. “My Guide to Thinking Critically About AI in Healthcare is now complete and available here. It takes the form of a themes annotated bibliography with brief summaries for each theme. A quick look at the themes:
1. What can AI be used for in healthcare? Pretty self-explanatory, but this theme covers: (a) four categories of use cases of narrow AI: Prediction, Classification, Association, and Optimisation; and (b) briefly touches on hypothetical uses of ‘General AI’.
2. Why is its use so appealing? This theme looks at the most common rhetorical arguments used to justify the development and implementation of AI inc. P4 Medicine and the Triple Aim.
3. How is it developed? Here is a very(!) brief overview of the AI model lifecycle, the steps involved, and the types of considerations that might come up at different stages.
4. What are its limitations? Most of the limitations are covered in detail in other themes, but this section provides a high-level list of some of the generalisable limitations related to robustness, reproducibility, and causality.
5. Does it actually work? This is a two-pronged theme – considering both how evidence of efficacy/accuracy is generated in the context of AI and the quality of the currently available evidence. SPOILER ALERT the existing evidence is pretty bad.
6. How easy is it to implement? This theme tries to summarise some of the complexities surrounding implementation e.g., data access and curation needs; system interoperability concerns; questions related to care pathways and workflows; and workforce/skills needs.
7. Will healthcare practitioners adopt it? Big question!! The main point of this theme is that willingness to adopt is conditional and shouldn’t be taken for granted.
8. Will patients and publics accept it? Like the above, this theme covers the conditions that influence attitudes towards the use of AI in healthcare, amongst patients and publics. Of note is the influence of understanding.
9. How is it Governed? Covered here both hard and soft governance i.e., the massive legal complexities (medical device, data protection, anti-discrimination, medical negligence, consent, consumer protection) as well as the need for policies, and standards governing the non-legal considerations related to safety and trustworthiness.
10. What about do no harm? My favourite theme – the ethical implications, broken down by ethical principle (autonomy, beneficence, non-maleficence, justice, and Explainability). Covering issues from bias to psychological harm, to overdiagnosis and the undermining of patient autonomy.”