Download PDF
Professor Jim Warren
PhD, FACHI
Professor of Health Informatics
School of Computer Science
University of Auckland
Perhaps we haven’t yet reached the point where computer-based
Artificial Intelligence (AI) has overtaken humanity as the masters of
the world, but there’s a groundswell of sentiment that AI can now ex-
ceed human performance for almost any specialised task. In 2015, AI
displaced a bastion of human mastery when a computer programme
beat a human champion in Go, a game that had long held out against
attempts to exceed the best human players, in part due to the size of
the board and many possible moves. The winning computer program
employed techniques that are hallmarks of the new wave of AI: using
a deep neural network (a system of nodes and weighted connec-
tions with multiple layers between the inputs and the outputs); ‘big
data’ (a comprehensive collection of transcripts of high-level human
Go games in this case), and massive computation (notably, to learn
from the data to recognise good moves and the value of board po-
sitions, reinforced by the equivalent of lifetimes of simulated games
against itself. 1 AI has been an active field of research ever since digital
electronic computers emerged after World War II, and while noto-
riously difficult to define, intertwined with concepts of rational and
human-like thought and action AI can be taken simply as the attempt
to build intelligent entities. 2 There’s a tendency to move the threshold
for what constitutes ‘real AI’ forward to exclude established innova-
tions. For instance, AI accomplishments of past decades, such as the
automated interpretation of electrocardiograms (ECGs), may now
be seen merely as useful technology without much regard to the
human-like nature of the task being accomplished. 3 But this new wave
of AI based on deep learning has re-ignited both the imagination of
the public in general and the health-care community, in particular in
terms of the potential of AI to change our lives.
Learning to beat humans at games is an important part of AI research,
not just as a publicity stunt, but for the insights learned in finding ways
to out think humans at tasks that have attracted individuals to dedi-
cate a lifetime of training to becoming experts (such as Grandmasters
in chess). But AI has always had its applied side as well, including
learning to imitate (or exceed) the performance of medical experts.
Hard on the heels of the breakthrough in Go, a deep learning system
was demonstrated to provide formidable sensitivity and specificity
for detecting diabetic retinopathy in fundus images as compared to a
panel of United States licensed ophthalmologists and ophthalmology
senior residents. 4 The authors themselves took some care to point
out limitations – the algorithm would not necessarily detect non-di-
abetic retinopathy lesions that were outside of its training data, nor
would it be a replacement for a comprehensive eye exam – yet there
is a temptation for the findings of this frequently-cited article (893
times on Google Scholar at 6 March 2019) to be consolidated simply as – with deep learning, AIs can now match specialists. A recent Jour-
nal of the American Medical Association editorial indicated that the
new wave of AI is one of a series of technology-based advances, and
makes a comparison to how computed tomography has become part
of the radiology toolkit. 5 Nonetheless, the concluding words, ‘artificial
intelligence and deep learning are entering the mainstream of clinical
medicine’ and, ‘physicians need to actively engage to adapt their prac-
tice’, set a tone that we have reached a tipping point for AI in medical
decision making. A medical student could be forgiven for feeling some
anxiety, wondering just what a future with AI making better decisions
than specialists implies for their role and the doctor-patient relation-
ship, or how they might be expected to engage this phenomenon.
We can expect that AI systems for health application will continue to
grow in diversity and effectiveness. ‘Super-computing’ is now readily
available: the graphics processing units in the video cards of our home
computers turn out to be superb number-crunchers for neural net-
work algorithms; or we can rent scalable computing power through
cloud computing services by Amazon, Google or others. Moreover,
the ever-increasing permeation of health-care systems with comput-
ing has as its natural by-product a growing archive of electronic med-
ical records ripe for analysis. While this AI boom is indeed likely to be
transformative to health-care delivery, there are reasons to take the
view that this change will be incremental, manageable, and (hopefully)
on balance positive.
