A comprehensive search was conducted across 4 databases—MEDLINE (Ovid), Web of Science, IEEE Xplore, and Scopus—to identify studies evaluating or predicting EHR usability. All searches were limited to English-language publications from January 1, 2009, to April 9, 2025. Keyword strategies incorporated terms related to usability, prediction or analytics, and EHRs.

The search strategy included combinations of controlled vocabulary terms and keywords related to usability (“usability,” “user experience,” “human factors,” “human-computer interaction”), prediction (“predict*,” “regression,” “data mining,” “machine learning”), and EHR (“EHR,” “electronic health record,” “EMR”).

Search strategies for each of the 4 databases were as follows:

This study was reviewed by the University of Victoria Human Research Ethics Board (UVic HREB) Chair and determined to be exempt from formal human research ethics review because the research does not involve human participants and is limited to the analysis of existing, publicly available documents and/or datasets. The determination was made in accordance with the Tri-Council Policy Statement: Ethical Conduct for Research Involving Humans (TCPS2 2022) and the University of Victoria Board of Governors Research Policy RH8100 and University of Victoria Board of Governors Regulations for Research Involving Humans RH8105. The UVic Human Research Ethics Chair concluded that the activities described in this study fall outside the scope of research requiring institutional ethics review because they do not involve human participants, human biological materials, or identifiable private information. The work is limited to the analysis of publicly available documentation and datasets.

The included studies were drawn from 4 major databases: Scopus (n=1526), Web of Science (n=1458), MEDLINE (n=590), and IEEE (n=428). Database searches yielded 428 papers in IEEE Xplore, 1458 in Web of Science, 590 in MEDLINE (Ovid), and 1526 in Scopus. After the removal of duplicates (n=1679), a total of 2323 records underwent title and abstract review in the first screening (Figure 1). A total of 98 full-text papers were assessed during the full-text screening. In the full paper screening, 51 records were excluded because of wrong outcomes (n=22), wrong study design (n=21), studies that did not address usability (n=7), and wrong intervention (n=1). Thus, a total of 47 studies met the full screening eligibility and were retained for analysis.

From 2323 records screened, 47 studies met the inclusion criteria. The majority were cross-sectional surveys that applied TAM, UTAUT, or ISS frameworks to evaluate EHR adoption and US. Common predictors included PU, PEOU, effort expectancy (EE), and facilitating conditions (FCs).

Cross-sectional designs were the most common (31/47), followed by survey-based studies (26/47) (Multimedia Appendix 2). Almost all studies were conducted postimplementation. Data collection methods most frequently included direct data capture (36/47) and questionnaires (31/47). Regression analysis was the most common modeling approach (26/47).

Across the 47 included studies, EHR adoption was most commonly predicted with usability measures that were predominantly subjective, relying on surveys and self-reported satisfaction. Objective indicators—such as task completion time, error rates, and audit log data—were rarely incorporated, and no study combined task-based HCI data with predictive analytics for longitudinal monitoring.

The reliance on survey-based methods, with limited integration of task-level or objective usability metrics, was a consistent finding. Acronym use across the studies (Table S1 in Multimedia Appendix 1) reflects a synthesis of 47 studies from 26 countries (2009‐2025), largely dominated by TAM, UTAUT, and D&M ISS frameworks. Analytic approaches included regression, SEM, and in a small number of studies, machine learning or decision trees. While nearly every study examined adoption or acceptance, very few explicitly focused on prediction even with the EHR adoption constructs with significant determinants.

TAM and UTAUT frameworks in many studies were evaluated using regression and SEM. Some studies used confirmatory factor analysis (CFA) and exploratory factor analysis to validate constructs. A small number integrated machine learning or neural networks with SEM, but no study applied AI to predict EHR usability.

Regression was the most commonly applied predictive analytic method, followed by SEM, which was the most frequently applied analytic methods, supplemented by ANOVA, multivariate ANOVA, CFA, and exploratory factor analysis in 13 studies [18,44-55].

