Whether for its application in facial recognition or driverless cars, artificial intelligence (AI) is making headlines. The field of medicine is no exception. In fact, medical sciences are estimated to receive more investment in AI than any other field. There is a good reason for this. AI has the potential to improve nearly all areas of medical practice on an unprecedented scale.

Broadly speaking, artificial intelligence is the ability of computers to carry out tasks that typically require human intelligence. However, while AI systems may aim to mimic human intelligence, replication is not necessarily the final goal. In many cases outperforming humans is well within its capability. The growth of AI has been facilitated by rapid advances in computing power and storage which allow more data than ever before to be collected and analysed. 

Particularly relevant to medicine, machine learning is a subclass of AI characterised by its ability to learn from previous experience of large amounts of data to, for example, discern patterns or make predictions. This sets it aside from traditional algorithms designed to make decisions by following rules explicitly stated to them by human designers. For example, a simple flowchart may be used to determine whether a patient should be discharged from hospital. In contrast, from a large structured data set, machine learning software could determine which parameters are important for deciding whether a patient should be discharged or not, and how to weigh such parameters to make an appropriate decision. Outside hospitals, this could manifest as a healthcare app recommending whether someone should schedule a GP appointment or not.

Go one step further, however, and you reach deep learning. A subsection of machine learning, deep learning works on unstructured data using layers of algorithms, each one building on the statistical model devised by the last. Much like the human brain, it passes data through a hierarchy of transformations to determine its final output. Automating this allows processing of vast quantities of complex health data, potentially in irregular forms such as patient notes, that a single individual could not put together on their own. From this, clinicians hope that previously invisible correlations can be identified and used to inform medical practice.

Diagnostics is an area in which AI has already shown significant promise. It is particularly suited to image analysis. With an ability to analyse images at the level of single pixels at rapid speed, AI programs such as LYNA, Google’s LYmph Node Assistant which analyses images taken from patients suspected of having metastatic breast cancer, can already diagnose with greater accuracy than clinicians. LYNA derived its approach by analysing a training set of images of lymph nodes labelled by clinicians depending on their characteristics, in a process called supervised learning.

Nevertheless, combining LYNA with clinical opinion halved the rates of missed small lymph node metastases and outperformed both clinicians and LYNA in isolation. However, it is less certain how these comparisons hold up when doctors are given access to patient-specific data along with images, a circumstance that better reflects the realities of clinical practice.

Further applications in imaging include improving treatment of snake bites which are estimated by WHO to cause between 81,000 and 138,000 deaths annually. Correctly identifying the snake responsible is crucial in determining which antivenom should be administered to victims, yet in many cases clinicians lack sufficient expertise or training in snake identification. AI-assisted identification algorithms could help, particularly as cheap smartphone cameras are becoming more ubiquitous. Anyone present nearby at the time could take a photo of the snake responsible and send its image for analysis.

Overall, AI algorithms have the potential to slash costs dramatically for high quality medical advice in areas cut off geographically, financially and socially from such care. This could narrow health disparities across the globe. Whilst it should aim to be, AI does not always need to be better than the global gold standard for care, but rather better than local alternatives which may vary vastly from region to region. This should not, however, negate the long-term responsibility of governments and healthcare providers to bring gold standard care to these regions. Rather, it is a recognition that the realities of healthcare available can inform the use of AI.

AI algorithms have the potential to slash costs dramatically

Another reality of clinical practice is the fact that much of the data generated is not formatted appropriately for algorithms to analyse. Natural language processing (NLP) can be used to get around this. NLP is a type of AI which can be used to transform unstructured language into structured data for machine learning to work upon. This allows AI to gain inferences from a wide range of data such as consultations notes and written questionnaires. Babylon, a UK start-up, is using AI to develop chat services which can be used in primary care to provide quick advice and direct patients to appropriate services. With artificial speech recognition rapidly advancing, it is not a far-fetched idea for software to listen to conversations in tandem with clinicians and give real time advice regarding therapeutics, diagnoses and follow-up questions. Combining increasing and more complex aspects patient’s history could help uncover patterns and relationships undetectable to the human eye and an individual’s clinical expertise. 

Looking forward again, AI may also give indicators of prognosis. This could allow better identification and monitoring of high-risk patients, for example those more likely to develop sepsis. On the other hand, it could also reduce inappropriate, expensive and potentially harmful precautionary interventions for patients at a lower risk. With the advent of wearable medical devices recording everything from heart rate to blood glucose levels, data collected by users at home could form an increasing part of their health record and inform healthcare decisions in the future. This stands in contrast to expensive and potentially biased data which is normally only collected upon visits to healthcare services.

Accurate AI-assisted decision making can also prevent mistakes in medical practice, such as the 237m prescribing errors estimated to occur in the NHS annually. According to a recent study conducted at John Hopkins University, medical errors are the third leading cause of death in the US, though such errors are still poorly reported. Whilst behind a minority of mistakes lies medical negligence, far more are the result of human failure in applying optimal care repeatedly in every circumstance under pressures including sleep deprivation and time. Predicting high-risk situations can facilitate better preparation for them. AI will not eliminate medical errors entirely, but rather can reduce their likelihood by offering support in dealing with increasingly complex patient data and a set of medical knowledge turning over at a higher rate than ever before. Increasing acknowledgement that mistakes are widespread in medical practice should make it easier to embrace AI-linked solutions to reduce them. Yet AI itself is also not infallible and it is uncertain how to manage accountability for mistakes made by algorithms.

Reducing medical workload could also allow healthcare professionals to spend more time focusing on personal interactions with patients and aspects of medicine which AI is not equipped to deal with. While radiologists may carry out less routine image scrutiny in the future, their role in interventional radiology and in managing complex cases, as highlighted by the President of the Royal College of Radiologists, makes the profession far from obsolete. Despite the potential for a more human brand of medicine with the advent of AI used in healthcare, it is likely that at least some of the productivity gains associate will be offset against increasing healthcare demands not met by increases in NHS funding and staffing. 

Moving outside clinical practice, another area in which AI could bring dividends is drug discovery. Analysis of patient data could facilitate disease modelling, for example by elucidating key genetic factors and pathways previously not thought to be associated with a condition. Algorithms can also be used to identify drug targets in the body and subsequently to engineer molecules that can appropriately interact with such targets whilst filtering out less promising candidates. With recent advances in language processing, AI can also read through scientific papers to identify relationships useful for drug discovery. Given a recent estimate that each approved prescription drug costs upwards of £2bn ($2.6bn) to bring to market, AI could bring huge rewards by improving the success rate of drugs taken into expensive clinical trials. In addition, even if a drug passed all such trials with flying colours, this does not necessarily mean that it could not have been improved further. This could also encourage the development of treatments for conditions previously riskier or less financially for the pharmaceutical industry. These could include central nervous system conditions for which treatments have some of the highest failure rates in clinical trials.

For the foreseeable future AI looks set to operate alongside those working in healthcare to improve decision making, reduce error rates and integrate vast quantities of data in a meaningful way. Such technology has the potential to improve the relationships that healthcare professionals have with their patients while giving them to tools to offer personalised approaches to their care. AI can also enhance efficiency and improve worldwide access to specialist care. Despite the challenges associated with fair, safe and effective use of such technology, its potential benefits are too great to ignore.