Many resource-limited countries, especially those in Africa, face health inequalities that have led to a triple burden of traditional, infectious, and chronic diseases [1]. The health inequality gap manifests in many ways, including the lack of medical and food accessibility, poor sustainable income leading to a lack of meeting daily needs, poor alignment of innovation with disease burdens, and underserved communities that limit the monitoring of patients who are chronically ill [2]. The literature indicates that chronic diseases, both communicable and noncommunicable, have contributed to many deaths and have caused a great deal of strain on health care systems in resource-limited countries [3,4]. In addition, patients with chronic conditions often become traumatized while facing an incurable illness. This is detrimental to their medical compliance and ultimately contributes to their death [1,5].

Many resource-limited countries face unemployment, which has led to migration from rural to urban areas [1]. With the increase in urbanization, more than 56% of the people in resource-limited countries now live in towns and cities. However, socioeconomic disparities and other factors such as work stress, environmental factors, and low income generation negatively impact the lives of these new town or city immigrants [4,6]. Because of the high cost of living in urban areas, people’s lifestyles have changed drastically. This has contributed to the increasing prevalence of both communicable and noncommunicable diseases as well as infectious diseases [2]. Diseases such as cardiovascular disease, chronic obstructive pulmonary diabetes, tuberculosis, and HIV have become more prevalent in resource-limited countries [5].

The World Health Organization report of 2020 indicated that diabetes and other noncommunicable diseases resulted in over 16% of all deaths in South Africa in 2016 [7]. This high prevalence was primarily due to the lack of awareness of diabetes, lack of access to adequate health care, and inability to follow medical prescriptions. It is pertinent to mention that this is taking place at a time when South Africa is spending a great deal on health care, with diabetes being on its political agenda [5]. The COVID-19 pandemic has strained the health care systems in multiple countries. The high prevalence of chronic diseases in resource-limited countries has further exacerbated. Moreover, many resource-limited countries, including South Africa, had already faced numerous burdens of disease before the emergence of COVID-19. For example, for decades, malaria and measles have been traditional health hazards in equatorial regions [1,6]. These multiple burdens of disease demand technological interventions such as mobile health (mHealth) to empower patients to take control of their own health.

Therefore, it is important to provide patients who consume medicines regularly with technology-based self-monitoring. An mHealth self-monitoring application system will serve as a reminder to timely attend to patients’ health needs. Those living with a chronic disease often give up trying because of anxiety [1,8]. The World Health Organization [7] showed that inconsistency in drug intake is a major cause of high mortality rates in people with chronic diseases. This is because routine and timely medication administration suppresses symptoms and other maladies that complicate diabetic conditions. Low immunity levels make these ailments and symptoms more likely to worsen the patient’s condition owing to poor adherence to medicine [9,10]. The use of technology to elicit behavioral changes for self-monitoring of complicated ailments such as diabetes has been deemed rewarding by several researchers [11,12].

Regardless of the country’s economic status, health care responsiveness is of paramount importance. To ensure equal access to health care [13], countries should find better ways of allowing all citizens access to health care, regardless of their location. A key component of this is providing health services closer to communities, especially those outside the constant reach of health care workers. According to the study by Achoki et al [14], increasing the accessibility of health care systems is crucial to bridging the urban-rural divide by solving health care inequality. Because of efforts to improve health care system accessibility and availability, information technology has been used to provide health care [10,15,16].

The use of technology has become ubiquitous and pervasive in facilitating the monitoring and provision of health care. According to the study by Islam et al [17], mHealth has gained traction as a powerful tool for monitoring both patients and medical personnel. These researchers have indicated that mHealth technology has been leveraged to manage chronic diseases such as diabetes, hypertension, and other cardiovascular ailments, as well as maternal health and psychological disorders. Researchers such as Abduo et al [18] and Lin et al [19] have also observed that mHealth has the potential to improve the efficiency and effectiveness of health care management for patients by health care workers and health system managers. This is accomplished by offering real-time guidelines, referral services, and procedures. When mHealth is used effectively, it can also serve as a reminder system for patients to self-manage and administer their medication, thereby improving medication adherence [9].

As a communication tool, mHealth has been used to distribute information via simple messages (SMS text messages) and alerts via other reminders, enabling patients to adhere to their treatment regimen. However, because medication adherence involves lifestyle changes, including behavioral adjustment, patients require positive incentives for such behavioral changes [20]. Hence, as self-management and administration of medical prescriptions are affected by behavior, an mHealth system that works as a reminder must include a component that considers human behavior [9]. It is imperative that technological applications such as mHealth be integrated with human behavior in the form of a persuasive model. This implies that mHealth alone may not be sufficient as a communication tool, and the behavioral aspects of patients must be considered. This also requires contextualization, thus considering patients’ cultural and social backgrounds [5,21].

