Searches were undertaken in MEDLINE, EMBASE, Cochrane Library, PsycINFO, and Web of Science to identify eligible studies. The search strategy comprised 3 blocks: eHealth, medication adherence, and randomized clinical trial (see Multimedia Appendix 1 for the MEDLINE search). Reference lists and citations of included studies were checked to ensure literature saturation. Titles and abstracts were screened and full-text articles were assessed by 2 researchers (BP and BvdB or BP and JV) independently based on the inclusion criteria below. Discrepancies between researchers were resolved through discussion or by reaching consensus with the third researcher (BvdB or JV).

Eligible studies fulfilled the following inclusion criteria: (1) randomized controlled trial with a usual care control group; (2) applying an interactive eHealth intervention aimed at the patient or patient’s caregiver; (3) medication adherence as primary outcome; (4) a total sample size of at least 50 adult patients using long-term medication as determined by Zwikker et al [18]; and (5) published between 2014 and July 2019. Only peer-reviewed English full-text articles were included. We considered all interventions solely applied over distance as eHealth interventions (eg, online portals, telephone calls). Blended care interventions, where face-to-face contact is combined with online components, were excluded. Interventions were considered interactive if there was tele-feedback regardless by whom on medication adherence (eg, bidirectional text messaging, sending adherence reports). Only validated medication adherence outcomes (ie, objective measures or subjective measures that have been compared to objective measures) were taken into account.

Two researchers (BP and JV) independently assessed the internal validity of included studies using the Cochrane Collaboration’s tool for assessing risk of bias [19]. Seven domains were scored as having low (+), high (–), or unclear (?) risk of bias. Because blinding of participants and personnel is hardly feasible in studies evaluating interventions aimed at adherence, this domain was considered high risk (–) for all studies. Studies with a positive score (+) on at least five domains were considered high-quality studies. If relevant information was not reported, the corresponding author was contacted to request additional information. When no additional relevant information was provided, the risk of bias domain was scored as unclear (?).

A standardized template was made to extract data on study characteristics, eHealth interventions, and medication adherence outcomes. Details of the eHealth interventions were extracted according to the Template for Intervention Description and Replication (TIDieR) checklist [20]. Additionally, the mode of adherence tele-feedback of each eHealth intervention was extracted. We distinguished the following modes of adherence tele-feedback: monitoring device, SMS text messaging, interactive voice response (IVR), mobile app, call, or e-training. Intervention strategies were categorized based on the strategies defined by Lowe et al [21] to support behavior change (ie, strategies focusing on adopting treatment behaviors); to inform and educate; to support (ie, strategies assisting consumers with their medicines use such as peer support); to teach skills; to minimize risk and harms (ie, strategies focusing on preventing or managing adverse events); to facilitate communication or decision making; and to improve health care quality (ie, strategies improving, coordinating, or integrating the delivered care). Only the adherence outcome at study endpoint was extracted where magnitude of the intervention effect would be most apparent. For continuous outcomes Cohen d for assessing effect size was calculated if means and SDs were provided [22]. Dichotomous outcomes were recalculated to odds ratio (OR) where OR ≥2.0 is deemed to be a strong OR [23]. Additionally, if the authors reported a statistical significance favoring the intervention group compared to the control group, this was scored positive (+). A negative score (–) means there was no statistically significant difference to report. Data were extracted by one researcher (BP) and checked for accuracy by a second researcher (JV).

Statistical data pooling was not feasible due to heterogeneity between studies and interventions. Therefore a best evidence synthesis was performed to examine the effectiveness of interactive eHealth interventions on medication adherence. The Cochrane Back Review Group defines 4 levels of evidence: strong, moderate, limited, and conflicting evidence [24]. Strong evidence reflects consistent (ie, 75% or more of the studies report results in the same direction) results among 2 or more high quality studies. Moderate evidence reflects consistent results of 1 high-quality study and 2 or more lower-quality studies. Limited evidence reflects the result of 1 lower-quality study. Conflicting evidence reflects inconsistent results among 2 or more studies. If there were 2 or more high-quality studies, the lower-quality studies were disregarded in the best evidence synthesis. A post-hoc sensitivity analysis was performed to examine the robustness of the best evidence synthesis using a more stringent cut-off score (6 out of 7 instead of 5 out of 7 domains graded as low risk of bias) for determining the quality of the studies.

Figure 1 shows a flow diagram of the literature search which provided a total of 9047 publications for screening and yielded 21 articles reporting on 29 interactive eHealth interventions [25-45]. One article, by Reese et al [34], reported on 2 studies. Five studies reported on more than 1 intervention [28,34,35,38,45].

