This was a single-center, retrospective study that was approved by the Institutional Review Board of National Taiwan University Hospital, Taipei, Taiwan. Informed consent was obtained from all participants.

The study was conducted from December 2009 to April 2013 at the Telehealth Center of the hospital by the Taiwan ELEctroHEALTH Study Group (TELEHEALTH Study Group). Patients older than 20 years with diagnosed chronic CVD and receiving the telehealth program at our telehealth center were enrolled as the study group. The decision of whether to receive the telehealth program depended on the patients and/or their caregivers. Chronic CVD included coronary artery disease, myocardial infarction, heart failure, peripheral artery disease, stroke, and hypertension.

The definition of CKD is based on the Kidney Disease Outcomes Quality Initiative guidelines published by the National Kidney Foundation [14]. Chronic kidney disease is defined as a GFR of less than 60 mL/min/1.73 m². The renal function status is classified into (1) normal: no kidney damage, with a GFR ≥60 mL/min/1.73 m²; (2) CKD: kidney damage, with GFR <60 mL/min/1.73 m²; or (3) end-stage renal disease (ESRD): patient on maintenance dialysis, including hemodialysis and peritoneal dialysis.

The fourth-generation telehealth program at our telehealth center is a synchronized and integrated remote management program for chronic conditions. The Internet-based platform was developed by the Graduate Institute of Biomedical Electronics and Bioinformatics, National Taiwan University, Taiwan. The details of this program have been reported previously [15]. Briefly, this telehealth program provides the following services: (1) biometric data, including single-lead electrocardiography, blood pressure, heart rate, and pulse oximetry, which are transferred from patients to our telehealth center daily and on-demand; (2) nurse case managers communicate with patients daily and on-demand by telephone to ensure compliance to medication and medical instruction; (3) full-time nurse case managers and cardiologists are in charge of telehealth care 24 hours a day; and (4) long-term medications and management are discussed with the patients’ primary care physicians if acute events occur. This telehealth program emphasizes compliance to medication and medical instruction, as well as prevention and early detection of clinical deterioration.

All demographic and clinical data were obtained from the electronic database of the hospital. The diagnosis of a chronic disease was based on the electronic database. The discharge diagnosis was used if there was disagreement between outpatient and discharge diagnoses. The endpoints of this study were all-cause and cardiovascular hospitalizations.

Statistical analysis was performed using the R 3.3.1 software (R Foundation for Statistical Computing, Vienna, Austria). In statistical testing, a two-sided P value ≤.05 was considered statistically significant. The distributional properties of continuous variables were expressed as mean and standard deviation, categorical variables were presented as frequency and percentage, and the survival outcome curves were estimated using the Kaplan-Meier method. In the univariate analysis, the differences in the distributions of continuous variables, categorical variables, and survival outcomes were examined among three groups of patients (normal renal function, CKD, and ESRD) using the Kruskal-Wallis test, chi-square test, Fisher exact test, or log-rank test as appropriate for the data type. Next, multivariate analysis was conducted by fitting the Cox proportional hazards models with time-dependent covariates (called the “Cox model” for simplicity) to estimate the adjusted effects of risk factors, prognostic factors, and predictors on survival outcomes.

Contract compliance rate was defined as the percentage of days covered by the telehealth service duration within a certain time period before hospitalization. Specifically, we defined four time-varying variables of contract compliance for the time periods of 4, 8, 12, and 24 weeks, respectively, and then used them in fitting the Cox models. We assumed that the “incubation period” of the telehealth service was approximately 28 days; that is, the effect of the telehealth service might not appear until 28 days after signing or renewing the contract. The time-dependent contract compliance rates for 4, 8, 12, and 24 weeks were calculated using the following equation: contract compliance rates in n weeks = [(number of days covered by the telehealth program contract within n weeks before a specific event time – 28 days) / (7 days × n)] × 100%, where n=4, 8, 12, and 24. As shown in Figure 1, “a specific event time minus 28 days” is a specific event time moved 28 days backward to account for the assumed incubation period of the telehealth service.

