2.1. Initial referral and diagnosis of cancer

The usual first step along a cancer patient's journey is their contact with primary care. General practitioners (GPs) are responsible for the critical decision as to whether or not a case warrants referral for urgent investigation. Gathering all of the relevant information and making such important judgements within a 10-minute appointment is a challenging task for any clinician.

Any potential use of AI in this process is complicated by the diverse and complex nature of patient data, which are often contained in multiple individual electronic silos within the care record, or even, indeed, on a piece of paper. The Topol Review argues that fully digitizing and integrating these records is necessary for the implementation of AI, but how can this be achieved? The openEHR Foundation works to develop and implement the infrastructure of improved electronic medical records in order to help centralize patient data and improve continuity of care. However, the more we rely on technological solutions to NHS data issues, the more many of us become concerned regarding security. The WannaCry ransomware attack in 2017 cost the NHS £92 million and brought this concern to the fore (Harding, Reference Harding2018). Clearly, developments in data protection must match the degree to which we become reliant on technology.

The Topol Review also proposes that AI will enable us to benefit from improved triage. For example, Babylon Health has already developed a chatbot tool that claims to do just this by asking questions about the patient's symptoms and then forwarding them to advice or a video call. However, many argue that its claims of accuracy and superiority over clinicians are unfounded. One publication concluded that there is limited evidence that Babylon is better than its human counterparts, and the Care Quality Commission's 2017 inspection reported that the company was failing to provide safe care (Fraser et al., Reference Fraser, Coiera and Wong2018).

Thus, to maximize the information-gathering and pattern-recognition benefits of AI, which could present the busy GP or oncologist with a clinically relevant overview of the patient in front of them to guide assessment, there are still huge challenges in data collection, collation and presentation that need to be overcome.

Following a referral, investigations are done to confirm or exclude a cancer diagnosis. These include biopsies, alongside histopathological, biochemical and radiological analyses. A proposed strength of AI would be the collation, assessment and pattern recognition of integrated datasets pertaining to these investigations to speed up and improve the accuracy of cancer diagnosis.

The Royal College of Pathologists is investing in the development of digital pathology (https://www.rcpath.org/discover-pathology/public-affairs/digital-pathology.html) and anticipates the future incorporation of AI into this. Likewise, radiology is also particularly ripe for integration of AI, such as through complex imaging pattern analysis (Hosny et al., Reference Hosny, Parmar, Quackenbush, Schwartz and Aerts2018). The Swiss start-up SOPHiA Genetics has developed a radiomics platform that aims to notice features in radiographic images that are missed by the human eye.

At present, cancer multidisciplinary team meetings contain a pathologist, radiologist and oncologist interpreting and discussing their various diagnostic tests. Proponents of AI suggest that the replacement of three experienced consultants with a single AI system may well be cost effective, as well as improving diagnostic accuracy, but proof of this in clinical practice remains elusive thus far.

2.2. Somatic and constitutional cancer genome analysis

Genomic analysis generates exceptionally large datasets. Ways of streamlining the analysis to allow easier recognition of clinically relevant variation are required, and such technology is already being used in real-world scenarios. Due to the fact that advances in genomic technologies generating wide-scale whole-genome data are a recent event, computational solutions for these analyses have always existed.

AI use in this type of analysis tends to use machine learning and pattern recognition to present input data in a format that lends itself to easier interpretation. The aforementioned SOPHiA Genetics developed a genetics AI that is the company's main platform and that uses knowledge-sharing between hospitals to help train its algorithms. This can locate and highlight mutations in cancer-linked genes within two hours of uploading raw genomic data, and 503 hospitals around the world are already using the technology, including several in the UK under the NHS (Macauley, Reference Macaulay2018).

This optimization of data processing equally serves the interest of patients. The waiting period following genetic testing is associated with anxiety and other distressing psychological symptoms, which many patients find difficult to cope with (Godino et al., Reference Godino, Turchetti and Jackson2015). Reducing the amount of time that this unpleasant process takes could help ease its mental toll.

