We applied ARM for the CKD population using the 2017 BRFSS data to generate rules for detecting lifestyle risk factors for CKD progression, including demographic information, lifestyle behaviors, clinical symptoms, and chronic disease conditions. Correlation analysis was performed to assess differences in lifestyle risk factors in the status of comorbidity-related CKD. To enrich the BRFSS data, SemMedDB was mined to identify lifestyle risk factors for CKD presented in publications. The workflow is shown in Figure 1. The arules package (version v1.6-4) for R (version 3.5.2) was used for ARM analysis.

We first selected 58 variables (involving 18 modules) related to behaviors from the BRFSS 2017 data by utilizing domain expert knowledge (from two nephrologists: S Peng and X Liu), with a focus on the presence of a condition or behavior rather than the questions about obvious feelings (as shown in Multimedia Appendix 1). If the given condition of interest was present in the patient, it was marked as 1, otherwise 0 or NA. For example, completion of the flu vaccine series was defined by a participant answering “yes” to the question: “During the past 12 months, have you had either a flu shot or a flu vaccine that was sprayed in your nose?” (possible answers were “yes,” “no,” “don’t know/not sure,” and “refused”). Only those who answered “yes” were annotated as 1 and included in the analysis. Records with responses of “no,” “unknown,” or “refused” were annotated as 0; those with missing data were completely excluded from the analysis to minimize underestimation. For each patient, we extracted all variables that were marked as 1 and prepared the input.

We then applied the Apriori algorithm [24] on the input data for 58 variables among 17,547 CKD patients. Apriori is a popular algorithm for mining association rules that is divided into two steps: (1) finding frequent itemsets and (2) constructing rules from frequent itemsets. An association rule is an implication between disjoint itemsets: m ⇒ n. The left-hand side of the rule is the antecedent and the right-hand side the consequent. An itemset containing k items is called a k-itemset. If T is a transaction, m is an itemset, and m ⊆ T, then T contains m. The support of the rule m ⇒ n is the fraction of transactions that contain both m and n (equation 1 in Figure 2). A frequent itemset is one whose support is at least some threshold, always denoted as minSup.

The rule m ⇒ n with confidence (equation 2 in Figure 2) means that the fraction of transactions in T containing m that also contain n is confidence. It measures how often items in m appear in transactions that contain n. Confidence can also be referred to as the strength of the rule. The threshold of confidence is always denoted as minConf.

Lift (equation 3 in Figure 2) is an index that indicates the relative magnitude of the probability of observing m under the condition of n, compared with the overall probability of observing m. When lift = 1, the two occurrences, m ⇒ n, are independent of each other. When the lift value is greater than 1, the two occurrences are dependent on one another; the higher the value, the greater the relevance of the interaction.

We used the following heuristic for generating the final association rules to be analyzed. We first selected itemsets with support value larger than the average of support values of all itemsets and lift value greater than 1. We then kept itemsets with lift value larger than the average lift values of those selected itemsets as our final association rules. We focused our analysis of the association rules of five NCDs as determined by the CCI: cardiovascular disease, chronic pulmonary disease, rheumatoid arthritis, diabetes, and non-skin cancer.

To assess differences of lifestyle risk factors in the status of comorbidity-related CKD, correlation analysis was performed. We retrieved five subcohorts of NCDs (cardiovascular disease, chronic pulmonary disease, rheumatoid arthritis, diabetes, non-skin cancer) from the CKD cohort and evaluated the contribution and correlation of lifestyle risk factors.

We mined SemMedDB for lifestyle risk factors that were not present in the BRFSS system by also using ARM. We first retrieved all relevant triplets (subject, predicate, and object) related to CKD. For example, <Diet Therapy, Treats, CKD>, <CKD, Coexists_With, Diabetes>, and <CKD, Causes, Hypertensive Disease> are some example triplets retrieved. We then selected terms from the triplets that were relevant to lifestyle behavior, symptoms, and diseases based on a list of relevant semantic types (see Multimedia Appendix 2). We filtered out 47 terms with generic meaning (eg, patients, agent, woman, child, author, disease).

We then applied the Apriori algorithm on the extracted pairwise terms to mine frequent itemsets and generate rules. Based on our previous work [25], our item matrix was very sparse, with a density of 0.00026. To mine sufficiently interesting rules, we set the minimum support and minimum confidence by making sure every rule was presented at least two times and selected itemsets with lift value greater than 1. We then kept itemsets with lift value larger than the average lift values of those selected itemsets as our final association rules. Specifically, we focused our analysis of the association rules on six specific semantic types (daily or recreational activity, food, hazardous or poisonous substance, individual behavior, mental or behavioral dysfunction, finding) to detect lifestyle risk factors present in publications that the BRFSS questionnaire does not mention.

