The search was carried out in August 2018. To ensure exhaustive search results, cross-sectional databases in the fields of medicine, nursing care, public health, science, engineering, and social science were covered. The following databases were searched from 2000 to August 2018: PubMed, MEDLINE, CINAHL, EMBASE, EmCare, ProQuest, Scopus, Web of Science, ScienceDirect, and IEEE Xplore. We also included the Cochrane Library and gray literature sources, including Google Scholar and Open Access Theses and Dissertations. The snowball method was used to manually search citations within the studies included. We also hand-searched all the RCTs from the Journal of Medical Internet Research (JMIR) journals. Owing to the variation in terminology used to describe the topic of interest, we employed a broad, inclusive search strategy that covered the concepts of medication administration, eHealth, and nonhospital settings. Considering the appropriate Medical Subject Headings (MeSH) terms, we developed a set of master search terms that were applied to electronic data sets (enclosed in the Multimedia Appendix 1).

The inclusion criteria are presented as follows. We included all RCT studies that examined the effect of an eHealth intervention involving medication administration in a nonhospital setting. The periods for intervention needed to be at least six months. eHealth covers a broad range of mHealth, digital health, telehealth, electronic messaging, and electronic reminder interventions. Purely telephone-based outreach was not considered as health technology. We included studies in which participants needed to take medication regularly under nonacute settings. Only studies that focused solely on oral drug administration were selected. If the study did not specify a medication route or the intervention was administrated through a variety of routes (eg, [24]), it was not included. Studies were evaluated only when they reported, either directly or indirectly, on drug medication adherence, health care outcomes in adverse drug events, patient satisfaction, costs, or quality of life. Cluster and pragmatic RCTs were also included. No limit was applied to the databases in terms of article language. Studies that focused on participants with mental health problems (including depression, stress disorder, psychosis, and schizophrenia), or those who may have suffered from complex psychological issues often associated with serious illnesses (such as HIV/AIDS), were excluded because of their potential to weaken the study population’s representativeness and affect the generalizability of the results.

Two reviewers (one with a medical and public health background [BS] and another one with health technology and informatics background [ZSYW]) independently carried out title and abstract reviews in the screening phase. Both were well familiar with the aforementioned inclusion and exclusion criteria and evaluated the full texts of articles independently. Their results were compared, and Cohen κ was measured in the assessment of eligibility stage in order to evaluate the inter-rater reliability between the reviewers. For conflicting decisions, a consensus was arrived at between reviewers by discussing rationales and concerns and reexamining how each compromised decision satisfied the principles of inclusion and exclusion as outlined. The Rayyan web application [25] was used to facilitate the double-blind evaluations and to maintain review records. We developed a data extraction sheet for the studies included (Multimedia Appendix 2) that complied with the minimum standards of the Data Extraction Template for Included Studies, as developed by the Cochrane Consumers and Communication Review Group [26].

The details recorded (wherever available) for each included paper were as follows: general review information, study population, study characteristics, outcome measures, and results. We reported the primary outcomes and all the available secondary ones. Both qualitative and quantitative materials were extracted. We were interested in the frequency of the measurement outcomes of the intervention group (health technology) when compared with control (usual care practice). Risk difference, which is one of the most useful ways to present RCT research results, was used as a quantitative reporting measure. The outcome measures with the largest possible intervention timeframe for results were reported. The study protocol was published in the PROSPERO registry on August 7, 2018 (Registration number: CRD42018096627).

Medication adherence [27] refers to the degree to which patients’ medication-taking behavior accords with appropriate medical advice [28]; it was set as the primary outcome measure in this study. For the secondary outcome measures, we included indications of adverse event (or safety outcome), quality of life, patient satisfaction, and health expenditure/spending, as eHealth studies often assess these health care measures and considering them would allow us to evaluate eHealth impact more comprehensively.

It is common for eHealth applications to have multiple features. Referring to [4] and the range of capabilities of drug application [21], we compiled a list of commonly used health technology functions, namely, mechanisms to communicate with caregivers, monitoring health features, reminding and alerting, providing information and education, dispensing treatment and administration support, personalized feedback, reporting and trending, dynamic treatment adjustment, social support, and setting goals and planning. Based on the eHealth intervention described in each study included, we tallied the occurrence of these functions to determine how often the functions were applied in eHealth interventions. Considering the presence of each health technology function and improved medication adherence at P<.1, P<.01, and P<.001 as variables, we constructed 2 × 2 contingency tables to examine whether the proportions for different health technology functions were different. Fisher exact test of independence was employed.

