94 © Royal College of Physicians 2019. All rights reserved.

FUTURE Future Healthcare Journal 2019 Vol 6, No 2: 94–8

DIGITAL TECHNOLOGY The potential for artificial intelligence in healthcare

Authors: Thomas Davenport A and Ravi Kalakota B

The complexity and rise of data in healthcare means that artificial intelligence (AI) will increasingly be applied within the field. Several types of AI are already being employed by payers and providers of care, and life sciences companies. The key categories of applications involve diagnosis and treatment recommendations, patient engagement and adherence, and administrative activities. Although there are many instances in which AI can perform healthcare tasks as well or better than humans, implementation factors will prevent large-scale automation of healthcare professional jobs for a considerable period. Ethical issues in the application of AI to healthcare are also discussed.

KEYWORDS : Artifi cial intelligence , clinical decision support ,

electronic health record systems

Introduction

Artificial intelligence (AI) and related technologies are increasingly

prevalent in business and society, and are beginning to be applied

to healthcare. These technologies have the potential to transform

many aspects of patient care, as well as administrative processes

within provider, payer and pharmaceutical organisations.

There are already a number of research studies suggesting that

AI can perform as well as or better than humans at key healthcare

tasks, such as diagnosing disease. Today, algorithms are already

outperforming radiologists at spotting malignant tumours, and

guiding researchers in how to construct cohorts for costly clinical

trials. However, for a variety of reasons, we believe that it will be

many years before AI replaces humans for broad medical process

domains. In this article, we describe both the potential that AI

offers to automate aspects of care and some of the barriers to

rapid implementation of AI in healthcare.

Types of AI of relevance to healthcare

Artificial intelligence is not one technology, but rather a collection

of them. Most of these technologies have immediate relevance

to the healthcare field, but the specific processes and tasks they

A B

ST R

A C

T

support vary widely. Some particular AI technologies of high

importance to healthcare are defined and described below.

Machine learning – neural networks and deep learning

Machine learning is a statistical technique for fitting models

to data and to ‘learn’ by training models with data. Machine

learning is one of the most common forms of AI; in a 2018

Deloitte survey of 1,100 US managers whose organisations

were already pursuing AI, 63% of companies surveyed were

employing machine learning in their businesses. 1 It is a broad

technique at the core of many approaches to AI and there are

many versions of it.

In healthcare, the most common application of traditional

machine learning is precision medicine – predicting what

treatment protocols are likely to succeed on a patient based on

various patient attributes and the treatment context. 2 The great

majority of machine learning and precision medicine applications

require a training dataset for which the outcome variable (eg onset

of disease) is known; this is called supervised learning.

A more complex form of machine learning is the neural

network – a technology that has been available since the 1960s

has been well established in healthcare research for several

decades 3 and has been used for categorisation applications like

determining whether a patient will acquire a particular disease.

It views problems in terms of inputs, outputs and weights of

variables or ‘features’ that associate inputs with outputs. It has

been likened to the way that neurons process signals, but the

analogy to the brain's function is relatively weak.

The most complex forms of machine learning involve deep

learning , or neural network models with many levels of features

or variables that predict outcomes. There may be thousands

of hidden features in such models, which are uncovered by the

faster processing of today's graphics processing units and cloud

architectures. A common application of deep learning in healthcare

is recognition of potentially cancerous lesions in radiology images. 4

Deep learning is increasingly being applied to radiomics, or the

detection of clinically relevant features in imaging data beyond

what can be perceived by the human eye. 5 Both radiomics and deep

learning are most commonly found in oncology-oriented image

analysis. Their combination appears to promise greater accuracy

in diagnosis than the previous generation of automated tools for

image analysis, known as computer-aided detection or CAD.

Deep learning is also increasingly used for speech recognition

and, as such, is a form of natural language processing (NLP),

Authors: A president's distinguished professor of information

technology and management, Babson College, Wellesley, USA ;

B managing director, Deloitte Consulting, New York, USA

FHJv6n2-Davenport.indd 94FHJv6n2-Davenport.indd 94 6/3/19 8:49 AM6/3/19 8:49 AM

© Royal College of Physicians 2019. All rights reserved. 95

Artificial intelligence in healthcare

described below. Unlike earlier forms of statistical analysis, each

feature in a deep learning model typically has little meaning to

a human observer. As a result, the explanation of the model's

outcomes may be very difficult or impossible to interpret.

Natural language processing

Making sense of human language has been a goal of AI

researchers since the 1950s. This field, NLP, includes applications

such as speech recognition, text analysis, translation and other

goals related to language. There are two basic approaches to it:

statistical and semantic NLP. Statistical NLP is based on machine

learning (deep learning neural networks in particular) and has

contributed to a recent increase in accuracy of recognition. It

requires a large ‘corpus’ or body of language from which to learn.

