0% found this document useful (0 votes)
183 views9 pages

Chen Decary 2019 Artificial Intelligence in Healthcare An Essential Guide For Health Leaders

This document provides an overview of artificial intelligence (AI) technologies and their applications in healthcare. It describes machine learning, natural language processing, and AI voice assistants as the main AI technologies currently available. Machine learning uses algorithms to identify patterns in large datasets and make predictions. The most common machine learning methods are supervised learning, unsupervised learning, and deep learning. While machine learning has achieved successes in areas involving medical imaging and large data, it is not a solution for all tasks and does not reveal causal relationships. The document recommends health leaders develop strategies to integrate different AI technologies to effectively solve problems faced by healthcare organizations.

Uploaded by

Theja Thushari
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
183 views9 pages

Chen Decary 2019 Artificial Intelligence in Healthcare An Essential Guide For Health Leaders

This document provides an overview of artificial intelligence (AI) technologies and their applications in healthcare. It describes machine learning, natural language processing, and AI voice assistants as the main AI technologies currently available. Machine learning uses algorithms to identify patterns in large datasets and make predictions. The most common machine learning methods are supervised learning, unsupervised learning, and deep learning. While machine learning has achieved successes in areas involving medical imaging and large data, it is not a solution for all tasks and does not reveal causal relationships. The document recommends health leaders develop strategies to integrate different AI technologies to effectively solve problems faced by healthcare organizations.

Uploaded by

Theja Thushari
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 9

ORIGINAL ARTICLE

Healthcare Management Forum


Artificial intelligence in healthcare: 2020, Vol. 33(1) 10-18
ª 2019 The Canadian College of
An essential guide for health leaders Health Leaders. All rights reserved.
Article reuse guidelines:
sagepub.com/journals-permissions
DOI: 10.1177/0840470419873123
Mei Chen, PhD1 and Michel Decary, MSc1 journals.sagepub.com/home/hmf

Abstract
Artificial Intelligence (AI) is evolving rapidly in healthcare, and various AI applications have been developed to solve some of the
most pressing problems that health organizations currently face. It is crucial for health leaders to understand the state of AI
technologies and the ways that such technologies can be used to improve the efficiency, safety, and access of health services,
achieving value-based care. This article provides a guide to understand the fundamentals of AI technologies (ie, machine learning,
natural language processing, and AI voice assistants) as well as their proper use in healthcare. It also provides practical recom-
mendations to help decision-makers develop an AI strategy that can support their digital healthcare transformation.

Introduction reasoning, and problem solving.3 As such, AI is not a single


technology but a range of intelligent processes and behaviours
Artificial Intelligence (AI) is evolving rapidly in healthcare due to
generated by computational models and algorithms. Recently,
its potential to unlock the power of big data and gain insight for
refined computational models and algorithms, coupled with
supporting evidence-based clinical decision-making and
powerful computers and the availability of massive data, have
achieving value-based care. It is crucial for health leaders to under-
accelerated the advancements of AI, particularly in ML, Natural
stand the state of AI technologies and the ways that such technol-
Language Processing (NLP), AI voice technology, AI assistants,
ogies can be used to improve the efficiency, safety, and access of
and robotics. New powerful solutions have been developed to
health services, supporting the digital transformation of healthcare.
solve complex real-world problems in image understanding,
It is evident that AI has begun to affect almost every aspect of
speech recognition, big data analytics, and healthcare. In the
healthcare, from clinical decision support at points of care,
subsequent sections, we will look at the AI technologies
patient self-management of chronic conditions at home, to
currently available and discuss their proper use in healthcare.
drug research in the real world. The development and
deployment of AI technology, however, is challenging and
Machine learning
costly. Health organizations need to overcome a series of
challenges in order to bring AI to success. Such challenges Machine learning represents the dominant approach in AI, and it
include the following: (1) a lack of understanding about what is responsible for most of the recent advancements in the field.
a particular type of AI technology can or cannot do; (2) a lack of Typically, ML refers to a system that trains a predictive model
clear strategies for integrating different AI technologies into the by identifying patterns of data from input, then uses such a
existing care systems to effectively solve the most pressing model to make useful predictions from new, never-before-seen
problems that health organizations currently face; (3) a data. Machine learning algorithms can automatically learn and
shortage of a well-trained workforce for AI implementation; improve from experience without being explicitly programmed,
(4) the incompatibility of AI technologies with legacy and such “learnability” represents a key feature of AI. Machine
infrastructure; and (5) a lack of access to good and diverse learning is widely used in other types of AI technologies, such as
medical data for training Machine Learning (ML) algorithms.1,2 NLP, voice technology, and robotics. Health leaders need to
In this article, we will address the above issues by first become familiar with the main ML algorithms as they are the
describing the state of AI technologies and the potential of foundation for understanding the potential and limitations of
such technologies for transforming healthcare. We will also various types of AI technologies. Briefly speaking, the most
discuss issues related to the selection, development, and common ML algorithms are supervised learning, unsupervised
implementation of AI technologies for improving care quality, learning, Reinforcement Learning (RL), and deep learning.
access, and cost. Finally, we will provide recommendations to Supervised learning uses a data set as input and some known,
help health leaders develop a strategy to support their AI inte- labelled outcomes as output, then identifies patterns that
gration and digital healthcare transformation. correlate the outcomes with input for making predictions. In

