You are here

  • Home
  • Archive
  • Volume 15, Issue 5
  • Triple cardiovascular disease detection with an artificial intelligence-enabled stethoscope (TRICORDER): design and rationale for a decentralised, real-world cluster-randomised controlled trial and implementation study

Article Text

Article menu

Download PDFPDF

Cardiovascular medicine
Protocol
Triple cardiovascular disease detection with an artificial intelligence-enabled stethoscope (TRICORDER): design and rationale for a decentralised, real-world cluster-randomised controlled trial and implementation study
  1. Mihir A Kelshiker1,2,
  2. Patrik Bächtiger1,2,
  3. Josephine Mansell1,2,
  4. Daniel B Kramer1,3,
  5. Saloni Nakhare1,2,
  6. Melanie T Almonte1,2,
  7. Abdullah Alrumayh1,
  8. Camille F Petri1,
  9. Alexei Peters1,
  10. Ceire Costelloe4,
  11. Emanuela Falaschetti5,
  12. Carys Barton2,
  13. Rasha Al-Lamee1,2,
  14. Azeem Majeed4,
  15. Carla M Plymen2,
  16. Nicholas S Peters1,2
  1. 1 National Heart and Lung Institute, Imperial College London, London, England, UK
  2. 2 Imperial College Healthcare NHS Trust, London, England, UK
  3. 3 Harvard Medical School, Boston, Massachusetts, USA
  4. 4 School of Public Health, Imperial College London, London, UK
  5. 5 Imperial Clinical Trials Unit, Imperial College London, London, England
  1. Correspondence to Dr Nicholas S Peters; n.peters{at}imperial.ac.uk

Abstract

Introduction Early detection of cardiovascular disease in primary care is a public health priority, for which the clinical and cost-effectiveness of an artificial intelligence-enabled stethoscope that detects left ventricular systolic dysfunction, atrial fibrillation and cardiac murmurs is unproven but potentially transformative.

Methods and analysis TRICORDER is a pragmatic, two-arm, multi-centre (decentralised), cluster-randomised controlled trial and implementation study. Up to 200 primary care practices in urban North West London and rural North Wales, UK, will be randomised to usual care or to have artificial intelligence-enabled stethoscopes available for use. Primary care clinicians will use the artificial intelligence-enabled stethoscopes at their own discretion, without patient-level inclusion or exclusion criteria. They will be supported to do so by a clinical guideline developed and approved by the regional health system executive board. Patient and outcome data will be captured from pooled primary and secondary care records, supplemented by qualitative and quantitative clinician surveys. The coprimary endpoints are (i) difference in the coded incidence (detection) of heart failure and (ii) difference in the ratio of coded incidence of heart failure via hospital admission versus community-based diagnostic pathways. Secondary endpoints include difference in the incidence of atrial fibrillation and valvular heart disease, cost-consequence differential, and prescription of guideline-directed medical therapy.

Ethics and dissemination This trial has ethical approval from the UK Health Research Authority (23/LO/0051). Findings from this trial will be disseminated through publication of peer-reviewed manuscripts, presentations at scientific meetings and conferences with local and national stakeholders.

Trial registration number NCT05987670

  • Artificial Intelligence
  • CARDIOLOGY
  • HEALTH ECONOMICS
  • PUBLIC HEALTH
  • Heart failure
  • Primary Health Care

Data availability statement

Data sharing not applicable as no datasets generated and/or analysed for this study.

https://creativecommons.org/licenses/by/4.0/

This is an open access article distributed in accordance with the Creative Commons Attribution 4.0 Unported (CC BY 4.0) license, which permits others to copy, redistribute, remix, transform and build upon this work for any purpose, provided the original work is properly cited, a link to the licence is given, and indication of whether changes were made. See: https://creativecommons.org/licenses/by/4.0/.

https://doi.org/10.1136/bmjopen-2024-098030

Statistics from Altmetric.com

    Request Permissions

    If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.

    STRENGTHS AND LIMITATIONS OF THIS STUDY

    • Use of real-world data for outcomes evaluation minimises the administrative impact on frontline primary care services and enhances the generalisability of results.

    • Integration of artificial intelligence predictions with routinely collected electronic medical record data allows sensitivity analyses at the level of patient, site and region.

    • Outcome assessment from routine medical coding lacks the granularity and fidelity of primary research form collection.

    Introduction

    Early detection of cardiovascular disease is a global public health priority.1–3 Heart failure (HF), atrial fibrillation (AF) and valvular heart disease (VHD) can have common pathophysiology and significant mortality and morbidity.4–7 Given the reported adult prevalence of HF of 2%, the recognised incidence of HF in Europe of 5 per 1000 person-years in adults8 9is likely to be an underestimate.10–13 HF is deadlier than common, serious cancers,14 profoundly impacting quality of life and costing the UK National Health Service (NHS) over £2 billion per year —or over 2% of its annual budget.15 16 Unacceptably, 70%–80% of all new HF diagnoses are only made after an emergency hospital admission, despite most patients having been seen in primary care with opportunities for detection and therefore earlier initiation of disease-modifying treatment, 17 18expected to translate to improved survival, quality of life and health service utilisation costs.18

    Notwithstanding the mixed evidence of the the effectiveness of screening programmes to improve mortality and stroke associated with AF, the substantial clinical and health economic costs of AF and VHD are similarly driven by late diagnosis.19 20

    Across these cardiac conditions, early detection reduces the risk of disease progression and improves survival, through the initiation of inexpensive, guideline-directed therapies.21–23 However, primary care providers face the challenge that symptoms (if present) are frequently non-specific,24 25 and there are no universally effective point-of-care screening tools to identify the diagnosis or diagnostic pathways.

