Designing AI tools to advance health equity in resource-constrained low- and middle-income countries

Abstract

In the resource-constrained settings, shortages of health workers and underdeveloped infrastructure hinder the delivery of equitable, high quality care. This perspective outlines strategic principles to support the design of AI tools that are genuinely beneficial in low-resource contexts: adopting problem-driven approaches, understanding the socio-technical context, selecting appropriate clinical tasks, and ensuring point-of-care accessibility, clinical comprehensibility, and actionable recommendations, ultimately improving clinicians’ decision-making and promoting health equity in low- and middle-income countries.

Keywords

artificial intelligence global health health equity digital health clinical decision support resource-limited settings

Introduction

Artificial Intelligence (AI) is increasingly being embedded into clinical decision support (CDS) tools to improve care delivery and patient outcomes.^1–3 Yet, adoption remains limited in environments where resources for high-quality healthcare are constrained.^1–3 Key barriers include restricted access to specialist doctors and underdeveloped infrastructure.^4–7 In such contexts, AI can be potentially transformative, expanding access to health services and strengthening systems that underpin health equity.^8,9 By supporting better decision-making and optimising resource allocation, AI has the potential to improve access and reduce disparities across both developing and developed nations. Even in high-income countries, resource-constrained settings persist especially in rural and regional areas,¹⁰ Indigenous health services¹¹ and safety-net organisations.¹²

While global health informatics has advanced in developed nations and tertiary centres, there remains limited literature on the design of clinician-facing AI tools in resource-constrained settings.¹³ A recent review found that 91% of AI evaluation studies in clinical settings were conducted in secondary or tertiary care environments, and all were undertaken in high-income countries.¹ This gap limits the contribution of AI to equitable health outcomes globally. Recent contributions have focused on human-AI interaction,¹⁴ frameworks for AI-based clinical trials,¹⁵ and guidance for translating AI into tertiary academic centres.¹⁶ This perspective aims to outline approaches to designing clinician-facing AI tools through a socio-technical lens, with the aim of advancing health equity in resource constrained low- and middle-income countries (LMICs). We outline eight principles to support safer and more effective AI tools in these contexts. These principles extend beyond the foundational requirements of high-quality datasets, model performance, and model transportability, which continue to challenge tool development across the AI lifecycle.⁸

Eight principles for designing AI tools for resource-constrained clinical settings

Take a problem-driven approach

AI system design frequently prioritises the availability of annotated datasets and model performance metrics—priorities that are often misaligned with the operational realities of resource-constrained settings. The approach to AI use in resource-constrained settings should be problem-driven, guided by care delivery needs. These settings typically face three interrelated challenges: inequitable access to care due to workforce shortages, inadequate health infrastructure, and the dual burden of infectious and chronic diseases. Together, these underscore the urgent need for stronger preventive approaches. This contrasts with high-income nations, where AI adoption has historically been top-down and technology-driven, supported by well-coordinated health systems, state-of-the-art information infrastructure, and interoperability standards.¹⁷

The challenges in LMICs are complex and interconnected, requiring multi-pronged solutions. Yet, AI may often be perceived as a silver bullet. A problem-driven approach helps refocus attention on specific clinical needs, treating AI as one of many possible interventions and enabling comparison with alternative approaches. When AI is identified as a viable option, model design and selection should be guided not by technology-driven performance but by alignment with the socio-technical context and the task at hand.

The problem-driven approach to public health challenges is evident in recent studies, particularly in contexts where decision-making is suboptimal. For instance, AI screening tools for infectious diseases (e.g. tuberculosis and HIV) in Malawi and chronic diseases (e.g. diabetic retinopathy) in Bangladesh.^18,19 Both AI tools were evaluated in primary care to improve the productivity of preventive care by 40%, address limited access to specialist doctors,¹⁸ enhance the timeliness of interventions, and increase cost efficiency.¹⁹

Examine the socio-technical context for introducing AI

Improving the design and implementation of new technologies, including AI, requires integrating socio-technical considerations to ensure social and technical elements are effectively aligned. As defined, “the socio-technical view attempts to understand the contribution of phenomena at the human social level to the performance of technical systems.”²⁰ In healthcare, this means AI must be designed and evaluated within the complex socio-technical context in which it operates, particularly in relation to the clinicians who use it. As illustrated in Figure 1, informatics interventions are embedded across multiple, interacting system layers,²¹ with clinicians occupying a central role within the socio-technical system.