We reorganized the original wide-form data into a long-form structure first, using the so-called “counting process style of input” for the survival outcome of interest, such as time to first hospitalization. Next, at each ordered event time of the survival outcome, we computed the values of the previously mentioned four time-dependent covariates for each of the patients at risk in the transformed long-form dataset. Finally, we fitted the Cox models to the long-form data with all relevant time-fixed and time-varying covariates using the coxph() function of the survival package in R.

To ensure the analysis quality, we used the model-fitting techniques for (1) variable selection, (2) goodness-of-fit assessment, and (3) regression diagnostics and remedies in our regression analyses. The stepwise variable selection procedure was applied to obtain the best candidate final regression model. The significance levels for entry and stay in the model were set to .15 to be conservative. The final regression model was identified manually by dropping the covariates with P>.05 one at a time until all regression coefficients were significantly different from zero.

Simple and multiple generalized additive models (GAMs) were fitted to detect nonlinear effects of continuous covariates and to identify appropriate cut-off point(s) for discretizing continuous covariates, if necessary, during the stepwise variable selection procedure. The vgam() function (with default values for the smoothing parameters) of the VGAM package was used to fit GAMs for continuous, binary, and count responses in R. Because GAMs were originally developed for smoothing the effects of continuous covariates in generalized linear models, we fitted GAMs of binary responses (eg, 1=hospitalization vs 0=no hospitalization) for our survival outcomes. In particular, we performed GAM analyses for the three groups of patients (normal, CKD, and ESRD) separately to detect heterogeneous nonlinear effects of continuous covariates in subgroups. Finally, the statistical tools of regression diagnostics for residual analysis, detection of influential cases, and check of multicollinearity were applied to discover any model or data problems. A variance-inflating factor value of 10 or greater in continuous covariates or 2.5 or greater in categorical covariates indicated the occurrence of multicollinearity problems among some of the covariates in the fitted regression model.

In total, 715 patients were enrolled in this study and divided into three groups according to the status of their renal function: normal (n=490), with CKD (n=178), or with ESRD and on maintenance dialysis (n=47). The mean age was 66.6 (SD 15.0) years and 66.0% (472/715) of the participants were male. The patients’ baseline characteristics, stratified by renal function status, are shown in Table 1.

The three groups were significantly different in terms of age and the proportion with hypertension, diabetes, heart failure, or peripheral artery disease, and the proportion receiving dialysis. The duration of telehealth use was different between the three study groups (longest in the group with CKD compared with the group with normal renal function and the group with ESRD on dialysis; Table 1). The median follow-up time was 694 days (IQR 338-1163). Because of the different follow-up times, events were divided according to follow-up time (days) in the subsequent analysis. During the follow-up period, there were 57 emergency room visits, 173 cardiovascular hospitalizations, and 350 all-cause hospitalizations. The Kaplan-Meier curves for time to first cardiovascular and all-cause hospitalization are shown in Figures 2 and 3, respectively.

We performed a univariate analysis using each variable associated with the first cardiovascular hospitalization (Table 2). Among all study participants, a history of diabetes (OR 1.74, 95% CI 1.20-2.54, P=.002), heart failure (OR 2.08, 95% CI 1.43-3.03, P<.001), peripheral artery disease (OR 2.7, 95% CI 1.36-5.34, P=.003), or hemodialysis (OR 3.13, 95% CI 1.37-7.18, P=.004) were associated with a higher risk of first cardiovascular hospitalization. The use of angiotensin-converting-enzyme inhibitors (OR 1.86, 95% CI 1.00-3.42, P=.03), metformin (OR 2.10, 95% CI 1.11-3.86, P=.01), spironolactone (OR 2.13, 95% CI 1.28-3.52, P=.002), or statin (OR 1.52, 95% CI 1.05-2.19, P=.02) were also significantly associated with the first cardiovascular hospitalization. Among patients with normal renal function, only heart failure (OR 1.84, 95% CI 1.09-3.08, P=.02) was associated with a higher risk of first cardiovascular hospitalization. The associations between medication use and hospitalization in patients with normal renal function were similar to the overall population, except in the case of statins. However, there was no association between clinical factors or the use of medication and the first hospitalization in patients with CKD or ESRD.