The state of AI technology 1


Cogilex R&D Inc., Montreal, Quebec, Canada.
Artificial intelligence, in a practical sense, refers to computer Corresponding author:
systems that simulate or exhibit a specific aspect of human Mei Chen, Cogilex R&D Inc., Montreal, Quebec, Canada.
intelligence or intelligent behaviour, such as learning, E-mail: [email protected]
Chen and Decary 11

this approach, the algorithm needs to know the conclusions it Machine learning has been applied to many data types (such
should come up with from a given data set. With enough data as images, speech, videos, and text) on complex tasks that
and correctly labelled answers, the algorithm eventually learns involve massive data, producing results that are comparable
to make predictions from input data it has never seen before. to and sometimes superior to human experts in terms of both
Supervised learning has been widely applied to healthcare, accuracy and efficiency. It has shown the potential to provide
providing data-driven clinical decision support for mapping data-driven, evidence-based clinical intelligence for advancing
input variables into discrete categories (eg, using medical medical diagnosis, treatment decisions, biomedical research,
imaging to diagnose cancer tumor, its subtypes, and severity) and service delivery across the full spectrum of healthcare.
and predictive analytics within a continuous output (eg, using Even though ML has achieved great successes in areas
Electronic Health Record [EHR] data to make predictions about where medical imaging and big data are involved, ML is not
the recurrences, prognosis, and mortality of a disease). an all-purpose solution. For tasks that require common sense
Unsupervised learning is used to discover the structure of reasoning or domain-specific knowledge, or situations that are
data and make predictions based on input alone. This learning outside of the ML training data set, ML is less applicable. This
algorithm is more applicable to situations in which the is because ML relies on computational power and massive data
outcomes are unknown or the labelling of data is too costly. to identify superficial patterns and correlations. As such, it does
It is mostly used in an exploratory way for clustering, anomaly not reveal the causal relations or clear understanding of the
detection, and pattern recognition in a variety of data types. In phenomenon under study. Consequently, it is difficult to
healthcare, unsupervised learning is particularly useful for explain the results of ML and fix the specific known errors
predicting individual disease risks using genetic biomarkers produced by the ML algorithms.
or for designing personalized treatments based on genomic
variations. As unsupervised learning can automatically Natural language processing
“learn” without human’s labelling of outcomes, it is closer to Natural language processing uses computational methods to
“true AI” in some sense. However, without human teaching, automatically analyze and represent human languages, mostly
unsupervised learning is more prone to errors because it may in text format. Recently, ML methods have been applied to NLP,
use trivial features of the data to make predictions. So, in achieving impressive results in speech recognition, machine
practice, supervised and unsupervised learning are often used translation, text classification, question answering, sentiment
in combination by making use of a large amount of unlabelled analysis, information extraction, and search engine. In
data for training with only a small proportion of data labelled. healthcare, there is a very large amount of unstructured textual
This is called semi-supervised learning which takes the data in the forms of doctors’ notes, test results, lab reports,
advantages of both learning algorithms. medication orders, and discharge instructions. Natural
Reinforcement learning is a more autonomous learning language processing tools can be used to extract critical
algorithm that allows a computer agent to take actions and information about patients from such rich descriptive data,
interact with the environment using rewards and errors as the helping improve diagnoses and treatment recommendations.
feedback to guide training. This can be seen as the ultimate self- The capacity for machines to digest huge amounts of imagery
learning approach because the agent learns from its own and textual data quickly through ML and NLP will enable
experience without either data or labelling; it produces very physicians to make timely diagnoses and treatment decisions,
good results for sequential decision-making tasks or tasks that which can have profound impact on health service delivery,
are well defined with clear rules and outcomes, such as abstract particularly on the ways that patients are treated.
strategy board games like Go. It is also used successfully in self-
driving cars and robotics. In healthcare, it can be used in Artificial intelligence voice technology and assistants
situations in which an agent needs to continuously interact Voice is the most intuitive, natural, and universal way for
with the environment and adjust its actions based on the humans to communicate. Artificial intelligence voice
feedback from the environment, for example, for optimizing technology is changing the nature of human-machine
treatment designs and robotic-assisted surgery. communication, making it much easier for people to obtain,
Deep learning discovers the intricate structure in large data understand, use, and store health information. Voice interface
sets by using a backpropagation algorithm operating on has the potential to optimize users’ experience, helping them
multiple levels of abstraction.4 It is aimed at increasing the overcome barriers existing in text-based information exchange
capacity of supervised and unsupervised learning algorithms or complex system operation.
for solving complex real-world problems by adding multiple Voice technology has been widely used in various industries
processing layers based on artificial neural networks, including and it has started to be incorporated into healthcare to address
“hidden layers” that play a key role in breaking down the some of the information challenges faced by both health
problem to be analyzed. Deep learning has brought professionals and patients. Current EHR systems are complex
breakthroughs in domains that rely heavily on feature and hard to use, so many EHR vendors and health providers are
detection and big data (eg, Go game, computer vision, speech incorporating voice technology into their EHR systems to
recognition, NLP, drug discovery, and genomics). simplify the clinical documentation process. On the consumer
12 Healthcare Management Forum

