Documente Academic
Documente Profesional
Documente Cultură
I
n the era of evidence-based medicine, a applied successfully to big data problems in vari-
vast amount of information is collected on ous sectors, with applications including speech
patients.1,2 This information has become recognition and search engine optimization.3 In
increasingly useful in guiding treatment and opti- medicine, the IBM Watson Health (International
mizing clinical outcomes in medical care. The Business Machines Corp., Armonk, N.Y.) cogni-
result is an ever-expanding volume of data con- tive computing system has used machine learning
taining complex patterns that may extend beyond approaches to create a decision support system
the physician’s ability to use traditional data pro- for physicians treating cancer patients, with the
cessing techniques such as regression and multi- intention of improving diagnostic accuracy and
variate analysis for interpretation.1,2 As innovators, reducing costs. Initially trained at Memorial Sloan
plastic surgeons must then adapt to the grow- Kettering Cancer Center using large volumes of
ing trend of “big data,” and find ways to tap its patient cases and over 1 million scholarly arti-
resources to deliver more efficient health care cles, the project now has 14 participating cancer
and improved surgical outcomes. centers.4,5 All of these centers contribute to an
The answer may lie in “machine learning.” ever-expanding corpus of information that helps
A subfield of artificial intelligence, machine
learning involves generating algorithms capa-
ble of knowledge acquisition through historical Disclosure: None of the authors has a financial
examples. Machine learning has already been interest in any of the products, devices, drugs or
procedures mentioned in this article.
From the Division of Plastic and Reconstructive Surgery,
Faculty of Medicine, McGill University; the Division of
Plastic and Reconstructive Surgery, Harvard University, the Supplemental digital content is available for
Division of Plastic and Reconstructive Surgery, Beth Israel this article. Direct URL citations appear in the
Deaconess Medical Center; and the Department of Biological text; simply type the URL address into any Web
Sciences, University at Albany. browser to access this content. Clickable links
Received for publication May 22, 2015; accepted December to the material are provided in the HTML text
22, 2015. of this article on the Journal’s Web site (www.
Copyright © 2016 by the American Society of Plastic Surgeons PRSJournal.com).
DOI: 10.1097/PRS.0000000000002088
890e www.PRSJournal.com
Copyright © 2016 American Society of Plastic Surgeons. Unauthorized reproduction of this article is prohibited.
Volume 137, Number 5 • Big Data and Machine Learning
Watson fine tune its ability to suggest optimal based on Current Procedural Terminology codes.8
treatment options for cancer patients based on Using data from 2005 to 2009, the support vector
the nature of their specific illness.4,5 machine was trained to determine the association
Although similar to data mining, which tradi- between Current Procedural Terminology and
tionally involves knowledge acquisition through mortality, morbidity, Clavien classification type
analysis of preexisting data sets, machine learning IV complications, and surgical-site infection to
places a greater emphasis on descriptive modeling produce an algorithm capable of generating pro-
and outcome prediction for novel data.5 Further- cedural risk scores.8 When tested using National
more, machine learning algorithms are capable Surgical Quality Improvement Program data
of improving or “learning” when exposed to more from 2010, the support vector machine approach
information.5,6 As the algorithm attempts to find achieved a greater level of discrimination for
the most appropriate hypotheses for a given data determining surgical complications compared
set, within the computational boundaries of the spe- with other measures of procedural complexity.8
cific machine learning approach used, it statistically Descriptive models, in contrast, fall under the
assesses how each model compares to each other category of “unsupervised learning.” Unsupervised
and models that have been assessed previously.6 The learning analyzes data that are unlabeled, and the
result of this process is the creation of data models system discovers structure in the data for interpre-
that are either predictive or descriptive in nature. tation (Fig. 2).6,9 (See Video, Supplemental Digital
Predictive machine learning models fall Content 1, which highlights the difference between
under the domain of supervised learning, where supervised and unsupervised machine learning
the algorithm has been trained using examples using hypothetical machine learning algorithms
of both inputs and desired outputs, allowing for capable of processing visual data for the detection
mapping of future inputs to outputs.6,7 The goal and differentiation of different types of craniosyn-
of this process is a unique mathematical model ostosis. Available in the “Related Videos” section of
capable of predicting desired target values from the full-text article on PRSJournal.com or, for Ovid
novel data (Fig. 1). For example, a recent surgi- users, available at http://links.lww.com/PRS/B706.)