    We have recently shown that artificial intelligence (AI) applied to ECG and phonocardiogram (PCG) waveforms captured by a ‘smart’ stethoscope can detect left ventricular ejection fraction ≤40%,26 the cause of heart failure with reduced ejection fraction (HFrEF, the most common form of HF), AF and cardiac murmurs–a cardinal examination finding in valvular heart disease.27 The statistical performance of these three AI algorithms has been shown to be high and consistent against international external validation studies.28 This takes a familiar clinical tool with an established workflow and augments examination findings with actionable additional insights, taking only 15 seconds.26 The clinical and health economic impact of deploying such a triple point-of-care AI detection tool in primary care practices is unknown. Furthermore, the determinants of successfully implementing29–31 such an innovation in routine clinical practice are also not known, given the historical failure of AI technology adoption in healthcare systems that is partly attributable to inattention to implementation science and a failure to overcome bias.32 33

    The objective of the TRICORDER study is therefore to conduct a real-world, decentralised cluster-randomised controlled trial and implementation study to determine if providing primary care teams with an AI stethoscope improves early diagnosis of HFrEF, AF and VHD. Secondary objectives include measurement of cost-effectiveness, examination of implementation strategies and identifying the determinants of uptake and utilisation of the AI stethoscope in routine clinical practice. This report describes the clinical and policy context for this publicly funded clinical trial, details of the protocol and plans for implementation and expected findings with a focus on implications for widespread adoption in health systems.

    Methods and analysis

    Setting

    TRICORDER is set within the primary care system of the NHS. General practitioners (GPs) within primary care teams are the gatekeepers to most specialist diagnostic and management services in the NHS through nationally agreed pathways recommended by the National Institute for Health and Care Excellence (NICE, figure 1). Each primary care practice (>8000 in UK) serves an average list of 8500 patients in the community.34 Typically, patients in whom the GP suspects HF are referred for natriuretic peptide (NP) testing,35 which may be via in-practice phlebotomy or via hospital appointment. The NP level determines the urgency of referral for echocardiography and specialist review. There is substantial geographical variation in access to community-based echocardiography services and lead times to hospital outpatient clinic review.36

    Figure 1
    Figure 1

    NICE diagnostic pathway for HF. There are substantial delays to confirmatory echocardiographic testing and specialist review for those who are referred for suspected HF through this pathway, such that only 10% of patients complete the pathway to time and target. HF, heart failure; NICE, National Institute for Health and Care Excellence; NT-pro-BNP, N-terminal pro-B-type natriuretic peptide.

    Funding

    TRICORDER is funded by the UK National Institute for Health and Care Research (NIHR) Invention for Innovation Challenge Award. This is a competitively awarded public funding source, following a multi-stage process involving internal, external clinical and academic peer review, in addition to patient and public scrutiny. Eko Health is not involved in the design, conduct, analysis, interpretation of data or reporting of TRICORDER.

    Study design

    TRICORDER is an open-label, two-arm cluster-randomised controlled trial. Taking a decentralised approach, primary care practices in the NHS North West London Integrated Care System (NWL ICS, 2.8 million residents, figure 2) and the Betsi Cadwaladr University Health Board (BCUHB, 670 000 residents) in North Wales, UK, will be enrolled for 18 months.

    Figure 2
    Figure 2

    Map of London, with the boroughs of North West London stratified by Index of Multiple Deprivation deciles

    Each individual primary care practice employing GPs serves as a cluster—the unit of participation and randomisation. Each participating cluster will variously employ GPs, practice nurses, physician associates and other healthcare professionals who constitute the care team. This unit of recruitment was chosen because the NHS' clinical, financial and information governance processes are standardised at this level.

    Sample size

    Interrogation of the Discover dataset37 (UK Secure Data Environment for TRICORDER endpoint analysis) suggests an incidence rate of HF of 0.62/1000 patients per year. Combining NWL (n=350) and North Wales (n=>50), there are at least 400 potentially eligible study sites. We estimate that each GP practice will have at least 60 eligible patients per year who will receive a stethoscope examination as part of their routine clinical care. Estimating a conservative recruitment ratio of only 50%, we anticipate 200 GP practices being recruited. This provides an eligible study population of 6000 patients per study arm. A review of prior literature in the field indicates an intracluster correlation coefficient of 0.01.38

    Utilising these estimates and assumptions, we would have over 80% power to detect a statistically and clinically significant mean difference if HF is diagnosed at a rate of 0.79/1000 patients in the intervention group, assuming a two-tailed test with a type I error of 0.05. This corresponds to a relative increase of 22%, which is deemed to be clinically meaningful (~17 to 22 HF diagnoses/GP practice).

    The power calculation refers to overall HF, which includes all relevant ICD-10 and SNOMED-CT codes (used in UK primary care), rather than the codes relating to HFrEF specifically. There are three reasons for such an inclusive approach.

    First, we aim to mitigate an artificially low incidence of HFrEF due to coding heterogeneity in primary care. Preliminary analysis of the Discover dataset revealed substantial heterogeneity in clinical coding of HF subtypes. This is well recognised in recent studies.39

    Second, this approach captures the effect on clinician behaviour from participation in the study—it is feasible that more of all HF subtypes may be detected during the study period.

    Finally, this approach enables rapid clinical and health economic endpoint analysis, avoiding the time and cost burden of manual data entry/data collection from chart arbitration, which would otherwise risk undermining the scalability of TRICORDER and set an impractical precedent for longer-term outcomes analyses.

