Telemedical monitoring of patients with chronic heart failure has a positive effect on total health costs

AOK-Curaplan Herz Plus is a telemedicine program for patients suffering from CHF, offered by the statutory health insurance AOK Nordost in Germany. The program consists of regular telephone coaching and counseling, provision of information leaflets about disease-related themes, and a telemedical scale for weight monitoring. Telephone coaching and counseling is conducted by trained nurses in a telemedicine service center, specialized on medical services. Telephone contacts include feedback to conspicuous weight increase, talks to patient-individual topics and standard themes (e.g. diet, exercise, adherence to medication) [9]. Telephone contacts are conducted every 4 to 12 weeks, dependent on the patients’ individual needs. If needed, patients can contact the telemedicine service center any time. In case of a deteriorating health situation, the nurses give concrete recommendations or contact the treating physician. The frequency of the telephone calls is individually tailored to each patient. The program is implemented in routine healthcare, additional to usual care, and has no defined duration. AOK-Curaplan Herz Plus started in 2006 and is still running in the German federal states of Berlin and Brandenburg.

Usual care for patients with CHF is based on German and international treatment guidelines.

The intervention was implemented in a regular care setting. Patients with a diagnosis of CHF (ICD-10 diagnosis codes I50.12, I50.13 or I50.14 and NYHA class > I) with a high risk for a heart failure related hospitalization were informed about the program and invited to participate. High risk was defined as a prior heart-failure related hospitalization. Exclusion criterion was the presence of any diagnosed mental disorder at the time of recruitment. The recruitment of patients occurred during two phases: from 2006 to 2009, eligible patients were included by their treating general practitioners and cardiologists in private practices. Since 2009, eligible participants were retrieved from the database of the statutory health insurance AOK Nordost and were contacted directly by the health insurance company. All participating patients provided written informed consent.

To analyze the effect of the intervention on total health costs compared to an appropriate control group (intention-to-treat analysis), a quasi-experimental design was implemented [11].

Data from 2606 intervention patients and from 205,738 other patients fulfilling the inclusion criteria for the telemedicine program from the time period between the last quarter year of 2006 until the second quarter year of 2012 were retrieved from the database of the statutory health insurance. The data was on a patient-individual level, but completely anonymized, and, except for information about the federal state, without any regional data, so that it was not possible to re-identify single patients. The database included data to all matching variables and outcomes.

The control group was compiled using a combination of exact matching and propensity score matching (see section “matching” for further details). The control patients were retrieved from the total database of the statutory health insurance. For patients of the intervention group, the quarter year of enrollment served as baseline. Figure 1 shows the study design with baseline and follow-up analyses.

In a subsequent sensitivity analysis, the subgroup of patients with a documented start of the intervention was examined (treated analysis). Because of the highly variable time periods between recruitment and start of the intervention throughout the patient group, we conducted a separate matching procedure for the treated analysis using the quarter year of the start of the intervention as baseline quarter year.

Both matching procedures and statistical analyses were based on reimbursement data provided by the AOK Nordost. To analyze the effects of the intervention multivariate regression models were fitted including all matching variables and an array of potential confounders [12].

The control group was retrieved from the total database of the statutory health insurance. The base population consisted of all patients with a diagnosis of chronic heart failure. We combined exact and propensity score matching using a nearest neighbor method (greedy algorithm) with restrictions in selected covariates (matching without replace, two controls per case at the maximum) [13].

Intervention patients were matched for their quarter year of enrolment (respectively, the quarter year of starting the intervention in the sensitivity analysis) and controls were drawn dynamically. Dynamically means that controls could be matched to an intervention patient every quarter year as long as they were insured, still alive and not matched in a prior quarter year to another intervention patient.