First, AI algorithms from deep learning are not so unlike comput-
ing capabilities that we have been using routinely in New Zealand
for years. For example, PREDICT is simultaneously decision-sup-
port software and an ongoing, prospectively designed, open cohort
study. 6 The PREDICT software integrates with the practice manage-
ment system to retrieve patient data, with any remaining required
data entered interactively to provide an individualised estimate of
the probability of a cardiovascular disease (CVD) event in the next
five years, along with treatment recommendations. Participant risk
factors captured by software that is regularly linked to national data-
bases included hospitalisations and deaths related to CVD, support-
ing ongoing research to improve the risk prediction – most recently,
based on over 400,000 patient encounters in New Zealand from
2002–2015. 7 At the heart of the risk prediction is a regression model
(specifically a Cox proportional-hazards model) that gives a particular
weight to each risk factor. The model is structurally much simpler
than a deep learning model, but has the advantage that the reasoning
behind the model’s recommendation is easily explained. Adding ex-
planation ability to deep neural networks is an active research area. 8
The experience, for patients and health-care professionals, in using a
deep learning AI (at least one that has been appropriately developed
and carefully tested) will be little different to that with PREDICT,
which has integrated smoothly with the existing health-care system
and professional roles.
Second, while AI will challenge the doctor-patient dynamic, informa-
tion technology (IT) challenging the doctor-patient dynamic is noth-
ing new. For over 25 years, the World Wide Web (the Web) has
been democratising access to information. Patients are at liberty to
bring into their consults printouts (or perhaps nowadays more likely
to brandish their cell phone or tablet) with the latest research find-
ings, as well as potentially questionable content biased by revenue
generation motives. As the Web has become more sophisticated and
IT reaches ever more intimately into our lives, so the diversity of ways
patients may bring IT into their health care has grown, now including
mobile apps, fitness trackers, and blog posts. An interesting example
is PatientsLikeMe, a Web-based network where patients connect to
others with the same disease and share experiences. Sharing of quan-
titative data is encouraged along with the organisation of research
studies, for example to test the effectiveness of off-label uses of
drugs. 9 In his book The Patient Will See You Now, Eric Topol describes
medicine as having reached a ‘Gutenberg moment’, where new free-
dom of information is enabling health consumers to take a revolu-
tionary degree of control of their health care. 10 Topol cites numer-
ous Web and IT-mediated trends, including sharing of big data and
direct-to-patient genetic test results (as exemplified by 23andme). 11
Meanwhile, mobile text-based services are slipping into the main-
stream of evidence-based medicine. For instance, a program including
motivational messages and behaviour-change techniques was shown
to significantly improve smoking cessation rates at six months. 12 The
package of intervention techniques and dialog strategies operation-
alised in this service in fact makes it a form of AI – one that can be
recommended to a patient by a doctor, or that a consumer can find
and download for themselves over the Web.
Third, health-care professionals can engage with, and encourage or
moderate, the advance of AI by routinely asking questions of prove-
nance. You may encounter AI-based decision support presented by
a patient, or integrated with the systems you use in your Primary
Health Organisation or District Health Board. In any event, you can
query where it comes from – who is endorsing and distributing it,
and what is their motivation (i.e, is it purely for profit through prolifer-
ation – licensing fees or banner-ad revenue – or is it publicly funded; is
it endorsed by a medical body?). Is it part of the new wave of AI based
on machine learning from big data? Or perhaps (as with the above
smoking cessation example) the capability is a product of ‘knowledge
engineering’, where techniques based on human experts have been
deliberately selected. If it is based on data, then data from where and
when? Does that data seem likely to be a good representation of your
own patient population, or would there be obvious gaps (e.g. lacking
Māori and Pacific cases)? Can the system be retrained on local data?
Can the system offer explanations for its recommendations, or is it
just a ‘black box’ that offers no specific insight for its assessment? Is
there evidence of the system’s effectiveness? If so, how has its per-
formance been evaluated: in what context, on what population, over
what duration, and particularly what was its performance compared
to? If the answers to these questions are hard to find, you should be
suspicious (or at least cautious); if the answers are unsatisfactory, you
should actively communicate about the system’s limitations.