Fifteen studies had a positive score on at least five domains and were regarded high-quality studies as shown in Figure 2. Two studies had the lowest score with 2 out of 7 domains scored as positive.

Over half of the studies (13/22) included long-term medication for cardiovascular disease, diabetes, or both. Seven studies focused on other, single long-term conditions, leaving 2 studies that looked at any long-term conditions where long-term medication was in use.

The smallest study reported on 70 participants at baseline and the largest study involved 21,752 participants. Because all studies were randomized, baseline characteristics of the different groups were generally the same.

Follow-up was short (ie, less than 6 months) in 11 studies and long (at least six months) in 11 studies. The primary medication adherence outcome of each of the studies was mainly assessed objectively using medication monitoring devices, pharmacy prescription data, and serum levels. The remaining 6 studies measured adherence subjectively with validated self-report questionnaires (eg, Immunosuppressant Therapy Adherence Instrument).

Twenty-nine different interactive eHealth interventions were evaluated as shown in Table 1. Most (21/29, 72%) interventions specified using a (mobile) phone for calling, SMS text messaging, or mobile apps.

Most (25/29) interventions were aimed at the patient, 3 interventions were aimed at the caregiver, and another was aimed at either patient or caregiver.

Sixteen interventions were provided through automated software without involvement of a health care professional: 6 mobile apps, 5 monitoring devices, 3 SMS text messages or IVR interventions, and 2 e-training modules through an online portal. Another 7 interventions were provided through automated software in combination with tele-feedback by a health care professional or caregiver: 4 monitoring devices, 2 IVR or SMS text message interventions, 1 e-training. The 6 remaining interventions were telephone calls performed by health care professionals.

Regarding intervention strategies, nearly all (23/29, 79%) interventions aimed at informing and educating patients and just over half (15/29, 52%) sought to support patients by providing assistance and encouragement. All other strategies (eg, teaching skills, facilitating communication or decision making) were less frequently applied (see Multimedia Appendix 2).

Overall, 17 interventions yielded a statistically significant improvement of medication adherence compared to the control group (Table 2). For 14 of these interventions an effect size (Cohen d) could also be calculated; 2 interventions reported a large effect size (Cohen d ≥ 0.8) [25,29], 4 had a medium effect size (Cohen d ≥ 0.5 < 0.8) [35,43,45], 3 had a small effect size (Cohen d ≥ 0.2 < 0.5) [26,32,35], and 5 interventions had a negligible effect size (Cohen d < 0.2) [27,30,37,38]. For the remaining 3 interventions an OR could be calculated which showed strong odds for becoming adherent in the intervention group (OR ≥ 2.0) [36,39,42].

Further details of the study, population, intervention, and outcomes can be found in the extraction database provided as Multimedia Appendix 5.

The best evidence synthesis (Table 3) showed strong evidence for a positive effect for SMS text messages or IVR, mobile apps, and calls as mode of adherence tele-feedback. The evidence for e-training was weak and for monitoring devices conflicting.

In the post hoc sensitivity analysis the criteria for a high-quality study were more stringent (6 out of 7 domains graded as low risk of bias). The sensitivity analysis showed that the strong evidence for a positive effect for SMS or IVR as mode of adherence tele-feedback remained, whereas the evidence turned to conflicting for interventions delivered through mobile apps and calls (see Multimedia Appendix 3).

The level of evidence of the intervention strategies was also assessed. There was strong evidence for a positive effect of strategies to teach skills, to facilitate communication or decision making, and to improve health care quality. For all other intervention strategies (eg, to support, to inform and educate) there was conflicting evidence (see Multimedia Appendix 4).

This systematic review examined the effectiveness of interactive eHealth interventions to improve medication adherence in patients using long-term medication published between 2014 and 2019. A majority, 17/29 interactive interventions, had a statistically significant (P<.05) effect on medication adherence. There was strong evidence for a positive effect for interventions using SMS or IVR, mobile apps, and calls as mode of adherence tele-feedback. Intervention strategies to teach skills, to improve health care quality, and to facilitate communication or decision making showed a strong evidence for a positive effect. Overall, this review shows that interactive eHealth interventions can be effective in improving medication adherence.

This study showed strong evidence for a positive effect on medication adherence of eHealth interventions across various channels, including SMS, IVR, mobile apps, and calls. Our findings add robustness to the positive effect of eHealth interventions provided by previous systematic reviews and meta-analyses [46-51]. Where those authors were cautious with interpreting their findings because of low-quality studies, small sample sizes, and short follow-up, many studies we included were of high quality (22/29), had sample sizes of 100 patients or more (19/29), and follow-up of at least six months (14/29). IVR interventions that included information about health consequences suggest a stronger behavioral change, including medication-taking behavior [51]. This review confirms these findings as the included IVR interventions all contained information on the consequences of (not) taking medication as prescribed. For call, mobile app, and SMS text messaging interventions it remains unclear whether there are intervention elements (eg, content, intervention design, or extent of tailoring) that contribute to increased medication adherence because most eHealth interventions are multicomponent and elements vary widely across interventions [46,48,50].

