This scoping review was conducted in accordance with the Johanna Briggs Institute methodology for scoping reviews [16,17] and reported following the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews; Checklist 1) and PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Search) [18,19]. We were guided by Arksey and O’Malley’s [20] framework with the additions of Levac et al [21]. A protocol was prospectively registered on the Open Science Framework (OSF). Amendments to the protocol were updated and uploaded to the OSF [22].

An electronic search was conducted in 6 databases (MEDLINE [OVID], Embase [OVID], CENTRAL [OVID], CINAHL [EBSCO], Compendex, and Web of Science) and 4 gray literature sites (OpenGrey, GoogleScholar, arXiv, and medRxiv) with the aid of a biomedical librarian and information specialist. Our search strategy was not peer-reviewed but was tested through an iterative process by the biomedical librarian to ensure that search strategies returned identified seed papers. Initial search strategies were adapted from previously published work [11,14]. We searched databases and gray literature from inception to September 18, 2025. Table 1 provides the population-concept-context framework for our search strategy and illustrates how we operationalized our search. The full MEDLINE search strategy is outlined in Multimedia Appendix 1, and all other search strategies are uploaded and available on OSF [22]. To supplement the search, we screened the reference lists of relevant reviews and included records. We also searched the Cochrane Database of Systematic Reviews, PROSPERO, OSF, and JBI Evidence Synthesis to identify any active systematic or scoping reviews on the topic. The search approach for identifying gray literature is detailed in Multimedia Appendix 2.

Studies of adults (aged 18 years and older) with musculoskeletal conditions (≥25% of the sample had to be musculoskeletal-related) that identified and reported a digital health tool designed specifically to triage or diagnose in primary, urgent, or emergency care settings were included. Textbox 1 describes the inclusion and exclusion criteria. We excluded studies that evaluated the effectiveness of virtual assessments. We also excluded studies that used digital health tools for secondary diagnoses (ie, patient had already seen a practitioner and given a diagnosis), as the tools were typically used to manage symptoms and not for primary triage or diagnosis.

Records were collated and uploaded into EndNote (version 20.3; Clarivate Analytics), and duplicates were removed before uploading to Covidence (Veritas Health Innovation) for screening. Pairs of independent reviewers screened all records by title and abstracts, and a third reviewer (CLA) resolved any discrepancies if consensus could not be reached.

At the full-text stage, we first conducted pilot screening, where all reviewers assessed the same 5 full texts. If major discrepancies were identified, we met to review and discuss how to apply the screening criteria in a standardized manner. All full-text papers were reviewed by pairs of independent reviewers. Reasons for exclusion during full-text screening were recorded. Any disagreements between the reviewers at each stage of the selection process were resolved through consensus or by an additional reviewer (CLA) as required.

Data were extracted from included records independently by pairs of reviewers using a custom data extraction tool designed in Microsoft Excel by the research team. Any disagreements were resolved via consensus. We extracted the following details where available: study characteristics (author, country, sample size, and study aim), participants’ demographics (sex, age, and musculoskeletal pain or diagnosis), type of digital tool (name, purpose of tool, and target users), design and development process, platform of tool, tool delivery (eg, clinician or patient self-access), context or care setting in which the tool was used, assessment of performance or accuracy results, and key findings relevant to the review question. Where relevant, authors were contacted once via email to request missing data and to clarify details about the digital health tool.

Studies were summarized by study characteristics, digital health tool details, and performance assessment. Descriptive data were summarized as proportions when appropriate. Digital health tools were classified according to the World Health Organization digital health category [10] (eHealth, mHealth, and AI or machine learning), function (ie, triage, diagnosis, or both), care setting in which the tool was used (ie, ED, primary or urgent care, or mixed), research setting (ie, urban, rural, or both), how the tool was administered (eg, self-access or clinician-delivered), technology interface (eg, web-based or app), and intended user (patient-facing, clinician-facing, or both).

Digital tools were rated using the technology readiness level (TRL) and associated technology stage by the first author (LKT) and verified by a second rater [24]. The TRL ranges from 1 to 9, with 1 representing tool conception and 9 representing that the tool is ready to be used in real-world settings [24]. We classified TRL across technology stages (fundamental research, research and development, pilot and demonstration, early adoption, and commercially available) [24].

