Although an extensive literature search was conducted, it is possible that some relevant studies may have been missed. However, such omissions are likely to have been minimal as the search included all identifiable publications in the grey literature (including contact with clinical experts in the field). The data were analysed by assuming a binomial likelihood function for the sample data. The statistical model acknowledged the fact that events accumulate over time by adjusting for the varying durations of
each study using a complementary log-log links function. Parameter estimates, including the between- study standard deviation, were estimated using Markov chain Monte Carlo (MCMC), which allows for uncertainty in the estimate of the between-study standard deviation; it also allowed the estimation of the predictive distribution of the effect of each intervention in a new study.
The clinical effectiveness findings had a number of limitations. In particular, the RM interventions were heterogeneous in terms of monitored parameters and selection criteria for HF. This was the case even within each of the four specific types of RM (STS HH, STS HM, TM with medical support during office hours, TM with medical support 24/7). Clear descriptions of the RM interventions were not provided in many of the studies included in the systematic review, making it difficult to understand exactly what was provided as part of the intervention. In addition, a number of trials were underpowered to detect the clinical outcome of interest and did not report blinding of outcome assessors. A limitation of the statistical model (as a consequence of having only one observation from each study) was that it assumed that the hazards and relative intervention effects were constant over time; nevertheless, this is better than assuming that study duration has no impact on the data. Moreover, because of the differences in the HF populations (e.g. definition of HF, LVEF inclusion criteria) of the included studies the true estimate of treatment effect may be unclear. However, the NMA analysis used a random-effects distribution together with 95% CrIs to reflect the uncertainty associated with the population mean. In addition, the predictive distribution of a randomly chosen study in the population was presented. This reflects not only uncertainty in the population mean but also the heterogeneity in treatment effects between studies. Unfortunately, it was not possible to model the heterogeneity between studies using a meta-regression technique because of the lack of suitable data on potential treatment effect modifiers.
The cost-effectiveness analysis has been undertaken assuming that the NMA results represent the best knowledge regarding the relative uncertainty between treatments. Therefore, although the treatment effects estimated from the NMA were statistically inconclusive, the joint uncertainty about these effectiveness parameters was used to populate the economic model. The expected values of costs and QALYs produced, which were used to estimate the cost-effectiveness of the RM interventions, thus are also aligned with the best knowledge on relative effectiveness. The uncertainty within the cost-effectiveness results was quantified by estimating the probability of each intervention being the most cost-effective at different WTP thresholds, and the EVPI was calculated to explicitly quantify the cost of reducing the decision uncertainty by undertaking further research.
Any limitations in the evidence base also manifest as limitations of the cost-effectiveness model. Most of the included studies in the NMA provided information on mortality and/or hospitalisation rates, which allowed synthesis using meta-analytical methods, but only a few studies reported any data about other potentially relevant states/events (such as stroke, having a pacemaker fitted), which did not extend to reporting any differences between the usual care and RM arms. Given the lack of evidence, it was deemed prudent to use a two-state Markov model even though it involved simplifications and assumptions that may not exactly reflect clinical practice. An advantage of using this simple model is that it can be easily updated to include other states or events should there be future evidence demonstrating differences between the usual care and RM arms.
A limitation of the cost-effectiveness model was that there was no age-specific analysis. Another limitation was that the constant hazards and relative intervention effects over time were applied to the time-dependent baseline mortality hazard (which is greatest in the early period after discharge after a hospitalisation for HF and subsequently declines over time) and constant risk of hospitalisation. If the studies reported observations at different time points, time-dependent effectiveness parameters can be estimated and used in the cost-effectiveness model. Furthermore, the optimal duration for each of the RM interventions can also be identified.
None of the studies identified in the review provided an estimate for the utility of the patients and whether or not there was a difference between the RM and usual care groups. Thus, in the economic model,
similar utility values were used for HF patients in both the RM and usual care groups; however, the validity of this assumption is unclear. Furthermore, the lack of detail provided in research studies concerning the components of RM packages and usual care (e.g. communication protocols, routine staff visits and resources used) made it difficult to estimate costs. Costing scenarios for different RM classifications were developed and their costs were estimated using microcosting methods. Although the users can decide which of these analyses is most representative of their setting, uncertainties still remain about the assumptions made in the estimation of these costs. This uncertainty in the costs is a limitation, especially as, given the small difference in QALYs between STS HH and TM during office hours, a small change in the difference between the cost of TM during office hours and the cost of STS HH can lead to marked changes in the ICER. A further limitation is that the effectiveness remained the same for the different cost scenarios whereas in reality there might be some correlation between the cost and the effectiveness of different RM strategies.