    Inclusion and exclusion criteria

    Eligible primary care practices must meet the following inclusion criteria:

    • Caring for adult patients ≥18 years old.

    • Able to request natriuretic peptide blood testing (a standard screening test for HF in patients with symptoms) and initiating the HF diagnostic pathway recommended by the NICE35.

    Exclusion criteria:

    • Poor WiFi and/or mobile data connectivity within primary care consulting rooms, prohibiting the use of the AI-enabled stethoscope.

    • Not providing face-to-face patient consultations.

    Patient and public involvement

    Patients were involved in the conception and trial design via a patient steering group and through online survey of over 10 000 patients at Imperial College Healthcare NHS Trust, UK. The study design was further developed in collaboration with the Pumping Marvellous Foundation, the UK’s largest patient-led heart failure charity and their Patient Educator group.

    Enrolment

    Potential participating primary care practices will be approached by their parent NHS organisations: the North West London (NWL) Integrated Care System or Betsi Cadwaladr University Health Board (BCUHB), in England and Wales, respectively. Additionally, in NWL, primary care practices will be approached by the National Institute for Health and Care Research Clinical Research Network, whose mandate is to widen access to and diversify research participation. Practices expressing an interest will be contacted by the study team with up to two further follow-up emails and a subsequent phone call if necessary. Written informed consent will be recorded from leadership at each participating practice. Clinicians at all practices will be asked to complete a baseline questionnaire measuring their confidence in detecting and managing cardiovascular disease. The study team will have no direct contact with patients.

    Randomisation

    Practices will be randomised 1:1 to the intervention or control arms using a validated, automated and audited randomisation tool (Sealed Envelope, London, UK)40 with allocation concealment. Practices will be notified of their treatment allocation by email.

    Intervention arm: AI-enabled stethoscope

    This study will investigate the impact of an AI-enabled stethoscope (henceforth referred to as AI stethoscope) with integrated sensor technology—electrodes and a microphone—for recording digital single-lead ECG and PCG (Eko DUO, Eko Health Inc, California, USA). The AI stethoscope works as a conventional stethoscope with diaphragm and detachable tubing that enables conventional auscultation, with other features including noise filtration and cancellation. The device connects to an app on the user’s smartphone (Eko: Digital Stethoscope+ECG, Eko Health, USA) using Bluetooth connectivity for waveform visualisation. Connectivity via cellular/Wi-Fi allows access to cloud-based AI algorithms for analysis of waveforms (no data is stored in the AI stethoscope or user’s smartphone).

    The AI stethoscope and associated AI algorithms are regulated by the UK Medicines and Healthcare Products Regulatory Agency (MHRA) and are UKCA-marked as Class IIa medical devices, respectively. This means that they are authorised for use in clinical care in the UK, in accordance with their regulated intended purpose.

    The AI stethoscope is provided as a replacement stethoscope for routine patient clinical examination. Use of the AI stethoscope will be at the discretion of the responsible clinician, with no patient-level inclusion or exclusion criteria. The additional functionality in the form of AI algorithms for HFrEF, AF and VHD detection will be used within its regulated intended purpose. As established by previous studies,26 the optimum unit of examination to derive AI insights entails placing the device–‘listening’–over the pulmonary position (upper left sternal border) and taking a 15-s recording.

    In an NHS first, the clinical guidelines for actioning AI outputs have been agreed by the Cardiovascular Executive Groups of the NHS North West London Integrated Care System and Betsi Cadwaladr University Health Board (online supplemental figure 1). Patients will be examined with the AI stethoscope in accordance with these guidelines, and/or where stethoscope examination is deemed clinically appropriate by the primary care clinician. Patients will provide verbal consent for examination with the AI stethoscope as per any physical examination performed by healthcare professionals for direct care purposes, in accordance with UK law and General Medical Council guidelines.41

    Supplemental material

    Primary care clinicians at practices in the intervention arm will be provided with one session of in-person training in the use of the AI stethoscope. The training includes:

    1. Delivery and setup of the AI stethoscope in each consultation room

    2. Smartphone application installation and login (for either iOS or Android systems) on each clinician’s smartphone for the use of the AI stethoscope

    3. Pairing of all clinician smartphones with all AI stethoscopes in the same practice

    4. Demonstration of patient examination, single-lead ECG capture from the ‘spot check’ auscultation position and AI recommendation

    5. Troubleshooting signal capture

    6. Visual summary sheet for patient examination and troubleshooting affixed next to AI stethoscope in each consultation room

    7. Update on the local clinical guideline and summary sheet affixed next to the AI stethoscope in each consultation room (online supplemental figure 1)

    If any clinicians (who can initiate diagnostic pathways for HF, AF or VHD) decline to install the application on their smartphone, this will be logged by the study team.

    GP practices in the intervention arm will receive six monthly updates (three total) from the clinical research team. These will be communicated via direct email to a nominated clinician from each practice team and include relevant data on numbers of positive screening results and subsequent clinician actions such as specialist referrals and implemented medical therapies. The reports will include monthly results of the following:

    1. Number of patients with positive screening results

    2. Proportion of patients with positive screening results referred for specialist review

    3. Proportion of patients with positive screening results receiving guideline-direct medical therapies

    At the end of the study, clinicians in the intervention arm will be asked to complete a validated usability questionnaire for the AI stethoscope (System Usability Scale).42

    Control arm

    Primary care practices in the control arm will continue with usual clinical care, with decision-making for consideration of HF, and initiation of the NICE diagnostic pathway based on clinical judgement alone.

    Participating practices can withdraw at any time. If a primary care practice ceases face-to-face consultations during the study, they will be withdrawn from the study by the research team.