When available, the performance of the digital health tool was reported by identifying appropriate triage referrals or recommendations or diagnoses compared to a reference standard (eg, physician diagnosis). Measures of performance included diagnostic test accuracy (area under the receiver operator characteristic curve, sensitivity, specificity, positive predictive value, negative predictive value, and likelihood ratio) or reliability measures (internal consistency, test-retest reliability, and intra- or interrater reliability).

We sought to use the continuous active learning on Covidence to support title and abstract screening [25]. During attempts to calibrate the algorithm, the algorithm did not perform well in identifying relevant papers for this review. We were not confident in screening titles and abstracts with only 1 human reviewer (LKT). Instead, all titles and abstracts and full text records were screened in duplicate by 2 independent human reviewers.

As scoping reviews are an iterative process and aim to assess and evaluate the available evidence, a broad research question often results in a highly sensitive search and less specific records. We made pragmatic decisions and minor amendments to the selection criteria as the review progressed. At the full-text screening stage, studies including a general population had to report ≥25% of the sample being musculoskeletal-related to be included. This cutoff was determined based on studies that indicated the prevalence of musculoskeletal presentations in ED was approximately 25% [26,27].

The titles and abstracts of 5695 unique records were screened, and 189 papers were reviewed in full (Figure 1). In total, 34 studies met the inclusion criteria (n=37,509 participants across 33 studies, 12,470/37,509, 33% female). The median age was 50 (range 18-91) years. One study did not report the sample size. Sex and age data distribution were missing in 12 and 13 studies, respectively. A list of the studies that were excluded at the full-text stage is presented in Multimedia Appendix 3. Figure 2 charts the data of all 34 studies included by condition, publication year, and sample size of the study, if reported.

In total, 30 of 34 (88%) studies focused on primarily musculoskeletal conditions, and 4 of 34 (12%) studies focused on general populations with a subset of the sample being musculoskeletal conditions. Full study details can be found in Multimedia Appendix 4 [28-61].

In total, 25 of 34 (74%) studies were peer-reviewed, 7 of 34 (21%) records were conference abstracts, and 2 of 34 (6%) were industry case reports. Studies were published between 2010 and 2025. Cross-sectional (10/34, 29%) studies were most common, followed by nonrandomized or quasi-experimental studies (8/34, 24%), randomized controlled trials (5/34, 15%), retrospective cohort (4/34, 12%), mixed or multimethods (3/34), prospective observational cohort (2/34), and case report (2/34, 6%). All studies were conducted in high-income countries (n=1 country not reported), with the majority being from Europe (21/34, 62%) or the United States (7/34, 21%).

Digital health tools were evaluated across various care settings, including ED or urgent care (6/34, 18%), physician-led primary care (6/34, 18%), physiotherapist-led primary care (1/34, 3%), patient self-access (19/34, 56%), and mixed (eg, primary care and ED) settings (2/34, 6%).

Of the 34 studies identified, 19 (56%) evaluated the performance of the digital health tool (Table 3). The most common definition and method to determine performance was measuring concordance to a clinician diagnosis or recommended triage pathway, often by evaluating sensitivity and specificity. The performance of these tools varied widely and was partly dependent on the context in which they were being used (eg, ED or primary care). Sensitivity ranged from 39% to 91%, and specificity from 23% to 80% [28,30,31,37,41,45,46,48]. The methods for measuring accuracy were poorly reported, often in the form of proportion of correct triage or diagnoses. Reported accuracy ranged from 33% to 98% across 12 unique tools (n=16 studies) [28,29,34,36-38,41,45,49,51,59,61]. The accuracy of tools used by patients in tertiary settings (eg, seeking care from a specialist such as an orthopedic surgeon or rheumatologist) was reported as higher than the accuracy of tools used in primary care settings.

For studies that compared digital health tools against each other, ADA was the most common tool used for comparison. When compared to rheumatologists and medical students, ADA was superior to clinician’s diagnosis of rheumatic and nonrheumatic conditions, ChatGPT, and Bechterew-check [37,41,49]. ADA was comparable to Rheport for diagnostic accuracy of rheumatic conditions [48]. ChatGPT performed similar to experienced rheumatologists for potential diagnostic accuracy for rheumatic conditions [49].

In total, 4 tools were available in multiple languages (ADA, ChatGPT, Rheumatic?, and WebMD Symptom Checker), and 8 tools were accessible to the public; however, 2 were designed for German speakers (Table 3). Based on the TRL, we classified 6 tools as being at the commercially available stage (ADA, Buoy Health, ChatGPT, Phio, Therapha, and WebMD Symptom Checker).

