. 2024 Mar 22;71:102555. doi: 10.1016/j.eclinm.2024.102555

Table 1.

Characteristics of included studies.

Authors	Study design and methodology/data analysis	Method of data collection	Study population and sampling strategy	Number of participants	Participant characteristics	Topic of interest	AI studied	Main findings
Abuzaid et al. (2021)²⁹	Mixed methods, thematic analysis	Focus groups	Magnetic Resonance Imaging (MRI) technologists, snowball sampling	98	NR	Radiology	Hypothetical use of AI in radiology	MRI technologists have an understanding of AI and believe it would improve MRI protocols, reduce scan time, enhance post-processing and play a role in image interpretation. They highlighted a need for bespoke for AI training and practical sessions.
Adams et al. (2020)³⁰	Qualitative, thematic analysis	Focus groups	Patient and family advisors from urban and rural Canada, NR	17	11 (55%) female	Radiology	Hypothetical AI used in radiology	Patient perceptions were described in four themes: 1. Fear of the unknown, 2. Trust, 3. Human connection, 4. Cultural acceptability. Patient priorities for AI were described in five themes: 1. Improving access to imaging and reducing wait times, 2. Reducing time to diagnosis, 3. Increasing diagnostic accuracy, 4. Improving communication 5. Empowering patients.
Bergquist et al. (2023)³¹	Qualitative, NR	Semi-structured interviews	Mixed group of clinicians, researchers, and healthcare leaders in Sweden, purposive sampling	25	5 (20%) female; median age 53 years; 8 radiologists, 6 other medical professionals, 7 management, 4 engineers/developers	Radiology	Hypothetical AI used in radiology	Trustworthiness of AI is related to: 1. Reliability, 2. Transparency, 3. Quality verification and 4. Inter-organizational compatibility
Buck et al. (2022)³²	Qualitative, grounded theory	Semi-structured interviews	General practitioners in Germany, convenience sampling	18	9 (50%) female, 9 (50%) male	General	Hypothetical AI used in medical diagnosis	Three determinants of GPs' attitudes towards AI: concerns, expectations, and minimum requirements of AI-enabled systems. Individual characteristics and environmental influences as the 2 conditional determinants of GPs' attitudes toward AI-enabled systems.
Carter et al. (2023)³³	Mixed-methods, NR	Focus groups	Female members of the public aged 50–74 years old, Australia	50	Age: 20 (40%) aged 50–59, 21 (42%) aged 60–69, 9 (18%) aged 70–74 Personal history of breast cancer: 12 (24%) yes Education: 26 (52%) university educated	Radiology	Hypothetical AI used in radiology (breast cancer screening)	There is broad acceptance of the use of AI in breast screening, conditioned on human involvement and AI performance
Cartolovni, Malesevic and Poslon (2023)³⁴	Qualitative, thematic analysis	Semi-structured interviews	Patients, clinicians, and healthcare leaders working in Croatia	38	Patients 11, clinicians 12, leaders 11; age range 18–65 years old	General	Hypothetical clinical AI	Four themes were developed: 1. The current state of healthcare and the patient-physician relationship, 2. Expectation of AI, 3. A synergistic effect between physicians and AI, 4. The future of healthcare and the patient-physician relationship
Chen et al. (2021)³⁵	Qualitative, thematic analysis	Semi-structured interviews and focus groups	Radiologists and radiographers working in five NHS organisations in England, snowball sampling/convenience sampling	Interviews: 18 FG: 8	Interviews: 12 (60%) radiologists, 8 (40%) radiographers FG: 8 (100%) radiographers	Radiology	Hypothetical AI used in radiology	Considered responses to the use of AI in radiology in 1. Knowledge and 2. Attitudes, finding differences in attitudes towards AI between professional groups.
Drogt et al. (2022)³⁶	Qualitative, NR	Semi-structured interviews	Mixed group of professionals working in the pathology department of two hospitals in the Netherlands, convenience sampling	24	15 (63%) pathologists, 17 (29%) laboratory technicians, 2 (8%) computer scientists	Pathology	Hypothetical AI used in pathology	Three recommendations for embedding AI in pathology: 1. Foster a pragmatic attitude toward AI development, 2. Provide task-sensitive information and training to health care professionals working at pathology departments, 3. Take time to reflect upon users' changing roles and responsibilities.