Figure 1. Human-machine partnership in healthcare.

front, AI assistants like Alexa, Siri, Cortana, and Google How can AI transform healthcare?
Assistant have gained the “skills” to perform specific routine
Humans and machines each have their unique strengths and
and simple tasks in the healthcare context such as reminding
weaknesses, and they can complement each other in
patients when to take their medication and scheduling
providing and optimizing healthcare. The American Medical
appointments. At this moment, such AI voice assistants have
Association recently defined the role of AI in healthcare as
limited capacity in terms of providing reliable answers to
health-related questions.5 It seems that text-based chatbots are “augmented intelligence,” stating that AI will be designed
more reliable, thus have enjoyed a greater commercial success and used to enhance human intelligence rather than replace
(eg, Babylon, Ada, and Buoy). However, the reliability of such it.8 American Medical Association’s view emphasizes the
text-based chatbots is often achieved at the cost of restricting partnership between man and machine, which has important
user input to predetermined words and phrases, not allowing implications for the use of AI in healthcare. Below are our
users to take the initiative in the dialogue.6 With the high perspectives on the role of AI and how AI should be
expectation of investors and the efforts from both big tech designed, implemented, and integrated to support human
companies and start-ups, it is certain that more advanced AI performance and foster digital healthcare transformation.
voice health assistants that are capable of unconstrained
language input and human-like natural conversations will soon Artificial intelligence as a powerful tool and partner
be developed. The AI technologies described above can be used as powerful
tools and partners to enhance, extend, and expand human
Medical robotics capabilities, delivering the types of care patients need, at
Medical robots showcase the capabilities of all AI technologies the time and place they need them. Figure 1 shows how
described previously. Medical robots can help with surgical humans and AI can form a partnership to improve clinical
operations, rehabilitation, social interaction, assisted living, effectiveness (ie, quality, safety, and efficiency), access,
and more. One of the most commonly used medical robots is and affordability of care.
AI-assisted surgical robots, which can analyze data from pre- It is important to indicate that a human-machine partnership
operative medical records to physically guide a surgeon’s by no means implies that the machine cannot be used alone.
instrument in real time during a procedure. Such surgical For tasks where a machine has surpassed human performance
robots are frequently used in neurologic, orthopedic, and (eg, screening cancer, diabetic retinopathy, and certain heart
laparoscopic procedures, being operated both locally and conditions), tasks where mistakes do not lead to serious
remotely. Compared to traditional surgery, robot-assisted consequences (eg, flagging an at-risk population group for
surgery is minimally invasive, and it can reduce hospital stay, vaccination), or for situations where human doctors are
complications, and errors. Robots have also been used to help the unavailable but a machine can do a good job (eg, using a
rehabilitation of patients with stroke, to assist in the care of the chatbot to show a patient how to give an insulin injection),
elderly individuals, and to deliver medical supplies and complete AI automation is possible. The key in human-
equipment.7 It is reasonable to anticipate that one day robots machine partnership is to keep the delicate balance between
will be able to monitor a patient’s vital signs and take proper the types of care we value and the levels of automation that AI
actions when needed. technologies offer (see Figure 1).
Chen and Decary 13