Figure 1.

Four layers of informatics systems nest within each other (algorithm> computer programs> human-computer interaction> sociotechnical systems), adapted from Coiera (2015). AI/ML algorithms are embedded in software tools. Doctors interact with AI tools within resource-constrained settings where access to specialist and medical equipment is limited.

“The technical systems have social consequences, and social systems have technical consequences.”²² While AI offers significant potential to enhance healthcare quality and safety, its integration into clinical workflows can introduce risks that may compromise patient safety.²³ The inseparability of social and technical elements makes socio-technical analysis essential in AI design and implementation. This includes careful consideration of the socio-technical context, particularly in settings with constrained resources and competing demands.²⁰ By explicitly incorporating socio-technical considerations, AI can evolve from a purely automated tool into a socio-technical system that mitigates patient safety risks and supports responses to emerging threats.

One way to understand the socio-technical context of AI tools and their impact within a specific health system is through systems engineering. Frameworks such as SEIPS (Systems Engineering Initiative for Patient Safety) provide structured methods for examining this context.²⁴ SEIPS has been widely applied in health-system research and is closely tied to patient safety, a core dimension of modern health systems. For example, a qualitative study to examine clinicians use of AI tools can apply SEIPS to describe the work system: people, environment, tasks, and technology; within specific processes and outcomes, such as AI-based cardiovascular disease prediction in resource-constrained clinical settings.²⁵

Automate the right clinical task

While AI aims to extend clinician capabilities beyond what is possible unaided,²⁶ a critical question remains: which tasks should be automated?²⁷ AI must target the right tasks for a given clinical problem and socio-technical context. Assessing the work system, processes, and intended outcomes determines a task’s suitability for AI automation. For instance, in LMICs, available inputs may be limited to clinical data and basic laboratory tests. A user-centred approach is essential to validate appropriate clinical tasks,²⁸ ensuring alignment with user needs and minimizing workflow disruption. At the same time, a meaningful workflow adjustment is usually required for AI tools to function reliably, for instance, introducing a more structured or systematic data entry. As such, there is a mutual adjustment where both AI and the traditional workflow evolve to achieve safe and effective socio-technical integration.

In clinical practice, the most common clinical tasks of AI include imaging interpretation and predictive risk assessment, particularly for early sepsis detection.^1,29 Among these, AI has been reported to achieve its greatest success in clinical documentation and risk assessment.²⁹ In their real-world evaluation, a limited number of AI tools have shown potential to automate imaging interpretation for screening in remote or underserved areas, including applications in hepatobiliary disease,³⁰ childhood cataracts,³¹ and skin lesion.³²

Ensure AI tools are easily accessible at the point of care

AI tools should be delivered via mobile devices as they are widely available in resource-constrained settings and operable by users with basic digital literacy.^25,33 In LMICs, clinical decision support tools are most commonly accessed via personal mobile devices.²⁵ Smartphones are widely used for clinician communication and consultation, with data entry frequently performed by patients or non-physician health workers prior to physician review.²⁵ A recent review reported that communication (90%) and clinical decision-making (70%) were the predominant uses of smartphones by clinicians, including in LMIC contexts.³⁴ Increasing mobile processing power now enables on-device AI, and mobile-based tools can be designed to function offline, reducing reliance on internet connectivity.