Faric et al. (2023)³⁷	Qualitative, thematic analysis	Semi-structured interviews	Mixed group of clinicians, healthcare leaders, and patients across 5 hospitals in the UK, Belgium, and the Netherlands	39	12 patients: 10 (83%) female; 3 (25%) aged 40–50 years old, 3 (25%) aged 50–60 years old, 3 (25%) aged 60–70 years old, 3 (25%) aged 70–80 years old 25 clinicians/healthcare leaders: 6 (24%) female 2 healthcare leaders: 2 male	Radiology	AI in use for diagnosing lung nodules	Four main themes were developed: 1. Perceived drivers and benefits, 2. Design of the tool and integration, 3. Appropriation of the tool by expert labour, 4. Clinical governance, quality assurance, maintenance, and post-market surveillance
Fazakarley et al. (2023)³⁸	Qualitative, thematic analysis	Semi-structured interviews	Clinicians and AI researchers in the UK	13	5 (38%) female; mean age 38 years old, SD 9.1 years; 9 (69%) White British; 2 (15%) mixed or multiple ethnicities; 1 (8%) Asian; 1 (8%) other 3 (23%) doctors, 4 (31%) nurses, 2 (15%) IT technicians, 4 (31%) AI developer/researcher	Cardiology	AI in use within a randomised-control trial, to diagnose coronary artery disease	Four themes were identified: 1. Positive perceptions of AI, 2. Potential barriers to using AI, 3. Concerns regarding AI use, 4. Steps needed to ensure the acceptability of future AI tools
Gillner (2024)³⁹	Qualitative, thematic analysis	Semi-structured interviews, 1 focus group	A mixed group of AI providers across Europe and clinicians	Interviews: 17 FG: 5	Interviews: 17 AI researchers/leaders FG: 5 clinicians	General	Hypothetical clinical AI	Two major themes were developed: 1. Subsystems of complex healthcare systems, 2. Emergent practices of AI providers in healthcare
Haan et al. (2019)⁴⁰	Qualitative, grounded theory	Semi-structured interviews	Patients attending the radiology department of a tertiary care, academic hospital in the Netherlands for CT chest and abdomen, purposive sampling	20	9 (45%) female; age 39–79 years old (mean age 64)	Radiology	Hypothetical AI used in radiology	Six key domains related to AI use in radiology: 1. Proof of technology, 2. Procedural knowledge, 3. Competence, 4., Efficiency, 5. Personal interaction, 6. Accountability
Hallowell et al. (2022)⁴¹	Qualitative, NR	Semi-structured interviews	Membership of the “Minerva Consortium” and personal contacts of the study authors, convenience/snowballing sampling	20	Expertise: 9 (45%) clinical genetics, 2 (10%) paediatric genetics, 5 (25%) bioinformatics, 2 (10%) commercial, 3 (15%) other	Rare disease	Hypothetical AI used in diagnosing facial dysmorphology	Discussion of the value of trust in using AI for dysmorphology, concluding that trust in AI is grounded in its reliability, competence and “intentions.”
Held et al. (2022)⁴²	Qualitative, thematic content analysis	Semi-structured interviews	Mixed group of clinicians in Germany, convenience sampling	24	10 (42%) female; average year of birth 1971; 16 (67%) general practitioner, 3 (13) medical assistant, 5 (21%) ophthalmologists	Ophthalmology	Hypothetical AI used to diagnose diabetic retinopathy	Main determinants of implementation have been identified: personal attitude, organisation, time, financial factors, education, support, technical requirement, influence on profession and patient welfare.
Helenason et al. (2024)⁴³	Mixed-methods, NR	Semi-structured Interviews	Primary care clinicians in Sweden	15	NR	Dermatology	AI used to diagnose skin lesions, proposed for use	Three major themes were identified: trust, usability and user experience and clinical context
Henry et al. (2022)⁴⁴	Qualitative, grounded theory	Semi-structured Interviews	Mixed group of clinicians in a 285 bed acute-care, US hospital, purposive sampling	20	13 physicians (4 emergency department, 4 critical care, 5 general ward) and 7 nurses (3 emergency department, 4 critical care)	Sepsis	AI used to diagnose sepsis, in use by the institution	Themes identified included: lack of understanding of the difference between ML-based and conventional CDSS; ML-based systems play a supporting role; an overall willingness to trust AI despite lack of full understanding. Barriers highlighted included over-reliance on AI leading to deskilling.
