Obesity as a Risk Factor for Developing ... - Wiley Online Library

Obesity

Original Article EPIDEMIOLOGY/GENETICS

Obesity as a Risk Factor for Developing Functional Limitation Among Older Adults: A Conditional Inference Tree Analysis Feon W. Cheng1, Xiang Gao1, Le Bao2, Diane C. Mitchell1, Craig Wood3, Martin J. Sliwinski4, Helen Smiciklas-Wright1, Christopher D. Still4, David D. K. Rolston5, and Gordon L. Jensen6

Objective: To examine the risk factors of developing functional decline and make probabilistic predictions by using a tree-based method that allows higher order polynomials and interactions of the risk factors. Methods: The conditional inference tree analysis, a data mining approach, was used to construct a risk stratification algorithm for developing functional limitation based on BMI and other potential risk factors for disability in 1,951 older adults without functional limitations at baseline (baseline age 73.1 6 4.2 y). We also analyzed the data with multivariate stepwise logistic regression and compared the two approaches (e.g., cross-validation). Over a mean of 9.2 6 1.7 years of follow-up, 221 individuals developed functional limitation. Results: Higher BMI, age, and comorbidity were consistently identified as significant risk factors for functional decline using both methods. Based on these factors, individuals were stratified into four risk groups via the conditional inference tree analysis. Compared to the low-risk group, all other groups had a significantly higher risk of developing functional limitation. The odds ratio comparing two extreme categories was 9.09 (95% confidence interval: 4.68, 17.6). Conclusions: Higher BMI, age, and comorbid disease were consistently identified as significant risk factors for functional decline among older individuals across all approaches and analyses. Obesity (2017) 25, 1263-1269. doi:10.1002/oby.21861

Introduction Several studies have reported the positive association between BMI and functional decline among older persons using logistic regressions. Regression models such as linear models and generalized linear models have been successful in detecting main effects. However, they rarely take account of higher-order polynomials or interactions of risk factors, as those terms dramatically increase the number of potential models and make the model selection difficult (1). Furthermore, traditional regression approaches are limited in their ability to translate results for meaningful clinical application (e.g., clinicians usually identify patients as high or low risk in contrast to the risk ratio as determined by regression methods) (1,2).

In this article, we sought to predict functional decline using a data mining tool—the conditional inference tree analysis (3,4). The observations are viewed as points in the p-dimensional feature space expanded by p distinct predictors. The conditional inference tree analysis recursively splits the feature space into smaller and smaller subsets, and each split is made by defining a cutoff value on one of the p predictors, e.g., whether the age is greater than 70. The optimal splits assign training data points with similar response status into the same subsets and the ones with opposite response status into different subsets. To make a prediction, instead of plugging the predictor values into some prediction function, we only need to identify the subset that the observation falls into and then predict the response status by the majority vote of training data within that

1

Department of Nutritional Sciences, Pennsylvania State University, University Park, Pennsylvania, USA. Correspondence: Xiang Gao ([email protected]) Department of Statistics, Pennsylvania State University, University Park, Pennsylvania, USA 3 Obesity Institute, Geisinger Health System, Danville, Pennsylvania, USA 4 Department of Human Development and Family Studies, Pennsylvania State University, University Park, Pennsylvania, USA 5 Department of Internal Medicine, Geisinger Health System, Danville, Pennsylvania, USA 6 University of Vermont College of Medicine, Burlington, Vermont, USA. 2

Funding agencies: This study was supported by the United States Department of Agriculture, Agricultural Research Service agreement #8050-51530-012-01A and the Global Health Engagement Network Pilot Funding. Disclosure: The authors declared no conflicts of interest. Author contributions: All authors designed and conducted research; FWC, XG, and LB analyzed data; and FWC, XG, LB, and GLJ wrote the paper. XG and GLJ had primary responsibility for final content. All authors read and approved the final manuscript. Additional Supporting Information is available in the online version of this article. Received: 3 February 2017; Accepted: 28 March 2017; Published online 24 May 2017. doi:10.1002/oby.21861

www.obesityjournal.org

Obesity | VOLUME 25 | NUMBER 7 | JULY 2017

1263

Obesity

Obesity and Functional Decline Among Older Adults Cheng et al.