According to Fogg [22], during any human-computer interaction, humans may respond to computers as though they are human. This could be because humans naturally respond to social presence by expressing empathy, being angry, or participating in social activities. Humans are hardwired to respond to signals in the environment in which they live; such responses are instinctive rather than rational. Nass et al [23] argued that computers serve as persuasive social actors by rewarding individuals with positive feedback. This is accomplished by modeling the desired behavior or attitudes and thereafter providing psychological support. Furthermore, Fogg [24] observed that humans also experience social presence, whether by empathizing or feeling angry or by performing social actions. Fogg [24] added that computing products trigger individual automatic responses because of the social cues provided by them. The author emphasized that factors such as motivation, ability, and prompting are required as simultaneous actions for a given behavior to occur.

Researchers such as Larbi et al [11] and Reidy et al [25] noted that for self-monitoring, a patient must be motivated to perform the actual behavior of taking medicine. The motivation may arise from the patient’s yearning to live a stress-free life and have a social connection. This implies that such an individual must have the ability to easily comply with whatever the mHealth app is suggesting. This could be termed as self-efficacy and ease of use. This signifies the importance of including the acceptance and use of technology factors in the mHealth self-monitoring model. In addition, Fogg [24] emphasized that persuasive technology involves the incorporation of insights from psychology into the design of products. This applies to mobile apps and wearables that modify people’s habits and beliefs; factors such as motivation and ability should be considered in the design process.

To inform the acceptance and use of technology, various theories and models have been developed, most of which are extensions and modifications of the old popular models of the theory of reasoned action [26,27], diffusion of innovation [28], the technology acceptance model [29], and the Unified Theory of Acceptance and Use of Technology (UTAUT) [30]. According to these models, although heterogeneous factors may lead to the success of technological innovation, actual behavior, such as acceptance and use, is generally preceded by behavioral intention (BI) [31,32]. In general, newer theories and models have been developed because of the shortcomings of their predecessors.

Several commonly used theories and models for predicting technology acceptance and use are extensions of previous models, and their constructs are essentially the same. Similarly, the behavioral change wheel model [33], which has been widely used for behavioral change interventions, was developed by unifying 19 frameworks, 9 intervention functions, and 7 policy categories for classifying behavioral change. In contrast, the theoretical domains framework (TDF) [34] recommended for understanding behavioral change processes during the implementation of evidence-based interventions was also developed from a synthesis of psychological theories [35,36]. The TDF has 14 domains: social influences (SIs); environmental context and resources; social and professional role and identity; beliefs about capabilities; optimism; intentions; beliefs about consequences; reinforcement; emotion; knowledge; cognitive and interpersonal skills; memory, attention, and decision processes; behavioral regulation; and physical skills. These domains have been widely used to study how behavior changes when evidence-based care is implemented, which cuts across those of acceptance and use theories relating to behavioral change interventions.

Both the behavioral change wheel and TDF, along with acceptance and use theories, have been offered as theoretical approaches in the design of tools for interventions aimed at changing behavior [37]. Therefore, a conceptual model for exploring implementation problems for enhancing health care practices should clarify a wide range of potential mediators of behavioral change. This study used an extended version of the UTAUT, underpinning the design of its conceptual model. The extended model of the UTAUT encompasses behavioral change intervention aspects, technology characteristics that promote self-efficacy and ease of use, as well as individual characteristics (IC) that support patients’ motivation. In addition, the conceptual model considers age, gender, experience, and motivation as moderating variables. The conceptual model is illustrated in Figure 1.

Data for this study were collected from district hospitals in South Africa. Indicator trends for districts and provinces in the 2018-2019 District Health Barometer provided a snapshot of health care services in the public sector of South Africa, including the prevalence of common disease burdens and health care delivery [45]. On the basis of the population of each province and district municipality, this report presents estimates of the prevalence of diabetes. This revealed wide differences between provinces in diabetes prevalence and treatment coverage. According to a report, in 2017, diabetes among adults aged ≥15 years was more prevalent in the Western Cape, KwaZulu-Natal, Eastern Cape, Free State, Northern Cape, Gauteng, Limpopo, Northwest, and Mpumalanga provinces. In contrast, treatment rates were higher in the Northern Cape, Western Cape, Free State, Limpopo, KwaZulu-Natal, Eastern Cape, Gauteng, Mpumalanga, and Northwest regions [45]. On the basis of this report, geographic locations for selecting the study population were identified.

