Heart failure algorithms for remote monitoring in people with cardiac implantable electronic devices

The committee noted that the heart failure algorithms needed to be considered independently of each other because they are each unique and have different alert rates and levels of accuracy. Each of the heart failure algorithms collects different data types to monitor decompensation or predict a person's risk status. The committee noted that CorVue collects only intrathoracic impedance data, while the other algorithms monitor additional factors (see sections 2.5 to 2.11). The committee noted that the prognostic accuracy of CorVue may be affected by the collection of only 1 data type.

For the heart failure event endpoint defined by the Framingham Heart Study, CorVue had a sensitivity of 68%. For endpoints related to hospitalisation, clinic visits and changes to treatment, sensitivity ranged from 20% to 61.9%. This suggests that people who are experiencing decompensation may not have an alert triggered using CorVue. Clinical experts noted that heart failure algorithms should have a high sensitivity. This is to ensure that people with early signs of a heart failure event can be identified, assessed and have treatment if necessary, and so people are not missed. False positives were also considered to be high in all studies reporting this outcome. All studies were also assessed by the external assessment group (EAG) as being at a high risk of bias, with a key concern being the analysis methods. The committee concluded that CorVue cannot accurately predict heart failure events.

Evidence from a single published study suggested that, at the nominal threshold of 4.5, HeartInsight had 65.5% sensitivity and 86.7% specificity for the endpoint of first post-implant heart failure hospitalisation. For the endpoint of heart failure hospitalisation, outpatient intravenous intervention or death, HeartInsight had 54.8% sensitivity and 86.5% specificity. The positive predictive value was reported as 7.7%, indicating that there is a high probability that an alert is a false positive. This study was also assessed by the EAG as being at a high risk of bias, with a key concern being the analysis methods. The committee concluded that it is uncertain whether HeartInsight can accurately predict heart failure events.

Using the study endpoint of worsening heart failure, HeartLogic showed sensitivity ranging from 70% to 100%, and specificity ranging from 61% to 93%. False positives and unexplained alert rates were generally low in 6 studies. Statistically significant associations were observed between being in alert and hospitalisations, length of hospital stay, rate of heart failure events and rate of emergency care visits. All studies reporting prognostic accuracy data for HeartLogic were assessed by the EAG as being at a high risk of bias with the analysis methods being a key concern. The committee also considered data from a study that was published after the EAG's review was complete. Singh et al. (2024) evaluated HeartLogic in 1,458 people. For the endpoint of detecting heart failure events, HeartLogic demonstrated sensitivity of 74.5% and false-positive rate of 1.48 alerts per patient-year. The committee concluded that while there are some concerns regarding the quality of the prognostic accuracy data, it is likely that HeartLogic can predict heart failure events.

Using the study endpoint of worsening heart failure in people with a 'high-risk' status, TriageHF demonstrated sensitivity ranging from 87.9% to 98.6% and specificity ranging from 59.4% to 63.4%. The false-positive rate was reported in 1 study and was considered to be low. Most studies reporting prognostic accuracy data were assessed by the EAG as being at a high or unclear risk of bias. A key concern was the analysis methods, and some of the studies were only available as abstracts that contained limited information. The committee concluded that while there are some concerns regarding the quality of the prognostic accuracy data, it is likely that TriageHF can predict heart failure events.

The committee noted that heart failure algorithms are intended to be used to support review of heart failure data by healthcare professionals, and should not be used in isolation to make treatment decisions. This is because events other than heart failure decompensation can sometimes trigger an alert. For example, viral respiratory illnesses can increase a person's intrathoracic impedance, which could cause an alert to be triggered even if the person has no decompensation. In a study this would be classed as a false-positive alert. The committee noted that these alerts and the follow up with a person would still be valuable because a non-heart-failure clinical event may still require intervention.

Some studies reported high rates of false-positive alerts. But the committee noted that all alerts would be reviewed alongside other clinical information and discussed with the person before a treatment decision is made. So the committee considered that unnecessary treatment arising from false-positive alerts is unlikely and so harms from over treatment when using heart failure algorithms are expected to be low.

The committee discussed the impact that the number of false positives could have on service burden. Specialist committee members indicated that this burden is low in their experience, because initial interaction following an alert is usually via phone call rather than in-person. The committee concluded that more research should be done on the rate of false-positive alerts.

