Decision problem
In 2021, a review was commissioned by the UK NSC on automated grading in the UK DESP, including a review question examining the social and ethical implications of implementing AI-based tools in screening programmes and its acceptability to health professionals and the public (Question 4; evidence map).9 For this question, primary studies evaluating the social and ethical impacts of AI-based technology in medical screening programmes or similar setting, as well as review and opinion papers, were included. A considerable number of papers looking at the social and ethical aspects of AI implementation in screening programmes were identified. More specifically, 19 primary studies and 38 relevant review and opinion papers were included. All primary studies were surveys of clinicians and/or the general public or patients. This question aimed to evaluate the volume and type of evidence on the social and ethical implications of implementing AI-based tools in health care and their acceptability to healthcare professionals and the public.
In the context of the previous review, and given the volume and type of evidence identified, an evidence review on the ethical issues of implementing AI-based technology in population screening was recommended by the UK NSC, to summarise existing evidence on this topic and identify gaps requiring further investigation. It was suggested that the experiences and perceptions of health professionals and relevant patient groups should be investigated. Since 2021, new evidence has emerged, and this review broadens the scope beyond diabetic retinopathy to include other screening contexts where relevant studies are available. The present research commissioned by the UK NSC aims to summarise existing evidence and identify ethical issues of implementing AI-based technologies for medical image classification in screening for any medical condition. AI-based technologies have a broad range of potential applications, and it is important that they are implemented ethically. This review aims to inform stakeholders on the advancement of ethically responsible AI innovation.
Interventions
The intervention of interest for this review was any AI-based technology applied to image classification in a screening context. While the primary focus of the review was on screening interventions for breast cancer and diabetic retinopathy, insufficient evidence on these conditions led to a broader scope that included AI applications in screening programmes for any medical condition. Accordingly, the systematic literature review (SLR) included studies where AI was applied to medical imaging and reported on ethical issues or principles relevant to screening regardless of the specific condition.
Population including subgroups
The population relevant to this review is the ‘screening population’, that is people undergoing screening for any health condition, with the primary focus being on people undergoing screening to detect breast cancer and diabetic retinopathy. The views of health professionals, care providers, patients, the users of screening programs and the general public on the AI implementation in population screening were also to be sought.
No population subgroups were relevant for this review, and similar views emerging from health professionals, patients and the general public will be synthesised where possible.
Relevant comparators
As the application of AI in medical image classification in screening was not compared to any other interventions within the context of this review, there were no relevant comparators.
Outcomes
The outcomes to be addressed in this review are detailed below.
Ethical issues
- To identify ethical issues specific to traditional machine learning and deep learning techniques, more specifically, issues associated with non-continual learning and continual learning, respectively.
- Differentiation of ethical issues may be needed by use case, for example fully autonomous AI image interpretation where there is no human involvement versus different levels of human interaction or AI-interpreted images double-checked by a human.
Themes will be derived from the evidence identified for this review and are not pre-specified. Examples of ethical issues may include, but not be limited to, safety and transparency, AI false-negative and false-positive cases, automation bias, data privacy and liability. Relevant quantitative data emerging from included studies will be extracted and presented alongside themes identified from qualitative analysis for illustrative purposes. Themes identified will be compared with the UK NSC ethics framework and mapped to the digital health technology product life cycle.
Key issues
Not applicable.
Overall aims and objectives of the assessment
The purpose of this review is to inform stakeholders about the ethical considerations that should guide the use of innovative AI technologies in medical screening. This is achieved by identifying and synthesising qualitative evidence addressing the ethical implications of the implementation of AI technology in screening programmes. This review is relevant to UK NSC screening criterion 12: ‘There should be evidence that the complete screening programme (test, diagnostic procedures, treatment/intervention) is clinically, socially and ethically acceptable to health professionals and the public’.