Three conditions were considered in the selection of the study population. These were selecting provinces with high prevalence but low treatment rates, provinces with high prevalence and high treatment rates, as well as provinces within a good proximity but with a relatively dense population and urbanization. On the basis of these 3 conditions, the provinces of KwaZulu-Natal, the Western Cape, as well as Gauteng and Mpumalanga were selected. However, this study opted for district hospitals; there is only 1 regional hospital in each province that normally acts as a referral. This was expected to make data collection somewhat slower because of the busy schedules of health care providers in these regional hospitals.

Aspects of privacy, confidentiality, and avoidance of stigmatization of patients led to the provincial ethical clearance committees, from which authorization was sought for data collection, deeming it inappropriate to collect data from patients. However, because the aim of this study was to model the behavior of patients with diabetes regarding medication adherence, using health care providers was sufficient for the study. This study acknowledges that personal traits and characteristics play a significant role in determining an individual’s behavior. However, medical personnel and health workers can effectively and efficiently monitor patients’ daily routines and observe any deviations. In addition, health care providers can recognize patients’ functional and cognitive decline in their daily lives [46]. This confirmed the eligibility of health care providers as respondents of the study. Therefore, only health care providers, such as physicians and social workers, who are regularly involved in counseling, medication prescriptions, and treatment of patients with diabetes, were included in the study.

The pre-exploratory study that had been conducted earlier revealed that many South African district hospitals do not have specific diabetic clinics and that patients with diabetes are treated in exactly the same way as other patients in general wards or outpatient clinics. This implies that all medical personnel and social workers were qualified to form the study population. These staff members regularly dealt with patients with diabetes. Information also revealed that there were 30 to 50 medical personnel and social workers at a district municipality hospital. On the basis of the number of district hospitals in the 4 sampled provinces, the population of the study was 660, giving a sample size of 248 respondents when using the Krejcie and Morgan [47] tool for determining the sample size of a finite population. Simple random sampling was then applied based on the inclusive criterion that a respondent should be either a medical personnel or a health worker.

On the basis of the sample size, 350 questionnaires were distributed, of which 257 were returned, giving a response rate of 73%. This was deemed sufficient for the analysis. This study set out to analyze the moderating effects of the demographic and situational variables, thus variance-based structural equation modeling using smart partial least squares was applicable for data analysis.

This study was conducted as part of a doctoral study at the University of South Africa. It is mandatory that all studies conducted within the institution receive ethical approval before the commencement of data collection, as no retrospective approval could be allowed. Therefore, this study was approved by the University of South Africa Instructional Ethics Board. Moreover, collecting any medical-related data, whether from patients or health care providers in South Africa, requires the researcher to register the research in the medical database and to receive ethics clearance from each province to participate in the study. The main researcher, who is the corresponding author, sought permission and scheduled presentations of the proposal at each of the 3 provinces’ regional health departments before ethical approval could be granted. This was done, and ethical approval was received only to collect data from health care providers. The questionnaire for data collection was accompanied by a cover letter, which the respondents of the study were reminded to read before filling the questionnaire. The letter explained to the respondents about their privacy and confidentiality protection, their rights, and freedom to drop out of the study at any time without such decision prejudicing their rights. Most importantly, the respondents were informed that the data were intended for academic purposes only and that none of their particulars or those of their health institutions would be included anywhere in the study. Moreover, respondents were reminded that there were no monetary benefits or compensation expected from filling out the questionnaire.

The collected data were screened and cleaned, and outliers were eliminated, leaving 158 samples in the final data sets. Two measurement models were built with both moderated and nonmoderated constructs. These measurement models were evaluated, modified, and adjusted to form the path models. Figure 2 shows the nonmoderated measurement model of the constructs.

From the measurement model in Figure 2, the correlated values between the observed variables and the latent variables, the R² of the model, and the coefficient of the linear path regression between the latent variables were determined. Figure 2 also provided results for the convergent validities obtained by observing the average variance extracted (AVE), the internal consistency values (Cronbach α) and the composite reliability (CR; Dillon-Goldstein þ—rho) as demonstrated in Table 1.