Figure 1 Flowchart of Geisinger Rural Aging Study participants.

subset. The subset identification process can be visualized by using a tree diagram in which the parent nodes (larger subspaces) are partitioned into child nodes (smaller subspaces) by various questions regarding the predictor values, which enhances the translatability of findings into practice (1). At the bottom of the tree, a subset is likely defined by a sequence of questions that involve multiple predictors or repeated use of the same predictor so that it easily incorporates the interactions and polynomials. In addition, some of the predictors might not be involved in any split, which means the treebased method automatically performs the variable selection. As noted in previous literature, “this process of reconciling the clinical and statistical relevance of variables in the data ultimately yields a more well informed and statistically informative model than either a singularly clinical or statistical approach” (2).

at the Geisinger Medical Center (Danville, Pennsylvania) and were active in the electronic medical record (EMR) Epic Systems (Verona, Wisconsin) between January 1, 2001, and December 31, 2004.

Therefore, we used conditional inference tree analysis to construct a risk stratification algorithm for risk of developing functional decline, based on BMI and other potential risk factors for disability (e.g., age, sex, lifestyle factors, disease burden) in a prospective study including 1,951 older adults with normal physical function at baseline.

Compared to participants who were excluded from the study because of incomplete baseline or follow-up questionnaires, those who were included were slightly younger (mean 73.1 vs. 73.9 years) and healthier (as suggested by a lower mean modified Charlson index of 1.1 vs. 1.4) and had lower BMI (mean 29.3 vs. 29.8 kg/m2), but they were of similar sex proportion (men: 38.1% vs. 39.6%).

Methods

The study protocol was approved by the Institutional Review Board at the Geisinger Health System, and an IRB-approved Data Sharing Agreement was in place with the Pennsylvania State University.

Study population The Geisinger Rural Aging Study (GRAS) comprises a cohort of northeastern and central Pennsylvanians at least 65 years of age who were enrolled in a Medicare-managed risk program offered through the Geisinger Health System (Danville, Pennsylvania) (5). There were 4,565 GRAS participants who received primary care services

1264


Between 1999 and 2003, we mailed a questionnaire to assess baseline functional status (i.e., activities of daily living [ADL] and instrumental activities of daily living [IADL]) (6,7) to all GRAS participants. The questionnaire was completed by 3,583 GRAS participants (78.5%). In this prospective analysis, we excluded those (n 5 261) who had baseline functional limitation (i.e., who indicated having any of the ADLs or IADLs). Between 2009 and 2011, we mailed the same questionnaire for follow-up. Of the remaining 3,322 participants, 773 died and 598 did not return the questionnaire, leaving 1,951 individuals for the final analysis (Figure 1).

Assessment of functional limitation We assessed functional status using the ADL and IADL questions “I usually or always need assistance with (check all that apply):


Original Article

Obesity

EPIDEMIOLOGY/GENETICS

bathing, dressing, grooming, toileting, eating, walking or moving about, traveling (outside the home), preparing food, and shopping for food or other necessities” (6,7) at baseline and again in 20092011. Functional limitation (yes or no) was defined as reporting any of the nine items.