Ensure use of AI can improve decisions

Given there is limited access to confirmatory examinations, clinical decision-making is largely subjective in resource-constrained settings, with clinicians relying on their own experience or peer consultation.²⁵ The implementation of AI tools in these contexts should evaluate their impact on improving clinical decision-making. As such an AI-assisted clinician must perform better than the same clinician unaided. Given that most contemporary AI tools are assistive, leaving the final decision to the clinician, the interaction between clinician and AI is fundamental to improved decision-making.^1,3,35 Consequently, comparing AI model performance against clinicians alone is an inadequate metric, as it fails to capture outcomes when both work together. Evaluating clinical decision-making is a necessary precursor to demonstrating positive impacts on care delivery and patient outcomes.³⁶

Evidence of enhanced clinical decision-making can first be evaluated in controlled settings and subsequently in real-world environments after deployment. To evaluate improvements in decision-making, randomised controlled trials (RCTs) should be used, comparing the accuracy of AI-assisted versus unassisted clinicians.²⁷ However, few RCTs focus on decision-making itself, with most measuring downstream effects on care delivery and patient outcomes.² For instance, a real-world RCT assessing AI support during emergency calls demonstrated no significant difference in operators’ ability to recognise cardiac arrest.³⁷ On the other hand, clinical vignette studies offer a way to evaluate decision-making in controlled settings and to show the safety and efficacy of AI-assistance prior to costly and potentially disruptive trials.³⁸ As risk- and patient-free designs, they mitigate potential harm from early AI experimentation.³⁹

Consider time costs: Expediting diagnosis and treatment

In resource-constrained LMICs, clinicians face high patient volumes and limited time per case, making time arguably the scarcest resource.²⁵ This challenge is compounded by manual, repetitive tasks arising from the lack of electronic medical records. Contemporary assistive AI tools require clinicians to review, verify, and approve or reject AI-generated recommendations.^1,3 To be beneficial, AI must not add to clinician workload or delay diagnosis and treatment. Instead, AI-assisted decision support should provide timely recommendations that expedite rather than hinder time-sensitive decisions.⁴⁰

Accordingly the time costs associated with the use of AI tools for clinical decision-making should be evaluated.⁴¹ For example, AI-assisted ECG triage that operate in parallel with cardiologist confirmation, could substantially reduce door-to-balloon time for patients with myocardial infarction.⁴² A previous RCT of childhood cataract screening reported faster diagnosis with AI support compared to clinicians alone (2.79 min vs 8.52 min, p<0.001).³¹ Time-sensitive AI has the potential to alleviate, rather than worsen, delays in diagnosis and treatment common in resource-constrained settings. As noted in the “ten commandments” for CDS, “speed is everything.”⁴³ Conversely, AI that slows clinical processes risks becoming a form of clumsy automation, leading to disuse.⁴⁴

Ensure AI recommendations are clinically comprehensible through explanation

AI tools should provide explanations to help clinicians contextualise their output and support decision-making. Clinicians typically integrate multiple sources of information through clinical reasoning while attempting to interpret the “black-box” outputs of AI.⁴⁵ This preserves their decisional discretion when considering AI recommendations. In high-resource settings, clinicians can contextualise AI outputs with additional diagnostic tests or subspecialist opinions. By contrast, in LMICs, objective data for confirmatory diagnosis are often limited. This makes explainability critical, AI outputs must be interpretable to support safe clinical decision-making.⁴⁶ Studies have shown higher clinician acceptance and more positive perceptions when AI results are clinically comprehensible.⁴⁷ For example, studies of AI tools for pulmonary embolism risk assessment have shown that model-agnostic local explanations highlighting individual risk factors help clinicians interpret and trust AI outputs, while also enabling them to explain results to patients.⁴⁸ AI should support clinicians in providing meaningful explanations, thereby facilitating shared decision-making.⁴⁹

Ensure AI recommendations are actionable to impact care delivery

This principle highlights the importance of designing AI solutions that both meet genuine needs and generate outputs that are actionable within the contextual constraints of their intended settings. For AI tools to be valuable, clinician decisions informed by AI must translate into concrete actions.⁵⁰ Changes in patient management can then serve as indicators of actionable care delivery.¹ For example, following AI-based risk assessment for atherosclerotic cardiovascular disease, actionable care may involve prescribing primary preventive therapies such as aspirin or statins. The actionability of AI-assisted decision-making can be further evaluated through net benefit analysis.^51,52

