Implications for service provision
In summary, the volume of published evidence on the ethical implications of the use of AI-based technologies for medical image classification in screening has increased compared to the 2021 evidence map.9 However, the overall evidence base remains limited, with relatively few qualitative studies available and variability in their quality and focus. Previously, studies primarily consisted of surveys that posed general questions without considering respondents’ experiences or interactions with AI. Although this review identified more qualitative studies than the 2021 evidence map, the evidence is still limited to draw firm conclusions.
Within these constrains, the current evidence review was able to summarise the existing evidence and identify a range of relevant themes related to the ethical application of AI-based technologies in screening. These can provide an indicative overview of the issues that may need to be addressed before implementing AI in screening programmes. In particular, these findings may help stakeholders anticipate concerns related to transparency, accountability, trust, data privacy, professional roles and the practical challenges of implementation. While not definitive, these themes offer a useful starting point for discussion about ethically responsible innovations in AI and may inform the development of governance and oversight frameworks as evidence continues to emerge.
Different ethical issues and considerations identified in this review were deemed to be applicable to different stages of the digital health technology life cycles. Table 15 maps these themes onto the life cycle, illustrating how key ethical concerns such as accountability, transparency and trust may need to be addressed at each stage.
| Digital health technology life cycle | Review findings ‒ applicability to each stage |
|---|---|
| 1. Product conceptualisation | To define the AI’s purpose and scope in ethically appropriate ways, ensuring issues such as human oversight, accountability and bias are addressed from the outsetReview finding 1: Human in the loop ‒ designing AI with human oversight from the outsetReview finding 2: Liability/accountability – defining who is responsible for AI-driven decisionsReview finding 3: Algorithmic bias/discrepancies of real-world vs. training data – addressing potential biases in data selection early |
| 2. Product development | To build and train AI systems while addressing key ethical concerns, including data privacy, fairness in training data, transparency of models and the impact on professional rolesReview finding 4: Data confidentiality, cybersecurity, and consent – ensuring ethical data collection and secure handlingReview finding 5: Transparency – developing comprehensive/explainable AI modelsReview finding 6: Risk to doctors’ competencies and impact on the profession – designing AI to support, not replace clinicians |
| 3. Pre-market assessment and performance validation | To validate AI effectiveness in ways that balance ethical trade-offs, such as sensitivity vs. false positives, tolerance for error and setting realistic performance expectations compared with human standardsReview finding 7: Trade-off between false positives and lesion detection – optimising sensitivity and specificityReview finding 8: Lower tolerance for errors made by AI vs. humans – evaluating AI performance expectations vs. human standardsReview finding 9: Cases of reduced effectiveness – identifying potential limitations in real-world testing |
| 4. Market entry | To ensure regulatory approval and initial deployment are informed by ethical requirements for accountability, transparency and communication of AI’s limitations to stakeholdersReview finding 2: Liability/accountability ‒ ensuring legal and ethical clarity in AI decision-makingReview finding 5: Transparency – communicating AI capabilities and limitations clearly to stakeholders |
| 5. Integration and pilot studies | To integrate AI into healthcare responsibly, addressing ethical concerns such as building user trust, ensuring adequate governance structures and safeguarding patient safetyReview finding 10: Lack of familiarity and trust as a potential barrier to implementation – addressing scepticism among users and professionalsReview finding 11: Systemic integration of AI – ensuring necessary infrastructure, training and regulatory compliance |
| 6. Post market | To monitor AI in real-world settings with a focus on ethical concerns such as sustaining human oversight, identifying cases of reduced effectiveness early and addressing bias across diverse populationsReview finding 1: Human in the loop ‒ continually assessing human-AI interactions in practiceReview finding 9: Cases of reduced effectiveness ‒ identifying and mitigating performance issuesReview finding 3: Algorithmic bias/discrepancies of real-world vs. training data ‒ updating AI models based on diverse patient data |
| 7. Product update | To update AI technologies responsibly by refining models to minimise errors, adapting to evolving professional roles and maintaining transparency with stakeholdersReview finding 7: Trade-off between false positives and lesion detection – refining AI models for better accuracyReview finding 6: Risk to doctors’ competencies and impact on the profession – adapting AI systems to support potentially evolving clinical rolesReview finding 5: Transparency – keeping stakeholders informed about AI modifications |
Suggested research priorities
This review highlighted several areas where further evidence is required to guide the responsible implementation of AI-assisted screening. Future research should prioritise:
- Condition-specific investigations: Most existing evidence was general, with few studies examining ethical issues in depth for particular conditions such as breast cancer or diabetic retinopathy. Research focusing on condition-specific challenges (e.g. differing patient populations, designed explicitly to examine ethical implications of the use of AI-based technologies in screening) would provide more actionable insights in relation to screening for those conditions in particular.
- Patient and public perspectives: While many captured professional views, fewer explored patient acceptability, trust and understanding of AI in screening. Further qualitative research on these perspectives is critical to inform socially acceptable implementation.
- Accountability and liability: Concerns about responsibility for errors remain unresolved. Research on legal, regulatory and organisational frameworks for allocating accountability in AI-assisted screening in the UK clinical setting is needed.
- Algorithmic bias: Future work should explore how training data sets can be designed to minimise bias, particularly in under-represented populations and assess potential implications of such bias for equity of access and clinical outcomes.
- Transparency and communication: Research should examine how best to communicate the role and limitations of AI to patients and clinicians, including strategies to mitigate the ‘black box’ issue and build trust.
Addressing these research priorities could generate condition-specific and overarching evidence to develop ethical frameworks and support policy decisions on the use of AI in screening.