CIEDs may fail to transmit data if there are technical problems or connectivity issues, including internet problems, or if they are out of range of the transmission device. The committee noted that failure rates in the studies appear to be high, but it is difficult to separate transmission failure due to technical issues from transmission failure due to connectivity issues. If data transmission is missed, each algorithm has built-in retry mechanisms that will attempt transmission again. Healthcare professionals will be notified if a person's data transmission is missed for a number of weeks. The companies commented that they pay stringent attention to device failures and always follow up on these. High failure rates reported in Debski et al. (2020) have been addressed by ensuring that devices are correctly programmed and that local protocols are in place. The committee concluded that they have no concerns regarding transmission failure, because systems are in place to manage and resolve this.

A pairwise analysis approach was taken to estimating cost effectiveness. This was because of the lack of data comparing algorithms and because the comparator for each algorithm is a brand-specific CIED that is not using the heart failure algorithm. The EAG explained that the comparator costs differ for each pairwise analysis because different data sources were used to derive model inputs. For HeartLogic and TriageHF, evidence on hospitalisation rates was available, so different rates were used for these 2 interventions and their comparators. Because of the lack of evidence for CorVue and HeartInsight, no difference in hospitalisation was assumed between these heart failure algorithms and their comparators. The hospitalisation rate for these 2 algorithms and their comparators was assumed to be an average of the rates used for HeartLogic and TriageHF.

The EAG used comparative hospitalisation data to model the impact of heart failure algorithms rather than using prognostic accuracy data and a linked evidence approach. False-positive alerts were indirectly captured in the model for HeartLogic and TriageHF because study results on the number of unscheduled visits would be impacted by false-positive alerts. For CorVue and HeartInsight, no published data was available on unscheduled visits, so it was assumed that there was no difference in the number of unscheduled visits between these heart failure algorithms and their comparators. This may underestimate the impact that false-positive alerts have on the cost-effectiveness estimates for CorVue and HeartInsight.

The EAG's economic model showed that HeartLogic and TriageHF were more effective and less costly than standard care. The first probabilistic sensitivity analysis used probability distributions around mortality and intervention costs. This analysis showed that the probability of HeartLogic being cost effective was 81% at a threshold of £20,000 per quality-adjusted life year (QALY) gained and 73% at a threshold of £30,000 per QALY gained. The probability of TriageHF being cost effective was 85% at a threshold of £20,000 per QALY gained and 76% at a threshold of £30,000 per QALY gained. The committee noted that intervention costs would not be higher than the list price and so uncertainty around intervention cost should not be included in the probabilistic sensitivity analysis. The committee acknowledged the importance of considering the uncertainty around mortality rates, but noted limitations with how this uncertainty had been modelled. So, an additional probabilistic sensitivity analysis that excluded uncertainty around intervention costs and mortality was done. This analysis showed that both HeartLogic and TriageHF have a 100% probability of being cost effective at thresholds of £20,000 and £30,000 per QALY gained.

There was no comparative evidence on hospitalisations for CorVue or HeartInsight that was considered suitable for inclusion in the EAG's economic model. A conservative deterministic base-case model assumption was made of no difference in hospitalisations between CorVue and HeartInsight and their comparators. The results of this deterministic model showed CorVue and HeartInsight to be more costly than standard care and equally as effective. Threshold analysis showed that only a small reduction in hospitalisations would make these heart failure algorithms cost effective.

The EAG's model used published hospitalisation rates for HeartLogic and TriageHF. For HeartLogic, the hospitalisation incidence rate ratio of 0.28 was calculated from Treskes et al. (2021), which indicates a 72% lower rate of hospitalisations in the intervention group. For TriageHF, the hospitalisation incidence rate ratio of 0.42 was taken from Ahmed et al. (2024), which indicates a 58% lower rate of hospitalisations in the intervention group. These inputs resulted in the technologies being dominant (less costly and more effective than standard care) in the deterministic model base case. The committee noted that post hoc or subgroup analyses from these publications show that reductions in hospitalisations were smaller in some subgroups than in the overall study populations. For example, an analysis by Ahmed et al. that was limited to data collected before the onset of the COVID-19 in the UK, gave a hospitalisation incidence rate ratio of 0.69, indicating a 31% lower rate of hospitalisations in the intervention group. The committee also recalled the concerns regarding the quality of the comparative evidence for HeartLogic (see section 3.13) and TriageHF (see section 3.14) and noted that the risk of confounding may impact the reliability of the results from these studies. But, only small reductions in hospitalisations are needed for HeartLogic and TriageHF to be cost effective in the EAG's model. So the committee concluded that HeartLogic and TriageHF are likely to be cost-effective uses of NHS resources.