CR prioritizes variables according to their reliability and is considered the most appropriate for partial least squares. In contrast, Cronbach α, which measures internal consistency, is sensitive to the number of variables in a construct [48]. CR and Cronbach α were used to determine whether the study sample was biased. Thus, values between .60 and .70 should be noted, as well as .70 to .90, respectively, to achieve good fit and satisfactory results [49]. However, as shown in Table 1, the cultural construct had an AVE of 0. 279 (<0.50), Cronbach α of .407 (<.6), and CR values of less than 0.7. These results suggest that either items linked to culture should be eliminated during adjustments and modification of the model or that the construct should not be included for further analysis. Because culture had only 3 measurement items and was moderated by gender and age, further analysis of the construct is needed. The moderating effects had to be evaluated to determine whether they affected the prediction. Table 1 further shows that most constructs had AVEs less than 0.5, indicating that model modifications were necessary.

Figure 2 was also used to compute discriminant validity, which measures the independence of one construct from the other. In this study, the Heterotrait-Monotrait ratio (HTMT) of correlations was applied [49]. The results are summarized in Table 2. Fornell and Larcker [50] indicated that values for HTMT not exceeding 0.9 could be accepted, although HTMT values near 1 indicated a lack of discriminant validity. If the value of the HTMT is higher than the threshold, then there is a lack of discriminant validity, which suggests that the model should be adjusted.

The results in Table 2 indicate that culture and skills constructs do not exhibit good discriminant validity. However, most constructs had acceptable discriminant validity values that were below the threshold. The results confirmed the need to adjust the model and exclude cultures from the final model. Observed and latent variables were also checked for cross-loadings. It is recommended that during cross-loading, an item’s component loading on its own construct should be higher than that on other constructs. Researchers such as Ringle et al [49] and Henseler et al [51] recommended that when an item loads higher onto a construct other than its parent construct, the discriminant validity is compromised. A difference between loads less than 0.10 indicates higher cross-loading onto the other construct, which compromises discriminant validity. With the exception of culture, skills, and environment constructs, most variables had higher cross-loadings on their parent constructs than on the other constructs. Again, this indicates that modifications to the model were necessary. As this study tested for the direct influence of constructs and for the influence when the relationship of the construct is moderated, constructs whose HTMT was beyond the threshold were not discarded, as there was a need to check if their influencing effects would change with the moderating effects.

Variance-based structural equation modeling (SEM) has no specific fit indices tested, unlike other measures; therefore, covariance-based SEM may be used [52]. Among these is the Bentler-Bonett Normed Fit Index, an incremental measure of goodness of fit that has not been affected by the number of variables in the model for fitness (Normed Fit Index >0.90). In addition, the root mean square residual threshold is 0.08; the exact fit measures the squared Euclidean distance and the geodesic distance, which should have a nonsignificant probability (P>.05); and there is the chi-square value (×2) [49,52]. Table 3 shows the fitness of the model.

After fitting the measurement model, a path model was constructed and used to determine the significance of the contribution of each construct. The Bootstrap result approximates the normality of the data for a 2-tailed test at different confidence levels. For a 95% CI, significance was obtained when T value was 1.96 at the .05 level [49,52]. Table 4 presents the significance of the independent variables.

As shown in Table 4, the 3 independent variables—culture, skills, and SI—contributed to the overall prediction of the model as being not significant. On the basis of Figure 1, it was theorized that a patient’s environment influences the use of mHealthSM. This theory was confirmed by the results of the study, which showed that the influence of environmental aspects was significant for all independent variables (all P<.001).

Using the T statistics provided in Table 4, the suggested hypotheses are tested at a level of .05. The results are as shown in Table 5.

The results from Table 5 indicate the reasons for all hypotheses except culture, IC of skills, and SI being accepted.

Four moderating variables—age, gender, experience, and motivation—were tested for their interactions. To read the adjusted coefficients, the structural model was redesigned, the model quality was determined, adjustments were made, and a path analysis was conducted to determine the moderating coefficients. Multimedia Appendix 1 presents the results.

The suggested hypotheses were tested by comparing the significance of the interacting effects of the moderating factors based on the T statistic and the P value for each relation. When the significance of the interaction was observed, it implied that the hypothesized interaction existed, and when it was not observed, it implied that the interaction did not exist. Table 6 presents the tested hypotheses.

This section discusses and interprets the findings of the moderating factors hypotheses.

Patients’ management of their own lives is essential as this helps patients to remain healthy, minimizing their readmissions to hospitals, thereby reducing health-related costs and ameliorating their quality of life. With the possibility of future pandemics, it is no longer optional that health care should move toward real-time data analytics to improve timely decision-making related to saving lives. The pervasiveness of technological innovations in the health care domain to reduce the increasing disease burden has made mHealth a much sought-after tool to be leveraged in the health sector. mHealth systems not only work as a reminder system for patients but also as a source of information. Such systems allow health care professionals to collect quantitative information related to patients’ health and behavior regarding medication adherence. This helps health care professionals make meaningful decisions related to health risk prediction while developing more interventions.