Joshi et al. (2022)⁴⁵	Qualitative, thematic content analysis	Semi-structured interviews	Hospital leaders in the USA	21	5 (24%) informatics leadership, 10 (48%) clinical leadership e.g., CMO, 6 (29%) other executive leadership, convenience sampling	Sepsis	AI for diagnosis of sepsis, in use by the institution	Identified several barriers and facilitators to implementation of sepsis-detection AI, identifies consideration of workflow integration, and clinician buy-in as two key approaches to overcome identified barriers.
Jussupow et al. (2021)⁴⁶	Qualitative, grounded theory	Semi-structured interviews and ethnography	Radiologists working in a hospital in Germany, with experience of using an AI system to diagnose stroke, snowball sampling	14	2 chief radiologists, 4 senior radiologists, 8 assistant radiologists; mean self-reported diagnostic confidence (1–10) ranging from 4.3–10.0	Radiology	AI for stroke diagnosis, in use at the institution	Described three patterns of AI use. “Sensedemanding”radiologists will evaluate AI results in both confirming and disconfirming AI, “Sensegiving” radiologists will reinforce use if AI confirms their findings. “Sensebreaking” radiologists find no benefit from AI.
Kim et al. (2024)⁴⁷	Qualitative, ethnography with abductive reasoning	Semi-structured interviews and ethnography	Mixed group of clinicians and healthcare leaders working at a hospital in the Netherlands	Ethnographic observation over 3 years; 18 interviews	NR	Radiology	15 individual AI pipelines in use, for cross-specialty diagnostic tasks	Three key themes were developed to inform AI implementation: 1. Technology level, 2. Workflow level, 3. People and organisational level
King et al. (2023)⁴⁸	Qualitative, framework approach	Semi-structured interviews	Pathologists employed in UK hospitals, purposive sampling	25	20 pathology consultants/attendings, 5 pathology trainees. 14 (70%) male, 11 (30%) female.	Pathology	Hypothetical AI used in pathology	Required features of AI identified by pathologists were trustworthiness and explainability, usability and workflow integration. Key contextual information and concerns about AI included the context of AI deployment, pathologists involvement with AI development, liability, evaluation and validation of AI and resources for AI.
Lebovitz et al. (2022)¹⁷	Qualitative, grounded theory	Semi-structured interviews and ethnography	Radiologists working in three departments utilising diagnostic AI, NR	33 + 500 h of ethnographic observation	NR	Radiology	AI in use for diagnosing breast cancer, classifying lung nodules and determining bone age	Only radiologists diagnosing lung cancer engaged with AI tools, despite high accuracy of all AI tools in the study. Explainability of AI is a necessary feature for clinician engagement, but on its own is permissive rather than sufficient.
Lombi and Rossero (2023)⁴⁹	Qualitative, template analysis	Semi-structured interviews	Radiologists working in a mixture of private and public hospitals in Italy, purposive sampling	12	1 (8%) female, age range 36–64 years, 5 (42%) employed by private hospitals	Radiology	Hypothetical AI used in radiology	Three themes were developd: 1. ‘It will take time’ 2. ‘This is what being a radiologist means’ 3. ‘Don't be a DIY diagnostician!’
Massey et al. (2023)⁵⁰	Mixed-methods, content analysis	Semi-structured interviews	Otolaryngologists, working in the USA, purposive sampling	19	11 (58%) general otolaryngologists, 8 (42%) subspecialty rhinologists; 11 (58%) practicing in an academic setting.	Radiology	Hypothetical AI used in radiology for sinus CT interpretation	Six themes were identified: 1. Conventional reporting was indispensable for extra-sinus analysis, 2. Relationship with radiologist dictates trust in reporting, 3. Clinicians were open to utilizing AI, 4. Standardization of reporting was valued, 5. Anatomical analysis was preferred over descriptive assessments, 6. Trust in AI could be improved with additional validation in the literature
Mosch et al. (2022)⁵¹	Qualitative, thematic analysis	Semi-structured interviews	Mixed group of participants with expertise in the field of AI in medicine, medical education, and training, purposive sampling	24	Professional background: 15 (63%) medical, nine (38%) computer science, 3 (13%) medical education, 8 (23%) other	General	Hypothetical clinical AI	Three themes were developed: 1. Specific tasks of physicians will be taken over by AI systems, 2. AI will not replace physicians, 3. Ways of work: AI will transform how healthcare is delivered.