Conversely, predictive AI-based risk assessments that lack actionable outputs provide little value. For instance, an AI tool distinguishing ischemic from haemorrhagic stroke showed strong performance in prioritising air ambulance transport, thereby avoiding treatment delays and unnecessary deployments.⁵³ Yet, in regions without emergency helicopter services, such a tool would have minimal impact on care delivery or outcomes. AI implementation in resource-constrained LMICs must be responsible and sustainable, ensuring it does not exacerbate existing burdens or introduce new ethical challenges.⁵⁴ Treatment options in these contexts are often limited by shortages of essential medicines or specialised equipment. Accordingly, AI output must be tailored to what is realistically achievable in the specific setting. At the deployment stage, AI tools should be governed by clear clinical consensus and locally grounded policies, supported by ethics evaluation to ensure transparency and accountability.⁵⁵ Such governance enhances actionability while reducing the risk of unintended or adverse effects on care delivery.

Case study: Cardiovascular disease risk assessment

Using the eight socio-technical principles provides a pragmatic, structured approach to design AI tools for resource-constrained settings. To synthesise these principles, we present an example case study in risk assessment for Atherosclerotic Cardiovascular Disease (ASCVD) which is a priority in global health (Table 1). The case study focuses on resource-constrained settings in Indonesia as the socio-technical context where primary care doctors play an important role in preventing ASCVD by assessing and managing cardiovascular risk factors. It was drawn from previous works including a literature review,¹ and semi-structured interviews which examined the socio-technical context of cardiovascular risk assessment with limited access to medical specialists and advanced equipment (principles 1-4).²⁵ These requirements were then used to construct a conceptual prototype of an AI tool that was demonstrated to improve risk assessment and statin prescriptions, and reduced decision-making time (principles 5-8).⁵⁶ We show that ASCVD risk assessment is a promising application for AI in resource constrained settings, and merits further evaluation to assess real-world impact on clinical decision-making and patient management.

Table 1.

Applying the eight principles to design and test an AI tool for ASCVD risk assessment in Indonesia.

Principles for designing AI for resource-constrained clinical settings		Case study in ASCVD risk assessment^1,25,56
1. Problem-driven	Take a problem-driven approach	Problem: ASCVD is the most common cardiovascular problem placing significant clinical, social, and financial burdens on the Indonesian health system. Potential users i.e. doctors working in resource-constrained settings were interviewed to discuss and confirm ASCVD risk prediction as a compelling use case for an AI tool to improve primary prevention.²⁵
2. Socio-technical context	Examine the socio-technical context for introducing AI	The clinical work system in resource-constrained settings was examined using SEIPS, a theoretical framework to describe complex sociotechnical systems including people, processes, tasks, and technology.²⁴ Doctors faced substantial patient volumes. Therefore, time was a significant constraint. They relied predominantly on paper-based medical records, often made clinical decisions based on subjective judgement due to limited access to confirmatory diagnostic investigations, such as specialised laboratory tests or imaging.²⁵
3. Right task	Automate the right clinical task	Risk assessment represents a key application for AI in healthcare (28%).¹ Doctors reported needing support to prioritise high-risk patients and deliver patient education tailored to risk level.²⁵
4. Easy access	Ensure AI tools are easily accessible at the point of care	A prototype AI tool was developed and tested.⁵⁶ It was designed to be accessible on a mobile device which is a likely real-world implementation given the widespread availability and use of mobile phones among doctors, including for communication and consultation with colleagues regarding clinical cases.²⁵
5. Improve decisions	Ensure use of AI can improve decisions	A randomised controlled clinical vignette study was undertaken with 102 doctors to compare the prototype tool with the standard care for risk assessment without decision support, please see detailed methods in the reference.⁵⁶ Use of the AI tool was shown to improve ASCVD risk assessment by 27% (p<.001), equivalent to one additional correct risk classification for every four patients assessed with AI support.
6. Time costs	Consider time costs: expediting diagnosis and treatment	Use of the prototype AI tool reduced the time required for ASCVD risk assessment by doctors by 12% compared with standard care.⁵⁶ Decision-making assisted by AI remained faster than heuristic risk assessment without decision support.
7. Explanation	Ensure AI recommendations are clinically comprehensible through explanation	Doctors emphasised the need for AI tools to provide explanations of input factors and risk scores, rather than only risk level recommendations.⁵⁶
8. Actionable	Ensure AI recommendations are actionable to impact care delivery	AI recommendations were considered actionable given the availability of different options for prescribing statins. Prescription of statins (p<.001) and referrals was also improved (p=.029).⁵⁶