Figure 3 provides a visualization comparing TRL and performance evaluation for the identified digital health tools (Only 15 of the 16 identified digital tools are reported in this figure. The “phone built in camera” was not graphed.). If the tool did not complete a performance evaluation of the tool, a 0 was given for reported performance (ie, accuracy) on Figure 3. Despite some tools being commercially available, there was a discrepancy in reported performance findings for musculoskeletal conditions.

We aimed to identify and describe the available tools for triaging and diagnosing musculoskeletal conditions in primary, urgent, and emergency settings. Based on a synthesis of 34 studies and data from 16 different digital health tools, there were no digital health tools with sufficient evidence to support effective triage and diagnosis of musculoskeletal conditions across these settings. Approximately half of these tools were available to the public. Not all tools were available in English, with 2 tools only available in German (Bechterew-check and Rheport). The most frequently studied digital health tool was ADA (n=5), followed by Rheumatic? (n=4), then ChatGPT (n=3). Only 2 tools (DART and Phio) were purposely developed for screening musculoskeletal conditions. Both tools are not currently available outside of the UK’s National Health Service. We were surprised to find so few digital health tools targeting musculoskeletal conditions, given the substantial global burden of musculoskeletal conditions [1]. Notably, rheumatological or inflammatory arthritis was the most prevalent musculoskeletal condition studied, despite low back pain being the most common musculoskeletal condition seen in ED and primary care settings [62]. We identified 4 studies that included digital health tools targeting low back pain, but only 1 of these reported which tool was used (Therapha). Our findings reflect the discordance of research across digital health technology and the current health landscape. Many tools were inaccessible or not designed for practical use in managing musculoskeletal pain, the most burdensome conditions seen in primary care.

Our secondary objective was to summarize the performance and accuracy of the included digital health tools. Approximately 50% of the studies evaluated the performance of a digital health tool. Apart from ChatGPT, most generic digital health tools (eg, ADA and WebMD Symptom Checker) reported poor accuracy (often less than 50% accuracy in identifying the correct diagnosis compared to clinicians) for musculoskeletal conditions [37,40,48]. Despite the use of ChatGPT by the public as a symptom checker [15], ChatGPT’s accuracy for diagnosing musculoskeletal and rheumatic conditions was variable, ranging from 33% to 98% [29,34,49]. We suggest that further research is needed before considering ChatGPT as an accurate diagnostic or screening tool. Tools that were designed to diagnose peripheral or spinal musculoskeletal conditions (eg, low back pain or knee injuries) appear to be more promising with high sensitivity (88%‐91%) [28,31]. Finally, tools designed specifically to triage (rather than diagnose) musculoskeletal conditions (ie, Phio and DART) demonstrated the best performance. Recent findings published on DART and Phio indicate that these tools have high agreement (>90%) with expert physiotherapist recommendations on next care pathways [63,64]. However, the heterogeneity across evaluation methods highlights the importance of standardized development and evaluation frameworks to ensure that digital triage tools for musculoskeletal conditions are accurate, transparent, and safe before integrated into clinical settings and workflows.

One of the key findings of our review is that some tools are commercially available and integrated into health systems for musculoskeletal screening without robust methodological evaluation or reporting. Premature implementation raises concerns, particularly given the risk of misdirecting patients or delaying appropriate care. Before being adopted at scale, digital triage tools must demonstrate value in real-world settings and meet minimum standards for safety, accuracy, and usability. However, many studies evaluating these tools lack transparent reporting, making it difficult to assess how performance claims were derived. Studies reporting high accuracy of their digital tools often had poor transparency or a lack of details on their tool evaluation. We suggest caution with interpreting these digital health tools as ready for public use without further evaluation. It is also unclear how tools that use LLMs operate. These networks are often termed “black boxes” due to the inability to explain how these systems achieve their output [65].

Our findings have been confirmed by a recent study that evaluated diagnostic accuracy and clinical reasoning using 6 different generative AI (LLMs) for rheumatic diagnoses [66]. Despite the LLMs reporting high diagnostic accuracy (~80%), all models reported subpar clinical reasoning quality (eg, explaining reasons for supporting diagnoses) [66]. These findings underscore the importance of digital health tools requiring both high diagnostic accuracy alongside transparent algorithms to help to explain the logic behind the tool’s decision. To improve transparency and enable reproducibility, it is important to establish standards for incorporating ethical AI in digital health. Without transparency in how tools were developed, or in the algorithms used, it is unclear whether the tools are safe for the public to use.