    Data sources

    In the NWL ICS, patient-level complete primary and secondary care clinical and cost data is pseudonymised and pooled within Discover-NOW, a UK Trusted Research Environment.37 AI stethoscope usage statistics and AI predictions will be integrated into the Discover-NOW and tagged to appropriate patient records. The study team will access this real-world data platform to measure population-level clinical and health economic outcomes, in addition to patient-level sensitivity analyses for NWL ICS.

    At BCUH, only the direct care team will access medical records and record population-level outcomes as part of a service evaluation (coprimary endpoints), using epidemiological methods only.

    Study outcomes

    Outcomes are summarised in box 1.

    Box 1

    Study outcomes

    Study outcomes comparing intervention and control groups

    Coprimary endpoints

    1. Incidence of coded new diagnoses of HF

    2. Ratio of coded diagnoses of HF via hospital admission-based versus community-based pathways

    Secondary endpoints

    1. New coded diagnoses of atrial fibrillation (AF)

    2. New coded diagnoses of valvular heart disease (VHD)

    3. Cost-consequence differential (HF, AF, VHD)

    4. Health service utilisation for diagnostics

    5. Prescription of guideline-directed medical therapy for HF, AF, VHD

    6. New implantation of cardiac resynchronisation therapy (CRT) and/or implantable cardiac defibrillator (ICD) devices

    7. Differential rates of uptake and utilisation of AI stethoscope in primary care

    8. Determinants of utilisation of AI stethoscope in primary care

    9. Quality of life–healthy days at home (HDAH)50 51

    Independent variables for sensitivity and subgroup analyses

    1. GP practice population social deprivation

    2. GP practice national target performance

    3. GP practice clinical staffing model

    Statistical analysis

    Baseline patient-level data in NWL will be extracted at the index date, defined as the date of first examination using the AI stethoscope at each primary care practice. Data will be summarised as mean (SD) or median (IQR) for skewed data. Comparisons between groups will be undertaken on an intention-to-treat basis, using two-sample t tests for continuous variables, and χ2 tests for binary and categorical variables. For each outcome, we will perform generalised mixed-effect logistic regression to compare the study arms, with the GP practice as a random effect.

    Subgroup analyses for patient-level outcomes will be performed, stratified by age, sex, ethnicity and geography, patient comorbidities and GP practice characteristics (mean age of clinicians, size of clinical team, number of registered patients).

    Patient-level sensitivity analyses will be performed for patients with abnormal AI-stethoscope predictions for HF to identify direct associations between AI stethoscope predictions and specific diagnostic codes for HF, AF and VHD.

    We will measure relationships between items in the clinician questionnaire, utilisation rates of the AI stethoscope, compliance with recommendations and the primary endpoints. These analyses will identify domains that are most strongly related to utilisation and compliance, which will help improve the technology, and guide future implementation and clinical pathways.

    Interim analysis

    An interim analysis will be performed 6 months after the first site has been randomised. The interim analysis will report the following selected outcomes and will be reviewed by the Trial Steering Committee, who will remain blinded to the treatment allocation of each group until the study end:

    Coprimary endpoints:

    1. Difference in the incidence of coded HF between groups;

    2. Difference in the ratio of coded HF incidence via hospital admission-based versus community-based diagnostic pathways between groups

    Secondary endpoints

    1. Healthcare diagnostics utilisation

    2. Primary care appointments

    3. Emergency department presentations

    4. Non-elective hospital admissions

    5. AI stethoscope utilisation rates

    6. Completion of site setup for all practices in intervention group

    7. Data fidelity—ascertainment

    The following interim outcomes will be considered significant and warranting intervention:

    1. Mean AI-stethoscope utilisation rate of less than five recordings per month per practice in the intervention arm

    2. Healthcare diagnostics utilisation differential affecting patient safety

    3. Incomplete site setup in intervention group

    In keeping with formative evaluations in implementation research29–31 and behavioural change frameworks,43 adaptive interventions will be considered prospectively to maximise the success of the implementation of AI stethoscope (given the established confidence for the technology ‘working’ on the basis of widely validated statistical accuracy). In addition, prespecified interventions will be available to the Steering Committee following the interim analysis.

    • Semistructured interviews of high- and low-use participants

    • Provision of repeat training in the use of the AI stethoscope

    • Redistribution of the AI stethoscope to new, prospectively enrolled GP practices

    • Provision of latest generation of the AI stethoscope hardware and software to intervention group

    • Integration of the AI stethoscope system with the electronic health record system

    To mitigate lead-time effects, interim analysis of the coprimary endpoints will not incur a stopping rule at 6 months. If at 6 months there is no statistically significant difference between treatment arms for the coprimary endpoints, then a further interim review will be performed at 12 months. If there remains no clinically significant difference in the coprimary endpoint at 12 months, the Trial Steering Group will have the option to stop the trial on the basis of futility.

    Ethics and dissemination

    TRICORDER will not require written/signed consent from individual patients since the AI- tethoscope has full regulatory (MHRA) approval for use in direct clinical care and will be used within its regulatory-approved intended purpose. The NHS Integrated Care Systems involved in TRICORDER have developed and approved a clinical guideline for the use of the AI stethoscope in direct care, through agreement between executive primary and secondary care stakeholders (online supplemental figure 1), which are anchored in best practice national guidelines from the UK National Institute for Health and Care Excellence (NICE). As with any clinical test, the decision to perform it and act on results in line with guidelines is ultimately underpinned by the clinical judgement of the responsible medical professional and involves informed consent of the patient. This protocol has been reviewed and received a favourable opinion from a UK Health Research Authority (HRA) Research Ethics Committee (reference: 23/LO/0051), including its cluster design and confirmation that patient-level research consent is not indicated. For the use of Discover-NOW, there is existing UK HRA approval for access procedures, pseudonymisation and use of this platform for outcomes research (reference: 18/WM/0323). A scientific manuscript with the primary outcomes of the study will be published in a peer-reviewed journal. Further manuscripts reporting secondary outcomes will also be published in peer-reviewed journals. Results will be presented at scientific meetings and conferences with local and national stakeholders, including the national funder (National Institute for Health and Care Research), NHS organisations and patient and public groups such as the Pumping Marvellous Foundation.