Nelson et al. (2020)¹⁶	Qualitative, grounded theory	Semi-structured interviews	Patients attending general dermatology clinics and melanoma clinics at a hospital in the USA, purposive sampling	48	26 (54%) female; mean (SD) age 53.3 (21.7) years old; 16 (33%) history of melanoma, 16 (33%) history of non-melanomatous skin cancer, 16 (33%) no history of skin cancer; 45 (94%) White, 2 (4%) American Indian or Alaskan Native, 1 (2%) African American	Dermatology	Hypothetical AI used in dermatology for skin lesion classification	Patients describe a preference for AI as an assistive tool, rather than a replacement for a clinician. Increased diagnostic speed, accuracy and healthcare access were commonly perceived benefits of AI, but perceived risks included increased patient anxiety, AI errors and loss of human interaction.
Ng et al. (2022)⁵²	Qualitative, Phenomenology/thematic analysis	Focus groups	Radiographers working in public institutions in Singapore, purposive sampling	22	11 (50%) female; age 23–42 years old (median 30.5 years); working experience 1–18 years (median six years)	Radiology	Hypothetical AI used in radiography	Four themes were developed from the data: 1. Knowledge of AI and its applications, 2. Perceptions on the use of AI in radiographic practice, 3. Patients' perceptions as viewed by radiographers, 4. Prospective applications and expectations of AI.
Pelayo et al. (2023)⁵³	Qualitative, framework analysis	Semi-structured interviews	Latinx patients with T2DM at a single health center in the USA	20	12 (60%) female; mean age 59.8, range 14	Ophthalmology	Hypothetical AI used to diagnose diabetic retinal disease	Patients strongly prefer human review rather than AI; if AI is integrated it should be as a tool rather than a replacement
Prakash et al. (2021)⁵⁴	Qualitative, thematic analysis	Netnography, semi-structured interviews	Radiologists working in India, purposive sampling	15	5 (33%) female; mean age 40.7 years old, range 28–62	Radiology	Hypothetical AI used in radiology	Themes were developed from qualitative data: 1. Perceived threat, 2. Medico-legal risk, 3. Performance risk, 4. Performance expectancy, 5. Trust, 6. User resistance
Pumplun et al. (2021)²⁴	Qualitative, directed content analysis	Semi-structured interviews	Mixed group of AI experts, with detailed knowledge of clinical processes and AI, theoretical sampling approach	22	5 (23%) clinicians, 8 (36%) clinicians with leadership roles, 9 (25%) managers or IT staff; between 3 and 40 years of work experience	General	Hypothetical AI used in diagnosis	Developed a maturity model to score the readiness of a clinic for AI adoption, spanning three dimensions: organisation, adopter system and patient data.
Rabinovich et al. (2022)⁵⁵	Mixed-methods, NR	Structured interviews	Mixed group of clinicians in a hospital in Argentina, with experience of using diagnostic AI, NR	6	3 (50%) emergency physicians and 3 (50%) radiology residents	Radiology	AI in use in the institution, for diagnosing pneumothoraces, rib fractures, pleural effusions, and lung opacities on chest radiographs	Participants in general had positive experiences with using the diagnostic AI. They describe using it as a second opinion, to reduce human error, and valued its use in diagnostic confirmation.
