Debate & Analysis

Artificial intelligence in medicine:

current trends and future possibilities

Artificial intelligence (AI) research within

medicine is growing rapidly. In 2016,
healthcare AI projects attracted more
“… these systems are able to learn from each
investment than AI projects within any other incremental case and can be exposed, within minutes,
sector of the global economy.1 However,
among the excitement, there is equal
to more cases than a clinician could see in many
scepticism, with some urging caution at lifetimes.”
inflated expectations.2 This article takes
a close look at current trends in medical
AI and the future possibilities for general
practice. efficacy, new technologies entering leave the primary responsibility of patient
the medical field must also integrate management with a human doctor. There is
WHAT IS MEDICAL ARTIFICIAL with current practices, gain appropriate an ongoing clinical trial using AI to calculate
INTELLIGENCE? regulatory approval, and, perhaps most target zones for head and neck radiotherapy
Informing clinical decision making through importantly, inspire medical staff and more accurately and far more quickly
insights from past data is the essence patients to invest in a new paradigm. than a human being. An interventional
of evidence-based medicine. Traditionally, These challenges have led to a number radiologist is still ultimately responsible
statistical methods have approached this of emerging trends in AI research and for delivering the therapy but AI has a
task by characterising patterns within adoption. significant background role in protecting
data as mathematical equations, for the patient from harmful radiation.7
example, linear regression suggests a ‘line AI excels at well-defined tasks
of best fit’. Through ‘machine learning’ Research has focused on tasks where AI supports poorly resourced services
(ML), AI provides techniques that uncover AI is able to effectively demonstrate its A single AI system is able to support a large
complex associations which cannot easily performance in relation to a human doctor. population and therefore it is ideally suited
be reduced to an equation. For example, Generally, these tasks have clearly defined to situations where human expertise is
neural networks represent data through inputs and a binary output that is easily a scarce resource. In many TB-prevalent
vast numbers of interconnected neurones validated. In classifying suspicious skin countries there is a lack of radiological
in a similar fashion to the human brain. lesions, the input is a digital photograph expertise at remote centres.8 Using AI,
This allows ML systems to approach and the output is a simple binary radiographs uploaded from these centres
complex problem solving just as a clinician classification: benign or malignant. Under could be interpreted by a single central
might — by carefully weighing evidence these conditions, researchers simply had to system; a recent study shows that AI
to reach reasoned conclusions. However, demonstrate that AI had superior sensitivity correctly diagnoses pulmonary TB with a
unlike a single clinician, these systems and specificity than dermatologists when sensitivity of 95% and specificity of 100%.5
can simultaneously observe and rapidly classifying previously unseen photographs Furthermore, under-resourced tasks where
process an almost limitless number of biopsy-validated lesions.4 patients are experiencing unsatisfactory
of inputs. For example, an AI-driven waiting times are also attractive to AI in the
smartphone app now capably handles the AI is supporting doctors, not replacing form of triage systems.3
task of triaging 1.2 million people in North them
London to Accident & Emergency (A&E).3 Machines lack human qualities such as AI is a very picky eater
Furthermore, these systems are able to empathy and compassion, and therefore Developing ML models requires well-
learn from each incremental case and can patients must perceive that consultations structured training data about a
be exposed, within minutes, to more cases are being led by human doctors. phenomenon that remains relatively stable
than a clinician could see in many lifetimes. Furthermore, patients cannot be expected over time. A departure from this results
This is why an AI-driven application is to immediately trust AI; a technology in ‘over-fitting’, where AI gives undue
able to out-perform dermatologists at shrouded by mistrust.6 Therefore, AI importance to spurious correlations within
correctly classifying suspicious skin commonly handles tasks that are essential, past data. In 2008, Google tried to predict the
lesions4 or why AI is being trusted with but limited enough in their scope so as to seasonal prevalence of influenza using only
tasks where experts often disagree, such
as identifying pulmonary tuberculosis on
chest radiographs.5 Although AI is a broad
field, this article focuses exclusively on “... AI commonly handles tasks that are essential,
ML techniques because of their ubiquitous
usage in important clinical applications. but limited enough in their scope so as to leave the
primary responsibility of patient management with a
WHAT ARE THE CURRENT TRENDS IN
MEDICAL AI? human doctor.”
Aside from simply demonstrating superior

British Journal of General Practice, March 2018 143

 ADDRESS FOR CORRESPONDENCE
“Integrating these systems into clinical practice Mahiben Maruthappu

necessitates building a mutually beneficial Cera Care, 219 Kensington High Street, Kensington,
London W8 6BD, UK.
relationship between AI and clinicians …” E-mail: [email protected]

