RELEASE-HF study: a protocol for an observational, registry-based study on the effectiveness of telemedicine in heart failure in the Netherlands

Clinical effectiveness

Descriptives are presented for all variables of the HF registry and telemedicine components. We will present the number of days without unplanned hospitalisation for HF during 365 days after baseline, the average duration of unplanned HF-related hospital stays, the number of deaths and the time until death.

Preparation of the statistical model

As a first step, we describe how the overall effect of offering any type of telemedicine versus not offering telemedicine will be estimated. This analysis will subsequently be used for a comparison of telemedicine effects on days alive without unplanned HF-related hospitalisation during 1 year across subgroups.

The cohort is restricted to individuals with newly diagnosed HF, that is, at most 3 months prior to baseline. The exposure contrast is defined as offering any type of telemedicine after first presentation with HF at the hospital outpatient clinic versus patients not offered telemedicine. The telemedicine status for the entire follow-up is defined by 6 months after baseline measurement of telemedicine, used as an observational analogue of an ‘intention to treat’ analysis in which the stopping of telemedicine is disregarded.

Because telemedicine administration is dependent on patient characteristics, confounding by indication will be accounted for. Based on clinical consensus, the following patient-level confounding variables measured at baseline will be adjusted for: age, sex, SES, BMI, NYHA, LVEF, aetiology of HF, systolic blood pressure, diastolic blood pressure, heart rhythm, heart rate, chronic respiratory disorders, OSAS, stroke, extra cardiac arterial vascular pathology, DM, hypertension, anaemia, CKD, malignancy, heart rhythm disorders, depression and thyroid disease (see table 3 and online supplemental material 1).

The primary outcome model is a linear mixed model with a random intercept for the participating centre and is specified as follows. The outcome is measured in days, and because the final measurement may be less or more than 365 days after baseline, we rescale the number of days of unplanned hospitalisation to the number of days under follow-up to 365 days. The outcome is computed by subtracting the days of unplanned hospitalisation or death from 365. The earlier described confounding variables are added to the primary outcome model to adjust for confounding. Continuous covariates are modelled using splines. Multicollinearity between confounding variables and the telemedicine status is evaluated. The coefficient for telemedicine in the primary outcome model represents the average treatment effect of offering any telemedicine within 3 months after HF diagnosis compared with not offering telemedicine on days with unplanned HF-related hospitalisation during a year in the population of patients with newly diagnosed HF. Robust SEs are estimated to compute CIs.

For the secondary outcome functional status, a similar approach is taken to estimate the average treatment effect of offering any telemedicine within 3 months after HF diagnosis compared with not offering telemedicine on functional status after 1 year in the population of patients with newly diagnosed HF. For this analysis, the outcome model is a multinomial regression model, and baseline functional status is added as a covariate. The risk difference relative to the reference outcome category is assessed using the predicted outcomes of the primary outcome model offering of telemedicine versus not offering, to obtain the marginal risk difference.38–40 Similarly, for the secondary outcomes health status and hrQoL, the outcome model is a probit regression model, with baseline measurements added as a covariate. Healthcare utilisation will be assessed using a Poisson regression model, and all-cause mortality will be assessed using binary logistic regression.

We plan to perform the following additional analyses. First, we plan to perform the analysis with the start of follow-up defined at the T1 (after 6 months) measurement and the outcome of unplanned hospitalisation days during 6 months. This is because telemedicine status can be misclassified when it is started shortly after baseline but is not registered in the data. Second, we plan to perform a per-protocol analysis of telemedicine use if detailed information on starting and stopping of telemedicine is available through linkage. Finally, we plan to perform all analyses in the entire cohort, rather than the cohort restricted on HF diagnosis, to a maximum of 3 months prior to T0 (baseline).

Research question 1: effectiveness of telemedicine in patient subgroups

Research question 1 investigates the effectiveness of telemedicine across subgroups of patients with HF. Patient-related subgroups were identified by a systematic literature review of randomised controlled trials (RCTs) of telemedicine: age, severity of HF (NYHA class at baseline), sex, SES, presence of depression, atrial fibrillation and type of HF (HF with preserved ejection fraction (HFpEF), HF with midrange ejection fraction (HFmrEF), HF with reduced ejection fraction (HFrEF)) (table 3).20–24 41–49 Stratified analyses of the primary outcome model are performed to estimate the above-defined average treatment effect in each of the subgroups. Heterogeneity in telemedicine effect across age is assessed on a continuous scale, where an interaction between age and telemedicine status is added to the primary outcome model, and the expected number of days out of hospital across age ranges from 50 to 90 years is predicted from this model under telemedicine offered versus not offered.