Embedding AI features into workflows to support clinical usually combine the following approaches in such development:
decision-making hiring external talent, building capabilities in-house, licensing
capabilities from large technology firms, buying AI-focused
The best AI integration in healthcare requires AI technologies start-ups or other companies, partnering with other institutions,
to be embedded into the workflows to support clinical decision- and retaining internal talent. Currently, there seem to be three
making at the point of care. By integrating AI into the trends in developing healthcare AI:
workflows, AI can help us accomplish the following:
1. Machine learning is a dominant approach, proven to be
 Unlock the power of big data and gain insight into patients; reliable for disease detection, diagnosis, and
 Support evidence-based decision-making, improving management. In recent years, ML algorithms have been
quality, safety, and efficiency, coordinate care and proven reliable for detecting and diagnosing diseases.
foster communication; Many such algorithms have received approvals from the
 Improve patient experience and outcomes; US Food and Drug Administration (FDA) for their safe use
 Deliver value and reduce costs; and in healthcare. Meskó11,12 has offered a good overview of
 Optimize health system performance. such FDA-approved algorithms and indicated that most of
As EHRs serve as the backbone of digital healthcare systems, them are developed in the areas of radiology, pathology,
the preferable approach to integrate AI in healthcare is to embed cardiology, oncology, endocrinology, and dermatology.
useful AI features directly into EHR systems. Currently, most Besides disease detection and diagnosis, many hospitals
EHRs are “systems of records and storage” with limited capabil- in United States and Canada have started using ML for
ities. Machine learning, NLP, and voice technologies and AI predictive analytics for hospital management purposes
assistants can transform EHRs from “systems of records” to (eg, predicting adverse events, mortality rates, the
“systems of intelligence” and “systems of engagement.” Regard- number of patients to the emergency department). Such
less of whether a health organization plans to purchase a new EHR predictability enables hospitals to take proactive measures
system, adds AI features to an existing EHR system, or builds a for the foreseen events days in advance.
brand-new AI-powered next-generation digital healthcare sys-
Following are some ML highlights worth mentioning:
tem, it is necessary to consider certain AI features. Table 1 shows
the different system capabilities supported by specific AI features.  Stanford researchers have developed an AI algorithm
that can diagnose up to 14 types of medical
Knowing the key platforms, products, and services for conditions simultaneously from medical images.13
developing AI in healthcare  Mayo Clinic neuroradiologists are using AI to
find molecular biomarkers in magnetic
There are three main types of companies that are providing a wide
resonance imaging scans instead of testing
range of AI platforms, products, and services for developing AI-
samples collected during surgery.14
enabled healthcare systems. The first type is EHR vendors such as
 Researchers from Memorial Sloan Kettering
Epic, Cerner, Allscripts, Athena, which have started to add some
Cancer Center have developed a deep learning
AI capabilities in their EHR systems,9 incorporating ML, voice
model and system that uses only the reported
dictation, and NLP to support clinical decision-making,
diagnoses as labels for training on whole-slide
workflows, and patient engagement. The second type is big tech
images, thereby avoiding manual annotation of
companies such as Google, Microsoft, Amazon, Apple, and IBM,
big data sets other deep learning models require.
which are providing AI cloud platforms, services, and ML
Their study shows that their system has the ability
algorithms for health organizations to build, manage, and
to train accurate classification models without
deploy various AI applications with massive data. They also
annotated data sets, thus overcome a great
offer specialized healthcare products that can be used by health
challenge for deploying computational decision-
organizations. The third type is specialized healthcare AI firms.
support systems in clinical practice.15
There is a fast-growing number of companies, particularly start-
 Massachusetts Institute of Technology (MIT) has
ups that are producing various kinds of AI healthcare
recently developed a new deep learning-based AI
applications. Such applications usually fall in one of four
prediction model that can anticipate the
categories: patient-facing, doctor-facing, research, and
development of breast cancer up to 5 years in
telehealth.10 Table 2 is a summary of the key AI players, their
advance. More importantly, this AI model is
platforms, products, and services available for developing AI
specifically designed to ensure it is equally
applications.
accurate for white and black women.16
 Several hospitals in Canada are using AI to
The development and implementation of AI in healthcare predict the number of patients to the emergency
Using a wide range of AI platforms, tools, and services department 2 or 3 days in advance, allowing the
available, many health organizations have been collaborating hospital to take proactive action in staffing and
with technology companies to build their AI capabilities. They resource allocation.17,18
14 Healthcare Management Forum

Table 1. Embedding AI features into EHRS to support clinical decisional-making and workflows

System capabilities AI features

Providing clinical decision support at the point of care to improve  Diagnostic analytics using medical imaging (eg, CT-scan, X-rays, MRI,
diagnostic accuracy and treatment recommendations ECG/EKG, pathological images) or genomic, behavioural, and other
clinical data (eg, symptoms, family history)
 Predictive analytics (eg, predicting high-risk patients and outcomes to a
treatment)
 Personalized treatment recommendations based on clinical, genomic,
and behavioural data
 Prediction and prevention of adverse events
 Medication safety and reconciliation
 Routine integration of medical imaging with other clinical data for triage
and critical care monitoring, diagnostic interpretation, and treatment
modification
 Precision medicine and drug discovery
Extending access and expanding services  Access to on-line services (on-line consultation, appointment booking,
drug renewal)
Providing patient engagement technology to support self-care  Patient empowerment via access to their health data
 Patient engagement tools (eg, chatbots, wearables, and mobile devices)
for supporting patient education, informed decision-making, self-
monitoring, and self-management of chronic conditions
 Channels for patients to interact with healthcare providers, systems, and
services
 Integration of crucial patient data from wearable, mobile devices
and health apps into EHR
Optimizing workflows and resource allocation, improving  Predictions of the number of patients during a specific period
operational efficiency and resources needed (staff, equipment, and facility)
 Integrated voice technology in EHR for clinical documentation,
data entry, voice interface, question asking and answering
 Integrated NLP capacity for processing narrative health data (doctor’s
notes, clinical reports) and providing critical summaries of key
patient information
 Smarter search algorithms
 Simplification of operational processes through AI automation
 Detecting fraud, waste, and abuse via machine learning
Facilitating population health monitoring and management,  Population health monitoring
improving wellness  Identification of high-risk population groups
 Prioritization of at-risk patient populations and management of proactive
interventions
 Investigation of social determinants on healthcare and management
of population wellness
Supporting real-world clinical research and evidence-based  Collection and storage of real-world data for clinical research and care
medicine improvement
 Precision medicine and clinical trial matching
Abbreviations: CT, computed tomography; ECG, electrocardiographic; EHR, electronic health record; MRI, magnetic resonance imaging; NLP, natural language
processing.

