Description of health problem
Brief statement describing the health problem
Artificial intelligence (AI) is a technology that enables computers and machines to mimic human abilities like learning, understanding, making decisions and acting independently.1 AI can identify objects, understand and respond to human language and can learn from new information and experiences. AI can make recommendations and sometimes act on its own, reducing the need for human intervention.
Artificial intelligence algorithms are built using different approaches and techniques. Machine learning, a subset of AI, has been a popular approach in current healthcare applications as it allows computational systems to learn from data and improve their performance over time.2 In practice, however, most systems still rely on human instruction, whether through supervised learning (where algorithms are trained on labelled data, e.g. medical images marked as showing cancer or not) or unsupervised learning (where algorithms identify patterns without explicit labels).3 AI offers important advances, including the ability to process large volumes of complex data and identify patterns that may not be visible to human readers (e.g. subtle imaging features linked to disease). Rapid improvement in the performance of AI algorithms has led to advances in the healthcare applications of AI. At the same time, there are notable limitations: the quality of outputs depends heavily on the quality and representativeness of the training data; algorithms can replicate or amplify existing biases; and their decision-making process may be difficult to interpret, raising issues of transparency. These strengths and weaknesses need to be considered when assessing the potential role of AI in medical image classification.
Artificial intelligence has many potential uses, including in public health. For example, AI-based technologies can help screen for medical conditions like breast cancer and eye problems in people with diabetes. The use of AI could improve medical imaging and diagnosis. However, it is crucial that these technologies are implemented ethically, as their use in health care may raise challenges around fairness, safety, privacy and trust in healthcare systems. Thus, it is important to determine the ethical issues that the use of AI-based technologies in screening interventions may raise.
Current policy: target conditions and screening programmes in the UK
Breast cancer is the most common type of cancer in women in the UK, accounting for 15% of all new cancer cases (2017–9) and is the fourth most common cause of cancer death in the UK, accounting for 7% of all cancer deaths.4 Early detection of breast cancers through screening significantly lowers overall mortality.5 For example, the NHS Breast Screening Programme (NHSBSP), which invites UK women aged 50–70 years for screening every 3 years, is estimated to save one life for every 400 women screened, reducing breast cancer deaths by 25%.5 Screening programmes such as this offer opportunities for AI implementation. Mammograms, the input data for an AI algorithm, have a standardised format. The initial decision a radiologist makes, which would be the algorithm’s output, is a simple ‘yes’ or ‘no’ to the question: Does this person require further assessment?5
Screening programmes such as the NHSBSP generate a large number of data, which presents an ideal situation to train and test AI algorithms based on data sets of medical images. Providing effective screening for the UK requires significant time, costs and resources, and it represents opportunities for AI-based technologies to make substantial efficiencies. In addition, breast cancer detection is a challenging task as breast tissue has features that are confounding to the human eye in terms of what may or may not look like a tumour, requiring radiologists or radiographers with significant experience to interpret images.5 In UK clinical practice, to achieve maximal accuracy, every mammogram is assessed by two expert radiologists or specialist trained clinical staff. Breast screening services undertake arbitration to make definitive decisions, with each service using its own local protocol, for example using a third mammography reader or a small panel of readers. Usually, arbitration takes place in cases where there is disagreement between the first two readers, or where they both report an abnormality and recall rates for the service are outside programme standards.6 Within this framework, there is the potential for an automated system such as AI to reduce costs and save processing time.5 In practice, this is more likely to involve supporting specific tasks within the screening pathway, such as acting as a second reader, assisting with arbitration when human readers disagree, or providing an initial triage of images to prioritise cases for review.
Diabetes is estimated to affect > 5.6 million people in the UK.7,8 One of the eye conditions associated with diabetes is diabetic retinopathy; it can lead to sight loss if left untreated and is one of the leading causes of blindness in the working-age population in the UK.9 Eye screening tests have the potential to detect eye problems before they start impacting vision and enable early intervention to reduce the risk of sight loss. In the UK, all patients with diabetes aged ≥ 12 years are invited to attend the UK Diabetic Eye Screening Programme (DESP). This occurs once every year, and once every 2 years, following two consecutive negative screens in England, Wales, Northern Ireland and Scotland. During the visit, images of each eye are taken by healthcare professionals, which are then studied by appropriately trained people known as level 1 graders. Subsequently, all images of patients with suspected diabetic retinopathy are sent for further assessment by level 2 graders, with level 3 graders required to make a final decision if there is disagreement between level 1 and level 2 graders. With the number of people with diabetes increasing annually leading to a greater need for trained health professionals, diabetic eye screening presents an opportunity for the application of AI technology to read images, with potential for workload reduction, similar to breast cancer screening.