Table S1 in Multimedia Appendix 1 presents 26 reported adoption-related determinants and model constructs such as TAM, tripolar model of technology acceptance, theory of planned behavior, UTAUT, and Systems Engineering Initiative for Patient Safety (SEIPS) that had usability aspects of EHR/EMR that included (1) PU, PEOU, performance expectancy (PE), user behavior (UB), and US; (2) behavioral intention (BI), EE, FCs, and job relevance; (3) IQ, IT support, organizational culture, SQ, and service quality (SRQ); (4) broader usability frameworks and scales, such as the NASA-TLX, National Usability–Focused Health Information System Scale (NuHISS), and SUS.

illustrates the frequency of reported usability metrics. Metrics were grouped into nine categories: (1) attitudes and social aspects, (2) behavioral and intention-related factors, (3) IQ, (4) organizational and environmental factors, (5) PU, (6) SRQ, (7) SQ, (8) US and perceived usability, and (9) work efficiency and productivity.

FC, defined as the degree to which an individual believes that an organization and its technical infrastructure exist to support the use of the system [56], was the most frequently measured determinant construct, or predictor variable, followed by PEOU, PU, and attitude. In total, more than 50 unique usability-related metrics were reported across the 15-year span of literature.

The studies spanned 16 years from 2009 to 2025, with 12 published before 2015 (Multimedia Appendix 3). Study duration was, on average, less than 1 year and ranged from 2 months to 14 years. No clear temporal trend was observed, though extending the screening into 2023 to 2025 revealed an increasing number of studies. However, databases searched in 2025 only included papers published in that year up to April 9.

Geographically, the 47 studies originated from 26 countries (Multimedia Appendix 4). The United States (10 studies) and Finland (5 studies) contributed the largest number of studies found. North American studies more often employed the D&M ISS model compared to other countries, while nursing-focused studies frequently adopted the UTAUT framework.

Most frequent keywords across studies were EHR, usability, TAM, and UTAUT (Multimedia Appendix 5).

Across all studies published between 2009 and 2025, no study directly predicted EHR usability using computational or longitudinal modeling techniques. Most research examined adoption outcomes, satisfaction ratings, or intention-to-use constructs rather than usability performance or workflow-based usability metrics. Psychometric determinants (eg, PU, PEOU, EE, and FC) were frequently evaluated across adoption frameworks.

Regression and SEM techniques predominated, supporting cross-sectional inferences but offering limited value for forecasting usability over time. Although FCs often emerged as a significant predictor of intention-to-use EHRs, they were rarely examined within dynamic contexts, task-based assessments, or sequential workflows that reflect actual clinical use.

There were 95 predictive techniques of analysis across 12 technique categories found, with many studies using multiple techniques (Multimedia Appendix 6). Regression-based methods dominated (40/95) [15,17,18,27,44,45,47-49,51,55,57-73] (Multimedia Appendix 7) followed by SEM (26/95) [50-55,74-77]. Furthermore, there were several regression and SEM variants in the studies found.

Only a limited number of studies incorporated AI-related techniques, but these were used for prediction of EHR adoption and not for usability. Almarzouqi et al [78] used neural networks with SEM to predict adoption, Alsyouf and Ishak [79] used decision trees with heuristic evaluation and UTAUT, and Sachdeva et al [74] used decision trees to discover behavior patterns. Cluster analysis was used to predict general BI [79] and evaluate EHR use in primary care providers [70]. Four studies used principal component analysis for dimensionality reduction and feature extraction [44,45,70,74].

Across 47 studies, regression and SEM dominated, indicating minimal methodological progression toward predictive, automated, or real-time usability modeling.

This scoping review found that EHR adoption models dominated studies identified using usability and prediction-related keywords. However, adoption is not synonymous with usability, despite often being treated as interchangeable in the literature. As noted by Shachak et al [80], models such as TAM and UTAUT reduce complex sociotechnical properties of health information technology to individual perceptions and intentions, privileging acceptance over system-level performance and postimplementation use. Consequently, usability was assumed through adoption intent rather than directly measured or predicted.

The predominance of EHR adoption models in the reviewed studies underscores the need to address postimplementation system success. EHRs are complex sociotechnical systems in which usability emerges from interactions among interface design, configuration, organizational context, and user roles, making post hoc evaluation insufficient for meaningful intervention.