The ability for the machine to digest huge amounts of infor- and clinics for its health records project, allowing
mation and provide needed insights quickly will have a pro- consumers to exchange their health data with healthcare
found impact on the ways that medical health services are providers. Similarly, IBM has developed partnerships
delivered and managed. It will also equalize healthcare by with many hospitals, enabling them to make cancer
paying attention to patients who are typically left out. diagnosis and treatment recommendations using Watson
Health. The collaboration across sectors provides a
2. Cross-sector collaborations are vital for advancing AI leverage for both AI firms and health organizations,
in healthcare. Currently, most major AI firms are all enabling them to manage the multi-faceted complexity
working with partners in life sciences or sensor of AI and healthcare. Such an approach will have a
manufacture to develop their healthcare AI capabilities. significant impact on the advancement of AI in
For instance, Apple has partnered with over 100 hospitals healthcare.
Chen and Decary 15

Table 2. Key AI players, their AI platforms, products, and services for healthcare

Technology providers AI platforms, tools, and services

EHR vendors (eg, Epic, Cerner, Allscripts, Athena, EHR vendors have started to add some AI capabilities in their EHR systems,9
and others) such as:
 Natural language processing
 Machine learning for clinical decision support
 Integration with telehealth technologies
 Automated imaging analytics
 Voice dictation
 On-line tools for engaging patients
Big tech companies (eg, Google, Microsoft, Amazon, Big tech companies are providing AI cloud platforms, services, and machine
Apple, IBM, NVIDIA, Nuance, OpenAI, and others) learning algorithms for health organizations to build, manage, deploy various
AI applications with massive data, including capabilities for:
 Image recognition
 Natural language processing
 Voice recognition
 Question-answer matching
 Predictive analytics
 AI voice assistants: Alexa, Siri, Cortana, and Google Assistant
These companies have also developed specialized products for healthcare:
 IBM Watson can learn from new data and make diagnosis and treatment
recommendations for cancer
 Nuance offers NLP tools that can be integrated into commercial EHRs to
support clinical documentation and data entry
 Apple Health Records app enables consumers to enter their health data,
import such data from mobile apps and devices, then exchange with their
healthcare providers
 Amazon offers its Comprehend Medical, a NLP tool for analyzing
unstructured clinical text
 Amazon’s Alexa HIPAA can be used for building compliant healthcare chatbot
“skills”
 Google DeepMind has built a system for making diagnosis and treatment
recommendations for over 50 different eye diseases
 Microsoft Azure API for FHIR enables health providers to connect their
existing EHR data for analytics, machine learning, and actionable intelligence
 Microsoft Healthcare Bot provides healthcare intelligence, including language
models to understand healthcare information and content from credible
sources
Specialized AI firms in healthcare A fast-growing number of companies, particularly start-ups, are producing both
patient-facing and doctor-facing AI applications:
(a) Patient-facing application:
Meskó has provided an overview of such chatbots,11 below is a list of some
well-known chatbots:
 Text-based chatbots:
Woebot, Babylon health, Buoy health, Your.Md, Molly, Eva, Ginger, Replika,
Florence, Izzy, Safedrugbot, Sensely, GYANT, Bots4Health, and others
 Voice-activated chatbots:
Ada health, Infermedica, Avaamo
 Platforms for developing voice chatbots in healthcare:
Orbita
(b) Doctor-facing AI applications:
 Voice dictation tools for clinical documentation:
Dragon Medical Practice by Nuance
 Medical NLP tools:
Optum by United Health Group, MetaMap by NLM, cTakes by Mayo Clinic,
Inguamatics, and CLAMP Toolkit by University of Texas
Abbreviations: API, application programming interface; EHR, Electronic Health Record; FHIR, fast healthcare interoperability resources; HIPAA, Health Insurance
Portability and Accountability Act; NLP, natural language processing.
16 Healthcare Management Forum