Redrup Hill et al. (2023)⁵⁶	Qualitative, NR	Focus groups	Mixed group of patients/clinicians, researchers and healthcare leaders	31	4 software developers, 7 pathologists, 11 leaders, 9 patients/clinicians	Pathology	Existing AI to diagnose Barett's oesophagus or adenocarcinoma from pathology specimens	Six themes were developed: 1. Risks and potential harms, 2. Impacts on human experts, 3. Equity and bias, 4. Transparency and oversight, 5. Patient information and choice, 6. Accountability, moral responsibility and liability for error
Richardson et al. (2021)²⁶	Qualitative, grounded theory	Focus groups	Patients who had a recent primary care visit at a large academic health centre in the USA, convenience sampling	87	49% female; average age 53.5 years old; 93% white and 94% non-Hispanic/Latino; 87% education level higher than a high school degree; 20% employment history in technology or computer science; 45% employment history in healthcare/health science	General	Hypothetical clinical AI, using case studies to ground discussion	Description of six themes: excitement about healthcare AI but needing safety assurances, and expectation for clinicians to ensure AI safety, preservation of patient choice and autonomy, concerns about healthcare costs and insurance coverage, ensuring data integrity, and the risks of technology-dependent systems.
Richardson et al. (2022)²⁷	As above	As above	As above	As above	As above	As above	As above	Developed a conceptual framework for understanding how patients evaluate healthcare AI, based on patient experiences (with illness, health technology, relationship with clinicians, social context and familiarity with technology), beliefs (about healthcare and technology) and attitudes towards AI in healthcare (attitude formation, perceived acceptability and support for development).
Robertson et al. (2023)⁵⁷	Mixed-methods, NR	Semi-structured interviews	Patients recruited from cardiac clinics in Tucson, Arizona; convenience sampling	24	16 (67%) female; age range 19–92 years old; 10 (42%) White, 8 (33%) Hispanic, 3 (13%) Black, 2 (8%) Native American, 1 (4%) Asian, 7 (29%) University education	General	Hypothetical clinical AI	Narrative overview of qualitative data; patients discussed fallibility of AI systems, trust related to healthcare systems, knowledge of AI in use, confidence in human physicians and religious belief
Sangers et al. (2021)⁵⁸	Qualitative, thematic content analysis	Focus groups	Members of the public who took part in a customer panel of a Dutch health insurer, and social media platforms; purposive sampling	27	18 (67%) female; mean age 37.3 years, range 19–73; all use a smartphone at least every half day; 20 (74%) no history of skin cancer, 4 (15%) personal history of skin cancer, 3 (11%) family history of skin cancer	Dermatology	Hypothetical AI used in diagnosing skin cancer	Barriers to using AI apps for skin cancer diagnosis were: perceived lack of value, perception of untrustworthiness, preference for humans, concerns about privacy, complex user interface and increased costs. The facilitators were high perceived value, transparent and trustworthy identity of AI developers, endorsement by clinicians and regulatory bodies, easy to use interface and low costs.
Sangers et al. (2023)¹⁸	Qualitative, grounded theory	Focus groups	Dutch dermatologists and GPs identified through social media and via specialty newsletters, purposive sampling	33	Mean age 35.6 years, range 31–62; 17 (52%) female; 17 (52%) general practitioner, 16 (49%) dermatologist	Dermatology	Hypothetical AI used in diagnosing skin cancer	Dermatologists and GPs described preconditions for implementation: adequacy of algorithms, sufficient usability and accessibility, validation and regulation/clear liability, national guidance; they described benefits including improved health outcomes, care pathways and education. They described perceived barriers as doubts about AI accuracy, exacerbation of health inequalities, fear of replacement by AI, extra time taken to use AI and commercialization and privacy concerns.
Satterfield et al. (2019)⁵⁹	Qualitative, thematic analysis	Semi-structured interviews	3 groups of researchers: diagnosis, AI, “Learning Health Systems”, NR	32	18 (56%) from the “improving diagnosis” research group, 6 (19%) from AI research, 8 (25%) from the “Learning Health Systems” group	General	Hypothetical AI used in diagnosis	There is limited collaboration between the research communities, and the authors emphasise the importance of forming a multi-disciplinary “learning community” to ensure uptake of AI in diagnosis.
Scheetz et al. (2021)⁶⁰	Mixed-methods, thematic analysis	Semi-structured interviews	Mixed group of clinicians, including doctors and AHP, with experience of using an AI tool to screen for diabetic retinopathy, in outpatient clinicians in Australia, convenience sampling	8	3 (37.5%) male doctors, 5 (62.5%) female AHP	Ophthalmology	AI to screen for diabetic retinopathy	Participants agreed that the AI tool was easy to use and interpret, but reported challenges in explaining findings to patients, and allocating enough time to use the tool. They reported the requirement for validation of any AI tool to increase trust, and the value of AI was felt to be reducing the burden on individual clinicians.