The only digital health tool with robust evaluation of its performance was the generic health app, ADA, which is a Conformité Européenne–certified medical product [46,48]. ADA’s performance was inconsistent across the studies, and ADA correctly identified the musculoskeletal condition or triage option in fewer than half of the cases [37,41,46,48,49]. Condition-specific digital health tools (Rheport [46,48], Rheumatic? [45], Virtual Knee Doc [30,31], Therapha [28], and ReumAI [38]) performed slightly better. The reported accuracy was higher in these tools, especially if these tools were implemented in tertiary care settings (outside of the ED or primary care). We are not aware of an acceptable threshold for performance (ie, accuracy) for digital health tools. However, we recommend implementing tools that are at least more accurate than flipping a coin and provide consistent results across different study contexts or musculoskeletal conditions.

AI-driven tools, like ADA or ChatGPT, may perform better than clinician decision support systems or physicians or rheumatologists in diagnosing rheumatic conditions [49,67]. Integrating digital health tools in tandem with other nonspecialist professions (eg, general practitioners and allied health professionals) could help guide patients to their next care steps as they wait for specialists (eg, rheumatologists) or avoid unnecessary visits to specialists or other care providers. AI-driven tools that have included diagnostic findings (eg, imaging, clinical symptoms or signs, and bloodwork) have superior diagnostic accuracy to other AI models [41,67,68]. Until robust stand-alone digital health tools are developed (ie, a symptom checker that can be used independently by patients), combining digital health tools and clinician feedback may be the best method to streamline diagnosis and care in complex cases while providing timely care for common musculoskeletal conditions.

Several frameworks for evaluating digital health tools have been proposed [69]. A recent scoping review identified 12 key domains—ranging from tool description and content to safety, clinical effectiveness, and efficacy—across 95 frameworks that developers and researchers can draw on [69]. However, the heterogeneity reflects a broader challenge: many digital health tools span multiple categories (eg, eHealth or mHealth tools incorporating AI), making classification inconsistent and evaluation difficult. Advancing this field requires standardized terminology, harmonized testing and evaluation frameworks, and clear reporting guidelines—crucial steps to ensure both progress and patient safety.

Through the process of screening studies for inclusion into our review, we found definitions of musculoskeletal conditions that were vague and varied widely. Definitions of “musculoskeletal” are often limited to orthopedic conditions or pain related to musculoskeletal structures [1,23]. However, musculoskeletal conditions are a complex category involving heterogeneous conditions, such as rheumatological or inflammatory arthritis or gout, that are not typically grouped as musculoskeletal in clinical practice. We relied on a broad definition to capture specific musculoskeletal conditions (eg, rheumatological conditions, arthritis, and gout) and pain related to musculoskeletal structures (eg, sprains and strains).

There is nuance in how triage would be conducted for acute versus chronic musculoskeletal conditions, including screening questions related to condition pathophysiology, subjective history, pattern of symptoms, and disability (eg, red flags), which might explain some of the variability in performance metrics of different digital tools [70]. Early diagnosis and treatment planning is often iterative for those with musculoskeletal conditions and varies depending on the condition. For example, targeted medication plays a vital role in managing rheumatological conditions [71], whereas some orthopedic conditions are managed with exercise and minimal pharmacological interventions [2]. This complexity will impact triage algorithms by influencing treatment recommendations (eg, who the patient should see) and timing of care (eg, urgent or wait-and-see). Therefore, tools that have high accuracy (ie, good performance) for triaging and diagnosing general health conditions may not necessarily have the same effectiveness when applied to musculoskeletal conditions.

Digital health tools may perform poorly at screening because of user error relating to symptom data entry and patient interaction with the tool. One solution to this is adding more key information (eg, diagnostic tests) to an AI-driven model to improve the diagnostic accuracy of the model [67]. We also suggest future work to involve patient end users to develop and refine digital health tools. Most digital health tool algorithms are derived from clinicians’ clinical reasoning, which may not follow the same thought process as a patient. In a recent qualitative study exploring how patients should be engaged in AI application to health care, patients felt that the priorities of researchers, particularly for AI tools, were to improve efficiency and effectiveness of care [72]. In contrast, patients were more interested in using AI to address issues related to accessing health care [72]. Patients should be involved early in the design and development phases to enhance the usability and understandability of digital health tools. However, patient perspectives are often included only after the digital health tool is designed. We argue that engaging patients early in the development process, such as developing the AI algorithms, may yield more acceptable and usable digital health tools.