To take the concept of engagement further, it is worth noting that
Health Informatics is an established interdisciplinary field and a grow-
ing profession – this is the field that deals with methods of information
processing and management in health care, including AI in health-care
delivery. Membership in Health Informatics New Zealand (HINZ) is
open to anyone with an interest in the field; HINZ events, particu-
larly the annual national conference, are a great way to learn more about the field and meet the Health Informatics community. Sev-
eral New Zealand universities offer postgraduate degrees in Health
Informatics, and there are numerous options to study online with
overseas universities (HINZ maintains a list of domestic and over-
seas study options: https://www.hinz.org.nz/page/EducationOptions).
One can apply to become a member or fellow of the Australasian
College of Health Informatics based on contribution to the field, and
there now exists a training pathway to fellowship (https://www.achi.
org.au/achi-fellowship-program/). While in this article I have taken
a particularly medical/doctor centred view of the impact of AI on
health-care delivery (given the nature of the journal), it is important
to understand that the field is concerned with the whole health-care
team; notably, nurses have been especially active in Health Informat-
ics throughout its history. AI will influence and expand the capabilities
of every type of professional associated with health care, as well as
the health consumer.
The growing application of AI will add new and diverse inputs into the
clinical context, but it will be just one more source of information to
be considered in medical decision making. If you approach decision
making as a shared process in partnership with patients, then they will
be less likely to use Google to replace you!
References
1. Silver D, Huang A, Maddison CJ, Guez A, Sifre L, van den Driessche
G, et al. Mastering the game of Go with deep neural networks and
tree search. Nature. 2016;529(7587):484–9.
2. Russell S, Norvig P. Artificial intelligence: a modern approach. 3rd
ed. Prentice-Hall; 2010. Chapter 1, Introduction; p. 1–33.
3. Yu KH, Kohane IS. Framing the challenges of artificial intelligence in
medicine. BMJ Qual Saf. 2019;28(3):238–41.
4. Gulshan V, Peng L, Coram M, Stumpe MC, Wu D, Narayanaswamy
A, et al. Development and validation of a deep learning algorithm for
detection of diabetic retinopathy in retinal fundus photographs. JAMA.
2016;316(22):2402–10.
5. Stead WW. Clinical implications and challenges of artificial
intelligence and deep learning. JAMA. 2018;320(11):1107–8.
6. Wells S, Riddell T, Kerr A, Pylypchuk R, Chelimo C, Marshall R,
et al. Cohort profile: the PREDICT cardiovascular disease cohort
in New Zealand primary care (PREDICT-CVD 19). Int J Epidemiol.
2017;46(1):22.
7. Pylypchuk R, Wells S, Kerr A, Poppe K, Riddell T, Harwood M, et al.
Cardiovascular disease risk prediction equations in 400 000 primary
care patients in New Zealand: a derivation and validation study.
Lancet. 2018;391(10133):1897–907.
8. Montavon G, Samek W, Müller K-R. Methods for interpreting and
understanding deep neural networks. Digit Signal Process. 2018;73:1–15.
9. Frost J, Okun S, Vaughan T, Heywood J, Wicks P. Patient-reported
outcomes as a source of evidence in off-label prescribing: analysis of
data from PatientsLikeMe. J Med Internet Res 2011;13(1):e6.
10. Topol E. The patient will see you now: the future of medicine is in
your hands. New York: Basic Books; 2015.
11. Stoekle HC, Mamzer-Bruneel MF, Vogt G, Herve C. 23andMe:
a new two-sided data-banking market model. BMC Med Ethics.
2016;17:19.
12. Free C, Knight R, Robertson S, Whittaker R, Edwards P, Zhou
W, et al. Smoking cessation support delivered via mobile phone
text messaging (txt2stop): a single-blind, randomised trial. Lancet.
2011;378(9785):49–55.