Sibbald et al. (2022)⁶¹	Qualitative, content analysis	Semi-structured interviews	Emergency department physicians with personal experience of using an AI tool to support differential diagnosis (EDS) at triage, purposive sampling	13	2 (15%) female; 5 (38%) <5 years of practice, 4 (31%) 5–10 years, 1 (7%) 11–20 years, 3 (23%) >20 years; 6 (46%) family medicine specialists with subspecialisation in emergency medicine, 7 (54%) emergency medicine specialists	Emergency medicine	AI in use to generate differential diagnosis for emergency medicine triage	Four themes were identified: 1. The quality of EDS was inferred from the scope and prioritization of the diagnoses, 2. Trusting EDS differential diagnoses was linked to varied beliefs around the diagnostic process and potential for bias, 3. Who benefits? Not me, 4. Information flow between EDS and the Electronic Medical Record.
Strohm et al. (2020)²⁵	Qualitative, NR	Semi-structured interviews	Mixed group of radiologists, managers, implementation consultants and data scientists with experience using an AI for automating bone maturity assessments (BoneXpert), sampling for maximal variation	24	20 (83%) radiologists, 5 of which have a dual role as data scientists/managerial, 4 (17%) managers	Radiology	Hypothetical AI used in radiology, with reference to BoneXpert, an AI developed by a commercial company (Visiana) that automated bone maturity assessments using paediatric hand X-rays	Using the NASSS framework, identified facilitating and hindering factors for AI implementation, with one of the most important barriers identified as the non-acceptance of AI by clinicians.
Townsend et al. (2023)⁶²	Mixed-methods, thematic analysis	Semi-structured interviews	Clinicians with current or previous emergency department roles, located in the UK	9	4 (44%) female; age range 20–59 years; experience in emergency medicine range 1 month–22 years	Emergency department	AI in use to generate differential diagnosis for emergency medicine triage	The overarching theme is ‘trust’, with five subthemes: 1. Social, 2. Legal, 3. Ethical, 4. Empathetic, 5. Cultural
van Cauwenberge et al. (2022)⁶³	Mixed-methods, thematic analysis	Think-aloud interviews	Physicians working in a large tertiary care academic hospital in the Netherlands, purposive sampling	30	16 (53%) female; 7 (23%) in training, 8 junior (27%), 15 (50%) senior	General	Hypothetical AI for general clinical and diagnostic support	Four themes were developed: 1. Transparency, 2. Obstructivity, 3. Type of problem, 4., Certainty of advice
Wenderott et al. (2024)⁶⁴	Qualitative,	Semi-structured interviews	Radiologists and radiology residents in a hospital in Germany, convenience sampling	12	8 (67%) 2–4 years of work experience; 5 (42%) worked in department <1 year, 4 (33%) worked in department 1–3 years	Radiology	AI in use to diagnose prostate lesions on MRI	Findings were categorised into AI benefits/risks, barriers/facilitators, external factors influencing AI adoption and contradictory statements
Winter and Carusi (2022)⁶⁵	Qualitative, thematic analysis	Focus groups	Mixed group of professionals involved developing AI for clinical use, and patients/carers with lived experience of pulmonary hypertension, NR	21, split into two FG (10, 11)	FG1: 4 (19%) computer scientists, 4 (19%) clinicians,2 (10%) researchers, 1 (4%) patient representative FG2: 6 (29%) patients, 4 (19%) carers, 1 (4%) patient representative	Respiratory	Hypothetical AI used to diagnose pulmonary hypertension	Four themes were developed: 1. AI can result in early diagnosis, 2. Early diagnosis outweighs data risks of privacy and reuse, 3. Responsibility lies with specialist clinicians, 4. AI will result in deskilling of professionals.
Winter and Carusi (2022)⁶⁶	Qualitative, thematic analysis	Semi-structured interviews and ethnography	Mixed group of professionals involved in the development of a screening algorithm for pulmonary hypertension, NR	3	2 (67%) researchers, 1 (33%) clinician	Respiratory	AI under development to screen for pulmonary hypertension.	Collaboration between clinicians and researchers is encouraged, particularly in 1. Querying datasets, 2. Building the software and 3. Training the model.

NR, Not reported.