There was a lack of evidence for the impact of heart failure algorithms on heart-failure-related mortality rates and health-related quality of life. In the EAG's deterministic base-case model, conservative assumptions were made that there is no difference in heart-failure-related mortality rates between heart failure algorithms and their comparators. The committee noted that if there was an improvement in mortality or if health-related quality of life is greater when using heart failure algorithms, then there would be gains in QALYs. These potential benefits could not be captured in the model because of the lack of evidence.

The base case modelled 2 scheduled visits per year, in the intervention and comparator arms. In clinical practice, people who have stable heart failure would likely only have 1 scheduled visit per year. Other people may also only have 1 visit because of capacity issues. The EAG modelled 2 additional different scenarios for the intervention arms: no scheduled follow-up visits per year and 1 scheduled follow-up visit per year. These scenario analyses did not impact the direction of the model results.

In the EAG's base-case model it was assumed that all alerts are reviewed and followed by an unscheduled, in-person follow-up visit. In practice, alerts may be followed by an initial remote interaction (such as a phone call) to determine whether an in-person visit is necessary. Scenarios were modelled in which it was assumed that 25% and 50% of alerts initially require non-face-to-face follow up. The committee agreed that the base-case scenario and scenario analyses were reasonable. Scenarios were also done in which the base-case number of interactions in the intervention arm was doubled and quadrupled. The committee agreed that this assumption was reasonable. This was because the overall number of interactions is likely to be increased in the heart failure algorithm arm because of healthcare professionals reviewing alerts and following up with people remotely or in-person. Unscheduled emergency visits to an emergency department or primary care settings were not modelled. The committee noted that the number of emergency visits is expected to be lower for people with heart failure algorithms because alerts are intended to be triggered before the person experiences symptoms.

There was very limited evidence in people who have a CIED and do not have a diagnosis of heart failure but are at high risk of new onset acute heart failure. One study for TriageHF reported that a proportion of people in the cohort did not have a prior diagnosis of heart failure. But results were not reported separately for each population and so could not be used by the EAG to model the population of people at risk of heart failure. The committee agreed that data is needed on the prognostic accuracy and clinical impact of using heart failure algorithms in this population.

The committee noted that the evidence available for HeartLogic and TriageHF suggests that they can accurately detect the signs of worsening heart failure that could lead to hospitalisation or an unscheduled clinic event. The evidence also suggests that HeartLogic and TriageHF can reduce the number of heart failure events compared with conventional remote monitoring. But there are some concerns about the quality of the evidence for these algorithms and the size of the effects that could be seen.

To confirm the extent of the benefit seen in the studies, companies should work with the NHS to collect registry data for HeartLogic and TriageHF on:

• hospitalisation rates

• heart-failure-related mortality rates

• rates of emergency department or primary care visits

• patient-reported outcomes.

Algorithm-based remote monitoring systems are ideally positioned to reduce inequalities in access to healthcare. The committee heard that many people, particularly those from ethnic minority groups and lower socioeconomic backgrounds, do not seek medical assistance until they need to attend emergency services. Heart failure algorithms could benefit these people, because signs of decompensation would be detected and healthcare professionals automatically alerted, before the person needs to seek emergency assistance. For this reason, people who are unable to advocate for themselves or who have less awareness of their symptoms would also benefit from heart failure algorithms. Algorithm-based remote monitoring could also benefit people who are less mobile or people who live in remote areas. This is because following an alert, initial follow up may be a phone call to determine if an in-person follow up is necessary. This will reduce the need for unnecessary travel to hospital appointments.

Apart from the technologies that can use a landline to send data, access to technologies for remote monitoring may be restricted in some populations because of internet requirements. This may mean that older people, people from lower socioeconomic groups and people in rural areas could be less able to use algorithm-based remote monitoring because they do not have access to a Wi-Fi connection. The committee noted that the technology is incorporated into the person's CIED and does not need the person to engage directly with the technology themselves.

All of the algorithms can transmit data using Wi-Fi, and some using a landline connection. If people are not in range of a connection, their data will not be transmitted until they are in range. This may cause anxiety for some people when leaving their home, because they do not want to risk transmission of an important alert being missed or delayed.

The committee discussed that for heart failure algorithms to be used effectively in clinical practice, they should be used as part of a multidisciplinary specialist heart failure service and specialist staff should be available to review and action alerts. Protocols should also be in place to ensure heart failure is properly managed and alerts are responded to in a timely manner. At present, the way centres manage heart failure and respond to alerts can vary. This could mean that some centres are unable to implement heart failure algorithms into their services, which could lead to inequity of access across the country. Specialist committee members indicated that directives and initiatives are in place to steer heart failure services in the right direction for providing equitable access. Protocols should detail how heart failure alerts fit within the clinical pathway and how they should be responded to.