We used Cox proportional hazards models with time-dependent covariates (ie, the Cox model) to estimate the effect of predictors on first cardiovascular hospitalization (Table 3). The results showed that previous emergency department admission, peripheral arterial disease, the use of spironolactone, the use of statins or metformin in patients with normal renal function, the presence of atrial fibrillation in patients with normal renal function, and the presence of ESRD were associated with a higher risk of first cardiovascular hospital admission. In addition, contract compliance rate of the telehealth program among patients with CKD was also associated with a higher risk of first cardiovascular hospital admission. We used a GAM plot to assess the nonlinear relationship between contract compliance rate within 24 weeks and risk of cardiovascular hospitalization (Figure 4). The results showed that there was a triphasic, or U-shaped, relationship: patients with very low or very high contract compliance rates within 24 weeks were associated with a higher risk of cardiovascular hospitalization.

The Cox model for estimating the effects of predictors on the first all-cause hospitalization yielded similar results (Table 4). The presence of cancer and the use of spironolactone were associated with a higher risk of first all-cause hospitalization, whereas the presence of normal renal function and the use of fenofibrate were associated with a lower risk of first all-cause hospitalization. We used a GAM plot to assess the nonlinear relationship between contract compliance rate within 24 weeks and risk of all-cause hospitalization. The results showed that there was a triphasic relationship: patients with very low and very high contract compliance rates within 24 weeks were associated with a higher risk of cardiovascular hospitalization in all three patient groups (normal renal function, CKD, and ESRD; Figures 5-7). Notably, patients with ESRD had a lower threshold for the right arm of the U-shaped relationship compared with the other two groups; in other words, their risk of all-cause hospitalization was higher if their contract compliance rate was greater than 44.4% compared with 89.1% and 91.8% for patients with normal renal function and CKD, respectively.

The major findings of this study were as follows: (1) higher contract compliance rate with the telehealth program was associated with a higher risk of first cardiovascular or all-cause hospitalization; (2) for patients receiving the telehealth program, the level of renal function was associated with hospitalization and patients with ESRD were associated with a higher risk of first cardiovascular or all-cause hospitalization after adjusting for comorbidities and medications; and (3) there was an interaction between the effect of contract compliance rate and renal function on the time to first hospitalization. Patients with ESRD had a higher risk of hospitalization at a lower contract compliance rate, compared with patients with CKD or normal renal function.

In our previous study, patients receiving the telehealth program were associated with lower all-cause mortality than those receiving usual care [5]. Based on that study and these findings, we suggest that one reason for this was timely hospitalization for disease management. Patients receiving the telehealth program and exhibiting high contract compliance rate may be more vulnerable because of complex underlying diseases and comorbidity, and if so would require comprehensive care, with which the telehealth program would assist. Our results highlight the effect of contract compliance rate in patients receiving the benefits of the telehealth program.

The analyses of contract compliance rate and risk of hospitalization revealed a triphasic relationship. Patients with a very low contract compliance rate were associated with a higher risk of hospitalization. This implies that duration of contract to the telehealth program of close to or less than 28 days has no effect on the prevention of acute deterioration in disease. Patients with a midrange contract compliance rate were associated with a lower risk of hospitalization. This highlights the usefulness of the telehealth program for most patients. Unexpectedly, patients with higher contract compliance were associated with a higher risk of hospitalization. This last finding seems counterintuitive because most compliance studies show that higher compliance to pharmacotherapy is associated with a lower risk of complications [16-19]. The issue of compliance, however, has not been formally explored in telehealth studies and should not be taken for granted [20]. The effect of compliance to a telehealth program is difficult, if not impossible, to evaluate using a randomized controlled trial. Therefore, careful analysis of observational data is needed to explore this issue. In randomized controlled trials, reported compliance is usually good, with up to 85% of participants still using the telehealth program until the end of the study [21,22]. However, the impact of contract compliance rate on the risk for hospitalization in the real world has not been reported. Our report shows that with a low to midrange contract compliance rate, a higher contract compliance rate is associated with a lower risk of hospitalization.