In healthcare, the dominant applications of NLP involve

the creation, understanding and classification of clinical

documentation and published research. NLP systems can analyse

unstructured clinical notes on patients, prepare reports (eg on

radiology examinations), transcribe patient interactions and

conduct conversational AI.

Rule-based expert systems

Expert systems based on collections of ‘if-then’ rules were the

dominant technology for AI in the 1980s and were widely used

commercially in that and later periods. In healthcare, they were

widely employed for ‘clinical decision support’ purposes over

the last couple of decades 5 and are still in wide use today. Many

electronic health record (EHR) providers furnish a set of rules with

their systems today.

Expert systems require human experts and knowledge engineers

to construct a series of rules in a particular knowledge domain.

They work well up to a point and are easy to understand. However,

when the number of rules is large (usually over several thousand)

and the rules begin to conflict with each other, they tend to break

down. Moreover, if the knowledge domain changes, changing the

rules can be difficult and time-consuming. They are slowly being

replaced in healthcare by more approaches based on data and

machine learning algorithms.

Physical robots

Physical robots are well known by this point, given that more than

200,000 industrial robots are installed each year around the

world. They perform pre-defined tasks like lifting, repositioning,

welding or assembling objects in places like factories and

warehouses, and delivering supplies in hospitals. More recently,

robots have become more collaborative with humans and are

more easily trained by moving them through a desired task.

They are also becoming more intelligent, as other AI capabilities

are being embedded in their ‘brains’ (really their operating

systems). Over time, it seems likely that the same improvements

in intelligence that we've seen in other areas of AI would be

incorporated into physical robots.

Surgical robots, initially approved in the USA in 2000, provide

‘superpowers’ to surgeons, improving their ability to see, create

precise and minimally invasive incisions, stitch wounds and so

forth. 6 Important decisions are still made by human surgeons,

however. Common surgical procedures using robotic surgery include

gynaecologic surgery, prostate surgery and head and neck surgery.

Robotic process automation

This technology performs structured digital tasks for

administrative purposes, ie those involving information systems,

as if they were a human user following a script or rules. Compared

to other forms of AI they are inexpensive, easy to program and

transparent in their actions. Robotic process automation (RPA)

doesn't really involve robots – only computer programs on

servers. It relies on a combination of workflow, business rules and

‘presentation layer’ integration with information systems to act

like a semi-intelligent user of the systems. In healthcare, they are

used for repetitive tasks like prior authorisation, updating patient

records or billing. When combined with other technologies like

image recognition, they can be used to extract data from, for

example, faxed images in order to input it into transactional

systems. 7

We've described these technologies as individual ones, but

increasingly they are being combined and integrated; robots are

getting AI-based ‘brains’, image recognition is being integrated

with RPA. Perhaps in the future these technologies will be so

intermingled that composite solutions will be more likely or

feasible.

Diagnosis and treatment applications

Diagnosis and treatment of disease has been a focus of AI since

at least the 1970s, when MYCIN was developed at Stanford for

diagnosing blood-borne bacterial infections. 8 This and other early

rule-based systems showed promise for accurately diagnosing and

treating disease, but were not adopted for clinical practice. They

were not substantially better than human diagnosticians, and they

were poorly integrated with clinician workflows and medical record

systems.

More recently, IBM's Watson has received considerable attention

in the media for its focus on precision medicine, particularly cancer

diagnosis and treatment. Watson employs a combination of

machine learning and NLP capabilities. However, early enthusiasm

for this application of the technology has faded as customers

realised the difficulty of teaching Watson how to address

particular types of cancer 9 and of integrating Watson into care

processes and systems. 10 Watson is not a single product but a set

of ‘cognitive services’ provided through application programming

interfaces (APIs), including speech and language, vision, and

machine learning-based data-analysis programs. Most observers

feel that the Watson APIs are technically capable, but taking on

cancer treatment was an overly ambitious objective. Watson and

other proprietary programs have also suffered from competition

with free ‘open source’ programs provided by some vendors, such

as Google's TensorFlow.

Implementation issues with AI bedevil many healthcare

organisations. Although rule-based systems incorporated within

EHR systems are widely used, including at the NHS, 11 they lack the

precision of more algorithmic systems based on machine learning.

These rule-based clinical decision support systems are difficult to

maintain as medical knowledge changes and are often not able to

handle the explosion of data and knowledge based on genomic,

proteomic, metabolic and other ‘omic-based’ approaches to care.

This situation is beginning to change, but it is mostly present

in research labs and in tech firms, rather than in clinical practice.

