Identifying Behavioral Phenotypes in Chronic Illness: Self-Management of COPD and Comorbid Hypertension

Abstract

Objectives:

To identify and characterize the constellation, or clusters, of self-management behaviors in patients with chronic obstructive pulmonary disease (COPD) and comorbid hypertension.

Methods:

Cluster analysis (n=204) was performed with standardized scores for medication adherence to COPD and hypertension medications, inhaler technique, and diet as well as self-reported information on physical activity, appointment keeping, smoking status, and yearly influenza vaccination for a total of eight variables. Classification and regression tree analysis (CART) was performed to further characterize the resulting clusters.

Results:

Patients were divided into three clusters based on eight self-management behaviors, which included 95 patients in cluster 1, 42 in cluster 2, and 67 in cluster 3. All behaviors except for inhaler technique differed significantly among the three clusters (P’s<0.005). CART indicated physical activity was the first differentiating variable.

Conclusions:

Patients with COPD and hypertension can be separated into those with adequate and inadequate adherence. The group with inadequate adherence can further be divided into those with poor adherence to medical behaviors compared to those with poor adherence to lifestyle behaviors.

Practice Implications:

Once validated in other populations, the identification of patient clusters using patient self-management behaviors could be used to inform interventions for patients with multimorbidity.

1. Introduction

Multimorbidity is increasingly common as adults age and impacts patient outcomes negatively.[1] In a study of participants in the National Health and Nutrition Examination Survey (NHANES) from 2005 through 2012, 67% of the adult population of the United States aged 65 years and older had multimorbidity, where functional limitations increased with each additional chronic disease.[2] A large portion of chronic disease management depends upon patients’ daily decisions and actions.[3] Given the complexity of disease burden on patients with multimorbidity, these functional limitations can prevent effective self-management of multiple diseases and related issues such as polypharmacy.[4]

Facilitating effective self-management of chronic illness is an essential task for primary care practitioners, yet no existing health behavior theory fully characterizes the self-management behaviors of patients with multimorbidity.[4] Furthermore, the most effective interventions are disease specific and thus have limited applicability for patients with multiple diseases.[4] Despite the historical emphasis on individual chronic conditions, there are common self-management behaviors across conditions, including adhering to prescribed medicines, engaging in physical activity, and consuming a healthy diet. In order to advance the care of patients with multimorbidity, self-management of co-existing chronic diseases ought to be considered under the lens of a single conceptual model.

Toward this objective, we sought to determine how self-management behaviors cluster among patients with multimorbidity. Identifying these behavioral phenotypes across multiple chronic conditions may provide a basic understanding of patients’ health behavior in those with multimorbidity. Understanding human behavior on the basis of their self-management abilities may help build a comprehensive model of self-management in the context of multimorbidity in order to inform interventions down the line to assist disease management.

We chose to study behavioral phenotypes in patients with chronic obstructive pulmonary disease (COPD) because they are greatly impacted by multimorbidity. In a study of 1,003 patients with COPD, nearly half had five or more comorbid conditions.[5] The most common conditions included hypertension (55%) and hypercholesterolemia (52%).[5] However, many studies in this population focused on an isolated health-related behavior and did not take patients’ other illnesses into account. The presence of multiple chronic diseases among individuals with COPD made them a critical population to study self-management behaviors as successful self-management of disease can lead to better outcomes.[6, 7]

2. Methods

2.1. Settings and Subjects

For this study, we analyzed baseline data from 204 patients enrolled in an ongoing longitudinal cohort study of self-management behaviors among adults with COPD and multimorbidity. Patients from primary care practices in New York City and Chicago, IL, were interviewed, in English or Spanish, every five months. Additionally, prescription medication, diagnosis and outcomes data were confirmed by review of electronic medical records. The cohort included English- or Spanish-speaking, community dwelling adults with a physician-confirmed diagnosis of moderate or severe COPD; a ≥15 pack-year smoking history; an active prescription for either inhaled corticosteroids (ICS), a long-acting beta-agonist (LABA), and/or a long-acting anticholinergic for treatment of COPD; and a physician diagnosis of comorbid hypertension and/or diabetes with at least one current prescribed medication for the treatment of each. Patients were excluded if they had other non-COPD chronic respiratory illnesses or dementia. Participating patients provided signed informed consent and the study protocol was approved by the institutional review boards of the participating institutions.

For the present study, we further restricted our analyses to individuals with comorbid hypertension who had complete baseline data for each of the eight self-management behavior measures.