Note. AI, artificial intelligence; ASCVD, atherosclerotic cardiovascular disease.

The case study was confined to ASCVD risk prediction in Indonesia. Further research applying diverse AI tools across LMICs with varying clinical governance structures is needed to assess the generalisability of these principles. Alternatively, these principles could inform a desktop simulation to evaluate whether a proposed AI tool would yield material benefit in resource-constrained settings, with Principles 1 and 8 serving as the primary evaluative criteria.

Conclusion

Resource-constrained settings face profound challenges in delivering equitable, high-quality healthcare. AI has the potential to alleviate some of these pressures, but only if designed with contextual sensitivity. This perspective highlights key principles for responsible AI in LMICs: grounding development in problem-driven approaches, embedding socio-technical understanding, prioritising appropriate clinical tasks, and ensuring tools are accessible, comprehensible, and actionable at the point of care. By aligning AI design with local realities, clinicians’ decision-making can be strengthened and health equity meaningfully advanced across LMICs.

Footnotes

ORCID iDs

Anindya Pradipta Susanto

David Lyell

Bambang Widyantoro

Shlomo Berkovsky

Farah Magrabi

Author contributions

A.P.S. and F.M. were responsible for conceptualisation. A.P.S., D.L., B.W., S.B., and F.M. drafted, edited and approved the final manuscript before submission.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This study was supported by the International Macquarie University Research Excellence Scholarship awarded to A.P.S. (iMQRES 20201869) and the NHMRC Centre for Research Excellence (CRE) in Digital Health (APP1134919). The funders played no role in the writing of the article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship and/or publication of this article.

Data Availability Statement

No datasets were generated or analysed during the current study.*

Appendix

References

Susanto

Lyell

Widyantoro

, et al. Effects of machine learning-based clinical decision support systems on decision-making, care delivery, and patient outcomes: a scoping review. J Am Med Inform Assoc 2023; 30(12): 2050–2063. https://doi.org/10.1093/jamia/ocad180

Han

Acosta

Shakeri

, et al. Randomised controlled trials evaluating artificial intelligence in clinical practice: a scoping review. Lancet Digit Health 2024; 6(5): e367–e373. https://doi.org/10.1016/S2589-7500(24)00047-5

Magrabi

Lyell

Coiera

. Automation in Contemporary Clinical Information Systems: a Survey of AI in Healthcare Settings. Yearb Med Inform 2023; 32(1): 115–126. https://doi.org/10.1055/s-0043-1768733

Kim

Fischetti

Guy

, et al. Artificial Intelligence (AI) Applications for Point of Care Ultrasound (POCUS) in Low-Resource Settings: A Scoping Review. Diagnostics (Basel) 2024; 14(15): 1669. https://doi.org/10.3390/diagnostics14151669

van Zyl

Badenhorst

Hanekom

, et al. Unravelling ‘low-resource settings’: a systematic scoping review with qualitative content analysis. BMJ Glob Health 2021; 6(6): e005190. https://doi.org/10.1136/bmjgh-2021-005190

Meliala

Hort

Trisnantoro

. Addressing the unequal geographic distribution of specialist doctors in indonesia: the role of the private sector and effectiveness of current regulations. Soc Sci Med 2013; 82: 30–34. https://doi.org/10.1016/j.socscimed.2013.01.029

Liaw

S-T

Godinho

. Digital health and capability maturity models—a critical thematic review and conceptual synthesis of the literature. Journal of the American Medical Informatics Association 2023; 30(2): 393–406. https://doi.org/10.1093/jamia/ocac228

Wahl

Cossy-Gantner

Germann

, et al.

