“Healthcare is an information industry that continues to think that it is a biological
industry.”
– Laurence McMahon at the AAHC Thought Leadership Institute meeting, August, 2016.
With the current imbroglio of health care and burden of work, some of us have been
bereft of the pure joys of being a physician or caretaker in pediatric cardiology.
Now imagine your future experience as a practitioner for children with heart disease
in the year 2040:
You are in the serene cardiac intensive care pod where real-time analytics are displayed
(rather than the de rigueur vital signs) and deep learning (in the form of recurrent
neural network) are now routinely used for personalized intensive care unit decision
support and to mitigate the stress of families and physicians/nurses. There is no
longer formal lengthy AM rounds as communication among team members is now continuous.
The old electronic record and computers are all no longer omnipresent, and patient
information is displayed on the video wall only upon activation or automatically in
the case of sudden changes. The conversations are recorded with a tiny microphone
on your shirt and analyzed with natural language processing (NLP) in order to capture
key information onto the electronic record. You receive notifications about recent
advances in diagnostics and therapy that pertain to your patients, and you read these
new notifications and answer questions to gain points for individualized continuing
medical education (as board examinations have been eliminated finally).
You then go to the outpatient area. Several of your patients are triaged to be seen
remotely with telepresence. A new child with heart failure is seen after the admission
process is completed automatically with robotic process automation, and the echocardiogram
is preliminarily read, aided by computer vision (by the use of convoluted neural network
[CNN]). A precision medicine protocol is promoted by individualized therapy options,
and pharmacotherapy is accompanied by the pharmacogenomic profile. A cognitive architecture
based on heart failure experts from numerous centers, published reports in this disease
condition, as well as the accumulated data from the entire pediatric heart failure
patient cohort is utilized for best clinical decision for this child. You use augmented
reality to explain the heart condition to the parents including illustration of a
micro-axial device for ventricular support (as orthotropic transplantation has been
supplanted by advances in miniaturized support devices and nanomedicine). The patient
is then set up for wearable technology with embedded artificial intelligence (AI)
to monitor her blood pressure and heart rate while she is on new medical regimen at
home.
For a subspecialty that is particularly rich with imaging and clinical data already,
and with more sources of data to come in the very near future (especially with electrocardiogram
[EKG or ECG] apps, implantable monitors, and biosensors), pediatric cardiology remains
relatively dormant in this burgeoning domain of AI.[1] A comprehensive review by Johnson
et al. discusses the many dimensions how AI can affect cardiology: from research to
clinical practice and even population health.[2] This review also well delineated
the basics of machine learning (supervised/unsupervised learning and even the abstruse
concept of reinforcement learning). Another excellent and concise review by Shameer
et al. focuses on the aspects of cardiovascular medicine in the context of machine
learning.[3] Finally, there is an excellent review that showed how effective AI can
be in the context of precision cardiovascular medicine, an aspect of cardiology that
will be the cornerstone of pediatric cardiology in the near future.[4]
First, there is relatively increased use of AI and data science in the domain of cardiac
imaging ranging from ECG to echocardiography for both diagnosis and prognosis. The
major contribution to medical image interpretation has been CNN [Figure 1]. For signal
modalities, an arrhythmia detection end-to-end deep-learning strategy (accomplished
with a 34-layer CNN) with a single-lead ECG (with >90,000 single-lead ECGs from over
50,000 patients with 12 rhythm disturbance categories such as atrial fibrillation
and ventricular tachycardia) accomplished interpretation at a cardiologist level.[5]
While a prior study described work on automated quantification in echocardiography
(chamber quantification for left ventricular [LV] function and assessment of valvular
disease),[6] more recent works delineated how CNN was used for automated real-time
standard view classification and image segmentation to improve workflow.[7
8] In addition, machine learning has been applied to the ever so challenging (1) heart
failure patients with preserved ejection fraction and helped to set up a new phenotypic
risk assessment system for heart failure[9] and also (2) patients with either hypertrophic
cardiomyopathy or athletes LV hypertrophy based on expert-annotated, speckle-tracking
of echocardiograms.[10] One study even described using NLP for large-scale, automated,
and accurate extraction of structured, semi-structured, and unstructured data from
echocardiography reports.[11] Deep-learning algorithms have also been applied to cardiac
magnetic resonance imaging (MRI) as a prognosis prediction tool in patients with pulmonary
hypertension and shown to be superior to clinicians’ assessment.[12] In short, while
ECGs and static images such as cardiac MRI are relatively straightforward for machine
and deep learning, more complicated cardiac imaging such as echocardiograms as well
as three-dimensional and four-dimensional images are also being studied with machine
and deep learning. There is even a future role for hybrid imaging: combination of
several modalities of imaging to maximize the advantages of each modality.[13]
Figure 1
Convolutional neural network, a type of deep learning, and its many layers for medical
image interpretation (courtesy of Tobias Heimann of Siemens engineering).