cal application involved the use of data from the This type of machine learning has been applied
American College of Surgeons National Surgical in the field of molecular genetics and genomics
Quality Improvement Program to train a support to organize and interpret vast amounts of genetic
vector machine to quantify procedural complex- information.10 Following the application of leu-
ity and risk associated with different procedures kemic blasts from pediatric acute lymphoblastic
Fig. 1. Graphic representation of supervised machine learning. In supervised learning, original prepro-
cessed data sets, containing known variables and targets, are divided into training data and test data.
(Above) During the training phase, the training data are used to train a learning algorithm in an attempt
to develop an accurate predictive model. (Center) To validate the model, the test data are then applied
to the model and predictive accuracy is assessed. (Below) Once validated, new data are input into the
model in an attempt to make new predictions.
891e
Copyright © 2016 American Society of Plastic Surgeons. Unauthorized reproduction of this article is prohibited.
Plastic and Reconstructive Surgery • May 2016
Fig. 2. Graphic representation of unsupervised machine learning. In unsupervised learning, raw data,
containing unknown patterns and targets, are presented to an algorithm. The algorithm attempts to
develop descriptive models for the data based on regularities detected. (Adapted from Hudson Legal.
Unsupervised learning. Available at: http://us.hudson.com/portals/US/images/blogs/legal/wp/2011/09/
Unsupervised-Learning1.jpg. Accessed October 6, 2014.)
892e
Copyright © 2016 American Society of Plastic Surgeons. Unauthorized reproduction of this article is prohibited.
Volume 137, Number 5 • Big Data and Machine Learning
serving as a proxy for the assessment of burn interobserver variability. By pairing images of
depth for surgical planning. Artificial neural net- burns to precise measurements of the percent-
works consist of input nodes, representing the age of body area affected, an algorithm could be
data to be used for prediction, intermediate or trained to rapidly and accurately predict the per-
“hidden nodes” that calculate predictions based centage of burned tissue (Fig. 4). From these mea-
on the inputs, and output nodes that represent surements, more accurate resuscitation protocols
the predictions themselves.6 During training, arti- could be generated in addition to surgical plan-
ficial neural networks are tuned through a pro- ning strategies for autografting or allografting.
cess of “back-propagation” where the accuracy of
the output values is compared to the actual target Microsurgery
values (Fig. 3).6 In this investigation, normalized Postoperative monitoring after microsurgery is
spectral data served as input nodes, whereas the crucial for achieving desired clinical outcomes. In
two output nodes distinguished between spec- light of this, researchers have recently developed
tra associated with burns that healed in less than a postoperative microsurgery monitoring applica-
14 days and those associated with burns that took tion, the SilpaRamanitor, capable of quantifying
more than 14 days to heal.11 After examining free-flap tissue perfusion.12 Using 60 smartphone
reflectance spectrometry data from 41 wounds, images of middle and index fingers exposed to
the investigators determined that their model had varying degrees of vascular compromise to mimic
an average predictive accuracy of 86 percent, sug- vascular occlusion, a k-nearest neighbor algorithm
gesting that it may serve as an effective screening was trained to categorize flap tissue into different
tool for assessing burn depth and a superior alter- classes: normal, venous outflow occlusion, and
native to direct visualization.11 arterial inflow occlusion.12 In cases of occlusion,
Another task within burn care that might lend the degree was further categorized as partial or
itself to machine learning is the accurate and pre- complete. The k-nearest neighbor algorithm is
cise quantification of burn size (total body sur- a non–parameter supervised learning approach
face area). Current methods, such as the “rule whereby the classification rules are generated
of nines,” are limited by the asymmetry of burns, by the training samples themselves and do not
surface area variations related to patient age, and require the input of additional information.13 The
Fig. 3. Graphic representation of an artificial neural network. Modeled after biological neural
networks, artificial neural networks use input nodes, representing data input into the model;
hidden nodes, responsible for making the predictions); and output nodes, representative of
the predictions being made (Oncologists partner with Watson on genomics. Cancer Discov.
