There is a global epidemic of diabetes mellitus, with an estimated 10% of the global population, or 422 million people, affected. Around 90% of these people have type 2 diabetes mellitus (T2DM) [1]. The personal and health care costs of T2DM are substantial. Recent estimates suggest that about 11% of the total global health expenditure is due to diabetes [2]. Most of these costs are due to preventable complications [3]. Also, diabetes also results in societal costs, including the cost of missed workdays [4], personal costs, including out of pocket costs [2], and the impact on employment and earnings potential [5,6]. Patient education and self-management support has been identified as a priority for global health in recent years [7] and has the potential to both improve outcomes and reduce costs [8]. However, internationally, uptake of self-management education remains low [9], partly due to logistical problems with attending courses [10].

Web-based self-management support has the potential to increase uptake by overcoming some of the logistical problems associated with other forms of delivery as it can be accessed at home, at the user’s convenience. We have developed a comprehensive, evidence-based, theoretically informed, Web-based self-management program for adults with T2DM called Healthy Living for People with Type 2 Diabetes (HeLP-Diabetes). Overall content was guided by the Corbin and Strauss model, which hypothesizes that patients must undertake medical, emotional and role management in dealing with a long-term condition [11]. If effective, Web-based interventions have the potential to be highly cost-effective, as they can be delivered at scale across large populations, with relatively low additional costs per additional user [12], unlike telephone-based or face-to-face education where labor costs account for a substantial proportion of total cost [13].

We undertook an individually randomized controlled trial in primary care to determine the effectiveness and cost-effectiveness of HeLP-Diabetes compared to a simple, text-based website, all with the access to usual care for people with T2DM [14]. The aim of this paper is to present the health economic analysis of this comparison based on the data collected in the trial and to examine the cost-effectiveness of facilitated access to HeLP-Diabetes. The analysis was undertaken from the collective perspective of the National Health Services (NHS) and personal social services (PSS), following the National Institute of Health and Care Excellence (NICE) guidance [15], as the two share the same resources. The results on clinical effectiveness are reported in a separate article previously published [16]. All costs are presented in pounds sterling (£) 2014 prices.

Ethics approval was obtained from Camden and Islington National Research Ethics Service (NRES) committee, reference 12/LO/1571.

The HeLP-Diabetes trial was a multi-center, two-arm individually randomized controlled trial carried out in primary care settings in England, United Kingdom (UK). The detailed trial design was fully reported in the published protocol [14], and the clinical effectiveness article [16]. There were no changes to the methods after the protocol was agreed and the start of the trial. Twenty-one general practices from across England participated, with a mix of urban, suburban, and rural practices. Adults, aged 18 years or over, registered with participating practices, and diagnosed with T2DM were eligible for inclusion in the trial. People who were unable to provide informed consent (eg, due to psychosis or cognitive impairment), unable to use the intervention (eg, due to physical, sensory or intellectual impairment, or inability to understand basic spoken or written English), terminally ill, or currently participating in a trial of an alternative self-management intervention, were excluded. There were no exclusions based on the duration of diagnosis, level of diabetes control, previous experience of self-management education, computer and internet experience, or access to the internet at home.

Eligible participants were briefed on the trial by a practice or research nurse (See the Patient Information Sheet in Multimedia Appendix 1) and then individually randomized to either the intervention or control group using Web-based randomization independently of the trial team. Randomization was conducted in a 1:1 ratio using random permuted blocks of sizes 2, 4, and 6, stratified by recruitment center. Participants were informed the trial compared two forms of Web-based support but were blinded as which was the intervention and which the comparator. Each participant had access to their allocated intervention for 12 months after randomization. Follow-up was undertaken at 3 and 12 months postrandomization.

Recruitment took place between September 2013 and December 2014. Of 374 participants randomized, 185 were allocated to the intervention and 189 to the control group. The average age at randomization was 64.9 (SD 9.5) years old in the intervention group and 64.7 (SD 9.1) years old in the control group. There were four participants missing time since diagnosis of diabetes, two in each group. Among the participants with this information, the mean duration was 7.8 (SD 5.7) years in the intervention group and 8.2 (SD 6.1) years in the control group. The proportion of male participants in the intervention group (127/185, 69%) was similar to that in the control group (131/189, 69%). Further description of the participants’ characteristics can be found in the clinical effectiveness article [16].

In the intervention group, there were 143/185 (77%) participants who completed the three months self-report questionnaire and 129/189 (70%) participants who completed the 12 months one. In the control group, there were 152/189 (80%) participants who completed the three months questionnaire and 135/189 (71%) participants who completed the 12 months questionnaire. The difference in completion rate was not significant (Pearson’s chi-squared test P=.459 at 3 months, P=.718 at 12 months). There were rare cases where participants completed the questionnaire with one or two items missing (see Multimedia Appendix 3). In general, the missing data on self-report were due to not responding. The percentage of missing data was low and did not exceed 30% for any variable. The number of imputations was therefore set to 30.

