated with PSD in 16 studies, but 4 reported a link (1 study showed that age <68 years was an independent predictor of PSD within the first year after stroke and 3 other studies revealed that older age was a significant risk factor for PSD). Older patients were more often depressed in the “acute” poststroke phase. In our study of community-dwelling stroke survivors, most participants were likely in the chronic phase. In addition, they found that medical history, predisposing illness, or comorbidity was associated with PSD in 5 of 10 studies,14 consistent with our finding of a higher number of comorbidities associated with PSD. A third systematic review of studies published prior to May 2013 (23 studies;
n = 18,374) revealed that only 10 variables were assessed for their relationship with depression after stroke in 5 or more of the 23 included studies.15 A personal history of depression before stroke was associated with later depression in 4 of 7 studies. Cognitive impairment was associated with depression in 2 of 4 studies. The most consistent associations were severe stroke (in 4 of 6 studies), early physical disability (in 4 of 5 studies), and later disability (mild-to-moderate in 5 of 5 studies, and major in 7 of 8 studies). Older age was associated with PSD in 3 of 16 studies and female sex in 8 of 18 studies.15 The significant variability in predictors of PSD identified in these studies reflects variations in study setting (acute inpatient, rehabilitation, chronic care facility, or outpatient), study populations, method used to diagnose stroke, time since stroke onset, inclusion/exclusion criteria, PSD screening method, threshold for PSD diagnosis, potential PSD predictors evaluated, and statistical methods. Of note, women are more likely to screen positively for depressive symptoms; therefore, higher prevalence of depression is generally seen among women in comparison to men.16 The higher frequency of PSD seen among women in this study may be partially attributed to the higher prevalence of depression among women, limiting the strength of sex as a predictor of PSD.1 This study also revealed a higher prevalence of smoking among individuals with PSD, a factor not highlighted in previous studies.1,8 Whereas PSD was more frequently observed in smokers, as well as those with a history of an MI, both smoking and heart disease are also wellestablished risk factors for stroke. The higher frequency of PSD may in part be due to the higher prevalence of stroke among these specific populations, limiting the predictive significance of these factors.17 This study has several limitations. First, NHANES relies on self-reported demographic, socioeconomic, and healthrelated information including history of stroke. Although
ARTICLE IN PRESS S. HIRATA ET AL.
6
NHANES has not validated self-reporting of stroke, selfreported stroke has a positive predictive value of 79%, a sensitivity of 80%, and a specificity of 99%, suggesting that self-report can reliably be used in epidemiological analysis.18 Second, NHANES does not collect information regarding stroke type (ischemic versus hemorrhagic), location, severity, symptoms, poststroke disability, or date of stroke onset—factors that may have an impact on—or have been shown to have an effect on—the development of PSD. Although one meta-analysis of individuals within 2 months of stroke onset showed that PSD was most often seen in left anterior, compared with left posterior lesions, and in left anterior, compared with right anterior lesions,19 a meta-analysis of 35 cohorts published prior to August 199920 and a subsequent systematic review and meta-analysis of 43 cohorts published prior to January 2014 (n = 5507)21 did not show an association between laterality or location and stroke. Third, the NHANES interview did not collect information regarding premorbid conditions prior to the stroke (such as a history of depression). Fourth, uncertainty remains regarding the ideal screening tool for diagnosing depression in individuals with prior stroke. The NHANES used PHQ8, a method with reasonable sensitivity and specificity.13 Finally, NHANES evaluates noninstitutionalized individuals; therefore, patients with severe strokes who reside in skilled nursing facilities were not captured in the interviews. The strengths of this study include the rigorous, validated, standardized, and consistent methods for collecting data in NHANES each year, the nationwide scope, with oversampling of minority populations, and the number of factors evaluated. Identifying sociodemographic and medical factors that are consistently associated with PSD may assist us in targeting PSD prevention efforts and determining which specific population subsets may benefit from systematic screening.
References 1. Teasell R, Foley N, Salter K, et al. Evidence-based review of stroke rehabilitation: executive summary. Top Stroke Rehabil 2009;16:463-488. 2. Altieri M, Maestrini I, Mercurio A, et al. Depression after minor stroke: prevalence and predictors. Eur J Neurol 2012;19:517-521. 3. Ellis C, Zhao Y, Egede LE. Depression and increased risk of death in adults with stroke. J Psychosom Res 2010;68:545-551. 4. Morris R, Jones J, Wilcox J, et al. Depression and anxiety screening after stroke: adherence to guidelines and future directions. Disabil Rehabil 2012;34:733-739.
5. Hackett ML, Yapa C, Parag V, et al. Frequency of depression after stroke a systematic review of observational studies. Stroke 2005;36:1330-1340. 6. Bushnell C. Depression and the risk of stroke in women: an identification and treatment paradox. Stroke 2011;42:2718-2719. 7. Bos MJ, Linden T, Koudstaal PJ, et al. Depressive symptoms and risk of stroke: the Rotterdam study. J Neurol Neurosurg Psychiatry 2008;79:997-1001. 8. Ouimet MA, Primeau F, Cole MG. Psychosocial risk factors in poststroke depression: a systematic review. Can J Psychiatry 2001;46:819-828. 9. Centers for Disease Control and Prevention. National Center for Health Statistics (NCHS): National Health and Nutrition Examination Survey questionnaire, examination protocol, and laboratory protocols. Hyattsville: Department of Health and Human Services., 2005-2006. 10. Breiman L. Random forests. Mach Learn 2001;24:123140. 11. Diaz-Uriarte R, Alvarez de Andres S. Variable section from random forests using both backwards. Am Stat Assoc 2005;92:548-560. 12. Ayerbe L, Ayis S, Wolfe CD, et al. Natural history, predictors and outcomes of depression after stroke: systematic review and meta-analysis. Br J Psychiatry 2013;202:14-21. 13. Turner A, Hambridge J, White J, et al. Depression screening in stroke: a comparison of alternative measures with the structured diagnostic interview for the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (major depressive episode) as criterion standard. Stroke 2012;43:1000-1005. 14. De Ryck A, Brouns R, Geurden M, et al. Risk factors for poststroke depression: identification of inconsistencies based on a systematic review. J Geriatr Psychiatry Neurol 2014;27:147-158. 15. Kutlubaev MA, Hackett ML. Part II: predictors of depression after stroke and impact of depression on stroke outcome: an updated systematic review of observational studies. Int J Stroke. 2014;9:1026-1036. 16. Salokangas RK, Vaahtera K, Pacriev S, et al. Gender differences in depressive symptoms. An artefact caused by measurement instruments? J Affect Disord 2002;68:215220. 17. Sacco RL, Benjamin EJ, Broderick JP, et al. American Heart Association Prevention Conference. IV. Prevention and Rehabilitation of Stroke. Risk factors. Stroke 1997;28:15071517. 18. Engstad T, Bonaa KH, Viitanen M. Validity of self-reported stroke: the Tromso study. Stroke 2000;31:1602-1607. 19. Robinson RG. The clinical neuropsychiatry of stroke. Cambridge; New York: Cambridge University Press, 2006. 20. Carson AJ, MacHale S, Allen K, et al. Depression after stroke and lesion location: a systematic review. Lancet 2000;356:122-126. 21. Wei N, Yong W, Li X, et al. Post-stroke depression and lesion location: a systematic review. J Neurol 2015;262:8190.