A similar U-shaped curve (the “three-phase terrain”) of hospital admission has been observed in lifetime analysis of heart failure [23,24]. In that report, 30% of all cardiovascular readmissions occurred within the first 2 months of hospital discharge, and 50% occurred within the 2 months before death, with much lower admission rates (15%-20%) observed in the intercurrent plateau phase [23]. To achieve a longer plateau phase, patient care should extend into the home to monitor and maintain stability, and into active intervention for ambulatory patients who have early signs of organ decompensation [25]. Given the predictive or preventive scenarios for heart failure admission, the relationship between high admission risk and compliance with telehealth demonstrated in this study merits further consideration. There are two possible explanations for this result. First, we consider the predictive scenario [25]. Because participation in the telehealth program is voluntary and self-funded, contract compliance rate is largely determined by the needs of the disease process and the sense that these needs have been met as a result of participation in the program. Therefore, patients with greater disease severity remain in the telehealth program for longer. However, greater disease severity is also associated with an increased risk of hospitalization. Second, to prevent further deterioration of the clinical situation, the telehealth program might partly increase the rate of hospitalization if a decline in organ function cannot be managed in outpatient or emergency departments. We have shown in our previous study that our program is associated with lower all-cause mortality rates. According to the preventive scenario, the program increases the rate of hospitalization so that crises can be managed earlier, rapid deterioration can be halted, and overall outcomes can be improved.

Telehealth programs in CKD [26,27] have been less well studied compared with other chronic conditions such as heart failure, chronic obstructive pulmonary disease, or diabetes [2,28-31]. Knowledge about the efficacy and cost-effectiveness of telehealth among patients with CKD is still lacking. In a recent study, researchers showed that a telehealth program provided by an interprofessional team is a feasible approach for this patient group, but that it did not result in less hospitalizations, fewer emergency department visits, or lower all-cause mortality [26]. In our previous report, we showed that CKD was a predictor for all-cause mortality among these patients [5]. These results show that among patients receiving the telehealth program, renal function status was still a predictor for first hospitalization. Our program is specialized for risk factor control and adherence to recommended management guidelines among these patients. Determining whether telehealth improves the outcome of these patients requires further effectiveness comparison studies.

We observed a significant interaction between the effects of renal status and contract compliance rate on hospitalization. Among patients with normal renal function and CKD, a contract compliance rate greater than 90% was associated with a higher rate of all-cause hospitalization, whereas among patients with ESRD, a contract compliance rate greater than 44% was associated with a higher rate of hospitalization. It is possible that contract compliance rate is a reflection of patient’s awareness about their disease severity. The significantly lower contract compliance rate associated with increased hospitalization rate in patients with ESRD might suggest that, among ESRD patients, their awareness may be lower than their real need for medical help. This could be due to factors such as low awareness of the severity and long-term complications of their disease, low awareness of the importance of compliance to recommended management guidelines, or accessibility of the health care system via alternative care providers (eg, the dialysis center). Whatever the case, noncompliance to recommended management guidelines in this patient group is associated with a higher risk of long-term complications [32,33]. From the perspective of telehealth care providers, it would be prudent to ascertain that patients with ESRD exhibit good compliance to standard management. Renal status, therefore, is a major determinant for risk evaluation of patients receiving telehealth care programs.

There are several limitations in our study. Firstly, this is not a randomized controlled study comparing patients with and without access to the fourth-generation telehealth care program. Therefore, causation cannot be determined from our results. Secondly, there were fewer patients with CKD and ESRD (especially the latter) than with normal renal function. The effect of contract compliance rate was not statically significant for patients with ESRD, which may be due to limited patient numbers. Thirdly, our analysis did not account for recurrent events. It may be important to differentiate patients who are more likely to undergo several admissions from those with only one admission. Finally, our total patient number was limited. We have only listed the factors that were significant for predicting the first cardiovascular hospital admission in our final model; some predictors in the initial analyses were not statistically significant in the more comprehensive analyses (when the patients were divided into three groups). Thus, the interaction between clinical factors and each group may not be clear due to our limited patient numbers.

In summary, we found that contract compliance rate to the fourth-generation telehealth program is associated with cardiovascular outcomes, especially in patients with CKD. We demonstrated a triphasic relationship between contract compliance rate and risk of hospitalization among patients with CKD. Patients with low and high contract compliance rate to the telehealth program were associated with higher risk for hospitalization. A large, prospective, randomized controlled study is needed in these high-risk CKD patients to determine whether the telehealth program is beneficial.