It is unlikely that a “one size fits all” digital health tool can effectively diagnose and triage all musculoskeletal conditions. Most patient-facing tools in our review were web- or app-based tools in the form of generic symptom checkers. ChatGPT has an accessible interface and is relatively easy to use [15]. Clinician-facing tools may benefit from greater complexity or condition specificity, depending on the context in which the tools will be implemented. Instead of an either-or—general or condition-specific—we advocate for designers to consider their design goal (ie, triage or diagnosis) and intended user (ie, patient or clinician), which may improve accuracy in digital health tools for musculoskeletal conditions.

The field of digital health is growing and changing rapidly. Many health systems have been forced to move toward implementing digital health, particularly AI-driven tools, without being afforded adequate time and resources to consider safety, effectiveness, or downstream consequences [13]. This may be in part due to social and political imperatives to set key performance (productivity) indicators, transition of health care services, and drive toward greater and faster innovation. We suggest that such a climate could be dangerous for health care, especially if digital health implementation continues without adequate evidence, as our findings highlight.

There is a place for digital health triage tools used by patients and clinicians in the current health care context. Self-referral and symptom checkers can be effective for musculoskeletal conditions and to support patients’ access to care, particularly when patients do not have a consistent primary care team or provider [11]. Acute care clinics using a self-referral form found that patients with musculoskeletal conditions were accurate at self-referring, used less health care, and incurred fewer costs [73]. Emerging evidence also indicates that patients are using LLMs such as ChatGPT to make health care decisions, and it appears that the general public is accepting of using AI for health care advice and psychological support [15]. However, more research is needed to ensure that patients presenting with musculoskeletal conditions have a safe, accurate, and well-designed tool to direct them to the best care for their situation. Digital health tools also need to be designed to suit diverse populations, including those with low health literacy and limited digital literacy.

While there is a breadth of studies available for digital health and digital triage, we identified the following knowledge gaps: (1) reporting and transparency on digital health tool development must improve, (2) evaluating digital health tools needs a standard approach, (3) studying the accuracy of triage recommendations requires robust prospective studies, and (4) implementing musculoskeletal-focused digital health tools for first point-of-contact care requires attention.

Despite the absence of digital health tools for triage of musculoskeletal conditions, we are aware of other tools in development, such as SupportPrim [74], which might fill some of the knowledge gaps for health care providers. Our findings do not provide conclusive evidence to support using digital health tools to accurately screen musculoskeletal conditions in many health settings. We recommend that clinicians use these digital health tools as an adjunct to help guide patients, particularly when used as a symptom checker, but to still defer to sound clinical judgment and help patients understand the limitations of the tools.

Although we used a thorough search of published and unpublished data, it is possible that we have missed relevant digital health tools or papers. We set a sample threshold of at least 25% of the sample population with musculoskeletal conditions, and this may have resulted in us missing some studies (eg, studies that were just below the threshold were excluded). The threshold was intended to maximize external validity [26,27]. Our goal was to identify tools that were primarily designed to triage or diagnose (vs manage) musculoskeletal conditions. Therefore, we excluded studies and tools that were designed for self-management, even if they included a symptom checker. This led us to exclude studies that used tools for secondary triage or diagnosis (ie, used by patients who had a diagnosis or had already been seen in a primary or emergency setting) as we wanted to capture tools that could be used at the first point-of-contact. We identified some potential musculoskeletal-specific digital health tools that could be used for secondary triage or diagnosis (Multimedia Appendix 5). While we attempted to report on the performance and accuracy of the tools identified, some tools pooled data from the entire population (ie, not musculoskeletal only). Therefore, the findings may under- or overestimate the accuracy of the tool for musculoskeletal conditions. This again points to the need to design musculoskeletal-specific tools and carefully evaluate their performance.

The rapid growth of AI and digital health solutions is transforming health care systems worldwide, with increasing interest in automating triage and diagnosis. However, our review shows that musculoskeletal conditions remain a blind spot: few tools were specifically designed for this purpose, and most performed poorly when applied to musculoskeletal populations. Despite commercial availability and implementation in some settings, the evidence base was weak, and tool performance was inconsistent and opaque. Health systems and clinicians should exercise caution before integrating these tools into care pathways. Musculoskeletal-specific digital tools developed through transparent, standardized processes are urgently needed to ensure safety, clinical value, and trustworthiness.