The outcomes and the incremental cost and effectiveness of the primary analysis are presented in Table 3. The mean costs of health resources use in the 12 months trial period were higher in the control group than in the intervention group (for detailed health resources use, see Multimedia Appendix 3). The unadjusted difference in mean total costs was £12, with a lower value in the control group. When adjusted for baseline health resource use 12 months before the trial, the mean total costs in the intervention group were £131 (SE £169) higher than the control group. After further adjusting for baseline variables (ie, age, sex, history of cardiovascular diseases, smoking status, time since diabetes diagnosis), the difference was reduced to £111 (95% CI –156 to 362).

The mean PAID score was higher in the control group than the intervention group throughout the trial. The unadjusted difference in PAID score at 12 months was 3.1, with a lower score in the intervention group. After adjusting for baseline PAID score, the mean PAID score was 1.9 (SE 1.3) lower in the intervention group than the control group. This remained similar when further adjusting for other baseline variables (ie, age, sex, history of cardiovascular diseases, smoking status, time since diabetes diagnosis).

Mean EQ-5D-3L utility in both groups increased at three months from baseline. At 12 months, the mean utility value fell in both groups, but the fall was greater in the control group. Regarding QALYs, during the 12-month trial period, the intervention group had a mean QALY of 0.802 (SE 0.016) compared with 0.764 (SE 0.023) in the control group, giving an unadjusted difference of 0.038. After adjusting for EQ-5D-3L utility values at baseline and other baseline variables (ie, age, sex, history of cardiovascular diseases, smoking status, time since diabetes diagnosis), the incremental QALY was 0.020 (95% CI –0.001 to 0.044), comparing the intervention group to the control group.

The fully adjusted incremental cost-effectiveness ratio (ICER) was £58 (95% CI –411 to 587) per unit improvement on PAID scale, and £5550 (95% CI –21,077 to 52,356) per QALY gained, comparing the intervention with the control (Table 3). For both health outcomes, the ICERs indicated a costlier and more effective intervention. Comparing with the recommended WTP threshold for QALY, the intervention presented a much lower ratio. The majority of the bootstrapped replicates fell in the north-east corner of the cost-effectiveness plane, indicating a costlier and more effective intervention (Figure 1). A lesser proportion of the replicates fell in the south-east corner, indicating a less costly but more effective intervention. A small group (153/5,000, 3%) of the replicates fell on the left side of the Y axis, indicating a less effective intervention. Overall, the majority of bootstrapped replicates fell under the WTP thresholds. The CEAC further demonstrated the conclusion was likely to be robust with an 87% probability that the intervention was cost-effective at a WTP threshold of £20,000 per QALY and 92% at £30,000 per QALY (Figure 1).

A complete case analysis (CCA) was performed on 96/185 (52%) in the intervention group and 101/189 (53%) in the control group who had complete data on baseline variables required for adjustment, costs and the two health outcomes. Comparing the results from CCA to the primary analysis, the mean costs were lower in the CCA than in the primary analysis, except for the mean costs in the 12 months before recruitment in the intervention group (Table 4). Similarly, EQ-5D-3L utility scores at all time points were higher in the CCA than in the primary analysis, except for the score at baseline in the intervention group. The results of the CCA showed a smaller incremental outcomes (PAID –1.6, 95% CI –5.1 to 1.4 and QALY 0.010, 95% CI –0.018 to 0.044) and bigger incremental costs (£185, 95% CI –145 to 504), with higher variation surrounding each estimate, in comparison with the primary analysis. The ICER for a one unit improvement on PAID score was £116 (95% CI –1299 to 1690). The intervention was still cost-effective (ICER=£18,500/QALY, 95% CI -203,949 to 190,267), compared with the WTP threshold, but with a higher level of uncertainty (probability of being cost-effective was 45% at £20,000 and 52% at £30,000).

Assuming the effectiveness and the impact on health resources use remained the same, the ICER declined rapidly with increasing numbers of users (Figure 2). Once there were 81 users, the ICER dropped to around £20,000/QALY. The ICER became negative upon the intervention reaching 363 users, thereafter dominating the control (ie, the intervention became less costly and more effective than the control).

Figure 2. One-way sensitivity analysis of ICER in relation to the number of users

In this within-trial economic evaluation, we found that HeLP-Diabetes plus usual care was highly likely to be cost-effective, compared with free information-only websites, such as that provided by NHS Choices or Diabetes UK, plus usual care. The ICER in the primary analysis was estimated at £5550 per QALY gained with 87% and 92% probability of being cost-effective for WTP thresholds of £20,000 and £30,000 per QALY, respectively. Once there were over 363 users, HeLP-Diabetes became dominant (ie, less costly and more effective) compared to free information websites, on top of usual care.

The strength of this study comes from the reliance on actual data from a trial. We collected a wide range of information on health and social service resource use, including community, primary and secondary care. Most of the service use data came from the GP electronic health care records, supplemented by self-report data from participants for services not recorded in the electronic notes. This approach reduced the potential bias coming from retrospective recall by participants. The completeness of self-report sections was mostly secured by the mandatory questionnaire procedure which prevented skipping individual questions, therefore reduced missing data. Due to the volume of medication taken in this population, however, only the current medication at the time of data collection was extracted, and the assumption was made that all medications were for chronic use. This could overestimate medication costs, but it should not introduce a bias towards either group. The comparison between CCA and the primary analysis showed that the imputation had a greater impact on the control group than on the intervention group. The results suggested that the participants who completed all tests and questionnaires were likely to have a lower level of health services use and better health.