    Trial registration

    TRICORDER is registered with the NIH National Library of Medicine (NCT05987670). The study protocol has been reported in accordance with the SPIRIT-AI Extension44 (online supplemental table 1), and outcomes will be reported in accordance with the CONSORT-AI Extension.45

    Supplemental material

    Discussion

    TRICORDER is the first cluster-randomised implementation trial to evaluate a real-time, AI-dependent point-of-care diagnostic in primary care. The study is designed to meet the specifications of the NICE Evidence Standards Framework for Digital Health Technologies46 and therefore underpin recommendations for subsequent NHS-wide commissioning. This protocol—and lessons from the execution of the trial—will inform a replicable blueprint for the evaluation of other similar digital health technologies.

    The study is designed to address the unacceptable reality that cardiovascular disease, heart failure particularly, is most frequently detected at a late stage, after disease progression precipitates a hospital admission.17 18 The NHS Long Term Plan prioritises the need to reverse this trend, emphasising the need for increased rates of diagnosis through community (primary care initiated) pathways.1 We have recently highlighted the £2500 saving unlocked by every patient diagnosed with heart failure through such community pathways.18 Similar models inform a compelling health economic case for community-based detection of AF and VHD.19 20

    By spanning primary care practices across urban and rural geographies and serving populations with varied sociodemographic and ethnic backgrounds, this study will encompass a uniquely representative patient population. This is intended to consider the well-established concerns around AI bias,47 the digital maturity for technology adoption48 and to demonstrate the AI stethoscope’s suitability for deployment across diverse patient and clinician populations. The research team will collate qualitative data via clinician surveys, ensuring the AI stethoscope’s efficient integration into primary care, preservation of the traditional clinician–patient interaction and high-level usability of the device.

    This study is both decentralised and principally uses ‘real-world’ data; there is no lead/central study site or logistics/travel outside each GP practice. Each will contribute data equally, but without the need to complete any study-specific (cumbersome) data collection instruments. Instead, our use of Discover-NOW enables outcomes measurement with routinely recorded (real world) data. The platform’s offer of comprehensive linked (primary and secondary care) clinical and cost data affords patient-level sensitivity analyses. This will allow robust measurement of associations between diagnostic outcomes and use of the AI stethoscope. Separate to such novel elements of this study, TRICORDER has set several other important precedents for translational AI research, including approval of the first regional NHS clinical guideline for the use of a primary care focused AI technology, and sector-wide data governance approval covering use of an AI technology across hundreds of disparate primary care sites.

    This research protocol is best interpreted in the context of its limitations. The pragmatic design, which aims to have a minimal impact on GP workflows, may in some cases limit sustained use of the technology by not setting any expectations or requirement for use—though this may somewhat mitigate any Hawthorne effect.49 This study is the first to accompany a novel AI technology with an NHS sector-approved clinical guideline for use, but variable adherence may limit the impact attributable to the technology. This will be addressed systematically, taking an implementation science approach to maximise uptake of the intervention. Finally, the examination of real-world data is universally limited by the inconsistency and variable fidelity of medical coding in capturing specific variables of interest. For example, for HF, this is rarely coded with a granularity that describes preserved, moderately reduced or reduced ejection fraction. However, the otherwise comprehensive data flows associated with this study will allow holistic scrutiny of a broad selection of outcomes that need to be understood to underpin any recommendations for system-wide uptake. Ultimately, this will serve to provide a realistic estimate of the potential clinical and health economic impact of the AI stethoscope technology in NHS primary care.

    Data availability statement

    Data sharing not applicable as no datasets generated and/or analysed for this study.

    Ethics statements

    Patient consent for publication

    Acknowledgments

    The authors are grateful for the infrastructural support from the National Institute of Health Research (NIHR) Biomedical Research Centre based at Imperial College Healthcare NHS Trust and Imperial College London. AM is supported by the NIHR Applied Research Collaboration NW London. The views expressed in this publication are those of the authors and not necessarily those of the NIHR or the Department of Health and Social Care. We are indebted to the codevelopment with and feedback from patients from the Pumping Marvellous Foundation heart failure charity.