2015;5:788). During training, artificial neural networks, in a fashion similar to biological neurons,
take part in a process called back-propagation, whereby the weight of the connections between
nodes is adjusted based on the difference between the artificial neural networks output values
and known target values. This process ensures that the output of the artificial neural network is
as close as possible to the desired target values. (Adapted from Meyfroidt G, Güiza F, Ramon J,
Bruynooghe M. Machine learning techniques to examine large patient databases. Best Pract Res
Clin Anaesthesiol. 2009;23:127–143.)
893e
Copyright © 2016 American Society of Plastic Surgeons. Unauthorized reproduction of this article is prohibited.
Plastic and Reconstructive Surgery • May 2016
algorithm assigns a test sample, which in this study craniosynostoses, a regularized linear discrimi-
was an image of a free flap, to a category, which nant analysis algorithm was trained to diagnose
was the type and degree of occlusion.12 This classi- craniosynostosis and distinguish between differ-
fication is based on features of the test sample that ent types using an index of cranial suture fusion
are the most similar, or the “nearest neighbors,” to along with deformation and curvature discrep-
those from the training set.12 The overall sensitiv- ancy averages across five cranial bones and six
ity and specificity of this application were found suture regions.14 Regularized linear discriminant
to be 94 percent and 98 percent, respectively, analysis is frequently used for high-dimensional
with an accuracy of 95 percent.12 Thus, through data when there are a small number of samples,
the accurate and rapid monitoring of free-flap as was the case in this investigation.15 The result
perfusion, the SilpaRamanitor application is an of this machine learning process is an automated
example of how machine learning can be used to classifier capable of differentiating types of cra-
potentially increase the success of detecting early niosynostosis based on computed tomographic
anastomotic failure or thrombotic issues and con- scans, with a sensitivity of 92.3 percent and a speci-
comitant free-flap salvage.12 ficity of 98.9 percent.14 With accuracy comparable
However, the spectrum of potential applica- to trained radiologists, the authors propose that
tions of machine learning for microsurgery are their algorithm may provide an objective standard
not only limited to postoperative monitoring. capable of reducing interobserver variability and
Machine learning could also benefit preoperative providing a quantitative measure of procedural
consultation and surgical planning for microsur- success.14
gery. Through the collection of detailed infor- Although the automation of this process has
mation such as the size and location of defects, many benefits, we propose that machine learn-
the type of flap used, patient age, body mass ing theoretically has the potential to enable
index, smoking status, and resultant complica- surgeons to bypass the use of computed tomo-
tions in large-scale databases (such as Tracking graphic imaging for routine diagnostics. Using
Operations and Outcomes for Plastic Surgeons), three-dimensional surface photographs of differ-
machine learning algorithms could be trained ent plagiocephalies, an algorithm could be devel-
to assess a particular defect in a selected patient oped to differentiate between cases of synostotic
and suggest the reconstructive approach with the and deformational plagiocephaly. Together with
highest chance of a favorable outcome. clinical examination, the goal would be to further
reduce the need for ionizing radiation in these
Craniofacial Surgery children.
Currently, machine learning is being explored Another potential machine learning applica-
to facilitate the automated diagnosis of non- tion for craniofacial surgery involves the identifi-
syndromic craniosynostosis. Examining com- cation of candidate genes in nonsyndromic cases
puted tomographic scans from 141 subjects, of of cleft lip and palate.16 Through a combination
which 50 had either sagittal, metopic, or coronal of genomewide association studies and other
894e
Copyright © 2016 American Society of Plastic Surgeons. Unauthorized reproduction of this article is prohibited.