3. The partnerships between governments and AI industry Its AI brain imaging system has beat human expert physicians in
become common. To overcome the healthcare challenges diagnosing brain tumors and predicting hematoma expansion.20
that they face, governments in many countries have Robots are commonly seen in hospital, greeting and directing
formed strategic coalitions with the private sector. For patients; many private firms are teaming up with hospitals to
instance, the UK National Health Service has offer experimental AI-powered diagnosis services. For instance,
developed several partnerships with AI firms to Tencent’s AI medical imaging software that has over 90%
improve its healthcare services. It is partnering with accuracy in diagnosing preliminary esophageal cancer is now
Babylon Health to provide on-line medical diagnosis used in more than 100 hospitals across China.21 Clearly, China
and tele-consultation to make healthcare more is gaining global leadership in AI research, technology
accessible; it is collaborating with DeepMind, the innovation, and AI-enabled healthcare services. With its newly
Google-owned UK AI research firm, to develop deep developed 5G connectivity, China surely will gain more
learning model for continuously predicting the advantages, showing a convergence of AI and other advanced
likelihood of patients developing life-threatening technologies in its healthcare.
conditions, and it has recently formed a partnership
with Amazon to provide people with reliable answers Discussion and recommendations
to their medical questions through Alexa.
The development and implementation of AI in healthcare is
complex and costly, so health organizations need to make
The Canadian government has also started to develop a smart decisions and develop strategic plans that enable them
coalition called the Digital Health and Discovery Platform to bring real value to their organizations. Below are some
(DHDP), with a network of 100 partners across Canada. The considerations for the successful development, deployment,
network engages the Canadian Association of Radiologists as and integration of AI in healthcare.
well as healthcare institutions, private firms, and universities
across Canada in developing a cutting-edge, pan-Canadian Considering both short-term and long-term goals of your
health data platform for advancing precision medicine. The organization
DHDP will integrate diverse sources of data, including geno-
mics, imaging, and electronic medical records from multiple As decision-makers, it is essential to consider both the short-
hospitals across Canada to discover new therapeutic interven- term and long-term goals when you develop an AI strategy for
tions, such as AI-derived biomarkers, for personalized diag- your organization. In the short term, you need to build a use
noses and treatments on cancer and eventually on other case by identifying the most pressing problems that your
diseases. Canada has a public-funded universal healthcare sys- organization is facing and determine how such problems can
tem, world-renowned ML scientists, and top-notch researchers be solved by cost-effective AI technologies and methods
in both medical science and health informatics, so it shall be available. In the long term, you need to envision the future of
able to develop a cutting-edge, AI-powered digital healthcare your organization, considering how your organization may
ecosystem to improve healthcare for all Canadians. evolve and how the existing and emerging AI technologies
The US federal government has been slow in forming a can be scaled to effectively transform your organization,
coalition with the private sector on healthcare AI. In February building a hospital of tomorrow.
2019, the US government announced the American AI Initiative, Many health organizations are currently focused on ML on
promoting a whole-of-government strategy in AI collaboration hospital EHR data. As the 5G super-fast connectivity becomes
targeting explainable healthcare systems, systems that are available, there will be a convergence of technologies of AI,
“reliable, dependable, safe, and trustworthy.” In spite of the sensors, voice chatbots, virtual/augmented reality, and other
slow action taken by the federal government, the United States interactive media. Real-time monitoring, diagnoses, and
still shows dominance in AI due to the great efforts from its treatment optimization based on historical and current data of
private sector (particularly large tech companies, EHR system both individuals and the population will become possible. This
vendors, and innovative start-ups), top-notch universities, and will allow the development of an intelligent, integrated, and
very active federal department agencies. So far, the majority connected nationwide digital health ecosystem that will not
of FDA-approved ML algorithms are developed by US only support medical decision-making and clinical research
but will also improve patient education, participation, and
companies or universities. Most AI-enabled EHR systems and
care at home. So, health leaders need to design their AI strate-
consumer AI applications are the products of the United States,
gies and infrastructure capacities with a view on both the pres-
some of which are listed in Table 2.
ent and the future.
China is becoming a global powerhouse in AI,19 thanks to its
comprehensive national AI strategies, decisive action to support
Establishing the leadership, team, culture, and
AI industry with generous central government funding, state-
owned hospital networks, big population data, less restrictive collaboration for successful implementation
data governance, and aggressive start-up culture. China now Technology alone will not transform healthcare; it needs
publishes more scientific papers on AI than any other country. people who derive value from AI and who create impact
Chen and Decary 17