Professor Jim Warren
PhD, FACHI
Professor of Health Informatics
School of Computer Science
University of Auckland
Perhaps we haven’t yet reached the point where computer-based
Artificial Intelligence (AI) has overtaken humanity as the masters of
the world, but there’s a groundswell of sentiment that AI can now ex-
ceed human performance for almost any specialised task. In 2015, AI
displaced a bastion of human mastery when a computer programme
beat a human champion in Go, a game that had long held out against
attempts to exceed the best human players, in part due to the size of
the board and many possible moves. The winning computer program
employed techniques that are hallmarks of the new wave of AI: using
a deep neural network (a system of nodes and weighted connec-
tions with multiple layers between the inputs and the outputs); ‘big
data’ (a comprehensive collection of transcripts of high-level human
Go games in this case), and massive computation (notably, to learn
from the data to recognise good moves and the value of board po-
sitions, reinforced by the equivalent of lifetimes of simulated games
against itself. 1 AI has been an active field of research ever since digital
electronic computers emerged after World War II, and while noto-
riously difficult to define, intertwined with concepts of rational and
human-like thought and action AI can be taken simply as the attempt
to build intelligent entities. 2 There’s a tendency to move the threshold
for what constitutes ‘real AI’ forward to exclude established innova-
tions. For instance, AI accomplishments of past decades, such as the
automated interpretation of electrocardiograms (ECGs), may now
be seen merely as useful technology without much regard to the
human-like nature of the task being accomplished. 3 But this new wave
of AI based on deep learning has re-ignited both the imagination of
the public in general and the health-care community, in particular in
terms of the potential of AI to change our lives.
Learning to beat humans at games is an important part of AI research,
not just as a publicity stunt, but for the insights learned in finding ways
to out think humans at tasks that have attracted individuals to dedi-
cate a lifetime of training to becoming experts (such as Grandmasters
in chess). But AI has always had its applied side as well, including
learning to imitate (or exceed) the performance of medical experts.
Hard on the heels of the breakthrough in Go, a deep learning system
was demonstrated to provide formidable sensitivity and specificity
for detecting diabetic retinopathy in fundus images as compared to a
panel of United States licensed ophthalmologists and ophthalmology
senior residents. 4 The authors themselves took some care to point
out limitations – the algorithm would not necessarily detect non-di-
abetic retinopathy lesions that were outside of its training data, nor
would it be a replacement for a comprehensive eye exam – yet there
is a temptation for the findings of this frequently-cited article (893
times on Google Scholar at 6 March 2019) to be consolidated simply as – with deep learning, AIs can now match specialists. A recent Jour-
nal of the American Medical Association editorial indicated that the
new wave of AI is one of a series of technology-based advances, and
makes a comparison to how computed tomography has become part
of the radiology toolkit. 5 Nonetheless, the concluding words, ‘artificial
intelligence and deep learning are entering the mainstream of clinical
medicine’ and, ‘physicians need to actively engage to adapt their prac-
tice’, set a tone that we have reached a tipping point for AI in medical
decision making. A medical student could be forgiven for feeling some
anxiety, wondering just what a future with AI making better decisions
than specialists implies for their role and the doctor-patient relation-
ship, or how they might be expected to engage this phenomenon.
We can expect that AI systems for health application will continue to
grow in diversity and effectiveness. ‘Super-computing’ is now readily
available: the graphics processing units in the video cards of our home
computers turn out to be superb number-crunchers for neural net-
work algorithms; or we can rent scalable computing power through
cloud computing services by Amazon, Google or others. Moreover,
the ever-increasing permeation of health-care systems with comput-
ing has as its natural by-product a growing archive of electronic med-
ical records ripe for analysis. While this AI boom is indeed likely to be
transformative to health-care delivery, there are reasons to take the
view that this change will be incremental, manageable, and (hopefully)
on balance positive.