We decided at the outset not to include the investment costs incurred in the development of the intervention in our analysis. This was because, being funded by research grants, these costs were unlikely to re-occur if the intervention was widely implemented by the NHS, therefore being irrelevant to the decision maker for future planning. We used the running costs occurred in the trial to estimate the operating costs in practice. However, there may be more maintenance costs over a longer period as technology changes, and there may be a need for software updates. One example would be that adapting this program to optimize user experience on a mobile phone screen would incur additional programming costs. Neither PAID nor QALYs are clinical outcomes, and their association with outcomes such as glycated hemoglobin (HbA_1c) is not always straightforward. Although necessary for comparability with other interventions, their clinical relevance is limited.

According to National Diabetes Audit 2016-17 [31], 56% of registered T2DM patients are male, and 47% are under age 65 in England. While the mean age is similar, the proportion of male patients in our study participants is higher. What impact this might have on the overall effects of HeLP-Diabetes is unknown at this point, but it is worth noting that many face-to-face self-management interventions appear to appeal more to women than men, so the gender imbalance in favour of men seen in our trial may suggest that providing both face-to-face and online self-management interventions may be one way to reach both men and women [32]. Furthermore, the participants in our study were well-managed in their condition from the beginning, with mean HbA_1c of 7.3% (56 mmol/mol) at baseline. In a related point, the participants in our study had a duration of diagnosis ranging from 0-34 years. However, the effectiveness data from the trial showed that neither duration of diabetes nor baseline HbA_1c impacted on the overall change in HbA_1c [16].

Population impact of an intervention is a product of effect size, reach and uptake–a highly efficacious intervention that is only used by a very small proportion of the population may have less impact than a less efficacious intervention which is widely used. These trial data cannot give us an estimate of the eventual reach and uptake of this intervention, as trial participants are known to differ from the total population of patients who may be targeted by an intervention [33]. In parallel with the trial reported here, we undertook an implementation study to explore issues around reach and uptake, along with factors which impacted on these. These data will be reported separately.

A potential disadvantage of a rigorous evaluation is the time and resource required. There may be questions about the value of such investment in a rapidly evolving landscape where the intervention in question may be rendered obsolete by the time the evaluation is completed. However, while the digital technology changes rapidly, the underlying principles do not. Another potential limitation to consider is the importance of taking into account the target population of the intervention. It is possible that benefits will not scale across the population equally. For instance, in 2017, 37% of the adults aged 65+ in Great Britain read online newspapers or magazines, in contrast to over 70% of the adults aged under 55, whilst 20% of adults aged 65+ shared self-created content online, compared with over 50% in the age groups under 55 [34]. Interventions that have a wide target population will need to reflect the heterogeneity in preferences of their target populations, and the use and impact of the interventions may vary, affecting cost-effectiveness.

Although one of the major drivers for research into digital health interventions, such as HeLP-Diabetes, is their expected cost-effectiveness [12,35], there has been relatively little published evidence. A Cochrane review of digital interventions for alcohol consumption published in 2017 identified only 7/42 (16.7%) qualified studies reported economic evaluations [36]. Several systematic reviews of economic evaluations on mental health-related digital interventions conducted in recent years identified 5 studies for anxiety disorder [37], 16 articles for mental health in general [38], and 12 studies for depression [39], respectively. Other economic evaluations of digital health interventions cover a wide range of target conditions, including irritable bowel syndrome [40], substance misuse [41], weight management [42], insomnia [43,44], eating disorders [45,46] and postoperative recovery [47]. Results of these studies were favourable regarding costs, especially when wider health care or societal costs were taken into account but did not show a significant impact on quality of life during follow-up periods. Our results are therefore broadly in line with previous findings and thus contribute to the growing but still insufficient evidence pool of economic evaluation of digital interventions.

As there are 3.5 million people diagnosed with T2DM in Great Britain in 2016 [48-50], and over 90% of the households have internet access in 2017 [34], HeLP-Diabetes has the potential of delivering an effective intervention on a wide scale with negligible marginal costs. Although we do not expect internet-based interventions to be suitable for everyone at the moment, with the internet further permeating our daily life and people adapting to the internet era, there is potential for digital health interventions to help alleviate the burden of chronic conditions on health care systems in the long run. Our findings supported the cost-effectiveness of the intervention once taken up by patients. Further research is needed on how digital health interventions such as HeLP-Diabetes can be delivered and maintained in a sustainable and cost-effective manner, with the focus on user experience outside of a study setting. The successful realization of this effect might lie in identifying the more susceptible user groups and engaging them at an optimal time. More empirical studies are needed to help plan the systematic incorporation of digital interventions in medical practice.