    References

    1. ↵
      1. NHS
      . NHS long term plan. 2019. Available: https://www.longtermplan.nhs.uk
    2. ↵
      2. Vos T ,
      3. Lim SS ,
      4. Abbafati C , et al
      . Global burden of 369 diseases and injuries in 204 countries and territories, 1990–2019: a systematic analysis for the Global Burden of Disease Study 2019. The Lancet 2020;396. doi:10.1016/S0140-6736(20)30925-9
    3. ↵
      1. Organisation WH
      . Global health estimates 2020: deaths by cause, age, sex, by country and by region, 2000-2019. World Health Organisation; 2020.
    4. ↵
      2. Kannel WB ,
      3. Abbott RD ,
      4. Savage DD , et al
      . Epidemiologic features of chronic atrial fibrillation: the Framingham study. N Engl J Med 1982;306:1018–22. doi:10.1056/NEJM198204293061703
      OpenUrlCrossRefPubMedWeb of Science
    5. ↵
      2. Benjamin EJ ,
      3. Levy D ,
      4. Vaziri SM , et al
      . Independent risk factors for atrial fibrillation in a population-based cohort. The Framingham Heart Study. JAMA 1994;271:840–4. doi:10.1001/jama.1994.03510350050036
      OpenUrlCrossRefPubMedWeb of Science
    6. ↵
      2. Krum H ,
      3. Gilbert RE
      . Demographics and concomitant disorders in heart failure. The Lancet 2003;362:147–58. doi:10.1016/S0140-6736(03)13869-X
      OpenUrl
    7. ↵
      2. Ehrlich JR ,
      3. Nattel S ,
      4. Hohnloser SH
      . Atrial Fibrillation and Congestive Heart Failure: Specific Considerations at the Intersection of Two Common and Important Cardiac Disease Sets. Cardiovasc electrophysiol 2002;13:399–405. doi:10.1046/j.1540-8167.2002.00399.x
      OpenUrl
    8. ↵
      2. Meyer S ,
      3. Brouwers FP ,
      4. Voors AA , et al
      . Sex differences in new-onset heart failure. Clin Res Cardiol 2015;104:342–50. doi:10.1007/s00392-014-0788-x
      OpenUrlCrossRefPubMed
    9. ↵
      2. Brouwers FP ,
      3. de Boer RA ,
      4. van der Harst P , et al
      . Incidence and epidemiology of new onset heart failure with preserved vs. reduced ejection fraction in a community-based cohort: 11-year follow-up of PREVEND. Eur Heart J 2013;34:1424–31. doi:10.1093/eurheartj/eht066
      OpenUrlCrossRefPubMedWeb of Science
    10. ↵
      2. Conrad N ,
      3. Judge A ,
      4. Tran J , et al
      . Temporal trends and patterns in heart failure incidence: a population-based study of 4 million individuals. The Lancet 2018;391:572–80. doi:10.1016/S0140-6736(17)32520-5
      OpenUrl
    11. ↵
      2. James SL ,
      3. Abate D ,
      4. Abate KH , et al
      . Global, regional, and national incidence, prevalence, and years lived with disability for 354 diseases and injuries for 195 countries and territories, 1990–2017: a systematic analysis for the Global Burden of Disease Study 2017. The Lancet 2018;392:1789–858. doi:10.1016/S0140-6736(18)32279-7
      OpenUrlCrossRefPubMed
    12. ↵
      2. Roger VL
      . Epidemiology of heart failure. Circ Res 2013;113:646–59. doi:10.1161/CIRCRESAHA.113.300268
      OpenUrlAbstract/FREE Full Text
    13. ↵
      2. Mosterd A ,
      3. Hoes AW
      . Clinical epidemiology of heart failure. Heart 2007;93:1137–46. doi:10.1136/hrt.2003.025270
      OpenUrlFREE Full Text
    14. ↵
      2. Mamas MA ,
      3. Sperrin M ,
      4. Watson MC , et al
      . Do patients have worse outcomes in heart failure than in cancer? A primary care-based cohort study with 10-year follow-up in Scotland. Eur J Heart Fail 2017;19:1095–104. doi:10.1002/ejhf.822
      OpenUrlCrossRefPubMed
    15. ↵
      2. Braunschweig F ,
      3. Cowie MR ,
      4. Auricchio A
      . What are the costs of heart failure? Europace 2011;13:ii13–7. doi:10.1093/europace/eur081
      OpenUrlCrossRefPubMedWeb of Science
    16. ↵
      1. Excellence NIfHaC
      . Resource impact report: chronic heart failure in adults: diagnosis and management (NG106). National Institute for Health and Care Excellence; 2018. Available: https://www.nice.org.uk/guidance/ng106/resources/resource-impact-report-pdf-6537494413#:~:text=2.3%20It%20is%20estimated%20that,into%20Living%20with%20Heart%20Failure
    17. ↵
      2. Bottle A ,
      3. Kim D ,
      4. Aylin P , et al
      . Routes to diagnosis of heart failure: observational study using linked data in England. Heart 2018;104:600–5. doi:10.1136/heartjnl-2017-312183
      OpenUrlAbstract/FREE Full Text
    18. ↵
      2. Bachtiger P ,
      3. Kelshiker MA ,
      4. Petri CF , et al
      . Survival and health economic outcomes in heart failure diagnosed at hospital admission versus community settings: a propensity-matched analysis. BMJ Health Care Inform 2023;30:e100718. doi:10.1136/bmjhci-2022-100718
    19. ↵
      2. Burdett P ,
      3. Lip GYH
      . Atrial fibrillation in the UK: predicting costs of an emerging epidemic recognizing and forecasting the cost drivers of atrial fibrillation-related costs. Eur Heart J Qual Care Clin Outcomes 2022;8:187–94. doi:10.1093/ehjqcco/qcaa093
      OpenUrl
    20. ↵
      2. Moore M ,
      3. Chen J ,
      4. Mallow PJ , et al
      . The direct health-care burden of valvular heart disease: evidence from US national survey data. Clinicoecon Outcomes Res 2016;8:613–27. doi:10.2147/CEOR.S112691
      OpenUrl
    21. ↵
      2. McDonagh TA ,
      3. Metra M ,
      4. Adamo M , et al
      . 2021 ESC Guidelines for the diagnosis and treatment of acute and chronic heart failure. Eur Heart J 2021;42:3599–726. doi:10.1093/eurheartj/ehab368
      OpenUrlCrossRefPubMed
    22. ↵
      2. Hindricks G ,
      3. Potpara T ,
      4. Dagres N , et al
      . ESC Guidelines for the diagnosis and management of atrial fibrillation developed in collaboration with the European Association for Cardio-Thoracic Surgery (EACTS): The Task Force for the diagnosis and management of atrial fibrillation of the European Society of Cardiology (ESC) Developed with the special contribution of the European Heart Rhythm Association (EHRA) of the ESC. Eur Heart J 2020;42:373–498. doi:10.1093/eurheartj/ehaa798
      OpenUrl
    23. ↵
      2. Vahanian A ,
      3. Beyersdorf F ,
      4. Praz F , et al
      . ESC/EACTS Guidelines for the management of valvular heart disease: Developed by the Task Force for the management of valvular heart disease of the European Society of Cardiology (ESC) and the European Association for Cardio-Thoracic Surgery (EACTS). Eur Heart J 2021;43:561–632. doi:10.1093/eurheartj/ehab395
      OpenUrl
    24. ↵
      2. Fonseca C ,
      3. Morais H ,
      4. Mota T , et al
      . The diagnosis of heart failure in primary care: value of symptoms and signs. Eur J Heart Fail 2004;6:795–800, . doi:10.1016/j.ejheart.2004.08.002