Assessment of predictors We extracted information for all potential predictors from the EMR and used the earliest plausible values between January 1, 2001, and December 31, 2004, as baseline. These variables included weight, height, age, sex, smoking status (never/former/current smoker), alcohol drinker (yes/no), blood glucose, systolic blood pressure, diastolic blood pressure, triglycerides, high-density lipoprotein (HDL) cholesterol, low-density lipoprotein (LDL) cholesterol, hypercholesterolemia medication (yes/no), hypertension medication (yes/no), diabetic medication (yes/no), and International Classification of Diseases, Ninth Revision (ICD-9) codes. Strict algorithms were implemented to detect for inaccurate and implausible values before selecting the earliest plausible measurements from the baseline period. Briefly, we first excluded implausible weights (< 24.9 kg [ 453.6 kg [>1,000 pounds]) and heights ( 228.6 cm [>90 inches]) and then further identified inaccurate measurements by comparing all available values for each individual person using preestablished criteria (8). As a result, approximately 0.23% and 2.39% of the weights and heights, respectively, were excluded. We also eliminated implausible laboratory and blood pressure measurements: blood glucose < 30 or > 600 mg/dL (< 1.67 or > 33.3 mmol/ L), triglyceride < 10 or > 2000 mg/dL (< 0.11 or >22.60 mmol/L), HDL cholesterol < 10 or > 120 mg/dL (< 0.26 or > 3.10 mmol/L), LDL cholesterol < 30 or > 300 mg/dL (< 0.78 or > 7.76 mmol/L), systolic blood pressure < 60 or > 250 mmHg, and diastolic blood pressure < 40 or > 140 mmHg (8-11). BMI (kg/m2) was calculated using weights and heights measured by trained health professionals according to standard outpatient clinic visit protocol (e.g., without overgarments and shoes). The definition of metabolic health status was based on a modified Adult Treatment Panel III guideline similar to that utilized by Wildman and colleagues (12,13). Metabolically healthy represented having 1 of the following conditions: fasting glucose 100 mg/dL ( 5.56 mmol/L) or diabetes diagnosis (ICD-9 250.XX); blood pressure 130/85 mm Hg or hypertension diagnosis (ICD-9 401.XX-405.XX); triglycerides 150 mg/dL (1.69 mmol/L); and HDL cholesterol for women < 50 mg/dL (0, and BMI>38.5 (38.3% functional limitation).


Original Article

Obesity

EPIDEMIOLOGY/GENETICS

Figure 3 Odds ratios of developing any functional limitation between risk groups identified by conditional inference tree analysis (n 5 1,951).

individuals in the lowest-risk group had lower BMI and modified Charlson index and were younger, while only higher BMI was consistently recognized as a characteristic of the highest-risk group (Supporting Information Table S4). Last, exclusion of those with amputation produced the same conditional inference tree (data not shown).

Discussion In our study, elevated BMI was consistently identified as a significant risk factor for developing future functional limitation across all approaches and analyses, including conditional inference tree analysis, logistic regression modeling, and also sensitivity analyses. This is consistent with previous studies based on conventional approaches (e.g., logistic regression), which reported an association between elevated BMI and higher risk for developing functional limitation among older persons (15,20-25). For example, the Established Populations for Epidemiologic Studies of the Elderly reported that BMI 30 was significantly associated with greater risk of developing ADL disability among 12,725 adults aged 65 years over a 7-year follow-up period (20). In the Medicare Current Beneficiary Survey Study of 20,975 older adults aged 65 years, higher BMI was associated with new or exacerbated ADL or IADL disability in a dose-response manner over a 2-year period (21). This study demonstrated the potential benefits of using a data mining approach (e.g., conditional inference tree analysis) in addition to traditional regression methods to better address risk classification questions. The conditional inference tree analysis generated a reliable risk stratification algorithm for developing functional limitation based on health and laboratory data that were available in the EMR. The overall percentage of people developing functional limitation for our cohort


was 11.3%, but this percentage differed by almost sixfold (from lowrisk group: 6.40% to high-risk group: 38.30%) based on age, modified Charlson index, and BMI. Furthermore, our findings highlight the strength of conditional inference tree analysis in modeling complex interactions. The subpopulation generated from three splits was a result of three-way interactions, which are rarely considered in traditional regression models. With 11 potential predictors, there would be 211 5 2,048 possible models, but there would be 2(11 1 551165) 5 3.45*1069 possible models if two-way and three-way interactions were allowed, which would greatly challenge model selection. Furthermore, traditional regression methods can provide researchers with an idea of how a predictor associates with an outcome while controlling for other covariates, but it is difficult to utilize such findings to generate a clinical decision rule. Our conditional inference tree analysis suggests that for an individual 75.7 years with a modified Charlson index > 0, the addition of having a BMI >38.5 can place that individual in a higher-risk group (i.e., high risk vs. intermediate risk I) for functional decline. Knowing this information could assist clinicians in identifying at-risk individuals and allocating healthcare resources. It is important to note that conditional inference tree analysis (and data mining approaches in general) also requires adequate sample size to reach the threshold for appropriate binary splits in the tree structure, similar to traditional regression methods when there is a lack of power to detect for significant results (i.e., wide 95% confidence interval) (1,18). Age and comorbidity have long been recognized as risk factors for functional limitation (22,24,26). The fact that BMI, in addition to age and comorbidity, was identified as a significant predictor further confirms its added value in the context of predicting future functional decline. Previous literature has also reported BMI as an independent risk factor for functional limitation and disability (15,20-