The continued use of TAM, UTAUT, and D&M ISS models highlights an enduring emphasis on explaining acceptance and satisfaction rather than modeling or forecasting usability performance in clinical practice. Developing predictive models could shift usability research from reactive post hoc assessment to proactive identification of inefficiencies, workflow burdens, and potential contributors to clinician burnout.

This scoping review highlights that although the TAM, UTAUT, and D&M ISS provide useful foundations for understanding adoption, they offer limited ability to capture evolving usability patterns with task complexity or changes in user interaction over time. Poor usability (clunky interface and workflow disruption) hinders adoption and satisfaction, causing burnout, whereas good usability (intuitive design and efficient workflows) drives adoption and better patient outcomes, even if a system is technically adopted. Thus, an EHR can be widely adopted but still have terrible usability that leads to frustration and workarounds. Therefore, predicting usability adds value to adoption and more insight as the system becomes operationalized.

The predictive focus distinguishes this study. Zhang and Walji [12] conceptualized EHR usability as a balance between intrinsic complexity (reflecting the work domain’s inherent demands) and extrinsic difficulty (stemming from interface and workflow design). Their TURF framework—task, user, representation, and function—offers a roadmap linking usability determinants such as usefulness, satisfaction, and efficiency. Despite its potential, none of the included studies applied TURF to predictive modeling of EHR usability. EHR metadata, including audit or task logs, could serve as the foundation for predictive models of usability and clinician workload or burnout [34]. Significant predictors, such as PEOU and PU, highlight opportunities for AI-driven approaches that integrate adoption constructs with task-level metadata to advance predictive usability research.

Across the 16-year period reviewed (2009‐2025), very few studies attempted to predict EHR usability directly, regardless of model sophistication. The facilitating conditions, such as the predictor variable of EHR adoption, appeared to be significant and impacted the intent to use EHRs. The FC determinant is defined as “the degree to which an individual believes that an organization and its technical infrastructure exist to support the use of the system” [56]. This psychometric determinant was deemed to be a useful metric for predicting adoption.

FCs have a direct positive effect on the intention-to-use, but this effect often diminishes after initial system use, suggesting its relevance may shift toward sociotechnical factors affecting learnability and decision-making. That is, the FCs would have linkages to sociotechnical aspects related to learnability, which is one of the factors in usability as defined by Nielsen [81]. The implications for understanding HIC and clinical decision-making suggest that survey-based constructs may insufficiently capture deeper usability determinants rooted in workflow, cognitive processes, and task structure, as mentioned by Kushniruk et al [82]. The review showed that several countries represented in their use of EHR adoption models over the past 15 years to find success or failure in the actual or intended use of EHRs postimplementation.

None of the reviewed models predicted EHR usability. In addition, no models incorporated detailed task-level inputs (eg, time-on-task, error rates, and interaction complexity) into predictions of intended or actual system use. Instead, many studies focused on PEOU or PU as attributes of usability and not tasks. Nevertheless, these attributes are important to establish (eg, easy to learn, efficient use, easy to remember, few errors, and subjectively pleasing) as recognized core usability attributes in the concept of usability engineering in the 1990s as defined by Nielsen [81]. Therefore, this result from the literature review makes it clear that predicting EHR adoption based on BI and actual use is still under investigation, with very limited modeling related to EHR usability. Importantly, the models used in the papers did not state that categories or aspects of usability and their usefulness could be dynamically utilized postimplementation.

Logistic regression or hierarchical regression via SEM with many predictive variables and hypotheses was studied in several of the papers. However, these approaches generally reflect relationships among perceptions rather than true indicators of usability performance. Aspects of task completion or failure appear directly related to usability, yet they were rarely used to establish predictive usability metrics. Furthermore, SEM in the studies [48,50-54,59,64,75,76,78,83-86] showed complex networks of hypotheses linked through latent constructs. This highlights the need for more evidence connecting specific usability issues (eg, coded problems from video-based testing [82]) with operational real-world EHR use, which takes time, and predictive models of EHR usability could reduce the burden of large and extensive usability testing.