2.2. Measures of Self-Management

The self-management measures included both direct disease management behaviors and preventative care behaviors. Direct disease management included adherence to medications for COPD and hypertension and correctly demonstrated metered dose inhaler use. Preventative care included a physical activity assessment, adherence to diet recommendations for hypertension, medical appointment keeping, smoking status, and frequency of influenza vaccination. The behaviors chosen are all addressed in at least one of three major medical societies’ guidelines regarding the management of COPD,[8] and many are likewise addressed in guidelines for the management of hypertension.[9, 10]

The direct disease management measures included the use of two validated measures. Medication adherence was measured using the Medication Adherence Rating Scale (MARS), a validated 10-item self-reported measure designed to minimize social desirability bias.[11] Each item is rated on a 5-point Likert scale, with higher scores indicating greater adherence. Participants with an average MARS score of 4.5 or greater for all 10 items were classified as having good adherence.[12, 13] Of note, although we set a cutoff of 4.5 for the MARS score in our previous analyses, continuous MARS scores were used in our clustering methods, since the Classification and Regression Tree modeling method we used (described below) determines the most influential cutoffs that divide each group. The metered dose inhaler (MDI) score was calculated using a validated checklist of 11 steps.[14] Each step is considered complete or incomplete. We used the percentage of steps that were completed correctly. Scores of ≥75% were considered adequate.[14]

The preventative care measures were all based on self-report except for diet, which was measured through a validated score. Physical activity was measured in response to the question: “How often do you participate in at least 30 minutes of physical activity?” The possible responses were coded as a continuous measure from one to six: daily (1), 2–3 times a week (2), once a week (3), 2–3 time a month (4), monthly or less (5), and never (6).The Hypertension Self-Care Activity Level Effects (H-SCALE) DASH-Q Score was used to assess diet adherence.[15] The scale consists of 11 questions that ask patients to report the number of days per week they typically eat particular food groups. The score ranges from 0–77 where a higher score indicates greater adherence to a recommended diet for hypertension management. Medical appointment keeping was assessed with the question “In the past 6 months, how many medical appointments have you missed or had to reschedule?” The responses were grouped into three categories: Never (1), 1–2 times (2), and greater than 2 times (3). Patients reported their smoking status as either current or former smoker. Current smokers were considered non-adherent to COPD management guidelines. Patients were asked “How often would you say you get a flu shot?” Individuals were categorized as “yes” if they responded they received a vaccine every year and “no” if they responded with one of the three following responses: a few times in your life, one time in your life, or never.

2.3. Other Measures

Other data were collected to characterize the cohort of patients and to control for factors that have been associated with self-management behaviors among COPD patients including age, sex, race, education, English proficiency, health literacy, COPD severity, forced expiratory volume at 1 second (FEV1), and depression.[16–18] Health literacy was measured using a single-item proxy, which asked how often the patient needs help reading healthcare-related materials.[19–21] The patient was considered to have inadequate health literacy if they responded all or most of the time. COPD severity was measured using the COPD Severity Score, which captures respiratory symptoms, systemic corticosteroid use, and use of other respiratory medications. Scores range from 0 to 35 with higher scores indicating more severe disease.[22] FEV1 scores were divided into four groups: ≥ 80% of predicted, 50% ≤ predicted FEV1 < 80%, 30% ≤ predicted FEV1 < 50%, and 0% ≤ predicted FEV1 < 30%.[23]

2.4. Statistical Analysis

In order to group patients based on the eight selected self-management behaviors, we used the KAy-means for MIxed LArge data (KAMILA) method, which is a type of cluster analysis that uses an iterative process to develop clusters where continuous and categorical variables are given equal weight.[24, 25] We scaled the continuous measures, which included all variables except smoking and vaccine status. For these two categorical variables, we created dummy variables. We determined the optimal number of clusters (k) based on the prediction strength measure included in the “kamila” package for R statistical software,[24] and visual examinations of the cluster separation based on the t-distributed stochastic neighbor embedding (t-SNE) technique,[26] provided in the “Rtsne” package in R.[27]

We compared the differences in patient characteristics across the clusters using the chi-square test for categorical measures and the Kruskal-Wallis test for continuous measures.