Scarcely a week goes by without a research lab claiming that it

has developed an approach to using AI or big data to diagnose

FHJv6n2-Davenport.indd 95FHJv6n2-Davenport.indd 95 6/3/19 8:49 AM6/3/19 8:49 AM

96 © Royal College of Physicians 2019 . All rights reserved.

Thomas Davenport and Ravi Kalakota

a patient does not follow a course of treatment or take the

prescribed drugs as recommended – is a major problem.

In a survey of more than 300 clinical leaders and healthcare

executives, more than 70% of the respondents reported having

less than 50% of their patients highly engaged and 42% of

respondents said less than 25% of their patients were highly

engaged. 21

If deeper involvement by patients results in better health

outcomes, can AI-based capabilities be effective in personalising

and contextualising care? There is growing emphasis on using

machine learning and business rules engines to drive nuanced

interventions along the care continuum. 22 Messaging alerts and

relevant, targeted content that provoke actions at moments that

matter is a promising field in research.

Another growing focus in healthcare is on effectively designing

the ‘choice architecture’ to nudge patient behaviour in a

more anticipatory way based on real-world evidence. Through

information provided by provider EHR systems, biosensors,

watches, smartphones, conversational interfaces and other

instrumentation, software can tailor recommendations by

comparing patient data to other effective treatment pathways

for similar cohorts. The recommendations can be provided to

providers, patients, nurses, call-centre agents or care delivery

coordinators.

Administrative applications

There are also a great many administrative applications

in healthcare. The use of AI is somewhat less potentially

revolutionary in this domain as compared to patient care, but

it can provide substantial efficiencies. These are needed in

healthcare because, for example, the average US nurse spends

25% of work time on regulatory and administrative activities. 23

The technology that is most likely to be relevant to this objective

is RPA. It can be used for a variety of applications in healthcare,

including claims processing, clinical documentation, revenue cycle

management and medical records management. 24

Some healthcare organisations have also experimented with

chatbots for patient interaction, mental health and wellness, and

telehealth. These NLP-based applications may be useful for simple

transactions like refilling prescriptions or making appointments.

However, in a survey of 500 US users of the top five chatbots

used in healthcare, patients expressed concern about revealing

confidential information, discussing complex health conditions

and poor usability. 25

Another AI technology with relevance to claims and payment

administration is machine learning, which can be used for

probabilistic matching of data across different databases. Insurers

have a duty to verify whether the millions of claims are correct.

Reliably identifying, analysing and correcting coding issues

and incorrect claims saves all stakeholders – health insurers,

governments and providers alike – a great deal of time, money

and effort. Incorrect claims that slip through the cracks constitute

significant financial potential waiting to be unlocked through data-

matching and claims audits.

Implications for the healthcare workforce

There has been considerable attention to the concern that AI

will lead to automation of jobs and substantial displacement of

the workforce. A Deloitte collaboration with the Oxford Martin

and treat a disease with equal or greater accuracy than human

clinicians. Many of these findings are based on radiological image

analysis, 12 though some involve other types of images such as

retinal scanning 13 or genomic-based precision medicine. 14 Since

these types of findings are based on statistically-based machine

learning models, they are ushering in an era of evidence- and

probability-based medicine, which is generally regarded as positive

but brings with it many challenges in medical ethics and patient/

clinician relationships. 15

Tech firms and startups are also working assiduously on the

same issues. Google, for example, is collaborating with health

delivery networks to build prediction models from big data to warn

clinicians of high-risk conditions, such as sepsis and heart failure. 16

Google, Enlitic and a variety of other startups are developing

AI-derived image interpretation algorithms. Jvion offers a ‘clinical

success machine’ that identifies the patients most at risk as well as

those most likely to respond to treatment protocols. Each of these

could provide decision support to clinicians seeking to find the best

diagnosis and treatment for patients.

There are also several firms that focus specifically on diagnosis

and treatment recommendations for certain cancers based on

their genetic profiles. Since many cancers have a genetic basis,

human clinicians have found it increasingly complex to understand

all genetic variants of cancer and their response to new drugs and

protocols. Firms like Foundation Medicine and Flatiron Health,

both now owned by Roche, specialise in this approach.

Both providers and payers for care are also using ‘population

health’ machine learning models to predict populations at risk

of particular diseases 17 or accidents 18 or to predict hospital

readmission. 19 These models can be effective at prediction,

although they sometimes lack all the relevant data that might add

predictive capability, such as patient socio-economic status.