1267

Obesity

Obesity and Functional Decline Among Older Adults Cheng et al.

25). Excess body weight is associated with greater risk of several conditions that afflict older persons, such as osteoarthritis of the knees and reduced muscle strength (i.e., in those with sarcopenic obesity), which may negatively impact functional ability (26-31). In our analysis, among adults aged 75.7 years, the algorithm further stratified individuals into risk groups based upon modified Charlson index and BMI but not among participants aged > 75.7 years. This may be due to small sample size in this older population, as 44.6% of the “older” participants had died between baseline and follow-up, compared to only 21.2% among individuals aged 75.7 years. In fact, when we used the same stratification rule identified in the “younger” participants with individuals aged > 75.7 years, the percentages of those who developed functional limitation in each of the three risk groups (low risk: Charlson index 5 0, intermediate risk: modified Charlson index > 0 1 BMI 38, and high risk: modified Charlson index > 0 1 BMI > 38) were 17.7% (31/175), 17.3% (49/283), and 33.3% (2/6), respectively. Our speculation was further confirmed by the findings from the propensity score models, which showed the added value of BMI even among the individuals aged > 75.7 years (data not shown). Several limitations should be noted. First, we relied on self-reported ADLs and IADLs, although those measures have been shown to correlate well with more objective physical performance tests (32-34). We also did not have the onset dates of when participants developed any functional limitation, which precluded us from using Cox proportional hazards regression models, in addition to multivariate stepwise logistic regression, for comparison. Second, we did not have any physical activity information. Physical activity is related to functional limitation, independent of BMI, age, sex, disease status, and other sociodemographic and psychosocial factors (24). Third, body composition data were not available. Fat and muscle mass measurements, along with waist circumference, have been shown to be significant predictors of disability (25,35,36). Sarcopenic obesity is associated with greater risk of functional decline (31,37,38). Future studies are warranted to explore whether including body composition data would allow for a more precise risk stratification of individuals in relation to functional limitation. Moreover, we did not have detailed information (i.e., how often and amount) on alcohol consumption. Our findings from the stepwise logistic regression model showed that those who drank alcohol were less likely to have functional decline. Previous studies (15,39) have reported a Ushaped association between alcohol intake and functional decline, with the lowest risk among moderate drinkers. One study (39) observed a substantial attenuation of this relationship with the adjustment of lifestyle-related factors. Thus, we speculate that the “alcohol” variable in our cohort may be a proxy for more favorable behavioral and lifestyle factors that could benefit functional ability. However, the anti-inflammatory effect of alcohol could also be another potential explanation (40). Although we were mainly interested in examining the associations between risk factors, specifically BMI, and functional decline among survivors, we constructed five propensity score models to impute functional limitation data for those who had died or did not complete the follow-up survey. Results remained consistent. Last, our results may not be generalizable to other populations because members of our cohort were 99.8% Caucasians who resided in northeastern and central Pennsylvania. However, our results are comparable to those reported by other studies with more diverse populations (20-22,24,25).

