There is no health care system without health professionals. The health outcomes of people rely on well-educated nurses, pharmacists, dentists, and other allied health professionals [1]. Unfortunately, these professionals are in short supply and high demand [2,3]. Almost 1 billion people are negatively affected by the lack of access to adequately trained health professionals, suffering ill-health or dying [4,5]. In many low- and middle-income countries (LMICs), this situation is further aggravated by the difficulties in implementing traditional learning programs; deficiencies in health care systems and infrastructure; and lack of essential supplies, poor management, corruption, or low remuneration [6].

Digital education also known as e-learning is an umbrella term encompassing a broad spectrum of educational interventions characterized by their tools, technological contents, learning objectives or outcomes, pedagogical approaches, and delivery settings, which includes, but is not limited to, online and offline computer-based digital education, massive open online courses (MOOCs), mobile learning (mLearning), serious gaming and gamification, digital psychomotor skill trainers, virtual reality, or virtual patient scenarios [7]. Digital education aims to improve the quality of teaching by facilitating access to resources and services, as well as remote exchange of information and peer-to-peer collaboration [8]; it is also being increasingly recognized as one of the key strategic platforms to build strong education and training systems for health professionals worldwide [9]. The United Nations and the World Health Organization consider digital education as an effective means of addressing the educational needs among health professionals, especially in LMICs.

This review focused on offline digital education. This refers to the use of personal computers or laptops to assist in delivering stand-alone multimedia materials without the need for the internet or local area network connections [10]. The educational content can be delivered via videoconferences, emails, and audio-visual learning materials kept in either magnetic storage, for example, floppy disks, or optical storage, for example, CD-ROM, digital versatile disk, flash memory, multimedia cards, external hard disks, or downloaded from a networked connection, as long as the learning activities do not rely on this connection [11].

There are several potential benefits of offline digital education such as unrestrained knowledge transfer, enriched accessibility, and significance of health professions education [12]. Further benefits include flexibility and adaptability of educational content [13], so that learners can absorb curricula at a convenient pace, place, and time [14]. The interventions can also be used to deliver an interactive, an associative, and a perceptual learning experience by combining text, images, audio, and video via combined visual, auditory, and spatial components, further improving health professionals’ learning outcomes [15,16]. By doing so, offline digital education can potentially stimulate neurocognitive development (memory, thinking, and attention) by enhancing changes in the efficiency of chemical synaptic transmission between neurons, increasing specific neuronal connections and creating new patterns of neuronal connectivity and generating new neurons [17]. Finally, health professionals better equipped with knowledge, skills, or professional attitudes as a result of offline digital education might improve the quality of health care services provision, as well as the patient-centered and public health outcomes, and reduce the costs of health care.

This systematic review was one of a series of reviews evaluating the scope for implementation and the potential impact of a wide range of digital health education interventions for postregistration and preregistration health professionals. The objective of this systematic review was to evaluate the effectiveness of offline digital education compared with various controls in improving learners’ knowledge, skills, attitudes, satisfaction, and patient-centered outcomes.

We searched the following databases (from January 1990 to August 2017): MEDLINE (via Ovid), Excerpta Medica dataBASE (via Elsevier), Web of Science, Educational Resource Information Center (via Ovid), Cochrane Central Register of Controlled Trials, The Cochrane Library, PsycINFO (via Ovid), and the Cumulative Index to Nursing and Allied Health Literature (via EBSCO). The search strategy for MEDLINE is presented in Multimedia Appendix 1. We searched for papers in English but considered eligible studies in any language. We also searched 2 trial registries (EU Clinical Trials Register and ClinicalTrials.gov), screened reference lists of all included studies and pertinent systematic reviews, and contacted the relevant investigators for further information.

Only randomized controlled trials (RCTs) and cluster RCTs (cRCTs) of postregistration health professionals except medical doctors—as they were covered in a separate review [19]—using either stand-alone or blended offline digital education with any type of controls (active or inactive) measuring knowledge, skills, attitudes, satisfaction, and patient-centered outcomes (as primary outcomes) as well as adverse effects or costs (as secondary outcomes) were eligible for inclusion in this review.

We excluded crossover trials, stepped wedge design, interrupted time series, controlled before and after studies, and studies of doctors (including medical diagnostics and treatment technologies) or medical students. Participants were not excluded on the basis of sociodemographic characteristics such as age, gender, ethnicity, or any other related characteristics.