But whether rules-based or algorithmic in nature, AI-based

diagnosis and treatment recommendations are sometimes

challenging to embed in clinical workflows and EHR systems. Such

integration issues have probably been a greater barrier to broad

implementation of AI than any inability to provide accurate and

effective recommendations; and many AI-based capabilities

for diagnosis and treatment from tech firms are standalone in

nature or address only a single aspect of care. Some EHR vendors

have begun to embed limited AI functions (beyond rule-based

clinical decision support) into their offerings, 20 but these are in the

early stages. Providers will either have to undertake substantial

integration projects themselves or wait until EHR vendors add

more AI capabilities.

Patient engagement and adherence applications

Patient engagement and adherence has long been seen as the

‘last mile’ problem of healthcare – the final barrier between

ineffective and good health outcomes. The more patients

proactively participate in their own well-being and care, the better

the outcomes – utilisation, financial outcomes and member

experience. These factors are increasingly being addressed by big

data and AI.

Providers and hospitals often use their clinical expertise to

develop a plan of care that they know will improve a chronic or

acute patient's health. However, that often doesn't matter if the

patient fails to make the behavioural adjustment necessary, eg

losing weight, scheduling a follow-up visit, filling prescriptions

or complying with a treatment plan. Noncompliance – when

FHJv6n2-Davenport.indd 96FHJv6n2-Davenport.indd 96 6/3/19 8:49 AM6/3/19 8:49 AM

© Royal College of Physicians 2019. All rights reserved. 97

Artificial intelligence in healthcare

Institute 26 suggested that 35% of UK jobs could be automated

out of existence by AI over the next 10 to 20 years. Other studies

have suggested that while some automation of jobs is possible,

a variety of external factors other than technology could limit

job loss, including the cost of automation technologies, labour

market growth and cost, benefits of automation beyond simple

labour substitution, and regulatory and social acceptance. 27 These

factors might restrict actual job loss to 5% or less.

To our knowledge thus far there have been no jobs eliminated

by AI in health care. The limited incursion of AI into the industry

thus far, and the difficulty of integrating AI into clinical workflows

and EHR systems, have been somewhat responsible for the lack

of job impact. It seems likely that the healthcare jobs most likely

to be automated would be those that involve dealing with digital

information, radiology and pathology for example, rather than

those with direct patient contact. 28

But even in jobs like radiologist and pathologist, the penetration

of AI into these fields is likely to be slow. Even though, as we

have argued, technologies like deep learning are making inroads

into the capability to diagnose and categorise images, there

are several reasons why radiology jobs, for example, will not

disappear soon. 29

First, radiologists do more than read and interpret images.

Like other AI systems, radiology AI systems perform single

tasks. Deep learning models in labs and startups are trained for

specific image recognition tasks (such as nodule detection on

chest computed tomography or hemorrhage on brain magnetic

resonance imaging). However, thousands of such narrow

detection tasks are necessary to fully identify all potential

findings in medical images, and only a few of these can be done

by AI today. Radiologists also consult with other physicians on

diagnosis and treatment, treat diseases (for example providing

local ablative therapies) and perform image-guided medical

interventions such as cancer biopsies and vascular stents

(interventional radiology), define the technical parameters of

imaging examinations to be performed (tailored to the patient's

condition), relate findings from images to other medical records

and test results, discuss procedures and results with patients, and

many other activities.

Second, clinical processes for employing AI-based image work

are a long way from being ready for daily use. Different imaging

technology vendors and deep learning algorithms have different

foci: the probability of a lesion, the probability of cancer, a nodule's

feature or its location. These distinct foci would make it very difficult

to embed deep learning systems into current clinical practice.

Third, deep learning algorithms for image recognition require

‘labelled data’ – millions of images from patients who have

received a definitive diagnosis of cancer, a broken bone or other

pathology. However, there is no aggregated repository of radiology

images, labelled or otherwise.

Finally, substantial changes will be required in medical

regulation and health insurance for automated image analysis to

take off.

Similar factors are present for pathology and other digitally-

oriented aspects of medicine. Because of them, we are unlikely to

see substantial change in healthcare employment due to AI over

the next 20 years or so. There is also the possibility that new jobs

will be created to work with and to develop AI technologies. But

static or increasing human employment also mean, of course, that

AI technologies are not likely to substantially reduce the costs of

medical diagnosis and treatment over that timeframe.

Ethical implications

Finally, there are also a variety of ethical implications around

the use of AI in healthcare. Healthcare decisions have been

made almost exclusively by humans in the past, and the use

of smart machines to make or assist with them raises issues of

accountability, transparency, permission and privacy.

Perhaps the most difficult issue to address given today's

technologies is transparency. Many AI algorithms – particularly

deep learning algorithms used for image analysis – are virtually

impossible to interpret or explain. If

Related Tags