1268


Conclusion In conclusion, higher BMI, age, and comorbid disease were consistently identified as significant risk factors for functional decline among older individuals across all approaches and analyses. Our study demonstrated the potential benefits of using a data mining approach in addition to traditional regression methods to address health and clinical questions, which may enhance the translation of findings into practice.O

Acknowledgments The authors would like to acknowledge Bethann Whilden for her role in data collection and management. C 2017 The Obesity Society V

References 1. Lewis RJ. An introduction to classification and regression tree (CART) analysis. Paper presented at: Annual Meeting of the Society for Academic Emergency Medicine; May 2000; San Francisco, CA. http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.95. 4103 2. Morgan J. Classification and regression tree analysis. Technical Report No. 1. Boston University School of Public Health. May 8, 2014. https://www.bu.edu/sph/ files/2014/05/MorganCART.pdf 3. Hothorn T, Hornik K, Zeileis A. Unbiased recursive partitioning: a conditional inference framework. J Comput Graph Stat 2006;15:651-674. 4. Tanadini LG, Steeves JD, Hothorn T, et al. Identifying homogeneous subgroups in neurological disorders: unbiased recursive partitioning in cervical complete spinal cord injury. Neurorehabil Neural Repair 2014;28:507-515. 5. Jensen GL, Kita K, Fish J, Heydt D, Frey C. Nutrition risk screening characteristics of rural older persons: relation to functional limitations and health care charges. Am J Clin Nutr 1997;66:819-828. 6. Katz S, Downs TD, Cash HR, Grotz RC. Progress in development of the index of ADL. Gerontologist 1970;10:20-30. 7. Lawton MP. Scales to measure competence in everyday activities. Psychopharmacol Bull 1988;24:609-614. 8. Littman AJ, Boyko EJ, McDonell MB, Fihn SD. Evaluation of a weight management program for veterans. Prev Chronic Dis 2012;9:E99. doi:10.5888/ pcd9.110267 9. Pedroza-Tobias A, Trejo-Valdivia B, Sanchez-Romero LM, Barquera S. Classification of metabolic syndrome according to lipid alterations: analysis from the Mexican National Health and Nutrition Survey 2006. BMC Public Health 2014;14:1056. doi: 10.1186/1471-2458-14-1056 10. Yuan G, Al-Shali KZ, Hegele RA. Hypertriglyceridemia: its etiology, effects and treatment. CMAJ 2007;176:1113-1120. 11. American Diabetes Association. Diagnosis and classification of diabetes mellitus. Diabetes Care 2010;33 (Suppl 1):S62-S69. 12. Grundy SM, Brewer HB, Cleeman JI, Smith SC, Lenfant C. Definition of metabolic syndrome: Report of the National Heart, Lung, and Blood Institute/American Heart Association conference on scientific issues related to definition. Circulation 2004; 109:433-438. 13. Wildman RP, Muntner P, Reynolds K, et al. The obese without cardiometabolic risk factor clustering and the normal weight with cardiometabolic risk factor clustering: prevalence and correlates of 2 phenotypes among the US population (NHANES 1999-2004). Arch Intern Med 2008;168:1617-1624. 14. Deyo RA, Cherkin DC, Ciol MA. Adapting a clinical comorbidity index for use with ICD-9-CM administrative databases. J Clin Epidemiol 1992;45:613-619. 15. Sulander T, Martelin T, Rahkonen O, Nissinen A, Uutela A. Associations of functional ability with health-related behavior and body mass index among the elderly. Arch Gerontol Geriatr 2005;40:185-199. 16. NHLBI Obesity Education Initiative Expert Panel on the Identification, Evaluation, and Treatment of Overweight and Obesity in Adults. Clinical Guidelines on the identification, evaluation, and treatment of overweight and obesity in adults. National Heart, Lung, and Blood Institute; 1998. 17. Hothorn T, Hornik K, Strobl C, Zeileis A. A Laboratory for Recursive Partytioning. 2015. https://cran.r-project.org/web/packages/party/index.html 18. Fonarow GC, Adams KF, Abraham WT, Yancy CW, Boscardin WJ. Risk stratification for in-hospital mortality in acutely decompensated heart failure: classification and regression tree analysis. JAMA 2005;293:572-580. 19. Chang C-CH, Zhao Y, Lee C-W, Ganguli M. Smoking, death, and Alzheimer disease: a case of competing risks. Alzheimer Dis Assoc Disord 2012;26:300-306.