There are also published reports of using AI for clinical decision support in various
settings in both adult and pediatric cardiology. Decision-making in cardiology is
often complex and is particularly vulnerable to many heuristics and biases.[14] A
deep-learning algorithm (called deep learning-based early warning system [DEWS]) that
used only four vital signs had a high sensitivity and a low false-positive rate for
the detection of patients with cardiac arrest in a multicenter study.[15] A machine
learning in the form of support vector machine devised an effective risk calculator
that was shown to be superior (less-recommended drug therapy with less adverse events)
to the existing accepted American College of Cardiology (ACC)/American Heart Association
(AHA) cardiovascular disease risk calculator.[16]
Overall, there are relatively few reports of AI-focused publications in children and
adults with congenital heart disease. An AI-assisted auscultation algorithm that performed
well in a virtual clinical trial may be difficult to become a routine approach given
readily available echocardiographic assessments; in short, an AI strategy for an older
technology (the stethoscope) may or may not be adopted by clinicians.[17] Machine
learning algorithms were deployed to train a large dataset of adults with congenital
heart disease to prognosticate and to facilitate management.[18
19] A similar study to the aforementioned DEWS study revealed that predictive models
created by AI can lead to earlier detection of patients at risk for clinical deterioration
and thereby improve care for pediatric patients in the pediatric cardiac intensive
care setting.[20] In addition, four AI-based algorithms were employed to facilitate
a clinical decision support system for estimating risk in congenital heart surgery.[21]
One innovative report described using machine learning and system modeling to facilitate
a multicentric collaborative learning project for rapid structured fact-finding and
dissemination of expertise; this forward-thinking approach can provide a complement
(and perhaps render less necessary to) the traditional multi-center, randomized clinical
trials that are sometimes challenging to execute.[22]
As cardiology is both a perceptual or image-intensive field and a cognitive or decision-making
subspecialty, AI is a particularly valuable technology for cardiology with potentially
very rich dividends that are vastly underexplored at present but has great promise.
Given the complexity of operative and interventional procedures in pediatric cardiology,
perhaps it will be decades before a robot will be performing a cardiac procedure in
its entirety. Nevertheless, AI is a much-needed resource as the cardiovascular disease
burden remains singularly the largest and continues to climb in an aging population
worldwide.
Overall, the cardiologist can be liberated from the long list of relatively mundane
tasks to higher level of medical decision-making with full deployment of the various
AI tools available. In the complex decision-making area, deep reinforcement learning
and cognitive architecture can be particularly useful for the ever-increasingly complex
nature of diagnostic and therapeutic precision cardiovascular medicine in both the
intensive care (“precision intensive care”) or hospital setting and the outpatient
arena. This type of individualized medicine will need the many layers of data and
information, all integrated into an AI-enabled strategy for the delivery of key information
for knowledge and treatment. With more research and development, more sophisticated
robotic procedures in the interventional catheterization laboratory, as well as in
the operating room, will reach clinical use. The administrative aspects of a heart
program can be better managed with some of the robotic process automation tools that
are already available. Overall, all of these AI efforts promulgate a new paradigm
in cardiovascular medicine of machine intelligence-derived phenotype profiling using
multiple sources of data to eventually derive new clinical insights.
It is more than 100 years since the Flexner report that shaped our present medical
school education strategy. With the advent of the aforementioned emerging technologies,
it is now more important than ever before to reassess our relatively banal educational
and training strategy in pediatric cardiology and cardiac surgery. The advent of AI
(and other emerging technologies) is a precious gift from our technological colleagues
and is the naissance of a special epoch in medicine. While AI is not necessarily going
to replace clinicians, it should be part of every medical student's educational curriculum
as well as every physician's clinical portfolio from this point forward.
So is AI in pediatric cardiology and cardiac surgery an irrational hype or a paradigm
shift? I have had the privilege of discussing the possibility of AI in pediatric cardiology
and cardiac surgery for the past three decades (even named a specialized cardiology
supercomputer “Leo” after Leonardo da Vinci, who first described a congenital heart
defect as well as a coronary artery lesion in the 1500s). The more than a dozen technologies
described in the aforementioned scenario are all now available; pediatric cardiology
is the singularly best subspecialty to benefit the most from the portfolio of AI methodologies
and emerging technologies.[23] My long-awaited arrival of AI for our field has finally
arrived; we cannot passively wait for the future, but we will need to create this
future paradigm shift for our younger colleagues and our beloved children and adults
with heart disease.