Through the use of a close-ended questionnaire, this study gathered insights from diverse health care personnel interacting with patients with diabetes daily. The collected data were analyzed using smart partial least squares SEM to develop a model for mHealthSM. As demonstrated in this study, it could be deduced that with the ubiquitous adoption of smartphones across racial, educational, and socioeconomic groups, it is possible to develop new models of health care delivery for patients with chronic conditions, such as diabetes to promote health equality and equity. Therefore, it is vital that patients receive timely reminders to take their medicines at the right time and in the right quantity. This avoids double doses and enables prescriptions to be renewed over time. In addition, data collected from mHealth self-monitoring systems could be used to reduce clinical trial costs, thereby enabling pharmaceutical companies to gain accurate information about medicines, leading to the production of new drugs.

This study acknowledges that modeling individual behavior involves understanding perceptions. Hence, it would have been more helpful to use patients as respondents to report their own behaviors, rather than from health care providers. As much as health care providers monitor and can tell patients’ behavior toward medication adherence, there could be limitations in explaining why such nonadherence occurs. It is therefore recommended that future studies form joint research teams with health care personnel or medical experts, so the data are collected directly from patients rather than secondhand from health care providers.

This study was concerned with the development of a self-monitoring system for medication adherence only, and there was no intervention to influence health outcomes. However, there are also various other ways in which patients with diabetes may be monitored, such as the rate of physical activity, weight gain or loss, and blood glucose levels. These other health-monitoring facilities were outside the scope of this study. Future research could develop mHealth systems that combine all of these health conditions that should be monitored into one integrated system. Furthermore, owing to the increasing number of patients with chronic conditions, data storage and network stability may impede the effective use of the mHealth self-monitoring system. Therefore, this study recommends that future mHealth apps and device integration be developed to be supported by a more comprehensive, cloud-based system. Cloud-based solutions will provide various benefits, including stability, availability, and security, in addition to health care personnel now being in a position to analyze patient data from a central platform.

This study followed a quantitative approach in which respondents only answered questions with predetermined answers. Behavior patterns are influenced by a number of individual factors, many of which may not be expressed in words but by feelings and expressions. Such feelings and expressions can only be interpreted if the researcher has a one-to-one interaction with a participant in the form of qualitative interviews. Therefore, as much as the quantitative approach analyzes causal and interacting effects of moderating factors that are essential for determining individuals’ continuance behaviors, this study recommends that future research should use qualitative or mixed methods for the model’s validation.

This study contributes to the development of a model that incorporates the interacting effects of users’ moderating factors that have been recommended in various research studies on health interventions as being essential for the self-monitoring of patients. As recommended by other researchers, technological innovation may be accepted, adopted, and used; however, its continued use may not be sustained. Moderating factors are essential in predicting future use; the influence of some factors may cease, leading others to become salient. Hence, including moderating factors to predict health care interventions is a significant contribution, as the developed model will be used to extend research on health care interventions with confidence in predicting future maintenance use. Moreover, the developed model will be implemented in a fully functioning mHealth self-monitoring system that will be practically used by both health care providers and patients.

The pervasiveness of the use of technological innovations in the health care domain and the increase in disease burden has made mHealth a sought-after tool in the health sectors of many countries, whether resource limited or high income. mHealth has been widely applied in different aspects of health care management, especially with chronic complications that require routine monitoring, making adherence a challenge [21]. The mHealth models, such as the one developed in this study, could be used to implement systems that not only work as a reminder system for patients but also allow health care professionals to collect quantitative information related to patients’ health and behavior toward medication adherence, which helps personnel to make meaningful decisions. Through the data generated, stored, and disseminated by mHealth systems, health care providers will be capable of gathering patients’ related data and making decisions, such as patients’ risk prediction, need for physical monitoring, or admission to intensive care.

Patients’ management of their own lives is essential, as this helps them to remain healthy and minimizes their readmissions to hospitals, thereby reducing health-related costs and improving their quality of life [21,46]. With the increasing world pandemics, it is no longer optional that health care moves toward real-time data analytics to improve timely decision-making related to saving lives. In an effort to do so, more technological interventions into health care and management are needed to remain abreast with the increasing globalization in the fourth industrial revolution era.