In an additional analysis, heterogeneity across time since diagnosis is explored for the entire cohort, without restricting to patients with a maximum of 3 months since HF diagnosis at baseline (diagnosed ≤3 months before baseline compared with >3 months before baseline).

The secondary outcomes functional status, health status and hrQoL are assessed using similar approaches but using multinomial, probit and probit regression model, respectively. Healthcare utilisation will be assessed using a Poisson regression model, and all-cause mortality will be assessed using binary logistic regression.

Research question 2: effectiveness of different telemedicine components

Research question 2 investigates the effectiveness of the different forms of telemedicine intervention. We assess this question by performing three separate analyses in which intervention aspects are contrasted in the subset of participants that received telemedicine. The presence or absence of a telemedicine component is determined at a hospital level. The components of telemedicine that are contrasted are presence versus absence of a service centre, presence versus absence of an educational module, and high-intensity versus low-intensity protocol (table 2).

The population of interest is patients with newly diagnosed HF (maximum 3 months prior to baseline) who received telemedicine. Offering of telemedicine is measured at 6 months after baseline.

In addition to the set of confounding variables at the patient level used in the analysis above, confounding variables at the hospital level are considered because indication for telemedicine is expected to differ across hospitals. Confounding variables include academic or non-academic hospital, presence of a specialised heart centre, presence of an outpatient ward, size of outpatient ward (number of patients/year), staffing of the HF outpatient clinic and full-time equivalent of outpatient-ward staff.

For each of the three contrasts, the primary outcome model is fitted to estimate the average treatment effect of a component relative to the corresponding reference component on days of unplanned hospitalisation during 1 year in the population of patients with newly diagnosed HF that initiated telemedicine.

The secondary outcomes functional status, health status and hrQoL are assessed using similar approaches but using multinomial, probit and probit regression model, respectively. Healthcare utilisation and all-cause mortality will be assessed using Poisson regression model and binary logistic regression, respectively.

Cost-effectiveness

Costs will be estimated by computing the average costs (using cost guideline provided by the National Health Care Institute) per individual based on the information collected by the NHR and in the interviews with hospitals. Since 2023, Dutch hospitals can claim costs of telemedicine using the National Health Care Institute’s Diagnosis Treatment Combination (DBC). This DBC can be used to estimate the average costs.37 A 95% credible interval (CE) will be computed using the percentiles of a Monte Carlo bootstrap analysis with 5000 resamplings. This average cost represents a sum of the costs of telemedicine use, inpatient days, days at intensive care unit, HF-related hospital procedures and outpatient visits. In a secondary analysis, we will re-estimate the average costs and 95% CE using data enriched with VEKTIS data (if linking is available), meaning that the costs of visits to the general practitioner, pharmacy and care at home are also taken into account.

Linkage with CBS may be incomplete for several participants because of missing values or incomplete data sets (eg, twins, different available time windows between registries of collected data). If the number of non-linked participants is below 10%, we will perform multiple imputation; if it is above 10%, we will perform HTA analysis on the complete subset.

Utility values will be estimated using the average hrQoL score collected in the HF registry and, if needed, from the literature. The difference in average QoL score between the group that uses telemedicine versus the group that does not represents the difference in disutility between the groups under the assumption that the collected hrQoL is a good representation of the utility of the particular health state that individuals were in.

Quality-adjusted life years (QALYs) are estimated by multiplying the observed number of follow-up years by the corresponding hrQoL score. Subsequently, the incremental cost-effectiveness ratio (ICER) is computed by taking the ratio between the difference of the costs of the average patient in the telemedicine group and those not in the telemedicine group, and the difference in QALY of both groups, thereby providing the cost per additional QALY gained. The 95% CEs are obtained using the Monte Carlo percentile methods described above. Both deterministic and probabilistic sensitivity analyses will be performed to completely outline uncertainty on individual and combined parameters in the model.

In addition, willingness-to-pay curves will be drawn to highlight the impact of different thresholds on cost-effectiveness outcomes.

Research question 3: cost-effectiveness of telemedicine in patient subgroups

For research question 3, costs, utility, QALYs, ICER and willingness-to-pay curves will be computed to estimate the difference in cost-effectiveness between users and non-users of telemedicine in general. Subsequently, the analyses will be repeated to estimate cost-effectiveness in subgroups, similar to the groups defined for research question 2 in the clinical effectiveness analysis.

Research question 4: cost-effectiveness of different telemedicine components

For research question 4, costs, utility, QALYs, ICER and willingness-to-pay curves will be computed to estimate the difference in cost-effectiveness between components of telemedicine, similar to the comparisons defined for research question 3 in the clinical effectiveness analysis.

This post was originally published on https://bmjopen.bmj.com