Artificial intelligence (AI) and global health: how can AI contribute to health in resource-poor settings?

BMJ Glob Health 2018; 3(4): e000798. https://doi.org/10.1136/bmjgh-2018-000798

Ciecierski-Holmes

Singh

Axt

, et al. Artificial intelligence for strengthening healthcare systems in low- and middle-income countries: a systematic scoping review. NPJ Digit Med 2022; 5(1): 162. https://doi.org/10.1038/s41746-022-00700-y

10.

Thomas

Wakerman

Humphreys

. Ensuring equity of access to primary health care in rural and remote Australia - what core services should be locally available? Int J Equity Health 2015; 14: 111. https://doi.org/10.1186/s12939-015-0228-1

11.

Cordes

Bak

Lyndon

, et al. Competing interests: digital health and indigenous data sovereignty. NPJ Digit Med 2024; 7(1): 178. https://doi.org/10.1038/s41746-024-01171-z

12.

Nong

Maurer

Dwivedi

. The urgency of centering safety-net organizations in AI governance. NPJ Digit Med 2025; 8(1): 117. https://doi.org/10.1038/s41746-025-01479-4

13.

Fraser

HSF

Zahiri

Kim

, et al. The Global Health Informatics landscape and JAMIA. J Am Med Inform Assoc 2023; 30(4): 775–780. https://doi.org/10.1093/jamia/ocad024

14.

Wekenborg

Gilbert

Kather

. Examining human-AI interaction in real-world healthcare beyond the laboratory. NPJ Digit Med 2025; 8(1): 169. https://doi.org/10.1038/s41746-025-01559-5

15.

You

Hernandez-Boussard

Pfeffer

, et al. Clinical trials informed framework for real world clinical implementation and deployment of artificial intelligence applications. NPJ Digit Med 2025; 8(1): 107. https://doi.org/10.1038/s41746-025-01506-4

16.

Poddar

Marwaha

Yuan

, et al. An operational guide to translational clinical machine learning in academic medical centers. NPJ Digit Med 2024; 7(1): 129. https://doi.org/10.1038/s41746-024-01094-9

17.

Magrabi

Cresswell

Fraser

. International dimensions of clinical decision support systems. In: Greenes

(ed) Clinical Decision Support and Beyond: Progress and Opportunities in Knowledge-Enhanced Health and Healthcare. Elsevier Science & technology, 2023, p. 145.

18.

Abramoff

Whitestone

Patnaik

, et al. Autonomous artificial intelligence increases real-world specialist clinic productivity in a cluster-randomized trial. NPJ Digit Med 2023; 6(1): 184. https://doi.org/10.1038/s41746-023-00931-7

19.

MacPherson

Webb

Kamchedzera

, et al. Computer-aided X-ray screening for tuberculosis and HIV testing among adults with cough in Malawi (the PROSPECT study): A randomised trial and cost-effectiveness analysis. PLoS Med 2021; 18(9): e1003752. https://doi.org/10.1371/journal.pmed.1003752

20.

Coiera

. Putting the technical back into socio-technical systems research. Int J Med Inform 2007; 76(Suppl 1): S98–S103. https://doi.org/10.1016/j.ijmedinf.2006.05.026

21.

Coiera

. Guide to Health Informatics. 3rd ed. Taylor & Francis Group, 2015.

22.

Coiera

. Four rules for the reinvention of health care. BMJ 2004; 328: 1197–1199. https://doi.org/10.1136/bmj.328.7449.1197

23.

Lyell

Wang

Coiera

, et al. More than algorithms: an analysis of safety events involving ML-enabled medical devices reported to the FDA. J Am Med Inform Assoc 2023; 30(7): 1227–1236. https://doi.org/10.1093/jamia/ocad065

24.

Holden

Carayon

. SEIPS 101 and seven simple SEIPS tools. BMJ Qual Saf 2021; 30: 901–910. https://doi.org/10.1136/bmjqs-2020-012538

25.