In order to better understand and differentiate cluster membership based on the eight self-management behaviors, we applied Classification and Regression Tree (CART) analysis using the “rpart” package in R.[28] CART is a non-parametric test that we used to determine which behaviors were most influential in creating each cluster of behaviors.[29] CART determines the continuous or categorical variable that creates or splits the most homogenous groups based on the dependent variable and continues to find the next best univariate split in an iterative process until the subgroups either possess the minimum possible number of individuals or until no further differentiation is possible.[30, 31] In other words, CART was used to find which self-management behavior created two groups that most closely resembled two out of the three clusters determined beforehand through cluster analysis. The complete tree is then pruned using (10-fold) cross-validation. We set a minimum of 20 observations per group.[28] In other words, splits would not occur if they resulted in a cluster with n<20. The criteria for the splitting index was information gain, and the pruning method used was cost-complexity pruning.[28] In addition, we used 78% of our subjects in the training set and 22% of our subjects in the test set, and reported their respective overall accuracy.

Significance was based on a two-sided alpha of <0.05. Analyses were performed using SAS statistical software version 9.4 (SAS Institute, Inc., Cary, NC) and R statistical software (R version 3.6.0).

3. Results

3.1. Subject Characteristics

The mean age was 67.3 (8.6) years, 42.7% were male, 40.9% self-identified as Black, and 32.8% were college graduates (Table 1). Almost all of the participants (95.1%) had good English proficiency. Based on patients who never needed help reading hospital materials, 70.4% had adequate health literacy. In addition to COPD and hypertension, 42.2% of patients had depression and 34.3% had diabetes. The mean COPD Severity Score for the cohort was 9.2 (standard deviation [SD]: 5.6). The majority of patients (48.5%) had an FEV1 of 50 ≤ FEV1 < 80.

Table 1.

Baseline Characteristics of Study Population

Variable	All, N (%)
Age, Mean (SD)	67.3 (8.6)
Male Sex	87 (42.7)
Race and ethnicity
White	77 (37.9)
Black	83 (40.9)
Hispanic	34 (16.8)
Other	9 (4.4)
Education
Some High school or less	35 (17.2)
High school graduate	45 (22.1)
Some college	57 (27.9)
College graduate	67 (32.8)
Good English Proficiency	194 (95.1)
Health Literacy: Never needs help reading hospital materials	144 (70.6)
Comorbidities
Depression	86 (42.2)
Diabetes	70 (34.3)
COPD Severity Score,a Mean (SD)	9.2 (5.6)
FEV1
FEV1 ≥ 80	34 (17.4)
50 ≤ FEV1 < 80	95 (48.5)
30 ≤ FEVl <50	48 (24.5)
0 ≤ FEV1 < 30	19 (9.7)

^aCOPD Severity Score: 0–35, higher means more severe

3.2. Self-management Behaviors

Overall, 52.0% of the 204 participants were adherent to their COPD medications and 70.6% to their hypertension medications. Roughly a quarter of patients (24.0%) had adequate inhaler technique. The level of physical activity was between once a week and 2–3 times a month (3.4, SD: 2.0), and 57.8% of patients never missed appointments while 30.9% missed an appointment one or two times. The mean diet score was 34.3 (SD: 13.8), which is moderate. Former smokers made up 74.0% of the population, and 84.8% of participants received a flu shot annually.

3.3. Clusters of Behaviors

The KAMILA procedure identified three clusters based on the eight self-management behaviors (Table 2). Cluster 1 had 95 patients (46.6%), cluster 2, 42 patients (20.6%), and cluster 3, 67 patients (32.8%). Performance of seven of the eight behaviors studied were significantly different among at least two of the three clusters (COPD medication adherence p<0.0001, hypertension adherence p<0.0001, physical activity p<0.0001, diet score p<0.0001, missed appointments p<0.0001, former smoker p<0.0001, flu shot p=0.003). The performance of good inhaler technique did not differ among clusters (p=0.7). Overall, patients in cluster 1 demonstrated good self-management behaviors in most domains, cluster 2 patients had generally poor self-management behaviors, and cluster 3 patients had a mix of good and poor behaviors.

Table 2.