The literature indicates an increasing sophistication in applying TAM, UTAUT, and D&M ISS constructs by researchers, often through SEM. Many predictive analytics applied within these models reflect the complexity of usability-related factors. Researchers frequently tested 10 to 15 hypotheses of TAM, UTAUT, or ISS in the analysis, demonstrating rigorous statistical approaches. However, the extensive statistical testing often highlighted variability in the acceptance or intention to use the system postimplementation rather than convergence toward models capable of predicting usability. A major trend was the use of TAM or UTAUT with SEM to evaluate multiple factors—even though many of these factors could, in principle, support predictive usability modeling, for example, by linking task-based observations with survey responses.

SEM’s prominence reflects its origins in psychology and social sciences, where latent constructs such as attitudes and intentions have long been modeled through survey-based items. The Task–Technology Fit framework has also been used in SEM to assess alignment between tasks and IT [87]. CFA was commonly employed to evaluate usability parameters within hierarchical regression models [48]. However, the reviewed studies modeled adoption, satisfaction, or BI, which are important and should be done in unison with usability testing to evaluate its efficiency with UB and acceptance.

The 47 studies, representing 26 countries, indicate global interest in EHR adoption research. Regional differences were minimal, although the D&M ISS model was used more frequently in North America. In 2024, Rihayha and Ismaili [52] proposed a modified human–organization–technology fit model for resource-limited settings, integrating SQ, IQ, and SRQ via SEM. Hyppönen et al [48] proposed adoption predictors, including user friendliness, perceived benefits, technical problems, and cooperation metrics, applying NuHISS in Finland. Welchen et al [88] applied NuHISS in Brazil through a star-schema confirmatory analysis. Even though NuHISS is very detailed, it lacks objective usability measures such as error rates and time-on-task. The use of NuHISS across multiple countries was a major finding, though no multicountry comparative studies were identified.

Studies have updated versions of the TAM [16,50,57,58,69,76,79,89], TAM2 [66,84], and TAM3 (a hybrid of TAM and UTAUT) [63,66,78,86]. Important TAM usability determinants included PEOU and PU [16,44,48,50,54,57,58,60,63,64,66-68,70,78,83,85,89,90]. PU and PEOU can significantly predict health care professionals’ intentions to use health information systems [13,14,58,73]. Additionally, subjective norms, relevance, and computer anxiety predict intentions beyond PU and PEOU [87,88]. One study showed that PEOU was influenced by computer self-efficacy, external control perceptions, anxiety, and enjoyment [55]. Chi and Ku [61] showed that voluntariness affected determinants of actual use, with EE increasing in importance as voluntariness decreased. Despite their importance, PU and PEOU may be undervalued in their significance for predicting BIs, especially when contextualized within newer constructs.

UTAUT consistently demonstrated stronger predictive performance than TAM. Shiferaw and Mehari [53] reported that UTAUT explains up to 70 % of the variance in the intention-to-use technology. However, Bagozzi [91] critiqued the model in that it coheres the “many splinters of knowledge” to explain decision-making and its subsequent extensions, stating “UTAUT is a well-meaning and thoughtful presentation,” but “it presents a model with 41 independent variables for predicting intentions and at least 8 independent variables for predicting behavior.” Furthermore, Li et al [92] argued that UTAUT’s reliance on moderators complicates model interpretation and may artificially inflate R² values. They also criticized construct labeling practices that combine disparate items into a single psychometric measure.

Nevertheless, UTAUT adds self-efficacy, social influence, and FCs, replacing TAM’s PU and PEOU with PE and EE. For example, one research paper found that self-efficacy was a significant predictor variable in comparing urban, rural, and remote settings in the Philippines [49], indicating that one’s ability to succeed or accomplish a task significantly and positively increased the intention to use an EHR in that setting. FC is a direct and commonly significant determinant of UB in the UTAUT framework with many subfactors [91-95].

Several tools and frameworks have been developed to assess EHR usability. A few papers use the SUS and NASA-TLX [69,79]. However, this review identified no studies that incorporated ISO-based frameworks [96] as evaluation or predictive factors for EHR usability or adoption.