the search terms entered into its search appointment of every diabetic patient in a
engine. Because people’s searching habits practice in real time.
change dramatically with every passing AI-based systems will also bring specialist REFERENCES
year, the model was so poorly predictive of diagnostic expertise into primary care. If 1. CB Insights Research. Healthcare remains
the future that it was quickly discontinued.9 an image of a skin lesion is sufficient to the hottest AI category for deals. 2017.
https://www.cbinsights.com/research/
Additionally, data that are anonymised and capably diagnose its aetiology, images could
artificial-intelligence-healthcare-startups-
digitised at source are also preferable, as be captured at a GP practice and sent to a investors/ (accessed 15 Jan 2018).
this aids in research and development. specialist dermatology AI system for instant 2. Chen JH, Asch SM. Machine learning and
analysis. Patients identified as low risk would prediction in medicine — beyond the peak
FUTURE POSSIBILITIES IN GENERAL receive instant reassurance while high-risk of inflated expectations. N Eng J Med 2017;
PRACTICE 376(26): 2507–2509.
patients would experience lower referral
AI will extract important information from waiting times because clinics would only be 3. Burgess M. The NHS is trialling an AI chatbot
to answer your medical questions. Wired 2017:
a patient’s electronic footprint. At first receiving selected cases. This concept is not 5 Jan: http://www.wired.co.uk/article/babylon-
this will save time and improve efficiency, limited to skin lesions, AI has shown potential nhs-chatbot-app (accessed 15 Jan 2018).
but following adequate testing it will also in interpreting many different types of image 4. Esteva A, Kuprel B, Novoa RA, et al.
directly guide patient management. Take data including retinal scans,10 radiographs,5 Dermatologist-level classification of skin
the example of a consultation with a patient and ultrasound.11 Many of these images can cancer with deep neural networks. Nature
2017; 542(7639): 115–118.
with type 2 diabetes; currently a clinician be captured with relatively inexpensive and
5. Lakhani P, Sundaram B. Deep learning at
spends significant time reading outpatient widely available equipment. chest radiography: automated classification
letters, checking blood tests, and finding Future AI research should be directed of pulmonary tuberculosis by using
clinical guidelines from a number of towards carefully selected tasks that convolutional neural networks. Radiology
disconnected systems. In contrast, AI could broadly align with the trends outlined in 2017; 284(2): 574–582.
automatically prepare the most important this article. Integrating these systems into 6. Oppenheim M. Stephen Hawking: artificial
risks and actions given the patient’s clinical intelligence could be the greatest disaster
clinical practice necessitates building a
in human history. Independent 2016; 20 Oct:
record. It could also automatically convert mutually beneficial relationship between http://www.independent.co.uk/news/people/
recorded dialogue of the consultation into a AI and clinicians, where AI offers clinicians stephen-hawking-artificial-intelligence-
summary letter for the clinician to approve greater efficiency or cost-effectiveness diaster-human-history-leverhulme-centre-
or amend. Both of these applications cambridge-a7371106.html (accessed 15 Jan
and clinicians offer AI the essential clinical
2017).
would save considerable time and could exposure it needs to learn complex clinical
7. Chu C, De Fauw J, Tomasev N, et al. Applying
be implemented very quickly because they case management. Throughout the process machine learning to automated segmentation
assist clinicians rather than replacing them. it will be critical to ensure that AI does of head and neck tumour volumes and
As these systems become better not obscure the human face of medicine organs at risk on radiotherapy planning CT
validated, they will be given more because the biggest impediment to AI’s and MRI scans. F1000Research 2016; 5: 2104.
responsibility. For the patient with widespread adoption will be the public’s 8. Hoog AH, Meme HK, van Deutekom H, et al.
High sensitivity of chest radiograph reading by
type 2 diabetes, the threshold of statin hesitation to embrace an increasingly clinical officers in a tuberculosis prevalence
commencement could be determined by AI controversial technology.12 survey. Int J Tuberc Lung Dis 2011; 15(10):
on an individualised basis given nuisances 1308–1314.
of the patient’s history rather than a rigidly Varun H Buch, 9. Lazer D, Kennedy R, King G, Vespignani A.
defined ‘one-size-fits-all’ algorithm. The Clinical Technology Lead, Cera Care, London. The parable of Google flu: traps in big data
analysis. Science 2014; 343(6176): 1203–1205.
research required for this ‘personalised’
medicine would only be possible through Irfan Ahmed, 10. Shead S. Google DeepMind is funding
GP Lead, Cera Care, London. NHS research at Moorfields Eye Hospital.
AI intelligently summarising enormous Business Insider 2017; 3 Aug: http://
quantities of medical information. uk.businessinsider.com/deepmind-is-
Furthermore, because AI is able to Mahiben Maruthappu, funding-nhs-research-2017-7 (accessed 15
simultaneously monitor millions of inputs, CEO, Cera Care, London. Jan 2018).
it will have a significant role in preventative 11. Chen H, Wu L, Dou Q, et al. Ultrasound
standard plane detection using a composite
medicine. AI could proactively suggest Provenance
neural network framework. IEEE Trans
consultations when it determines that the Freely submitted; externally peer reviewed.
Cybern 2017; 47(6): 1576–1586.
patient’s risk of developing a particular 12. Naughton J. Giving Google our private NHS data
diabetic complication warrants intervention. Competing interests is simply illegal. Guardian 2017; 9 Jul: https://
In contrast, it would be impractical to task The authors have declared no competing interests. www.theguardian.com/commentisfree/2017/
jul/09/giving-google-private-nhs-data-is-simply-
a human being with the responsibility of
illegal (accessed 15 Jan 2018).
closely monitoring every test result and DOI: https://doi.org/10.3399/bjgp18X695213

144 British Journal of General Practice, March 2018