By now, it’s almost old news that artificial intelligence (AI) will have a transformative role in medicine. Algorithms have the potential to work tirelessly, at faster rates and now with potentially greater accuracy than clinicians.

In 2016, it was predicted that ‘machine learning will displace much of the work of radiologists and anatomical pathologists’. In the same year, a University of Toronto professor controversially announced that ‘we should stop training radiologists now’. But is it really the beginning of the end for some medical specialties?

AI excels in pattern identification in determining ’pathologies that look certain ways,’ according to Elliot Fishman, a radiology and oncology professor at Johns Hopkins University and a key proponent of AI integration into medicine. Ultimately, specialties that rely heavily on visual pattern recognition – notably radiology, pathology, and dermatology – are those believed to be at the greatest risk. With the advent of virtual primary care services, such as Babylon, General Practice may also have to adapt in the future.

Pattern recognition functions

In January of this year, an article in Nature reported that AI systems ‘outperformed’ doctors in breast cancer detection. This was carried out by an international team, including researchers from Google Health and Imperial College London on mammograms obtained from almost 29,000 women. Screening mammography currently plays a critical role in early breast cancer detection, ensuring early initiation of treatment and yielding improved patient prognoses. False negatives are a significant problem in mammography. The study found AI use was associated with an absolute reduction of 9.4% and 2.7% reduction in false negatives, in the USA and UK, respectively. Similarly, use of the AI system led to a reduction of 5.7% and 1.2% in the USA and UK respectively for false positives. The study suggested that AI outperformed the six radiologists individually, and was equivalent to the current double-reading system of two doctors currently used in the UK. These developments have already had perceptible consequences in practice: algorithms eliminate the need for a second radiologist when interpreting mammograms. However, critically, one radiologist remains responsible for the diagnosis.

“AI can also be deployed to predict the cognitive decline that leads to Alzheimer’s disease... allowing early intervention and treatment”

Earlier studies have also yielded similar results: a 2017 study published in Nature examined the use of algorithms in dermatology. The study, from Stanford University, involved an algorithm developed by computer scientists using an initial database of 130,000 skin disease images. When compared to the success rates of 21 dermatologists, the algorithm was almost equally successful. Likewise, in a study conducted by the European Society for Medical Oncology, it was found that AI exceeded the performance of 58 international dermatologists. A system reliant on a form of machine learning known as Deep Learning Convolutional Neural Network (CNN) missed fewer melanomas (the most lethal form of skin cancer), and misdiagnosed benign moles (or nevi) as malignant less often than the group of dermatologists.

Further applications in medicine

However, the prospects of AI technology extend beyond the clear applications in cancer diagnosis and radiology: recent studies have also demonstrated that AI may be able to detect genetic diseases in infants by rapid whole-genome sequencing and interpretation. Considering that time is critical in treating gravely ill children, such automated techniques can be crucial in diagnosing children who are suspected of having genetic diseases.

In addition, AI can also be deployed to predict the cognitive decline that leads to Alzheimer’s disease. Such computational models can be highly valuable at the individual level, allowing ‘early intervention and treatment planning’. FDA approval has also been granted to a number of companies for such technologies; these include Imagen’s OsteoDetect, an algorithm intended to aid wrist fracture detection. In addition, algorithms may have functions in other specialties such as anaesthesiology in monitoring and responding to physiological signs.

Limitations of AI

Despite the benefits that AI integration into clinical practice can provide, the technology is not without limitations. Machine learning algorithms are highly dependent on the quality and quantity of the data input, typically requiring millions of observations to function at suitable levels. Biases in data collection can heavily impact performance; for instance, racial or gender representation in the original data set can lead to differences in diagnostic abilities of the system for different groups, consequently leading to disparities in patient outcomes. Considering that certain pathologies, including melanoma, present differently between races and with different incidences, this can often lead to both later diagnoses and poorer outcomes for racial minorities, as found in a number of studies. ‘Volunteer bias’ of the data collected is also a pertinent consideration; for example, although lactate concentration is a good predictor of death, this is not routinely measured in healthy individuals.

“Considering the magnitude of what is at stake raises the question of whether it is appropriate to rely solely on machines without any human input.”

Other key problems which may arise include how algorithms ‘overfit’ predictions based on random errors in the data, resulting in unstable estimates which vary between data samples. In addition, clinicians may take a more cautious approach when making a diagnosis. Therefore, it may appear that a human underperforms compared to an algorithm since their actions may yield a lower accuracy in tumour identification, however this approach could lead to a lower number of critical cases missed.

Ultimately, the tendency for humans to favour propositions given by automated systems over non-automated ones, known as ’automation bias’, may exacerbate these problems.

Attempts to replace GPs with AI have been unsuccessful

The success of AI integration into clinical practice crucially depends on the receptiveness of patients. Babylon, a start-up company based in the UK, was developed to give medical advice to patients using chat services. Although Babylon has been referred to as ‘the biggest disruption in medical practice in years’ and a ‘game-changer’ in UK media – as quoted on Babylon’s website – it is questionable how successful the service has been so far Babylon has been slow in recruiting patients and this month, it came under fire for data breaches. The fact that patients lose access to their regular GP if they sign up to Babylon is perhaps a key contributing factor for Babylon’s slow take-off. Therefore, it appears that human contact is highly valued by patients, after all, at least for some medical specialties.

Potential effect of COVID-19

The COVID-19 pandemic, with its requirements for social distancing, could potentially accelerate the use of AI. COVID-related restrictions could change the perception of patients about remote medical consultations, paving the way for increased receptiveness to primary healthcare apps including Babylon. The pandemic has also highlighted the inadequacies in fast internet access throughout the country. This may encourage increased government investment into broadband infrastructure, which may, in turn, facilitate broader penetration of AI technology. The increased pressure on the NHS may also encourage greater use of algorithms to delegate menial tasks as seen in specialties such as radiology already.

The future

AI will likely become an indispensable tool in clinical medicine, facilitating the work of professionals by automating mundane, albeit essential tasks. By reducing the medical workload, this could allow healthcare professionals to dedicate greater efforts to other aspects of their work, including patient interaction. As emphasised by the President of the Royal College of Radiologists, radiologists can instead focus more of their time on interventional radiology and in managing more complex cases to a much greater extent. Indeed, innovation may aid clinicians and augment their decision-making capabilities to improve their efficiency and diagnostic accuracy, however it remains doubtful whether technology can fully replace these roles. After all, considering the magnitude of what is at stake – human life – raises the question of whether it is appropriate to rely solely on machines without any human input. Therefore, it remains likely that human involvement will need to continue across medical specialties, although this may be in a reduced or adapted form.