We found a lack of convincing evidence for interventions applying an electronic monitoring device or e-training. By contrast, van Heuckelum et al [52] found a positive effect for interventions using monitoring device feedback. In our review all interventions coupled their electronic monitoring devices to the same (Way To Health) communication platform which could be a possible explanation. Yet, van Heuckelum et al [52] also included interventions that gave face-to-face feedback on adherence data collected by monitoring devices. They showed that these interventions were effective, whereas those that applied tele-feedback were not. This suggests that feedback on tele-monitoring of medication adherence is best given face-to-face.

To describe intervention strategies among effective interactive eHealth interventions we used Lowe’s taxonomy as it is specific for adherence interventions with clear examples for each strategy. Although other taxonomies (eg, Abraham and Michie [53], Demonceau et al [54], Kini and Ho [7]) could have been used, they show many conceptual similarities with Lowe’s taxonomy. Following Lowe’s taxonomy, we provide evidence for interactive adherence interventions aimed at teaching skills such as self-management programs, improving health care quality by coordinating medication adherence care between professionals, and facilitating communication or decision making between patient and health care professional. These results should be interpreted with caution because interventions were multifaceted and combined different strategies. Furthermore, the strategies with the highest level of evidence were also those that were less used. It is not possible to assign success to a single strategy within a multifaceted intervention. Nonetheless, the effective strategies we identified in this review suggest to be good starting points for development or selection of interventions.

Noteworthy, the included studies in our review using eHealth interventions to address medication adherence reflect 2 distinct patient populations, namely, the large patient population (eg, metabolic and cardiovascular disease) and the population where optimal medication adherence is critical (eg, immunosuppressants, antiretroviral therapy). Applying eHealth to address medication adherence can be advantageous for both populations albeit for different reasons. eHealth interventions can be accessible for large patient populations, giving health care professionals a large outreach with limited resources. For populations where optimal medication adherence is critical, eHealth interventions can be tailored to patients’ specific needs and provide continuous support.

Where others found a lack of high-quality studies and stressed the importance of improving study quality [3,9,15], this review included 15 (out of 23) high-quality studies. The increase in quality presumably is a direct consequence of better reporting and study designs. We chose the Cochrane risk of bias tool (version 1) to assess study quality. This tool focusses on internal validity and does not cover all aspects of study design. We found flaws in study design that were not covered by the Cochrane risk of bias tool such as absence of sample size calculation, selection bias, and disputable (adherence) outcomes. This could have (negatively) affected the implications of the results.

We clustered evidence of various long-term conditions in our best evidence synthesis to provide a comprehensive overview. This overview is based on the statistically significant effects (P<.05) found by the authors and supported by the effect sizes we calculated. Of the 17 statistically significant (P<.05) interventions, 9 showed at least a small effect size (Cohen d ≥ 0.2) and 3 interventions showed strong odds (OR > 2.0) for effect in the intervention group.

The synthesis was limited to medication adherence and did not consider other clinical outcomes. As a result, our findings may not be applicable one-on-one to specific conditions. Next step is to study the identified effective interventions/strategies in specific long-term conditions to ascertain that this may lead to improved medication adherence and other clinical outcomes.

Only postintervention effectiveness on medication adherence was assessed in this review. Whether the found beneficial effects will be maintained over a longer period (>12 months) remains unclear. However, 12/17 effective interventions in our review had a follow-up of at least six months which is considered the shortest period to accurately assess long-term medication adherence [4].

We were surprised to find many interactive eHealth interventions that use technologies published in the 20th century. Although technology changes, applied techniques are very similar. To be able to build upon data and lessons learnt from older technologies, crosslinks between similar techniques need to be made (eg, between SMS text messaging and chat services such as WhatsApp or WeChat).

Technological developments are very fast paced and eHealth interventions continuously change. This high turnaround speed creates a need for study designs that allow continuous evaluation of interventions over a period of at least six months.

In this review intervention exposure ranged from a single call to daily messages for months. To establish a relation between exposure and medication taking behavior change, dose–response studies are called for.

We found that a majority of interactive eHealth interventions are effective in improving adherence to long-term medication. Intervention strategies that improve patient’s treatment involvement and their medication management skills are most promising. While most interactive eHealth interventions were multifaceted, even simple eHealth technologies such as SMS text messaging and telephone calls can be effective in promoting medication adherence in a wide variety of patient populations.