Two reviewers (BS and ZSYW) also performed quality assessment for the studies included using the Cochrane Collaboration’s Risk of Bias (RoB) assessment tool via RevMan version 5.3 software [29]. Following the RoB assessment guidelines [30], the reviewers assessed the RoB signaling questions as either Low risk, High risk, or Unclear RoB, based on the evidence accessible from each of the studies included. A third-party opinion was sought from another coauthor (KDSL), as and when needed.

Figure 1 presents the flow diagram of the literature search (JMIR hand-searched result is appended in Multimedia Appendix 3). After removing duplicates, our initial search yielded 9909 records, of which 92 were reviewed for full-text assessment and 24 satisfied the inclusion criteria and were included in this review. Cohen κ between the reviewers was 0.846, which is equivalent to a strong level of agreement.

Table 1 illustrates the characteristics of the studies included. All the studies included were published after 2008. Of the 24 studies included, 11 (46%) [2,31-40] were published in the last 5 years (since 2014) and 20 (83%) [2,3,31-48] were published in the last 10 years (since 2009). All the studies included were published in English. As many as 10 [2,3,33-35,39,42,45,46,49] out of 24 studies were undertaken in the United States and 8 [31,36,40,41,47,50-52] were carried out in European countries. Other countries studied included Canada [43,44], South Africa [32], Iran [37], South Korea [38], and Taiwan [48]. Further, 19 [2,3,31,32,34,35,37-41,43-46,48,49,51,52] studies focused on the adult population, and only 3 [33,42,47] specifically targeted older patients (aged 65 years and older). All the health technology interventions were primarily designed for home-based usage. The relevant clinical conditions covered were asthma [41,48,50], heart disease [31,34,39,40,43,45,47,49], hypertension [32,35,37,44,46,51,52], diabetes mellitus [38], post-transplantation (renal [36] and lung [2]), comorbidities of diabetes mellitus and hypertension [3], and other unspecified diseases [33,42]. Eight studies [2,3,32,38,41,43,45,46] reported incorporation of clinical decision support or advanced computational algorithms to aid the self-management of diseases, while five [2,3,32,43,45] offered technology-led (instead of caregiver-led) dynamic feedback tailored to suit the patients’ conditions.

Table 2 presents the definitions of different medication adherence measures and summarizes each of these measures as reported by the studies included. Among 19 studies [2,3,31,32,34-40,42-45,47,49,51,52] that evaluated and reported the impact of health technology on medication adherence, 5 [3,31,32,39,43] adopted questionnaires and scales, which are the most commonly used measures. Further, 15 [2,3,31,32,34,37-40,43-45,47,49,52] explicitly compared improvements in medication adherence as a result of the intervention with the control arms. As many as 12 [2,3,31,32,37-40,43-45,52] reported that the intervention arms had seen improvements in medication adherence at the significance level of P<.1. A total of 4 studies [32,38,40,43] showed significant improvement in medication adherence at the P<.01 level and 3 studies [32,38,43] were significant at the P<.001 level. These results indicate that eHealth can improve patients’ medication adherence in nonhospital settings. Multimedia Appendix 2 provides details of the synthesized outcomes, risk differences, and P values between intervention and control.

Table 2 also presents the definitions of the secondary outcome measures and summarizes each of them as reported by the studies included. In all, 11 studies [32-34,38,41-43,46,48-50] reported secondary outcome measures including those that are associated with adverse events [38,43], quality of life [41,42,48,50], patient satisfaction [32,41,43,49], and health expenditure [33,34,46]. Among the 2 studies [38,43] reporting adverse events relating to the interventions (or safety outcome), no statistically meaningful difference between the intervention and control was found. Two [48,50] of the 4 studies reporting quality of life of the patients [41,42,48,50] revealed that there were significant differences between the intervention and control arms at the significance level of P<.01. Among the 4 studies measuring patient satisfaction [32,41,43,49], 1 [41] observed a difference when compared with the control. Three studies (all conducted in the United States) [34,42,46] reported medical spending obtained from various sources, including insurance claims [33,34] and inpatient and outpatient costs [46], and none of these showed any significant difference between intervention and control. The above evidence indicates that eHealth for self-administration of medication can improve the quality of life of patients.