Three Clusters of Behaviors

	Cluster 1	Cluster 2	Cluster 3	p-value
Count (%), total n = 204	95 (46.6)	42 (20.6)	67 (32.8)	N/A
Behaviors
Adherent: COPD, n (%)	69 (72.6)	3 (7.1)	34 (50.7)	< 0.0001
Adherent: HTN, n (%)	79 (83.2)	9 (21.4)	56 (83.6)	< 0.0001
Good Inhaler Technique (≥ 75%), n (%)	22 (23.2)	12 (28.6)	15 (22.4)	0.74
Physical Activity,a mean (SD)	1.9 (1.0)	3.9 (2.0)	5.1 (1.5)	< 0.0001
Diet Score,‡ mean (SD)	41.1 (11.6)	34.1 (13.4)	24.7 (11.3)	< 0.0001
Missed Appointments,b mean (SD)	1.4 (0.6)	2.1 (0.8)	1.4 (0.5)	< 0.0001
Missed Appointments,c n (%)
Never	64 (67.4)	11 (26.2)	43 (64.2)	< 0.0001
One or Two Times	26 (27.4)	15 (35.7)	22 (32.8)
Greater than Two Times	5 (5.3)	16 (38.1)	2 (3.0)
Former Smoker, n (%)	84 (88.4)	20 (47.6)	47 (70.1)	< 0.0001
Flu Shot, n (%)	82 (86.3)	29 (69.0)	62 (92.5)	0.003

^aScale from daily (1) to never (6)

^bHypertension Self-Care Activity Level Effects (H-SCALE) DASH-Q Score: 0–77, higher means better diet

^cIn past 6 months, missed appointments never (1), One or two times (2), or Greater than two times (3)

In cluster 1, 72.6% of patients were adherent to their COPD medications and 83.2% to their hypertension medications. The patients were physically active with an average score 1.9 (SD: 1.0) on the physical activity scale, indicating almost daily physical activity. Additionally, patients scored 41.1 (SD: 11.6) out of 77 on the diet scale, indicating moderate adherence, and 67.4% of patients never missed appointments in the past 6 months. The majority (88.4%) reported that they did not currently smoke, and most patients received their annual flu shot (86.3%). A minority of patients (23.2%) had good inhaler technique.

Cluster 2’s self-management behaviors tended to contrast with those of cluster 1, with 7.1% of members adhering to their COPD medications and 21.4% to their hypertension medications. A small portion (28.6%) had good inhaler technique. The patients were not physically active with a self-reported 3.9 (SD: 2.0) on the scale, which corresponds closest with two to three times per month. Patients missed one to two appointments on average in the past six months. They had a diet score of 34.1 (SD: 13.4), indicating moderate adherence, roughly half of the group had quit smoking (47.6%), and 69.0% received their annual flu vaccine.

Cluster 3 was the least physically active of the three clusters with a score of 5.1 (SD: 1.5), and had the poorest diet score at 24.7 (SD: 11.3). Half (50.7%) of cluster 3 patients were adherent to COPD medications. Cluster 3 had a similar percentage (22.4%) of patients with good inhaler technique as both clusters 1 and 2. Comparable to cluster 1, 83.6% were adherent to hypertension medications, most patients (64.2%) did not miss appointments, no longer smoked (70.1%), and received a flu shot in the past year (92.5%).

3.4. Use of CART to Determine Behaviors Most Influential in Creating Clusters

The training set was comprised of 160 subjects (78%) and had an accuracy of 84% (Table 3). The test set included 44 subjects (22%) and had an accuracy of 86%. Physical activity determined the initial split between 66% of the cohort as more active than the remaining 34% (Figure 1). Of the 66%, the majority of the subjects (70%) were in cluster 1. From the 66%, those with a poorer diet (CART-determined score <33) but greater adherence to hypertension medications (CART-determined score ≥4.4), were more likely to be in cluster 1 (19% of the total) while those with a hypertension MARS score <4.4 were more likely to come from cluster 2 (6% of the total).

Table 3.

Patient Characteristics across Three Clusters

	Cluster 1	Cluster 2	Cluster 3	p-value
Count (%), total n = 204	95 (46.6)	42 (20.6)	67 (32.8)	N/A
Age, Mean (SD)	69.6 (8.8)	65.2 (7.6)	65.3 (8.1)	0.002
Male Sex	47 (49.5)	15 (35.7)	25 (37.3)	0.2
Race and ethnicity				0.1
White	43 (45.7)	8 (19.1)	26 (38.8)
Black	35 (37.2)	20 (47.6)	28 (41.8)
Hispanic	13 (13.8)	11 (26.2)	10 (14.9)
Other	3 (3.2)	3 (7.1)	3 (4.5)
Education				0.03
High school or less	10 (10.5)	11 (26.2)	14 (20.9)
High school graduate	17 (17.9)	10 (23.8)	18 (26.9)
Some college	28 (29.5)	14 (33.3)	15 (22.4)
College graduate	40 (42.1)	7 (16.7)	20 (29.9)
Good English Proficiency	93 (97.9)	38 (90.5)	63 (94.0)	0.2
Health Literacy: Never needs help reading hospital materials	64 (67.4)	27 (64.3)	53 (79.1)	0.1
Comorbidities
Depression	27 (28.4)	25 (59.5)	34 (50.8)	0.0007
Diabetes	22 (23.2)	18 (42.9)	30 (44.8)	0.007

Figure 1. Classification and Regression Tree (CART) of Clusters as Determined by Behaviors.