Overall, a total of 17,547 participants were reported have CKD in the BFRSS 2017 data; 80.09% (14,053/17,547) were white and 60.13% (10,551/17,547) were men. The mean age was 64.42 (SD 13.81) years. The characteristics of the CKD cohort are presented in Table 1.

For heuristics, we set a lower bound of 0.1 for support and computed the average for all selected support values. As a result, we set the average support (0.150) as a threshold and selected 12,141 frequent itemsets. Among the 12,141 frequent itemsets, we then picked the average lift 1.094 as the threshold to finalize 7677 association rules. Figure 3 shows the curve between ranked associations and interestingness metrics (support and lift). The threshold was also marked on the curve.

Among the 7677 association rules, we retrieved subsets that related to five adverse conditions included in NCDs from CCI, including cardiovascular disease, chronic pulmonary disease, rheumatoid arthritis, diabetes, and non-skin cancer. For each of the input conditions, we then selected the top 10 association rules with the highest lift score regardless of whether the disease appeared on the left or right side.

From the top rules of each comorbidity, we determined that (1) CKD patients with comorbidity of cardiovascular disease have symptoms of high blood pressure, high cholesterol, asthma, function limitation, and lower aerobic and strengthening level; (2) CKD patients with a comorbidity of chronic pulmonary disease tend to have clinical manifestations of being overweight, hypertension, unhealthy diet (french fries or fried potatoes, less consumption of fruit and vegetables), and lower aerobic and strengthening level; (3) CKD patients with a comorbidity of rheumatoid arthritis are associated with hypertension, overweight, asthma, difficulty walking/doing errands alone, and less leisure-time physical activities; (4) CKD patients with a comorbidity of diabetes have a variety of clinical manifestations, including hypertension, high cholesterol, overweight, less leisure-time physical activities, and lower aerobic and strengthening level; and (5) CKD patients with non-skin cancer are associated with age (older than 65 years), asthma, less muscle strengthening, and lower aerobic level.

Examples of the top rules with the highest lift scores are shown in Table 2 (see Multimedia Appendix 3 for details on the results of the top 10 rules for subsets).

We conducted a correlation analysis using variables present in the top 10 rules for the five NCDs and CKD with a total of 25 variables. Figure 4 shows the heatmap of the correlation coefficient values of those variables with the five NCDs, CKD, and CKD with the comorbidities. Spearman rank correlation test was used for the analysis.

The correction analysis showed that people with NCDs including CKD have less physical activity in their leisure time and consume fewer fruits and vegetables. Hypertension, high cholesterol, age older than 65 years, male sex, difficulty walking, and difficulty concentrating had positive correlations with cardiovascular disease and rheumatoid arthritis but negative correlations with CKD comorbid conditions.

We set the support threshold to 0.00047 and the confidence threshold to 0.0001 to ensure every rule was presented at least two times, and then expressed the results with lift value greater than 1 by descending order to finalize important association rules. Among all 1323 association rules, 140 keywords from six specific semantic types were selected as a lifestyle word list (daily or recreational activity, food, hazardous or poisonous substance, individual behavior, mental or behavioral dysfunction, finding) to detect novel rules related to lifestyle risk factors present in publications. Multimedia Appendix 4 shows the top 20 rules.

Associations found using this method indicated that iron deficiency, depressed mood, sedentary lifestyle, and malnutrition were associated with anemia, hyperparathyroidism, obesity, and atherosclerosis, respectively, which the BRFSS questionnaire does not mention (Table 3).

The association rules indicated that CKD patients comorbid with cardiovascular disease are more likely to have symptoms of high blood pressure, high cholesterol, asthma, function limitation, and lower aerobic and strengthening levels.

The proper assessment of overall progressive risk in patients with CKD requires an adequate assessment of the presence and severity of other major risk factors. CKD is an independent risk factor for the development of cardiovascular disease; CKD is considered a cardiovascular disease risk equivalent [26,27]. Damaged kidneys may release too much renin, which helps to control blood pressure but increases the risk for heart attack, congestive heart failure (CHF), and stroke. CHF is responsible for up to 50% of deaths in patients with renal failure [3,28]. The signs and symptoms of heart failure include shortness of breath (dyspnea), fatigue, and weakness, consistent with our findings. The Physicians’ Health Study and other observational studies suggest that increased physical activity, higher cardiorespiratory fitness, and lower sedentary time are associated with reduced incidence of CHF [29]. Evidence shows that exercise training results in improved physical performance and functioning in patients with CKD [30]. Hence, the highlight of lifestyle modification of CKD with cardiovascular disease is to increase aerobic capacity by improving muscle strength or functional ability.