Susanto

Lyell

Akbar

, et al. Applying AI to support cardiovascular risk prediction in resource-constrained settings: Qualitative insights from Indonesian doctors. Applied Artificial Intelligence (under review).

26.

Friedman

. A “fundamental theorem” of biomedical informatics. J Am Med Inform Assoc 2009; 16(2): 169–170. https://doi.org/10.1197/jamia.M3092

27.

Coiera

. A New Informatics Geography. Yearb Med Inform 2016; 25(1): 251–255. https://doi.org/10.15265/IY-2016-018

28.

Seneviratne

Schreier

, et al. User-centred design for machine learning in health care: a case study from care management. BMJ Health Care Inform 2022; 29(1): e100656. https://doi.org/10.1136/bmjhci-2022-100656

29.

Poon

Lemak

Rojas

, et al. Adoption of artificial intelligence in healthcare: survey of health system priorities, successes, and challenges. Journal of the American Medical Informatics Association 2025; 32(7): 1093–1100. https://doi.org/10.1093/jamia/ocaf065

30.

Xiao

Huang

Wang

, et al. Screening and identifying hepatobiliary diseases through deep learning using ocular images: a prospective, multicentre study. The Lancet Digital Health 2021; 3(2): e88–e97. https://doi.org/10.1016/S2589-7500(20)30288-0

31.

Lin

Liu

, et al. Diagnostic Efficacy and Therapeutic Decision-making Capacity of an Artificial Intelligence Platform for Childhood Cataracts in Eye Clinics: A Multicentre Randomized Controlled Trial. EClinicalMedicine 2019; 9: 52–59. https://doi.org/10.1016/j.eclinm.2019.03.001

32.

Kim

Han

Yang

, et al. Prospective, comparative evaluation of a deep neural network and dermoscopy in the diagnosis of onychomycosis. PLoS One 2020; 15(6): e0234334. https://doi.org/10.1371/journal.pone.0234334

33.

Erku

Khatri

Endalamaw

, et al. Digital Health Interventions to Improve Access to and Quality of Primary Health Care Services: A Scoping Review. Int J Environ Res Public Health 2023; 20(19): 6854. https://doi.org/10.3390/ijerph20196854

34.

Lee

Bin Mahmood

ABS

Lee

, et al. Smartphone and Mobile App Use Among Physicians in Clinical Practice: Scoping Review. JMIR Mhealth Uhealth 2023; 11: e44765. https://doi.org/10.2196/44765

35.

Lyell

Coiera

Chen

, et al. How machine learning is embedded to support clinician decision making: an analysis of FDA-approved medical devices. BMJ Health Care Inform 2021; 28(1): e100301. https://doi.org/10.1136/bmjhci-2020-100301

36.

Coiera

. Assessing Technology Success and Failure Using Information Value Chain Theory. Stud Health Technol Inform 2019; 263: 35–48. https://doi.org/10.3233/SHTI190109

37.

Blomberg

Christensen

Lippert

, et al. Effect of Machine Learning on Dispatcher Recognition of Out-of-Hospital Cardiac Arrest During Calls to Emergency Medical Services: A Randomized Clinical Trial. JAMA Netw Open 2021; 4(1): e2032320. https://doi.org/10.1001/jamanetworkopen.2020.32320

38.

Jabbour

Fouhey

Shepard

, et al. Measuring the Impact of AI in the Diagnosis of Hospitalized Patients: A Randomized Clinical Vignette Survey Study. JAMA 2023; 330(23): 2275–2284. https://doi.org/10.1001/jama.2023.22295

39.

Comeau

Bitterman

Celi

. Preventing unrestricted and unmonitored AI experimentation in healthcare through transparency and accountability. NPJ Digit Med 2025; 8(1): 42. https://doi.org/10.1038/s41746-025-01443-2

40.

Carayon

Hoonakker

Hundt

, et al. Application of human factors to improve usability of clinical decision support for diagnostic decision-making: a scenario-based simulation study. BMJ Qual Saf 2020; 29(4): 329–340. https://doi.org/10.1136/bmjqs-2019-009857

41.