Volume 137, Number 5 • Big Data and Machine Learning
investigations, several genetic factors have been For example, using information from previous
elucidated for this condition for which the cause is cases and radiographic images of fractures, an
poorly understood.16 However, our understanding algorithm could be developed to anticipate the
of the molecular pathogenesis of nonsyndromic positioning of Kirschner wires, plates, and screws
cleft lip and palate remains far from comprehen- in preoperative planning for hand surgery. Fur-
sive. Using methods similar to those used previ- thermore, using databases containing informa-
ously in the field of genomics, machine learning tion derived from sensorial mapping following
has the potential to uncover previously unknown peripheral nerve repair, patterns of regrowth
candidate genes and regulatory sequences for could be used to develop an algorithm capable of
nonsyndromic cleft lip and palate, allowing for an prognosticating the degree of sensory and motor
improved understanding of the pathogenesis of restoration based on location, mechanism of
this condition.16 nerve injury, and physical examination findings.
Hand and Peripheral Nerve Surgery Aesthetic Surgery
Research in the field of hand and peripheral Machine learning also has potential applica-
nerve surgery has the potential to benefit from tions in more subjective areas of plastic surgery,
machine learning. As an example, investigators such as aesthetics. Using a form of supervised
have recently developed an artificial neural net- learning, an automated classifier for facial beauty
work capable of predicting the outcome of differ- was trained using extracted facial features from
ent tissue-engineered peripheral nerve grafts used 165 images of attractive female faces that were
in research applications.17 Using over 30 inde- also independently graded by human referees.20
pendent variables to describe tissue-engineering Decision tree algorithms assess a set of descrip-
materials, artificial neural networks were trained tive attributes, which in this particular investiga-
to predict the success of various grafts.17 The suc- tion included different facial ratios, and attempt
cess of the grafts, placed in a rat model, was quan- to determine attractive facial features most
tified using the critical regeneration length, along
closely related to postoperative target variables—
with a unitless parameter, the ratio of the actual
the human classification of facial beauty.20 When
length to the critical length.17 After application
subjected to the testing set of images, the auto-
of the validation data, the predicative accuracy of
mated classifier was shown to have a high accu-
artificial neural networks was 92.59 percent and
racy at approximating human referee scores.20 In
90.85 percent for the ratio of the actual length to
the critical length and the critical regeneration light of these results, this application may serve
length, respectively.17 Although preliminary, the as a predictive tool for estimating a patient’s per-
results of this investigation highlight the poten- ceived beauty following aesthetic surgery, provid-
tial role for machine learning in the analysis and ing a quantitative measure to set expectations
development of tissue-engineering strategies for and possibly discourage patients from undergo-
peripheral nerve repair. ing procedures that offer marginal improvement.
The biomechanics of the upper extremities In conjunction with optical head-mounted
are particularly complex, involving multivari- display technology, machine learning also has
ate nonlinear relationships that are theoretically the potential to facilitate intraoperative visual-
amenable to modeling by machine learning. In ization of surgical outcomes. One such applica-
light of this, artificial neural networks have been tion could be in reconstructive breast surgery,
used to develop automated controllers for a vari- whereby machine learning software incorporated
ety of neuroprostheses, including those that are into a head mount display could predict what a
used to restore hand grasp and wrist control breast might appear like in three-dimensional
along with more proximal upper extremity func- space based on potential changes in implant posi-
tion in patients with C5/C6 spinal cord injury.18,19 tion. The system could be trained to identify fea-
Although the results of these initial investigations tures of breast aesthetics such as symmetry, nipple
were mixed, they highlight the potential for arti- position, superior pole fullness, and degree of
ficial neural networks, along with other machine ptosis, optimizing aesthetic results while mini-
learning techniques, in the development of neu- mizing trauma and operating time. Although
roprosthetic controllers for the hand and wrist. these potential tools may not replace the trained
Along with the examples presented above, human eye, they would aid the plastic surgeon by
machine learning has the potential to provide providing a higher degree of objectivity to aes-
additional innovation in hand and nerve surgery. thetic surgery.
895e
Copyright © 2016 American Society of Plastic Surgeons. Unauthorized reproduction of this article is prohibited.
Plastic and Reconstructive Surgery • May 2016
896e
Copyright © 2016 American Society of Plastic Surgeons. Unauthorized reproduction of this article is prohibited.
Volume 137, Number 5 • Big Data and Machine Learning
897e
Copyright © 2016 American Society of Plastic Surgeons. Unauthorized reproduction of this article is prohibited.