Findings also point to similar risk factors for CKD with chronic pulmonary disease or rheumatoid arthritis. The relationship between rheumatoid arthritis and chronic pulmonary disease (especially for chronic obstructive pulmonary disease) was found recently, in which people with rheumatoid arthritis were at 47 percent greater risk of hospitalization for chronic obstructive pulmonary disease than those in the control group [31]. Our finding supported evidence in the CKD cohort, the mechanisms that link CKD comorbid with the two diseases are speculative at present, which might be inflammation, autoimmunity, or genetic predispositions shared between them.

The lifestyle risk factors of these two comorbidities include hypertension, overweight, unhealthy diet (french fries or fried potatoes, less consumption of fruit and vegetables), and physical inactivity. Because the items did not determine exactly when symptoms of CKD or other NCDs originated, there are two possible interpretations of the result. One possible interpretation is that participants began reducing their physical activity and intake of fruits and vegetables when they developed CKD or chronic conditions. Symptoms of chronic conditions, such as hypertension, bone pain, peripheral neuropathy, side effects from medicines and fluid retention, itch, or sleep disturbance, can all negatively affect daily physical activity level, especially for CKD patients. Fruits and vegetables are a rich source of carbohydrates, vitamins, potassium, magnesium, and dietary fiber, whereas legumes and dried beans are important vegetable proteins. However, the limitation of potassium, fructose [32,33], or dietary protein intake has been common practice to control uremia. Despite the known benefits of fruit and vegetable consumption, intake remains poor in both the general and CKD populations [34].

An alternative interpretation is that lower vegetable and fruit consumption contributes to the development or maintenance of CKD or other NCDs. This interpretation has greater plausibility because it is consistent with other epidemiological studies and existing biological knowledge. However, fruits and vegetables should not be omitted from the everyday diet; this practice may lead to nutrient deficiency and low fiber-related constipation, which contribute to further accumulation of uremic toxins. The national “2 fruits and 5 vegetables” campaign guides Australians toward healthy fruit and vegetable consumption, which is applicable to CKD [35]. Also, regular participation in moderate-intensity exercise may enhance certain aspects of immune function and exert anti-inflammatory effects. Therefore, the lifestyle modification of CKD with chronic pulmonary disease or rheumatoid arthritis should be high dietary fiber intake and participation in moderate-intensity exercise to decrease inflammation and oxidative stress.

CKD is associated with insulin resistance and, in advanced CKD, decreased insulin degradation. In the association rules for CKD with diabetes, the results pointed to hypertension, high cholesterol, overweight, less leisure-time physical activities, and lower aerobic and strengthening as lifestyle risk factors rather than an unhealthy diet. Hence, the lifestyle modification of CKD with diabetes is consistent with the prevention of type 2 diabetes (predominantly exercise and weight loss), which can successfully decrease the development of CKD with diabetes.

CKD is recognized as a disease that may complicate cancer and its therapy (eg, immunotherapy). Cancer can cause CKD either directly or indirectly through the adverse effects of therapies; conversely, CKD may be a risk factor for cancer [36,37]. We found that age older than 65 years and physical inactivity were associated with CKD with non-skin cancer. The BRFSS questionnaire does not incorporate the therapeutics of cancer; therefore, the lifestyle risk factors of CKD with cancer cannot be evaluated in our research.

The BRFSS does not specifically target CKD or NCDs; therefore, many clinical manifestations were not considered, including potentially relevant items such as anorexia, nausea, vomiting, fatigue, anemia, and bone disease. To enrich the questionnaire, we used the SemMedDB to find lifestyles that related to the clinical manifestations specifically with CKD from publications. The results indicated that iron deficiency, depressed mood, sedentary lifestyle, and malnutrition are associated with anemia, hyperparathyroidism, obesity, and atherosclerosis, respectively, which the BRFSS questionnaire did not mention. CKD can affect a patient’s health-related quality of life in many ways. The diagnosis alone might cause fear or anxiety. Anemia, frailty, coexisting comorbidities, and depression are also major contributory factors to quality of life in CKD. Meat and meat alternatives are the main source of protein in the CKD diet. Healthy choices include lean cuts of meat, skinless poultry, eggs, fish, seafood, and plant-based protein foods such as legumes, dried beans, nuts, and seeds. The questionnaire of the BRFSS does not contain the variables of meat or protein consumption, nor does it contain information on micronutrient deficiency.