Salwei

Carayon

Hoonakker

PLT

, et al. Workflow integration analysis of a human factors-based clinical decision support in the emergency department. Appl Ergon 2021; 97: 103498. https://doi.org/10.1016/j.apergo.2021.103498

42.

Lin

Liu

W-T

Chang

C-H

, et al. Artificial Intelligence–Powered Rapid Identification of ST-Elevation Myocardial Infarction via Electrocardiogram (ARISE) — A Pragmatic Randomized Controlled Trial. NEJM AI 2024; 1(7): AIoa2400190. https://doi.org/10.1056/aioa2400190

43.

Bates

Kuperman

Wang

, et al. Ten commandments for effective clinical decision support: making the practice of evidence-based medicine a reality. J Am Med Inform Assoc 2003; 10(6): 523–530. https://doi.org/10.1197/jamia.M1370

44.

Parasuraman

Riley

. Humans and automation: Use, misuse, disuse, abuse. Human Factors 1997; 39: 230–253. https://doi.org/10.1518/001872097778543886

45.

van Baalen

Boon

Verhoef

. From clinical decision support to clinical reasoning support systems. J Eval Clin Pract 2021; 27(3): 520–528. https://doi.org/10.1111/jep.13541

46.

Arbelaez

Starke

Lorenzini

, et al. Re-focusing explainability in medicine. Digit Health 2022; 8: 20552076221074488. https://doi.org/10.1177/20552076221074488

47.

Schmidt

Biessmann

Teubner

. Transparency and trust in artificial intelligence systems. Journal of Decision Systems 2020; 29(4): 260–278. https://doi.org/10.1080/12460125.2020.1819094

48.

Diprose

Buist

Hua

, et al. Physician understanding, explainability, and trust in a hypothetical machine learning risk calculator. J Am Med Inform Assoc 2020; 27(4): 592–600. https://doi.org/10.1093/jamia/ocz229

49.

Badal

Lee

Esserman

. Guiding principles for the responsible development of artificial intelligence tools for healthcare. Commun Med (Lond) 2023; 3(1): 47. https://doi.org/10.1038/s43856-023-00279-9

50.

Kawamoto

Houlihan

Balas

, et al. Improving clinical practice using clinical decision support systems: a systematic review of trials to identify features critical to success. BMJ 2005; 330(7494): 765. https://doi.org/10.1136/bmj.38398.500764.8F

51.

Ehrmann

Joshi

Goodfellow

, et al. Making machine learning matter to clinicians: model actionability in medical decision-making. NPJ Digit Med 2023; 6(1): 7. https://doi.org/10.1038/s41746-023-00753-7

52.

Liu

Bates

Wiens

, et al. The number needed to benefit: estimating the value of predictive analytics in healthcare. J Am Med Inform Assoc 2019; 26(12): 1655–1659. https://doi.org/10.1093/jamia/ocz088

53.

Uchida

Kouno

Yoshimura

, et al. Development of Machine Learning Models to Predict Probabilities and Types of Stroke at Prehospital Stage: the Japan Urgent Stroke Triage Score Using Machine Learning (JUST-ML). Transl Stroke Res 2022; 13(3): 370–381. https://doi.org/10.1007/s12975-021-00937-x

54.

Alami

Rivard

Lehoux

, et al. Artificial intelligence in health care: laying the Foundation for Responsible, sustainable, and inclusive innovation in low- and middle-income countries. Global Health 2020; 16(1): 52. https://doi.org/10.1186/s12992-020-00584-1

55.

Wang

Freeman

Magrabi

. Governance for safe and responsible AI in healthcare organisations: A scoping review of frameworks. PREPRINT (Version 1) available at Research Square 2025.

56.

Susanto

Lyell

Widyantoro

, et al. AI-Assisted Cardiovascular Risk Assessment by General Practitioners in Resource-Constrained Indonesian Settings Using a Conceptual Prototype: Randomized Controlled Study. J Med Internet Res 2025; 27: e73131. https://doi.org/10.2196/73131