across your organization. Senior leaders can make a difference Health organizations need to understand such limitations and
in their AI projects by providing the funding, talent, and provide adequate, balanced, diverse, and representative data
resources required. In addition, it is crucial to build a team of from its population for retraining and validating ML models
people who possess the diverse expertise required for AI when they deploy them. Decision-makers need to be aware that
development, technology integration, data migration, and most AI technologies are not “out-of-the-box” products that
medical service integrations. It is equally important to you can simply plug into your digital system for it to work.
develop a corporate culture for organization-wide participation Small-scale on-site pilot testing is a good way to validate any
in AI innovation. Health organizations should be prepared AI application.
to collaborate with partners across the industry, working
with partners to make smart decisions and bring the AI Determining the context and protocols for the safe use of
implementation and integration to success. AI technology
Selecting the right AI platform, tools, and approaches for Ensuring the safety, privacy, and well-being of patients requires
one to conduct a hazard analysis, evaluate the consequences of
implementing your AI strategy potential false positive and false negative, and develop hazard
Healthcare providers vary in their sizes, types, challenges, prevention protocols. For crucial clinical processes that can lead
priorities, and resources. For providers that have already to serious consequences (eg, making medical diagnosis and
installed a sound EHR system, adding AI capabilities into the treatment decisions), a dual safety mechanism is required. In
EHR system is possible as many EHR vendors have opened such cases, doctors are the ones who make the calls, using
their platforms to allow data exchange and system connection. data-generated insights as references. Furthermore, for any AI
In addition, many vendors are adding AI features into their product to be deployed, an on-site pilot implementation and
EHR systems. For most hospitals, working with the EHR validation is needed. Finally, it is important to collect real-
vendor and other AI technology firm to develop the solutions world evidence and develop a mechanism to continuously
they need is perhaps the best option. For organizations that monitor system performance, ensuring the safety and
have the expertise and resources to build their own AI effectiveness of the deployed AI product on an ongoing basis.
capabilities or that want to become an AI player in the In all these processes, it is important to establish policies and
healthcare industry, they can do so by using commercial AI protocols to ensure the privacy, security, and ethics of AI use.
cloud platforms and service currently available (see Table 2). However, we must keep a balance between patient privacy and
To keep business as usual, they can build their new AI data sharing, and between regulation and innovation. Artificial
infrastructure and process independently, then link it to the intelligence professionals need to work with a large volume of
old infrastructure. This gives health organizations complete real patient data to ensure the accuracy and safety of the ML
control in instantiating a new process while avoiding models, so patients need to know that AI can only be advanced
interfering with the ongoing operations. when they share data more freely, and this can be done in a
secure environment. Meanwhile, health organizations need to
Forming a good data strategy to derive patient insights ensure that their AI approaches are lawful, ethical, and robust,
Successful ML relies on the access to large volumes of quality showing complete transparency about what they do with patient
data; the source, size, and quality of data can dramatically data.
impact the ML models developed. Collecting large-scale data
that are complete, accurate, up-to-date, and representative of Establishing performance standards to measure AI
typical populations is a big challenge for analytics success
professionals.22 Part of the bias in AI is due to the lack of
Assessing AI approaches takes time but it will enable health
diverse data available for training the algorithms. So the
organizations to discover problems and fix them before it is too
capacity to collect, store, and learn from data is crucial for
late. Before implementation, it is essential to define
AI success and often AI workers spend a large part of their
performance evaluation metrics, then measure AI success
time to clean up data in order to ensure the quality of the ML
accordingly at different stages of the development and
models they are developing. Some data scientists believe that
implementation (eg, pilot testing, scaled implementation, and
collecting new data that meet the current data standards is
validation). Such performance metrics should reflect the
better than cleaning up old messy data. This is a valid
values, priorities, and vision of your organization. There are
perspective because the old data in EHR systems often
many ways to assess AI technologies. Generally speaking,
contain noises, biases, errors, and unusable data.
things to consider in the evaluation should include improved
clinical effectiveness (quality, efficiency, and safety), extended
Retraining ML algorithms and validating AI applications
access and expanded services to patients, improved patient
with data and patients from the local organizations experience and outcomes, optimized operational processes,
Not including enough meaningful and representative data improved staff satisfaction with the work environment, and
during training and validation is a common problem in ML. reduced costs and increased revenue.
18 Healthcare Management Forum