The search results from the different electronic databases were combined in a single EndNote (X8.2) library, and duplicate records of the same reports were removed. In total, 2 reviewers independently screened titles and abstracts to identify studies that potentially meet the inclusion criteria. The full text versions of these articles were retrieved and read in full. Finally, 2 review authors independently assessed articles against the eligibility criteria, and 2 reviewers independently extracted the data for each of the included studies using a structured data extraction form and the Covidence Web-based software (Veritas Health Innovation, Melbourne, Australia). We extracted all relevant data on the characteristics of participants, intervention, comparator group, and outcome measures. For continuous data, we reported means and SDs and odds ratios (ORs) and its 95% CIs for dichotomous data. For studies with multiple arms, we compared the relevant intervention arm to the least active control arm, so that double counting of data does not occur. Any disagreements were resolved through discussion between the 2 authors and if no consensus was reached, a third author acted as an arbiter.

In total, 2 reviewers independently assessed the risk of bias of the included studies using the Cochrane Collaboration’s Risk of Bias tool [18]. Studies were assessed for risk of bias in the following domains: random sequence generation; allocation concealment; blinding of participants or personnel; blinding of outcome assessment; completeness of outcome data (attrition bias); selective outcome reporting (reporting bias); validity and reliability of outcome measures; baseline comparability; and consistency in intervention delivery. For cRCTs, we also assessed and reported the risk of bias associated with an additional domain: selective recruitment of cluster participants. Judgments concerning the risk of bias for each study fell under 3 categories: high, low, or unclear risk of bias.

Data were synthesized using Review Manager version 5.3. In cases where studies were homogeneous enough (in terms of their population interventions, comparator groups, outcomes, and study designs) to make meaningful conclusions, we pooled them together in a meta-analysis using a random-effects model and presented results as standardized mean difference (SMD). We assessed heterogeneity through a visual inspection of the overlap of forest plots and by calculating the chi-square tests and I² inconsistency statistics [18].

We prepared the Summary of Findings (SoF) tables to present the results for each of the primary outcomes. We converted results into absolute effects when possible and provided a source and rationale for each assumed risk cited in the table(s) when presented. A total of 2 authors (PP and MS) independently rated the overall quality of evidence as implemented and described in the GRADEprofiler (GRADEproGDT Web-based version) and Chapter 11 of the Cochrane Handbook for Systematic Reviews of Interventions [20]. We considered the following criteria to assess the quality of evidence: limitations of studies (risk of bias), inconsistency of results, indirectness of the evidence, imprecision and publication bias, and downgraded the quality where appropriate. This was done for all primary outcomes reported in the review.

Details of each trial are presented in Table 1 or Multimedia Appendix 3; a summary is given below. The included trials were published between 1991 and 2016, and originated from Brazil 3.7% (n=1), Hong Kong 3.7% (n=1), Iran 7.4% (n=2), Korea 3.7% (n=1), the Netherlands 11.1% (n=3), Norway 3.7% (n=1), Taiwan 11.1% (n=3), Turkey 3.7% (n=1), the United Kingdom 7.4% (n=2), and the United States 44.4% (n=12). A total of 5 trials employed cluster design [21-25], whereas the remaining studies used a parallel group design. The majority of studies (51.8%) were conducted in nurses [23,24,26-37] followed by pharmacists (14.8%) [38-41], mental health therapists (11.1%) [21,25,42], dentists (7.4%) [43,44], midwives [22], physical therapists [45], patient care personnel [46], and substance abuse counselors [47]. The evaluated interventions included blended learning [35]; CD-ROM and emails [43]; computer-assisted instruction (CAI), computer-based training, or computer-mediated training [22,23,26,28,29,34,36,38, 41,42,44-46]; CD-ROM [24,30-33,37,47]; PowerPoint presentation [39,40,46]; and software [21,25,27]. The duration of the intervention ranged from 50 min [28,38] to 3 months [21,25,31,46]. The intensity ranged from 15 min [44] to 2.4 h [37]. Comparison groups included no intervention [22,26,32,33,37,43-45,47], blended learning [21,25], and traditional learning [23,24,27-31,34-36,38-42,46]. Primary outcomes included knowledge in 20 studies [21-24,26-48], skills in 9 studies [22,24,31,35,37,38,41-43], attitudes in 7 studies [26,30,35,40,41,44,45], satisfaction in 9 studies [23,25,28, 33,35,36,38,40,42], and patient-centered outcomes in 2 studies [22,25].

We present our judgments about each risk of bias item for all included studies as (summary) percentages in Figure 2.

Figure 3 shows separate judgments about each risk of bias item for each included study.

The risk of bias was predominantly low for random sequence generation (55.5% of the studies), selective reporting, baseline comparability, and consistency in intervention delivery. The risk of bias was predominantly unclear for allocation concealment blinding of participants, personnel, or outcome assessors. A total of 12 studies (44.4%) had a high risk of attrition bias; 6 studies (22.2%) had a high risk of bias for validity and reliability of outcome measures; and 5 studies (18.5%) had a high risk of bias for baseline comparability. In total, 3 studies (11.1%) had a high risk of performance bias; and 1 study (3.7%) had a high risk of detection bias. For cRCTs, all 5 studies had a low risk of bias for selective recruitment of cluster participants.