In recent years, there has been a significant development in medical AI, particularly with respect to breast cancer detection and screening.3 For example, the Digital Mammography DREAM Challenge aimed to generate algorithms that can reduce false positives without impeding cancer detection.10 Algorithms have reported 80.3–80.4% accuracy, while algorithms that can fully match the accuracy of expert radiologists are being developed.3 Moreover, Google DeepMind Health, NHS Trusts, Cancer Research UK and universities have been developing machine learning technologies for mammogram reading.5 In addition, the Mammography Screening with Artificial Intelligence trial has evaluated the integration of AI in mammography screening, demonstrating the potential of AI to triage cases based on risk scores and assist radiologists by highlighting suspicious findings, offering a hybrid approach to improve efficiency and accuracy.11 Alongside these research developments, commercially available AI systems for breast imaging have also received regulatory approval in the USA and Europe [e.g. Food and Drug Administration (FDA) or Conformité Européenne (CE)-marked platforms such as Transpara or ProFound AI], and prospective studies have begun assessing their performance in routine clinical practice.12–14 Furthermore, AI systems using digital fundus photography instruments have been developed for diabetic retinopathy screening. For example, the US FDA (Silver Spring, MA, USA)-approved, fully autonomous, EyeArt® AI Eye Screening System (Eyenuk, Inc, Los Angeles, CA, USA) has been shown to detect eyes with ‘more than mild’ diabetic retinopathy, with 96% sensitivity and 88% specificity.15
Ethical framework
Overall, AI has the potential to profoundly change the screening and diagnosis of medical conditions, including breast cancer and diabetic retinopathy, speeding up processing times, reducing medical costs and potentially eradicating human errors.5 AI may become widely implemented and an integral part of medical image classification and screening for medical conditions. There has been a wealth of high-level governmental, professional and industry statements on the ethical, legal and social implications of AI, conveying both excitement about the potential benefits of AI and also highlighting concerns about the potential risks and harms.3
The ethical principles of health screening are encompassed within the UK National Screening Committee’s (UK NSC) criteria for assessing the viability, effectiveness and appropriateness of a screening programme. To further clarify and elucidate the overarching goals of screening, the UK NSC developed an ethics framework for screening, which is comprised of four broad ethical principles that are outlined below.16
- Improve health and well-being; that is screening programmes should aim to enhance population’s health and well-being. A screening programme should only be implemented if its potential benefits outweigh its potential harms. The primary focus should be on the individuals offered screening, considering their potential benefits and harms. Broader societal impacts can also be considered. Benefits may include preventing death and disease, improving mental and physical health and enhancing quality of life. Potential harms involve unnecessary tests or treatment, ambiguous results, false reassurance and heightened anxiety. Efforts should be made to minimise these risks.
- Treat people with respect; that is individuals’ rights, wishes and feelings must be respected. This means allowing people to make informed screening decisions that align with their personal values, recognising the influence of family and other relationships. People’s choices about screening must be honoured and supported. For those unable to make their own choices, those deciding on their behalf should be informed of the benefits and harms to the person being screened. Screening programme policies should consider the perspectives of those affected, and the rationale behind these decisions should be clearly communicated.
- Promote equality and inclusion; that is screening programmes should not worsen health inequalities and should strive to minimise them. Access to and the provision of screening should be as fair and inclusive as possible. When starting and carrying out screening, any potential broader, short- and long-term effects on society should be taken into account.
- Use public resources fairly and proportionately; that is the total cost of a screening programme should involve the fair and appropriate use of available public funds. Decisions about screening should be based on evidence from cost-effectiveness analyses.
However, currently, there are no ethical frameworks applicable to developing and using AI specifically in a screening context. Nevertheless, several broad ethical frameworks apply to developing and using AI in a general healthcare context that could potentially be adapted to the screening context. For example, the World Health Organization (WHO) Ethics and Governance of Artificial Intelligence for Health guidance outlines six key principles for the ethical use of AI in health care.17 These include: protecting human autonomy, promoting human well-being and safety and the public interest, ensuring transparency, explainability and intelligibility, fostering responsibility and accountability, ensuring inclusiveness and equity, and promoting AI that is responsive and sustainable.
In addition, there is guidance for understanding AI and safety from the Alan Turing Institute that serves as a comprehensive guide for the ethical and responsible design and implementation of AI systems in the public sector.18 It emphasises integrating ethical principles at every stage of AI development to ensure fair, safe and reliable systems that benefit society, and it identifies potential harms of AI such as bias, discrimination, lack of transparency and invasion of privacy. It includes SUM Values (respect, connect, care and protect) and FAST Track Principles (fairness, accountability, sustainability and transparency) as foundations for ethical AI innovation, aiming to foster trust, prevent harm and ensure public benefit. Within this framework, the importance of establishing governance, accountability and transparency in enabling oversight and trust is emphasised, while implementation considerations such as stakeholder impact assessments, robust training for implementers and a proactive approach to identifying and mitigating risks of bias or harms are highlighted.