Developing nationwide AI-powered digital healthcare 9. Davenport T, Hongsermeier T, Alba K. Using AI to improve
ecosystems electronic health records. Harvard Business Review. 2018(12).
Available at: https://hbr.org/2018/12/using-ai-to-improve-
Given the high complexity and costs involved in developing electronic-health-records. Accessed July 1, 2019.
the various types of AI technologies needed for improving the 10. Kuo E. AI in healthcare: industry landscape. Techburst Report.
effectiveness, access, and affordability of healthcare, each 2017. techburst.io/ai-in-healthcare-industry-landscape-c433
country needs to have a national AI strategy for building a 829b320c. Accessed July 1, 2019.
nationwide AI-powered digital healthcare ecosystem that 11. Meskó B. The top 12 health chatbots. The Medical Futurist. 2018.
benefits both health organizations and patients. Currently, the Available at: https://medicalfuturist.com/top-12-health-chatbots.
majority of funding is used for developing ML on big EHR Accessed June 6, 2019.
data, mostly for the benefits of health professionals. As the total 12. Meskó B. FDA approvals for smart algorithms in medicine in one
health can only be achieved through jointed efforts between giant infographic. The Medical Futurist. 2019. Available at:
health professionals and patients, patients need to have https://medicalfuturist.com/fda-approvals-for-algorithms-in-
AI-powered tools for self-monitoring and self-managing their medicine. Accessed August 8, 2019.
chronic conditions. 13. Armitage H. Artificial intelligence rivals radiologists in screening
ORCID iD X-rays for certain diseases. 2018. Available at: https://med.
stanford.edu/news/all-news/2018/11/ai-outperformed-
Mei Chen, PhD https://orcid.org/0000-0003-4414-8835
radiologists-in-screening-x-rays-for-certain-diseases.html.
References Accessed July 1, 2019.
14. Korfiatis P, Erickson BJ. Deep learning can see the unseeable:
1. Chen M, Decary M. AI in healthcare: from hype to impact.
predicting molecular markers from MRI of brain gliomas.
Workshop presented at ITCH 2019: Improving Usability, Safety
Clin Radiol. 2019;74(5):367-373. doi:10.1016/j.crad.2019.
and Patient Outcomes with Health Information Technology,
01.028.
Victoria, British Columbia, Canada. 2019. Available at: https://
15. Campanella G, Hanna MG, Geneslaw L, et al. Clinical-grade
www.slideshare.net/MeiChen39/ai-in-healthcarefrom-hype-to-
impact. Accessed June 6, 2019. computational pathology using weakly supervised deep learning
2. Mckinsey and Company. AI adoption advances, but foundational on whole slide images. Nat Med. 2019;25(8):1301-1309. doi:10.
barriers remain. Survey Report. 2018(11). Available at: https:// 1038/s41591-019-0508-1.
www.mckinsey.com/featured-insights/artificial-intelligence/ai- 16. Yala A, Lehman C, Schuster T, Portnoi T, Barzilay R. A deep
adoption-advances-but-foundational-barriers-remain. Accessed learning mammography-based model for improved breast cancer
June 6, 2019. risk prediction. Radiology. 2019;292(1):60-66.
3. McCarthy J. The Dartmouth summer research project on artificial 17. Ubelacker S. From bionic arms to predicting patient surges in ER,
intelligence. Artificial intelligence: past, present, and future. AI is reshaping patient care. CBC News. 2019. Available at:
1956. Available at: http://www.dartmouth.edu/*vox/0607/ https://www.cbc.ca/news/canada/edmonton/bionic-arms-
0724/ai50.html. Accessed June 6, 2019. artificial-intelligence-patient-care-alberta-1.5090172. Accessed
4. Esteva A, Robicquet A, Ramsundar B, et al. A guide to deep June 6, 2019.
learning in healthcare. Nat Med. 2019;25(1):24-29. 18. Cuttler M. Transforming healthcare: how artificial intelligence is
5. Bickmore TW, Trinh H, Olafsson S, et al. Safety first: reshaping the medical landscape. CBC News. April 26, 2019.
conversational agents for healthcare. In: Moore RJ, Szymanski Available at: https://www.cbc.ca/news/health/artificial-intel
MH, Arar R, Ren GJ, eds. Studies in Conversational UX Design, ligence-healthcare-1.5110892. Accessed August 8, 2019.
Human–Computer Interaction Series. Berlin, Germany: Springer; 19. Lee KF. AI Superpowers: China, Silicon Valley, and the New
2018. World Order. Boston, MA: Houghton Mifflin; 2018.
6. Laranjo L, Dunn AG, Tong HL, et al. Conversational agents in 20. Yamei. China Focus: AI Beats Human Doctors in Neuroimaging
healthcare: a systematic review. J Am Med Inform Assoc. 2018; Recognition Contest. Xinhuanet. 2018. Available at: http://www.
25(9):1248-1258. xinhuanet.com/english/2018-06/30/c_137292451.htm. Accessed
7. Crawford M. Top 6 robotic applications in medicine, the American June 6, 2019.
Society of Mechanical Engineers. 2016(9). Available at: https:// 21. Yan NT, Lu C, Feng W. China Focus: Tech Giants Tap into AI
www.asme.org/engineering-topics/articles/bio engineering/top- Healthcare Market. Xinhuanet. 2018. Available at: http://www.
6-robotic-applications-in-medicine. Accessed June 6, 2019. xinhuanet.com/english/2018-06/19/c_137265090.htm. Accessed
8. The American Medical Association. AMA passes first August 8, 2019.
policy recommendations on augmented intelligence. 2018. 22. Househ MW, Kushniruk AW, Borycki EM. Big Data, Big
Available at: https://www.ama-assn.org/press-center/press- Challenges: A Healthcare Perspective: Background, Issues,
releases/ama-passes-first-policy-recommendations-augmented- Solutions and Research Directions. 1st ed. Berlin, Germany:
intelligence. Accessed July 1, 2019. Springer; 2019.

You might also like