We identified 4 studies from upper middle-income countries (Brazil, Iran, and Turkey), and the remaining studies were conducted in high-income countries (Hong Kong, Korea, the Netherlands, Norway, the United Kingdom, and the United States). Only 4 studies (15%) were conducted during the 1990s and the remaining studies were from 2000 onward. In 15 studies (55.5%), information about the frequency of the interventions was missing, thereby often making it difficult to analyze in depth and interpret the findings. Similarly, economic evaluations of the interventions were missing in 26 (96%) studies.

Overall, the quality of evidence was low or very low. We assessed the quality of evidence using the Grading of Recommendations, Assessment, Development, and Evaluations system and presented the findings in SoF Multimedia Appendix 4 for all comparisons. The reasons for downgrading the evidence most commonly pertained to the high risk of bias. For instance, only 13 (48.1%) of the studies reported complete outcome data. Reducing the dropout rate might reduce the risk of attrition bias and further improve the quality of the studies. Only 8 studies (29.6%) had a low risk of bias for validity and reliability of outcome measures. This issue of nonvalidated measurement tools has repeatedly been raised and is paramount to advance the field [49]. Only 2 studies (7.4%) adequately described blinding of participants and personnel. As with many educational interventions, blinding of participants or personnel might prove challenging. However, we highlighted the need for more adequate descriptions of masking to further reduce the risk of performance bias and allow clearer judgments to be made. We also downgraded the overall quality of evidence for inconsistency (where there was a high level of heterogeneity, ie, I²>50%). Overall, there was a moderate-to-considerable level of heterogeneity of meta-analyses (I² range 41% to 88%); and 4 (out of 5) meta-analyses had I²>50%. More reasons for downgrading included indirectness (we downgraded once for 1 outcome only—where there were differences in the population used). Participants were not homogeneous and ranged from nurses, pharmacists, mental health therapists, dentists, midwives and obstetricians, physical therapists, patient care personnel to substance abuse counselors. Other sources of indirectness also stemmed from heterogeneous interventions (their duration, frequency, and intensity), comparison groups, and outcome assessment tools ranging from multiple choice or single choice questionnaires, tests, observations, checklists, scales, surveys, visual analogue scales, and simulations. Finally, we also downgraded for imprecision where the sample size was small. The included studies also failed to provide details of sample size and power calculations and may have therefore been underpowered and unable to detect change in learning outcomes.

This systematic review has several important strengths that include comprehensive searches without any language limitations, robust screening, data extraction and risk of bias assessments, and a critical appraisal of the evidence. However, some limitations must be acknowledged while interpreting the results of this study. First, we considered subgroup analyses to be unfeasible because of the insufficient number of studies under the respective outcomes and professional groups. However, we minimized potential biases in the review process and maintained its internal validity by strictly adhering to the guidelines outlined by Higgins et al [18].

A review by Al-Jewair [50] found some evidence to support the effectiveness of computer-assisted learning in improving knowledge gains in undergraduate or postgraduate orthodontic students’ or orthodontic educators’ knowledge, but no definite conclusions were reached; and future research was recommended. Rosenberg [51] concluded that computer-aided learning is as effective as other methods of teaching and can be used as an adjunct to traditional education or as a means of self-instruction of dental students. Based on 4 mixed-results RCTs, Rosenberg [52] was unable to reach any conclusions on knowledge gains and recommended more high-quality trials evaluating the effectiveness of computer-aided learning in orthodontics. However, we are familiar with newer technologies being currently evaluated for the same outcomes; MOOCs or mLearning can play a very important role in health professions education such as improving clinical knowledge and promoting lifelong learning [53-54]. We are also aware of recent reviews, which reached similar conclusions [55-62]. For example, digital education seems to be at least as effective (and sometimes more effective) as traditional education in improving dermatology, diabetes management, or smoking cessation–related skills and knowledge [58,61,62]. Most of these reviews, however, stressed the inconclusiveness of overall findings mainly because of the low certainty of the evidence.

Offline digital education may potentially play a role in the education of health professionals, especially in LMICs, where there is a lack of access to Web-based digital education for a variety of reasons, including cost; and there is some evidence to support the effectiveness of these interventions in improving the knowledge of health professionals. However, because of the existing gaps in the evidence base, including limited evidence for other outcomes; lack of subgroup analyses, for example, CD-ROM or PowerPoint; low and very low quality of the evidence, the overall findings are inconclusive. More research especially evaluating patient-centered outcomes, costs, and safety (adverse effects); involving those subgroups; and originating from LMICs is needed. Such research should be adequately powered, be underpinned by learning theories, use valid and reliable outcome measures, and blind outcome assessors.