Training Tree (n=160). The first row of each box is titled with the cluster from which the majority of the patients in the group are drawn from. The second row of each box represents the percentage represented from cluster 1, cluster 2, and cluster 3, respectively. The third row of each box is the percentage of patients from the total test group. The scale for physical activity is from daily (1) to never (6). The range for diet score is 0–77, where higher means better diet. The Medication Adherence Rating Scale for hypertension and COPD (HTN MARS HTN and COPD MARS) is an average score from 10-item self-reported measures for a total of 5 points. In our previous analyses, participants with a MARS score of 4.5 or greater were classified as having good adherence, but the CART algorithm independently determines the most influential cutoffs that divide each group. The same is true for the metered dose inhaler (MDI) score, where ≥ 0.75 indicates adequate technique, but the CART algorithm independently determines the most influential cutoff.

Furthermore, patients with a physical activity score ≥4.5 made up 34% of the total, 71% of whom were from cluster 3. Hypertension medication adherence was split across this subgroup of patients with poor physical activity: 18% of the total had a CART-determined hypertension MARS score <4.8, indicating poorer adherence, and 17% of the total had a hypertension MARS score ≥4.8. The majority of patients with a lower hypertension MARS score were from Cluster 2 (50%) while the majority of patients with better adherence, or a score ≥4.8, were from Cluster 3 (93%).

3.5. Association of Clusters of Behaviors and Patient Characteristics

A larger portion of patients in cluster 1 were college graduates (42.1%) compared to cluster 2 (16.7%) and cluster 3 (29.9%, p=0.03, Table 3). Regarding comorbidities in cluster 1, 28.4% had depression and 23.2% had diabetes compared to 59.5% and 42.9% in cluster 2, and 50.8% and 44.8% in cluster 3 (Depression, p=0.0007, Diabetes, p=0.007).

4. Discussion

4.1. Discussion

In our preliminary analysis, we identified three self-management behavior clusters among patients with COPD and hypertension Overall, patients in cluster 1 had good self-management behaviors for the all of the eight behaviors aside from inhaler technique. Cluster 2 had the poorest performance in the majority of behaviors. Notably, Cluster 2 had better inhaler technique than Cluster 1 and Cluster 3. Cluster 3 had a mixture of optimal and suboptimal behaviors. With respect to self-management behaviors, none of the three clusters were completely homogenous with all good or all bad behaviors.

Drawing upon our preliminary cluster analysis, the data from Clusters 2 and 3 reveal two distinct profiles among patients with low adherence. Those in Cluster 2 had poor adherence to healthcare-related behaviors, such as medication adherence, missed appointments, and receipt of the flu vaccine, whereas those in Cluster 3 had poorer adherence to lifestyle behaviors, such as inhaler technique, physical activity, and diet. These distinctions are not perfect as those in Cluster 3 had a higher percentage of former smokers than those in Cluster 2; however, all patients in this study have a history of smoking.

Previous research has focused on patient characteristics to cluster patients and determine how health outcomes differ among clusters. Burgel, et al. used cluster analysis to group patients with COPD into five subgroups based on variables such as body mass index (BMI), FEV1, and number of COPD exacerbations in the last year as well as cardiovascular comorbid diseases and diabetes.[32] The clusters were then compared using each group’s all-cause mortality at three years.[32] Similar to Burgel, Ahlqvist, et al. studied Finnish and Swedish type 2 diabetic populations using five health-metric variables, including age, BMI, hemoglobin A1C, beta-cell function, and insulin resistance, to differentiate the participants into five clusters with distinct characteristics and varying levels of complication risk.[33] Burgel and Ahlqvist used patient characteristics to identify clinically meaningful clusters of patients, which ultimately risk stratifies patients based on inherent patient data.