The following variables had to match exactly: sex, 5-year-age group, NYHA class, number of hospital admissions in the 12 months prior to enrolment, cost category (23 categories) and the presence of any mental or behavioral disorders (ICD-10: F00 – F99). Patients with missing information in any of these variables were excluded from the matching procedure. The following parameters were included as propensity score matching criteria: cardiovascular medication in 6 groups of active agents in the baseline quarter year (angiotensin-converting enzymes, beta-blockers, renin-inhibitors, cardiac glycosides, diuretics, and AT1 receptor blockers) and mental and behavioral disorders in 11 diagnoses groups in the baseline quarter year). Mental illness was considered as a disabling factor for successful program participation. However, a considerable number of patients with mental illness were enrolled. Accordingly, mental disorders had to be included in the matching procedure.

The statutory health insurance uses an algorithm to exclude patients with extreme costs and to cap costs exceeding a defined threshold in their internal analyses to prevent that single outliers bias the results. The underlying assumption is that very high costs are caused by events that are not associated with the disease (e.g. accidents). We followed this algorithm and excluded patients with health costs over 100,000€ per year. Matching partners of high cost patients were also excluded. Additionally, in all remaining patients, inpatient health costs were capped to a maximum of 20,000€ per year, costs for home health care to 15,000€ per year and medication costs to 7000€ per year.

The primary analysis was an intention-to-treat analysis which included all matched patients. After matching, all matched patients received a statistical weight to account for the fact, that for some patients only one matching partner could be assigned (1 for 1943 matched intervention patients, 0.957 for 3552 patients of the control group who were matched to two partners in the intervention group and 1.914 for 167 control patients for whom only one matching partner could be identified).

In the baseline analysis, group differences in continuous variables were compared using the t-test, the chi-square test was used to compare categorical variables.

Primary endpoint of the analysis was the average of the total health costs over the four quarter years of the first year after recruitment. If patients survived less than four quarter years after baseline, we calculated the average of the quarter years when patients were still alive. The analysis of the two-year follow-up proceeded the same way with a maximum of eight quarter years.

The analyses were conducted using multivariate linear regression models. All models were adjusted by all baseline matching variables and by the residential area of the patients (dichotomized as Berlin (= urban) and other (= rural)). An interaction term “study group x residential area” was included in the models because bivariate analyses indicated an effect modification of the intervention of the place of residence on the total health costs.

Statistics were calculated using R statistical software, version 3.0.0 (R Foundation for Statistical Computing, Vienna, Austria). We used the R packages “MatchIt” [14] for data matching, “survey” [15] for weighted descriptive analyses, and “stats” [16] for regression models.

Patients were excluded from the regression analyses for the following reasons: 1) missing follow up data (health cost data were available until the fourth quarter year of 2011), 2) patients changed their health insurance company during follow-up time, 3) extreme health costs above 100,000€/year either of a patient (see methods for details) (Fig. 1).

One thousand nine hundred forty three intervention patients (74% of the patients in the intervention group) could be matched to 3719 control patients. Most frequent cause of failing matching was missing information about the NYHA-class (60% of the 679 not matched intervention patients).

Table 1 shows the distribution of parameter-values for the quarter year of enrolment in the intervention program for patients of the intervention group, and the quarter year of matching for patients of the control group (baseline quarter year). 54% of the patients in both groups were males; the mean age was 74.4 years. NYHA II and III were the most frequent NYHA classes. The total health costs per quarter year in the year prior to the baseline quarter year averaged 3747€ (SD: 3007€) in the intervention group and 3750€ (SD: 3021€) in the control group. 743 patients of the intervention group (38%) had at least one diagnosed mental disorder (ICD-10 Chapter V) in the baseline quarter year.

Intervention and control group were similar in most variables. Significant differences concerned the medication with AT1 receptor blockers, the number of emergency hospitalizations, the residential area, the level of care, and the proportion of patients living in a nursing home.

Figure 1 shows the flowchart of the analysis; 5433 of 5662 patients could be included in the one-year-follow-up analysis. The processed mean average quarterly health costs in the first year following the recruitment the total health care costs differ by 8€ overall between the intervention (2921€) and control (2929€) groups. Comparing the health costs of various subgroups of patients reveals considerable differences. As expected, costs are much higher in the group of patients who died in the year after recruitment. Rural patients caused lower health costs in both groups compared to the patients in Berlin. The crude difference between intervention and control patients was − 64€ in Berlin and − 259€ in rural regions, indicating effect modification which needed to be appropriately addressed in the fully adjusted model. Average health costs in the first year after baseline were broken down to cost segments. Inpatient costs contribute the largest part of the total health costs in both groups (Table 2).