      OpenUrlCrossRefPubMedWeb of Science
    25. ↵
      2. Yancy CW ,
      3. Jessup M ,
      4. Bozkurt B , et al
      . 2013 ACCF/AHA guideline for the management of heart failure: a report of the American College of Cardiology Foundation/American Heart Association Task Force on Practice Guidelines. J Am Coll Cardiol 2013;62:e147–239. doi:10.1016/j.jacc.2013.05.019
      OpenUrlFREE Full Text
    26. ↵
      2. Bachtiger P ,
      3. Petri CF ,
      4. Scott FE , et al
      . Point-of-care screening for heart failure with reduced ejection fraction using artificial intelligence during ECG-enabled stethoscope examination in London, UK: a prospective, observational, multicentre study. Lancet Digit Health 2022;4:e117–25. doi:10.1016/S2589-7500(21)00256-9
      OpenUrlCrossRef
    27. ↵
      2. Davidsen AH ,
      3. Andersen S ,
      4. Halvorsen PA , et al
      . Diagnostic accuracy of heart auscultation for detecting valve disease: a systematic review. BMJ Open 2023;13:e068121. doi:10.1136/bmjopen-2022-068121
    28. ↵
      2. Chorba JS ,
      3. Shapiro AM ,
      4. Le L , et al
      . Deep Learning Algorithm for Automated Cardiac Murmur Detection via a Digital Stethoscope Platform. J Am Heart Assoc 2021;10:e019905. doi:10.1161/JAHA.120.019905
    29. ↵
      2. Bauer MS ,
      3. Kirchner J
      . Implementation science: What is it and why should I care? Psychiatry Res 2020;283:112376. doi:10.1016/j.psychres.2019.04.025
      OpenUrlCrossRefPubMed
    30. ↵
      2. Elwy AR ,
      3. Wasan AD ,
      4. Gillman AG , et al
      . Using formative evaluation methods to improve clinical implementation efforts: Description and an example. Psychiatry Res 2020;283:112532. doi:10.1016/j.psychres.2019.112532
      OpenUrl
    31. ↵
      2. Handley MA ,
      3. Gorukanti A ,
      4. Cattamanchi A
      . Strategies for implementing implementation science: a methodological overview. Emerg Med J 2016;33:660–4. doi:10.1136/emermed-2015-205461
      OpenUrlAbstract/FREE Full Text
    32. ↵
      2. Parikh RB ,
      3. Teeple S ,
      4. Navathe AS
      . Addressing Bias in Artificial Intelligence in Health Care. JAMA 2019;322:2377–8. doi:10.1001/jama.2019.18058
      OpenUrlCrossRefPubMed
    33. ↵
      2. Challen R ,
      3. Denny J ,
      4. Pitt M , et al
      . Artificial intelligence, bias and clinical safety. BMJ Qual Saf 2019;28:231–7. doi:10.1136/bmjqs-2018-008370
      OpenUrlFREE Full Text
    34. ↵
      2. Digital N
      . Patients registered at a gp practice. 2023.
    35. ↵
      1. Excellence NIfHaC
      . Chronic heart failure in adults: diagnosis and management. National Institute for Health and Care Excellence; 2018. Available: https://www.nice.org.uk/guidance/ng106/resources/chronic-heart-failure-in-adults-diagnosis-and-management-pdf-66141541311685
    36. ↵
      2. Cowie MR
      . The heart failure epidemic: a UK perspective. Echo Res Pract 2017;4:R15–20. doi:10.1530/ERP-16-0043
      OpenUrlCrossRef
    37. ↵
      2. Bottle A ,
      3. Cohen C ,
      4. Lucas A , et al
      . How an electronic health record became a real-world research resource: comparison between London’s Whole Systems Integrated Care database and the Clinical Practice Research Datalink. BMC Med Inform Decis Mak 2020;20:71. doi:10.1186/s12911-020-1082-7
    38. ↵
      2. Yao X ,
      3. Rushlow DR ,
      4. Inselman JW , et al
      . Artificial intelligence-enabled electrocardiograms for identification of patients with low ejection fraction: a pragmatic, randomized clinical trial. Nat Med 2021;27:815–9. doi:10.1038/s41591-021-01335-4
      OpenUrlCrossRefPubMed
    39. ↵
      2. Kahn M ,
      3. Grayson AD ,
      4. Chaggar PS , et al
      . Primary care heart failure service identifies a missed cohort of heart failure patients with reduced ejection fraction. Eur Heart J 2022;43:405–12. doi:10.1093/eurheartj/ehab629
      OpenUrl
    40. ↵
      2. Ltd SE
      . Simple randomisation service. Sealed Envelope Ltd; 2022. Available: https://www.sealedenvelope.com/simple-randomiser/v1
    41. ↵
      2. Council GM
      . Good medical practice. General Medical Council, 2013.
    42. ↵
      2. Brooke J
      . SUS: A quick and dirty usability scale. Usability Eval Ind 1995.
    43. ↵
      2. Michie S ,
      3. van Stralen MM ,
      4. West R
      . The behaviour change wheel: a new method for characterising and designing behaviour change interventions. Implement Sci 2011;6:42. doi:10.1186/1748-5908-6-42
      OpenUrlCrossRefPubMed
    44. ↵
      2. Rivera SC ,
      3. Liu X ,
      4. Chan A-W , et al
      . Guidelines for clinical trial protocols for interventions involving artificial intelligence: the SPIRIT-AI Extension. BMJ 2020;370:m3210. doi:10.1136/bmj.m3210
    45. ↵
      2. Liu X ,
      3. Cruz Rivera S ,
      4. Moher D , et al
      . Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension. Nat Med 2020;26:1364–74. doi:10.1038/s41591-020-1034-x
      OpenUrlCrossRefPubMed
    46. ↵
      2. Unsworth H ,
      3. Dillon B ,
      4. Collinson L , et al
      . The NICE Evidence Standards Framework for digital health and care technologies - Developing and maintaining an innovative evidence framework with global impact. Digit Health 2021;7:20552076211018617. doi:10.1177/20552076211018617
      OpenUrl
    47. ↵
      2. Obermeyer Z ,
      3. Powers B ,
      4. Vogeli C , et al
      . Dissecting racial bias in an algorithm used to manage the health of populations. Science 2019;366:447–53. doi:10.1126/science.aax2342
      OpenUrlAbstract/FREE Full Text
    48. ↵
      2. Greenhalgh T ,
      3. Wherton J ,
      4. Papoutsi C , et al
      . Beyond Adoption: A New Framework for Theorizing and Evaluating Nonadoption, Abandonment, and Challenges to the Scale-Up, Spread, and Sustainability of Health and Care Technologies. J Med Internet Res 2017;19:e367. doi:10.2196/jmir.8775
    49. ↵
      2. Chiesa M ,
      3. Hobbs S
      . Making sense of social research: how useful is the Hawthorne Effect? Euro J Social Psych 2008;38:67–74. doi:10.1002/ejsp.401
      OpenUrl
    50. ↵
      2. Greene SJ ,
      3. O’Brien EC ,
      4. Mentz RJ , et al
      . Home-Time After Discharge Among Patients Hospitalized With Heart Failure. J Am Coll Cardiol 2018;71:2643–52. doi:10.1016/j.jacc.2018.03.517
      OpenUrlFREE Full Text
    51. ↵
      2. Lam MB ,
      3. Riley KE ,
      4. Zheng J , et al
      . Healthy days at home: A population-based quality measure for cancer patients at the end of life. Cancer 2021;127:4249–57. doi:10.1002/cncr.33817
      OpenUrl