2.3. Acute cancer management

2.4. Management of future cancer risk

Screening, prevention and early detection for cancer are both clinically preferable and cost effective (Turnbull et al., Reference Turnbull, Sud and Houlston2018). From a public health perspective, the media can have a large influence on public attitudes to hereditary cancer risk perception – so much so that the phenomenon has been termed the ‘Angelina Jolie Effect’, referring to the surge in enquiries and referrals following the actress's decision to undertake BRCA1 testing and a risk-reducing mastectomy (Evans et al., Reference Evans, Barwell and Eccles2014). However, when the phenomenon gives rise to inaccurate risk perception amongst the public, it can lead to patient anxiety and misuse of resources. This was a controversial area that most recently generated mainstream media debate following the personal reflections of the current UK Secretary of State for Health and Social Care, Matt Hancock, on his polygenic risk scores relating to his lifetime risk of prostate cancer.

2.5. A threat to jobs?

The potential threat that AI poses against the utility of human employees is an argument at the forefront of its opposition, and the field of medicine is no exception to this. The Topol Review argues that there will be ‘significant changes to the roles and responsibilities of current NHS staff’ thanks to AI, but what form could these changes take?

Physical contact is needed to perform any sort of medical procedure, and even the most intelligent of digital software is, of course, very obviously not able to do this. However, robots have been propositioned as the hardware that could be controlled by AI to make this strange and intangible concept a reality. Most current healthcare robots, such as those in laparoscopic surgery, must be controlled by humans, and others have only a low level of AI influencing their physical actions. Nonetheless, the Topol Review anticipates that robots will be developed to exploit more advanced AI algorithms, enabling them to perform both manual and cognitive tasks.

However, the general public appears less receptive than Topol to the idea of a robotic coup of medical procedures, with only 23% of UK adults reporting that they are comfortable with the idea of intelligent machines performing medical operations (Timson, Reference Timson2017). The importance of ensuring that the direction AI development takes aligns with the desires of the patients and public cannot be understated, meaning that AI-using robots should be developed with the end goal of supplementing, rather than replacing, the work of human professionals. Topol does also recommend that more employment opportunities should be generated to help develop these technologies, meaning that AI's growing role could actually generate more jobs in the NHS rather than threaten them.

2.6. Clinical governance

This review highlights the promising scope of innovation and enterprise in AI as a tool to improve the field of cancer genomics. Equally, however, issues and limitations that reach beyond the technology itself continue to persist. These issues must be addressed if AI is to be successfully integrated into the patient pathway.

The major increase in the use of genetic testing over the past decade has highlighted the inevitable fact that medical professionals are not infallible. In a review of one US legal database, more than 50 cases were identified in which healthcare providers were sued for failing to appropriately utilize genetic testing (Lindor & Marchant, Reference Lindor and Marchant2011).

If a patient were to seek legal action following such an incident caused by an artificially intelligent machine, then who would be deemed responsible in this instance? Would it be the doctor who referred them, the software developer or the AI itself? It is often the case that the inputs and outputs of an algorithm can be seen, but the way in which the output was decided cannot. This is referred to as the ‘black box phenomenon’, and it further complicates these questions.

Indeed, a solution has not yet been developed, but there are multiple avenues that the issue of liability could take in such an instance. There is debate as to what extent an AI platform could be allocated the status of legal personhood, similar to how corporations can be allocated rights associated with human persons. Further, the people who manage the AI could be held responsible through vicarious liability, in the same way in which an employer is responsible for their employees’ actions.

Platforms under the spectrum of eHealth in the UK are included in the classification system of the Medicines and Healthcare Products Regulatory Agency. However, the growth of AI is already beginning to expose weaknesses in their framework. For example, the Babylon chatbot is counted as a Class I medical device, placing it in the same category as a wheelchair or stethoscope. This means that it does not need to pass a conformity assessment and is under minimal regulation.