Figure 2 presents the functionalities in the studies included by inclusion frequency. It appears that all the studies included incorporated an established mechanism to communicate with caregivers. Monitoring health features [2,3,35-39,41,43,44,46,48-52], and reminding and alerting [2,31,33,35,36,39,40,42,46,47,51,52] are the second- and third-most commonly used functions, respectively. Of the 24 studies, 6 [3,32,35,38,43,52] used the function of providing information and education to users and five studies [33,34,36,39,51] incorporated dispensing treatment and administration support, such as pill organizer automatic opening devices [34]. Providing information and education and dispensing treatment and administration support are significantly different in the medication adherence (at P<.001 level) group (Fisher exact test, P=.02). This indicates that these health technology functions are effective in improving medication adherence (Details provided in Multimedia Appendix 4).

Figures 3 and 4 present a summary of the RoB assessment and judgments of each assessment domain for the studies included. Of the 24 studies, 12 [2,3,31-34,36,39-42,50] reported on the employment of random sequence generation methods that sufficiently produced comparable study arms, while 6 [2,3,36,39,43,49] utilized methods to conceal the allocation sequences before or during enrollment. These studies exhibited a low risk of selection bias. All the studies included had a high risk of performance bias, which refers to the lack of blinding of participants. It was found that some studies [2,32,34,37] managed to introduce blinding to the investigator or personnel engaged in the research. In terms of detection bias, only 2 studies [2,39] were classified as having low bias with an indication of postallocation assessors blinding and 20 studies were assessed as unclear due to unspecified information. The attrition rates were generally low among the studies and only 4 [37,44,45,48] were grouped as high risk in this domain. A total of 9 studies had high selective reporting bias owing to the limitation of the chosen outcome measures [31,39,45,49] or the design [37,42,47], or the inability to measure the control group’s outcomes [36,51], etc. Apart from the above, there were a large number of studies that did not provide sufficient details on how to handle various domains of RoB in the publications and that inevitably increased the challenges involved in assessing various biases in the RCT studies included.

All the studies included were designed for patient home-based care applications and chronic conditions. Chronic disease management and home health care are important health issues in the context of a rapidly aging population. Only 3 [33,42,47] of the studies included were targeted at older patients aged 65 or above, and their findings were either insignificant [33,47] or inconclusive (no control group outcome was measured) [42]. Furthermore, the conclusions drawn from interventions in the adult population may lack generalizability in explaining the impact upon an older population. This is due to potential divergence in acceptance of technology and ability to use technology-assisted services between the older population and the general adult population. To address this issue, future studies may consider specifically designing and evaluating patient-centered eHealth interventions that cater to the special needs of older patients.

All the studies integrated more than 2 functions into the health technology intervention, indicating that multimodal intervention tends to be a commonly adopted model for eHealth design. We found that studies with significant improvements in medication adherence were all equipped with 3 or more integrated functions, except [37]. For instance, Jeong et al. [38] presented a multifunctional eHealth intervention comprising heterogeneous functions such as web-enabled videoconferencing that helped connect with the caregiver, a blood glucose monitoring device, a body composition organizer, an automated short message feedback system, and access to care center education program, which achieved significant medication adherence improvement (P<.001). This finding was in line with the non-eHealth medication nonadherence studies that showed that multiple components incorporated into the intervention tend to be successful [6]. This may be because medication nonadherence is usually multifactorial [6,53] and is a complex behavioral issue that involves socioeconomic and therapy-, patient-, condition-, and health system/health care team-related factors [56]. Single-function interventions thus tend to be insufficient when it comes to tackling such a complex problem

All the studies included provided a common eHealth function, that is, a mechanism to communicate with the caregivers. This allowed regular patient status updates, sending alerts, remote coaching, and interactive treatment plan adjustment. Considering the sociotechnical aspect of health informatics applications, future studies should carefully consider the interplay between the health system and eHealth interventions to tackle patient self-management of diseases effectively.

Our results indicated that functions of providing information and education, and dispensing treatment and administration support tend to favor improved medication adherence outcomes. This may be because dispensing treatment and administration support offer a mechanism to regulate and track medication activities [57], and to support a smooth self-medication process. It is also important to educate patients so that they understand the related disease characteristics and the benefits of following the medication regime, through the function of providing information and education.