First, AI algorithms from deep learning are not so unlike comput-
ing capabilities that we have been using routinely in New Zealand
for years. For example, PREDICT is simultaneously decision-sup-
port software and an ongoing, prospectively designed, open cohort
study. 6 The PREDICT software integrates with the practice manage-
ment system to retrieve patient data, with any remaining required
data entered interactively to provide an individualised estimate of
the probability of a cardiovascular disease (CVD) event in the next
five years, along with treatment recommendations. Participant risk
factors captured by software that is regularly linked to national data-
bases included hospitalisations and deaths related to CVD, support-
ing ongoing research to improve the risk prediction – most recently,
based on over 400,000 patient encounters in New Zealand from
2002–2015. 7 At the heart of the risk prediction is a regression model
(specifically a Cox proportional-hazards model) that gives a particular
weight to each risk factor. The model is structurally much simpler
than a deep learning model, but has the advantage that the reasoning
behind the model’s recommendation is easily explained. Adding ex-
planation ability to deep neural networks is an active research area. 8
The experience, for patients and health-care professionals, in using a
deep learning AI (at least one that has been appropriately developed
and carefully tested) will be little different to that with PREDICT,
which has integrated smoothly with the existing health-care system
and professional roles.
Second, while AI will challenge the doctor-patient dynamic, informa-
tion technology (IT) challenging the doctor-patient dynamic is noth-
ing new. For over 25 years, the World Wide Web (the Web) has
been democratising access to information. Patients are at liberty to
bring into their consults printouts (or perhaps nowadays more likely
to brandish their cell phone or tablet) with the latest research find-
ings, as well as potentially questionable content biased by revenue
generation motives. As the Web has become more sophisticated and
IT reaches ever more intimately into our lives, so the diversity of ways
patients may bring IT into their health care has grown, now including
mobile apps, fitness trackers, and blog posts. An interesting example
is PatientsLikeMe, a Web-based network where patients connect to
others with the same disease and share experiences. Sharing of quan-
titative data is encouraged along with the organisation of research
studies, for example to test the effectiveness of off-label uses of
drugs. 9 In his book The Patient Will See You Now, Eric Topol describes
medicine as having reached a ‘Gutenberg moment’, where new free-
dom of information is enabling health consumers to take a revolu-
tionary degree of control of their health care. 10 Topol cites numer-
ous Web and IT-mediated trends, including sharing of big data and
direct-to-patient genetic test results (as exemplified by 23andme). 11
Meanwhile, mobile text-based services are slipping into the main-
stream of evidence-based medicine. For instance, a program including
motivational messages and behaviour-change techniques was shown
to significantly improve smoking cessation rates at six months. 12 The
package of intervention techniques and dialog strategies operation-
alised in this service in fact makes it a form of AI – one that can be
recommended to a patient by a doctor, or that a consumer can find
and download for themselves over the Web.
Third, health-care professionals can engage with, and encourage or
moderate, the advance of AI by routinely asking questions of prove-
nance. You may encounter AI-based decision support presented by
a patient, or integrated with the systems you use in your Primary
Health Organisation or District Health Board. In any event, you can
query where it comes from – who is endorsing and distributing it,
and what is their motivation (i.e, is it purely for profit through prolifer-
ation – licensing fees or banner-ad revenue – or is it publicly funded; is
it endorsed by a medical body?). Is it part of the new wave of AI based
on machine learning from big data? Or perhaps (as with the above
smoking cessation example) the capability is a product of ‘knowledge
engineering’, where techniques based on human experts have been
deliberately selected. If it is based on data, then data from where and
when? Does that data seem likely to be a good representation of your
own patient population, or would there be obvious gaps (e.g. lacking
Māori and Pacific cases)? Can the system be retrained on local data?
Can the system offer explanations for its recommendations, or is it
just a ‘black box’ that offers no specific insight for its assessment? Is
there evidence of the system’s effectiveness? If so, how has its per-
formance been evaluated: in what context, on what population, over
what duration, and particularly what was its performance compared
to? If the answers to these questions are hard to find, you should be
suspicious (or at least cautious); if the answers are unsatisfactory, you
should actively communicate about the system’s limitations.