Different from the research above, we sought to cluster patients with COPD and hypertension based on their self-management behaviors. While there are data on subgrouping patient populations based on health-related behaviors, there is little research on patient behaviors in the context of multimorbidity. Clustering with behaviors is complimentary to research that clusters patients by characteristics. Each type of study helps craft a holistic view of patient populations.

Ghanbari, et al. used latent-class analysis (LCA) to separate Iranian patients with hypertension into three classes based on four behaviors: diet, exercising, smoking status, and high blood pressure control.[34] The classes were labeled as low, intermediate, and high risk lifestyles.[34] Kino, et al. also used LCA to study patient behaviors and identified three classes in adult patients above the age of 18 from 27 European countries: healthy, moderate, and risky.[35] These classes were similar to those in Ghanbari’s study but were based on five health-related behaviors: smoking, alcohol consumption, fresh fruit consumption, physical activity and dental check-ups.[35] Of note, there was no specific disease of interest in this study.

Our research adds to the previous literature on clusters of health-related behaviors by examining behaviors in the context of multimorbidity. This is an important step toward identifying methods for better supporting people with multiple chronic illnesses. By identifying and targeting behaviors patients apply to more than one of their chronic conditions, researchers and clinicians may ultimately be able to achieve greater efficiency and success in supporting those with complex chronic illness.

While we did not cluster based on patient characteristics, we assessed whether there were differences in a small number of characteristics between our clusters. Clusters 1 and 3 had significantly larger proportions of college graduates, which may have selectively improved certain health-related behaviors. Clusters 2 and 3 were alike in that a greater portion of patients in these groups had comorbid diabetes and depression than patients in cluster 1. It is worth noting the combination of higher education with increased prevalence of diabetes and depression in cluster 3, which made it unique from both clusters 1 and 2. More research is needed to understand the role that these demographic factors play, especially the varying characteristics between those with poor health maintenance adherence versus those with poor lifestyle behavior adherence.

To our knowledge, the results presented above are the first attempt to identify clusters of self-management behaviors among patients with multimorbidity. Current behavioral research in patients with COPD is not as extensive as self-management behavior research in diseases such as cardiovascular disease, diabetes, and cancer,[36] and the relationship between COPD and comorbidities still needs to be refined.[37]

4.2. Limitations

Strengths of this analysis include the study of patients with multimorbidity using eight self-management behaviors. The cluster analysis is further enhanced with the use of CART. There are also some weaknesses that warrant discussion. The study approach was not theory-driven as no existing behavioral model was available to guide analysis. There is precedent, however, for data-driven analysis as evidenced by many of the studies referenced in the discussion. The study population only included patients with two comorbid diseases, COPD and hypertension, and it was limited to patients in an urban hospital setting. Therefore, the results may not apply to individuals with a different combination of comorbidities or to patients in sociodemographically discrete populations. We believe, however, that this is still an important population to study. Self-management activities may be unequally weighted in that certain behaviors, such as medication adherence, may have a larger impact on outcomes than other behaviors, such as receiving the flu shot annually. The use of weighting may promote increased homogenization of the clusters. Finally, our study is not validated in a second group of patients, but we believe our study provides an important precedent for future studies.

4.3. Conclusion

Patients with COPD and comorbid hypertension can be separated into those with adequate and inadequate adherence. The group with inadequate adherence can further be divided into those with poor adherence to medical behaviors compared to those with poor adherence to lifestyle behaviors. This paper provides an informative method with which to continue to explore an important aspect of research around managing multimorbidity.

4.4. Practice Implications

Our findings begin to uncover behavioral patterns in patients with multimorbidity. We believe our preliminary analyses need to be validated in other cohorts before we can form practical interventions. It is our hope that these studies may help physicians better characterize their patients for risk stratification in order to inform interventions and other applications.

Table 4.

Classification and Regression Tree Confusion Matrices

Training Set Accuracy: 84% (n=160)
		Predicted
		Cluster One	Cluster Two	Cluster Three
Actual	Cluster One	67	4	3
	Cluster Two	1	26	7
	Cluster Three	4	6	42
Test Set Accuracy: 86% (n=44)
		Predicted
		Cluster One	Cluster Two	Cluster Three
Actual	Cluster One	16	1	4
	Cluster Two	0	8	0
	Cluster Three	0	1	14

Highlights:

• Patients with multimorbidity create three distinct behavioral clusters

• A majority of patients perform self-management behaviors effectively

• Inadequate adherence can be differentiated by medical versus lifestyle behaviors

• Physical activity and medication adherence influenced cluster differentiation