To represent the modification of the intervention effect by place of residence of the patients (urban or rural) we added the interaction term “study group x residential area” into the regression model. Table 3 shows the regression model for total health costs in the first year of follow up. This model was adjusted for all matching variables, the patients’ residential area and the interaction between study group and residential area. Statistically significant determinants for lower health costs were the intake of a beta-blocker (p = 0.0014). Significant predictors for higher health costs were higher baseline health costs (p < 0.0001), the intake of diuretics (p < 0.0001) a diagnosis in the group of organic, mental disorders (F00 – F09, p = 0.0018).

Compared to the matched controls, total health costs in the first year after baseline were lower in the intervention group. The difference between both groups was − 276€ in rural regions and − 18€. Neither the parameters for the study group nor for the residential area and the interaction between them were statistically significant.

Table 4 shows the regression model for the two-year follow up (sensitivity analysis planned a priori). The model was adjusted for the same parameters as the model for the one-year-follow up. Statistically significant determinant for lower health costs was the intake of beta blockers (p = 0.0007). Higher baseline health costs (p < 0.0001) and the intake of diuretics (p = 0.0003) were significant determinants for higher health costs. The intervention group showed lower health costs in rural regions (− 299€ per quarter year) but higher health costs in Berlin (+ 109€, compound effect of study group (− 299€) and interaction of group and residential area (+ 407€)) compared to the control group.

In a sensitivity analysis, the total health costs per quarter year of patients with a documented start of treatment were compared to a separately matched control group. Here, only the one-year follow-up was analyzed because data of a large part of the participants were not available for two-year follow-up. The intervention group showed lower health costs in rural regions (− 551€ per quarter year) and in urban regions (− 267€, compound effect of study group (− 551€) and interaction of group and residential area (+ 284€)) as well.

A strength of this study is the large number of participating patients and its closeness to regular health care. All data used for the analysis were retrieved from original reimbursement data. Hence, there was no systematic difference in the data structure or data assessment between the patients in the intervention, and in the control group. The documentation of total direct health costs was complete and valid comparisons could be calculated. Since the intervention was conducted in a routine care setting, the results are likely transferable to the entire non-institutionalized CHF population of this statutory health insurance.

The study has several limitations. By applying propensity score matching we can only approximate the advantages of randomized trials. Data on social demographic variables were not available and despite a comprehensive two step matching process some residual confounding may have persisted due to unobserved variables. We could only match patients whose latest diagnose of NYHA class was within the timeframe of the available data (but before the date of matching).

A large part of the patients did not meet the inclusion and exclusion criteria because they had a diagnosed mental disorder at baseline. Possible causes were 1) participating cardiologists were not aware of diagnoses of mental disorders with their patients, 2) exact dates of diagnoses are not available in the data base, the data is stored on the basis of quarter years, and 3) data is reported retrospectively by the providers of medical services.

The large number of patients with mental disorders may have had an influence on the effect size of the intervention. On the other hand, including patients with mental disorders meets the real composition of the patient group and increases the external validity of the results.

The follow-up time is relatively short. Results might be inflated [22, 23] since long term effects like regression to the mean could not be analyzed. Interventions to maintain good adherence for long-term lifestyle changes can be very challenging [24]. The health insurance could not provide information on the number of telephone contacts per patient. As part of the program some patients had received additional services (such as provision of a modem-connected scale), but this was unknown for the individual patients. Finally, only data of one health insurance were available, therefore the results can only be representative for the member population of this statutory health insurance, not for the entirety of CHF patients.

The analysis of the telemonitoring program of the AOK Nordost and the results of other programs show that telemonitoring programs do not increase costs of treatment. This opens good opportunities to integrate telemonitoring programs in the treatment and monitoring of patients with CHF. Especially in rural regions with large distances to providers of medical services, this kind of intervention can ensure or improve the quality of healthcare.