    Footnotes

    • X @Azeem_Majeed

    • MAK and PB contributed equally.

    • Contributors MAK, PB and NSP were involved in conception and trial design. CFP, CC and EF provided statistical expertise. RAL provided expertise on patient and public involvement. MAK and PB were involved in drafting this manuscript. JM, DBK, SN, MTA, AP, AA, CB, AM, CMP and NSP provided critical revision of the manuscript and intellectual content. All authors were involved in the final approval of the article. MAK and PB contributed equally and are cofirst authors of the work. NSP is the guarantor.

    • Funding This work was supported by the National Institute for Health and Care Research Invention for Innovation (i4i) Challenge Award, Accelerated Access Collaborative and NHSX via the AI Award in Health and Social Care; Imperial Health Charity; British Heart Foundation. Funders were not involved in the study design, writing of the report or decision to submit the article for publication. All authors had full access to the protocol and take responsibility for the integrity of the manuscript.

    • Map disclaimer The depiction of boundaries on this map does not imply the expression of any opinion whatsoever on the part of BMJ (or any member of its group) concerning the legal status of any country, territory, jurisdiction or area or of its authorities. This map is provided without any warranty of any kind, either expressed or implied.

    • Competing interests MAK has nothing to declare. PB has nothing to declare. JM has nothing to declare. CFP has nothing to declare. MTA has nothing to declare. AA has nothing to declare. DBK has nothing to declare. AP has nothing to declare. CC has nothing to declare. EF has nothing to declare. RAL has nothing to declare. AM has nothing to declare. CMP has nothing to declare. NSP is an advisor to Eko Health Inc.

    • Patient and public involvement Patients and/or the public were involved in the design, conduct, reporting or dissemination plans of this research. Refer to the Methods section for further details.

    • Provenance and peer review Not commissioned; externally peer reviewed.

    • Supplemental material This content has been supplied by the author(s). It has not been vetted by BMJ Publishing Group Limited (BMJ) and may not have been peer-reviewed. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by BMJ. BMJ disclaims all liability and responsibility arising from any reliance placed on the content. Where the content includes any translated material, BMJ does not warrant the accuracy and reliability of the translations (including but not limited to local regulations, clinical guidelines, terminology, drug names and drug dosages), and is not responsible for any error and/or omissions arising from translation and adaptation or otherwise.