There was mention of the Health Information Technology Usability Evaluation Scale. This scale is specifically designed for evaluating the usability of health IT applications, including EHRs. It includes subscales for measuring PU, PEOU, and user control, among others [37]. Additionally, Carayon et al [97] proposed the SEIPS model, which provides a framework for understanding the relationships between technology, work processes, and patient outcomes in health care settings. This review identified a study by Alsyouf and Ishak [79] that used SEIPS to predict and evaluate the usability of EHR systems by identifying potential mismatches between the technology and the users’ workflows.

The regression model by Butler and Johnson [60] did not predict EHR usability but rather productivity, which may be related to usability. Butler and Johnson [60] used an estimation framework for a productivity index based on physician characteristics, EMR functions, and vendors. Another study by Sulley and Ndanga [71] predicted US among physicians and different EHR system types or vendors and found that the EHR system type was significant. In addition, this study suggests that some EHR vendors can have higher satisfaction levels, which could be linked to the Butler and Johnson estimation of productivity using EHR. Moreover, Yin et al [73] used a logistic regression model and found that participants who are more likely to consider EHR portals as valuable tools were those who found EHR portals useful for information.

Therefore, a regression model of productivity with usability factors and EHR components across vendors could be an alternative for predicting usability instead of adding external usability variables to more common EHR adoption frameworks, such as constructs of TAM, UTAUT, and D&M ISS models. Furthermore, cognitive aspects, such as the theory of reasoned action, theory of planned behavior, and a user’s motivation to use IT, are influenced by the system features and capabilities in TAM, UTAUT, or D&M ISS model constructs. Further investigation is needed to predict EHR usability based on productive HICs, the incorporation of aspects of EHR adoption frameworks, and other related models, such as Health Information Technology Usability Evaluation Scale, human factors engineering, SEIPS, and many more.

Emerging technologies, such as AI, machine learning, and advanced data mining, offer opportunities to extend these models and utilize large datasets. There is almost no research using AI with TAM, UTAUT, and D&M ISS.

The EHR adoption models initially emerged several decades ago. AI algorithms could add feedback loops or feedforward recursive networks to adjust to different data over time, for example, sequences of tasks, usability, and actual use of EHRs under different scenarios. Aspects of EHR usability could be added to the TAM and UTAUT models with training and test datasets via machine learning potentially integrated within SEM and CFA frameworks. Machine-learning techniques have been applied to the usability of health information systems [98].

Integrating audit log data, clickstream analyses, and HCI measures can enable predictive modeling of usability performance, informing adaptive design and proactive system improvements. There is evidence of increasing use of EHR audit log files as evidence for the macrostructure of EHR work completed [99]. Furthermore, there is a viewpoint that there are emerging domains for measuring health care delivery with EHR metadata: (1) team structure and dynamics, (2) workflows, and (3) cognitive environment [100]. These domains contain modified index or scoring, continuity index, comprehensiveness index, conformity score, undivided attention, and task switching [100].

Furthermore, Rule et al [34] stated that metadata from EHR from audit logs can be used to predict physician burnout, which could be considered a usability metric. Moreover, even if there are more recent studies that utilize metadata from EHR system logs, a major limitation is the absence of explicit task-level variables via large datasets to predict usability over time. There is, nevertheless, capacity to include such task measures in SEM-based models as determinants of PE, self-efficacy, EE, and traditional TAM constructs such as PU and PEOU.

Therefore, researchers should consider combining established adoption frameworks with continuous usability monitoring and predictive analytics. Predictive approaches—leveraging design attributes, certification data, usage patterns, and organizational indicators—could enable early identification of usability risks, when redesign and mitigation remain feasible. Furthermore, predicting EHR usability could be used for analyzing large datasets now being collected on EHR usability, anticipating usability problems with specific systems and supporting selection and procurement decisions for acquiring more usable EHRs. The work in this paper could form the groundwork for establishing future efforts in analyzing and predicting usability. Future work could develop feedback loops that link UB, system tasks, and usability outcomes, providing actionable insights for EHR vendors and health care organizations.

No study in this review developed a predictive model for EHR usability postimplementation only, and very few studies used AI to predict adoption. To advance both EHR adoption and EHR usability sciences, researchers could bridge gaps between static adoption models and dynamic human factors engineering by integrating AI, audit log analysis, and real-time usability testing.