The results of the correlation analysis found that hypertension, high cholesterol, age older than 65 years, male sex, difficulty walking, and attention deficit disorder were positively correlated with cardiovascular disease and rheumatoid arthritis, but negatively correlated with corresponding CKD comorbidities (CKD with cardiovascular disease/rheumatoid arthritis). The ARM results suggest that patients with CKD older than 65 years are more likely to have signs or symptoms of hypertension, asthma, and difficulty walking, which is inconsistent with the aforementioned findings. It was caused by the differences between the two algorithms: a correlation is the relationship that exists between two or more variables in which a change in one variable causes a change in the other variable when the two variables are said to be correlated. Association rules are of the form {X₁, ..., X_n} → Y, meaning that if you find all signs or symptoms of X₁, ..., X_n in a disease it is possible to find another sign or symptom (Y). Epidemiological studies and existing domain knowledge are inconsistent with the result of correlation analysis but consistent with the results of ARM.

A wide range of disorders may develop as a consequence of the loss of renal function with CKD. These include disorders of fluid and electrolyte balance, as well as abnormalities related to hormonal or systemic dysfunction. Treatment strategies should be modified based on the needs of the individual patient. Variations and inconsistencies are inevitable in clinical practice; therefore, recognizing modifiable risk factors in medical interventions are important for providing effective chronic disease management. ARM has several applications in the medical domain, and it has been used for detecting risk factors for diabetes and cardiovascular disease [38,39]. This study illustrates how ARM approaches could be used in risk factor detection of CKD and provides the potential effectiveness of the method of ARM analysis for NCDs. ARM methods, such as Apriori, have also been used on electronic health record data to identify associations among clinical concepts. The strength of the ARM approach compared with a more conventional correlation analysis is that it has identified sizeable groups that can easily be defined and identified for intervention at a practice level in real time to allow more focused and immediate correction of bias in chronic disease management.

This research used a large representative sample, was based on items that asked about diagnosed disease, and included a number of relevant covariates; however, there are some aspects of this study that should be noted as limitations.

First, from 17,547 CKD patients, only 15,348 completed the interview. The dataset was skewed toward the white race and male gender, which may affect the generalizability of lifestyle interventions to other races and females. Other research found similar results [40] in which lower response rates (<40%) were associated with the underrepresentation of racial/ethnic minorities (eg, Hispanics), women, and younger individuals in the BRFSS survey.

Second, CKD and cancer can influence each other either directly or indirectly through the adverse effects of therapies. Since the BRFSS questionnaire was based on self-reporting, we cannot connect enough information. The lifestyle risk factors of CKD with cancer could be confirmed in further research using direct physical examination or biochemical indexes.

Third, the semantic predications consist of UMLS metathesaurus concepts as arguments, so we cannot tell whether “sedentary lifestyle” and “depressed mood” can be treated as “leisure-time physical activity calculate variables” or “ever been told you have depressive disorder.” As such, whether these differences are involved in observed associations for CKD needs to be considered in further epidemiological research. More robust observational or quasi-experimental studies would be needed to fully support the long-term impact of interventions for modifiable risk factors.

Finally, for the semantic predication triples extracted from the SemMedDB, we ignored the semantic meaning of the predicates and only kept subjects and objects as pairwise associations. However, we also found some predications with negative meanings. For example, the triples <Diet; Protein-Restricted, Neg_Treats, CKD>, <Dietary intake, Neg_Associated_With, CKD>, <End-stage renal failure, Neg_Coexists_With, CKD>, and <Ferritin level, Neg_Manifestation_Of, CKD> contain predicates with negative meaning, like not treats, have no association with, does not coexist with, or does not manifest. The reason we did not completely remove those triples is that we found inconsistency because both positive and negative relationships for the same factor may be reported. For example, according to a 2015 study conducted by Wong [41], a positive relationship between abnormal blood pressure and CKD was found in SemMedDB; however, in a 1992 study conducted by Taniguchi et al [42], a negative relationship between the same two items was detected. The SemMedDB only maintains information contained in the title and abstract; therefore, it is difficult to address inconsistencies without reading through the full text. In the future, we will count positive and negative associations for each pairwise term and assign weights for different predications for a better semantic representation.

This study related both lifestyle risk factors and CKD with five other comorbid chronic conditions using the largest national US survey available and provided a suggestion for BRFSS questionnaire enrichment. Various lifestyle risk factors result in the presence of different comorbid conditions for CKD patients, and different signs and symptoms may be observed. The findings illustrate how ARM approaches could be used in risk factor detection of chronic diseases to allow more focused and optimized chronic disease management.