Original Article

Obesity

EPIDEMIOLOGY/GENETICS 20. Al Snih S, Ottenbacher KJ, Markides KS, Kuo Y-F, Eschbach K, Goodwin JS. The effect of obesity on disability vs mortality in older Americans. Arch Intern Med 2007;167:774-780. 21. Wee CC, Huskey KW, Ngo LH, et al. Obesity, race, and risk for death or functional decline among Medicare beneficiaries: a cohort study. Ann Intern Med 2011;154:645-655. 22. Stuck AE, Walthert JM, Nikolaus T, B€ ula CJ, Hohmann C, Beck JC. Risk factors for functional status decline in community-living elderly people: a systematic literature review. Soc Sci Med 1999;48:445-469. 23. Jensen GL, Friedmann JM. Obesity is associated with functional decline in community-dwelling rural older persons. J Am Geriatr Soc 2002;50:918-923. 24. Nusselder WJ, Looman CWN, Mackenbach JP. Nondisease factors affected trajectories of disability in a prospective study. J Clin Epidemiol 2005;58:484-494. 25. Chen H, Guo X. Obesity and functional disability in elderly Americans. J Am Geriatr Soc 2008;56:689-694. 26. Villareal DT, Apovian CM, Kushner RF, Klein S. Obesity in older adults: technical review and position statement of the American Society for Nutrition and NAASO, The Obesity Society. Am J Clin Nutr 2005;82:923-934. 27. Anandacoomarasamy A, Caterson I, Sambrook P, Fransen M, March L. The impact of obesity on the musculoskeletal system. Int J Obes (Lond) 2008;32:211-222. 28. Felson DT. Obesity and knee osteoarthritis. Ann Intern Med 1988;109:18. 29. Hart DJ, Spector TD. The relationship of obesity, fat distribution and osteoarthritis in women in the general population: the Chingford Study. J Rheumatol 1993;20:331-335. 30. Cicuttini FM, Baker JR, Spector TD. The association of obesity with osteoarthritis of the hand and knee in women: a twin study. J Rheumatol 1996;23:1221-1226. 31. Jensen GL, Hsiao PY. Obesity in older adults: relationship to functional limitation. Curr Opin Clin Nutr Metab Care 2010;13:46-51.


32. Wang C-Y, Sheu C-F, Protas E. Construct validity and physical performance of older adults in different hierarchical physical-disability level. J Aging Phys Act 2007;15:75-89. 33. Kivinen P, Sulkava R, Halonen P, Nissinen A. Self-reported and performance-based functional status and associated factors among elderly men. J Clin Epidemiol 1998; 51:1243-1252. 34. Van Den Brink CL, Tijhuis M, Kalmijn S, et al. Self-reported disability and its association with performance-based limitation in elderly men: a comparison of three european countries. J Am Geriatr Soc 2003;51:782-788. 35. Visser M, Harris TB, Langlois J, et al. Body fat and skeletal muscle mass in relation to physical disability in very old men and women of the Framingham Heart Study. J Gerontol A Biol Sci Med Sci 1998;53:M214-M221. 36. Visser M, Langlois J, Guralnik JM, et al. High body fatness, but not low fat-free mass, predicts disability in older men and women: the Cardiovascular Health Study. Am J Clin Nutr 1998;68:584-590. 37. Schaap LA, Koster A, Visser M. Adiposity, muscle mass, and muscle strength in relation to functional decline in older persons. Epidemiol Rev 2013;35:51-65. 38. Baumgartner RN, Wayne SJ, Waters DL, Janssen I, Gallagher D, Morley JE. Sarcopenic obesity predicts instrumental activities of daily living disability in the elderly. Obes Res 2004;12:1995-2004. 39. Maraldi C, Harris TB, Newman AB, et al. Moderate alcohol intake and risk of functional decline: the Health, Aging, and Body Composition study. J Am Geriatr Soc 2009;57:1767-1775. 40. Volpato S, Pahor M, Ferrucci L, et al. Relationship of alcohol intake with inflammatory markers and plasminogen activator inhibitor-1 in well-functioning older adults: the Health, Aging, and Body Composition study. Circulation 2004;109: 607-612.


1269