To take the concept of engagement further, it is worth noting that
Health Informatics is an established interdisciplinary field and a grow-
ing profession – this is the field that deals with methods of information
processing and management in health care, including AI in health-care
delivery. Membership in Health Informatics New Zealand (HINZ) is
open to anyone with an interest in the field; HINZ events, particu-
larly the annual national conference, are a great way to learn more about the field and meet the Health Informatics community. Sev-
eral New Zealand universities offer postgraduate degrees in Health
Informatics, and there are numerous options to study online with
overseas universities (HINZ maintains a list of domestic and over-
seas study options: https://www.hinz.org.nz/page/EducationOptions).
One can apply to become a member or fellow of the Australasian
College of Health Informatics based on contribution to the field, and
there now exists a training pathway to fellowship (https://www.achi.
org.au/achi-fellowship-program/). While in this article I have taken
a particularly medical/doctor centred view of the impact of AI on
health-care delivery (given the nature of the journal), it is important
to understand that the field is concerned with the whole health-care
team; notably, nurses have been especially active in Health Informat-
ics throughout its history. AI will influence and expand the capabilities
of every type of professional associated with health care, as well as
the health consumer.
The growing application of AI will add new and diverse inputs into the
clinical context, but it will be just one more source of information to
be considered in medical decision making. If you approach decision
making as a shared process in partnership with patients, then they will
be less likely to use Google to replace you!
References
1. Silver D, Huang A, Maddison CJ, Guez A, Sifre L, van den Driessche
G, et al. Mastering the game of Go with deep neural networks and
tree search. Nature. 2016;529(7587):484–9.
2. Russell S, Norvig P. Artificial intelligence: a modern approach. 3rd
ed. Prentice-Hall; 2010. Chapter 1, Introduction; p. 1–33.
3. Yu KH, Kohane IS. Framing the challenges of artificial intelligence in
medicine. BMJ Qual Saf. 2019;28(3):238–41.
4. Gulshan V, Peng L, Coram M, Stumpe MC, Wu D, Narayanaswamy
A, et al. Development and validation of a deep learning algorithm for
detection of diabetic retinopathy in retinal fundus photographs. JAMA.
2016;316(22):2402–10.
5. Stead WW. Clinical implications and challenges of artificial
intelligence and deep learning. JAMA. 2018;320(11):1107–8.
6. Wells S, Riddell T, Kerr A, Pylypchuk R, Chelimo C, Marshall R,
et al. Cohort profile: the PREDICT cardiovascular disease cohort
in New Zealand primary care (PREDICT-CVD 19). Int J Epidemiol.
2017;46(1):22.
7. Pylypchuk R, Wells S, Kerr A, Poppe K, Riddell T, Harwood M, et al.
Cardiovascular disease risk prediction equations in 400 000 primary
care patients in New Zealand: a derivation and validation study.
Lancet. 2018;391(10133):1897–907.
8. Montavon G, Samek W, Müller K-R. Methods for interpreting and
understanding deep neural networks. Digit Signal Process. 2018;73:1–15.
9. Frost J, Okun S, Vaughan T, Heywood J, Wicks P. Patient-reported
outcomes as a source of evidence in off-label prescribing: analysis of
data from PatientsLikeMe. J Med Internet Res 2011;13(1):e6.
10. Topol E. The patient will see you now: the future of medicine is in
your hands. New York: Basic Books; 2015.
11. Stoekle HC, Mamzer-Bruneel MF, Vogt G, Herve C. 23andMe:
a new two-sided data-banking market model. BMC Med Ethics.
2016;17:19.
12. Free C, Knight R, Robertson S, Whittaker R, Edwards P, Zhou
W, et al. Smoking cessation support delivered via mobile phone
text messaging (txt2stop): a single-blind, randomised trial. Lancet.
2011;378(9785):49–55.