The results should be interpreted with caution. Many studies that we included did not explicitly report RoB assessment items, particularly in the domains of blinding of outcome assessment (83% [20/24] unclear risk) and method for allocation concealment (58% [14/24] unclear risk). This increased the difficulties in evaluating the quality of the RCTs in full in order to make comparisons and draw appropriate conclusions. DeVito Dabbs et al. [2] and Young et al. [39] are relatively high-quality RCTs that evince minimal risk in selection, detection, and attrition biases.

Owing to the disparities in population characteristics, intervention design, functionality, and reporting adherence measures, it is difficult to draw inferences and definitive conclusions on some potential hypotheses, such as the combination of eHealth functions that contribute to improvements in medication adherence for a particular disease. Evidence on the effectiveness of eHealth on medication adherence and health care outcomes improvement exists, but is not compelling enough. Larger-scale RCTs with greater sample sizes are necessary to draw inferences among those infrequently observed measures such as safety outcomes in adverse events, hospitalization, and emergency visits.

Subjective and objective measures have their pros and cons [53]. To increase measurement sensitivity, future studies can consider employing a combination of measures to assess medication adherence. When designing eHealth interventions, it is important to be aware of how the selection of outcome measures can contribute to the trustworthiness of a study. The choice of medication adherence measures can affect who is assessing the outcome and the objectivity of the assessment. In general, where subjective outcomes are concerned, blinding is particularly important. For instance, medication adherence via questionnaire survey can be subjectively assessed by the participants, which is considered as low-quality adherence measurement. Such a design can influence the blinding of outcome assessment (ie, increasing detection bias).

This study attempted to examine how eHealth for patients’ medication management improves drug adherence and other health care outcome measures. Our study has a number of strengths. First, we comprehensively searched cross-sectional databases in the fields of medicine, nursing care, public health, science, engineering, and social science. We focused exclusively on studies that employed RCT, which is considered the highest standard of eHealth evaluation. Our search strategy was broad and involved a large number of studies for screening (eg, 9909). We successfully retrieved a solid body of evidence that indicated improvement in drug adherence through the application of eHealth. We also evaluated the quality of the studies included through the state-of-the-art Cochrane Collaboration tool. Our study provides practical insights into the future of eHealth design and applications for patient self-medication management.

Our study has a few limitations, as well. First, a publication bias may exist owing to the inherent tendency to publish positive results with significant findings. Furthermore, the studies included reported multiple outcome measures rather differently. Owing to this heterogeneity, it was not possible to carry out a meta-analysis. We synthesized review outcomes from studies that largely centered on chronic diseases. However, our search strategy was designed to consider a broad scope of studies. Based on our study framework, future studies can investigate how eHealth impacts a specific targeted disease or patient group (such as those who suffer from mental health issues or HIV/AIDS).

Furthermore, owing to the small sample size, we were unable to use more accurate parametric methods, and thus unable to ascertain the impact of a combination of multiple functions. Large-scale evaluation studies in the future can consider examining the impact of the integration of multiple functionalities into eHealth; studying the social and behavioral differences across targeted populations in reacting to health technology; and carefully evaluating other important health outcome measures including adverse events owing to health technology introduction, and cost and economic evaluation.

Understanding the nature of the eHealth intervention, one must be aware that some limitations in the study design are inevitable. For instance, all the studies included were unable to blind the intervention group from the control group (ie, the usual care group without using technology intervention). At present, there is no way for researchers to avoid performance bias in the studies reviewed. However, this situation may change in the future when successful health technology becomes the gold standard and replaces the current usual care system which does not employ health technology. Future RCTs for evaluating eHealth interventions should follow best practice guidelines, which include observing the limitations of the study design; blinding of participants, personnel, and assessors involved; selecting the most objective outcome measures; employing fair assessment methods; and avoiding selective outcome reporting.

This study investigated how eHealth interventions could affect patient medication adherence and health outcomes and identified the eHealth functions that are most effective in improving medication adherence. The evidence reviewed shows that eHealth can improve patients’ medication adherence and quality of life in nonhospital settings. Integrating multiple functions into health technology tends to be effective in achieving enhanced medication adherence. eHealth functions of providing information and education, and dispensing treatment and administration favored an improved medication adherence outcome. However, the literature base remains small, diffuse, and inconclusive at this time. Medication-taking behavior may vary tremendously based on the patients’ medical conditions, the population studied, and the specific medications assessed. Many interesting potential medical–social–behavioral research hypotheses are yet to be posed and answered in the existing literature.