[ Contemporary Reviews in Critical Care Medicine ]

Big Data and Data Science in Critical Care

L. Nelson Sanchez-Pinto, MD; Yuan Luo, PhD; and Matthew M. Churpek, MD, PhD

The digitalization of the health-care system has resulted in a deluge of clinical big data and has
prompted the rapid growth of data science in medicine. Data science, which is the field of study
dedicated to the principled extraction of knowledge from complex data, is particularly relevant
in the critical care setting. The availability of large amounts of data in the ICU, the need for
better evidence-based care, and the complexity of critical illness makes the use of data science
techniques and data-driven research particularly appealing to intensivists. Despite the
increasing number of studies and publications in the field, thus far there have been few
examples of data science projects that have resulted in successful implementations of data-
driven systems in the ICU. However, given the expected growth in the field, intensivists
should be familiar with the opportunities and challenges of big data and data science. The
present article reviews the definitions, types of algorithms, applications, challenges, and future
of big data and data science in critical care. CHEST 2018; 154(5):1239-1248

KEY WORDS: big data; critical care; data science; machine learning; prediction models

The digitalization of the health-care system is clinical realities of the ICU.5 Critical care
changing the way we practice medicine and research requires an integrative approach that
conduct clinical research.1,2 The widespread embraces the complexity of critical illness and
implementation of electronic health records the computational technology and algorithms
(EHRs) is paving the way for big data research that can make it possible.7,8 Moreover, the
and is bringing the world of data science to the data required to follow this approach are
patient’s bedside.2-4 Within the health-care being generated and digitized in troves. EHRs,
system, the ICU presents a particularly bedside monitors, medication pumps, and
convincing case for using data science to ventilators are continuously generating new
improve patient care.5 The evidence minable data, and soon the advancement of
supporting many of the interventions modern molecular diagnostics will result in a
performed in the ICU is scarce, and practice deluge of “omics” data derived from the
variability is abundant.5,6 In addition, the genome, transcriptome, microbiome, and a
complexity of critical illness makes the long list of other “-omes” (Fig 1).
traditional reductionist approach to medical
research insufficient; that is, single-drug As big data and data science gradually
intervention trials or single pathway infiltrate most aspects of clinical research
biomarker studies are unlikely to satisfy the and, ultimately, clinical care in the ICU, it is

ABBREVIATIONS: AUC = area under the receiver-operating charac- Department of Medicine (Dr Churpek), The University of Chicago,
teristic curve; EHR = electronic health record; MIMIC = Multipa- Chicago, IL.
rameter Intelligent Monitoring in Intensive Care; NLP = natural CORRESPONDENCE TO: Matthew M. Churpek, MD, PhD, Pulmonary
language processing and Critical Care, University of Chicago, 5841 S Maryland Ave, MC
AFFILIATIONS: From the Department of Pediatrics (Critical Care) (Dr 6023, Chicago, IL 60637; e-mail: matthew.churpek@uchospitals.edu
Sanchez-Pinto) and the Department of Preventive Medicine (Health Copyright Ó 2018 American College of Chest Physicians. Published by
and Biomedical Informatics) (Drs Luo and Sanchez-Pinto), North- Elsevier Inc. All rights reserved.
western University Feinberg School of Medicine, Chicago, IL; and the DOI: https://doi.org/10.1016/j.chest.2018.04.037

chestjournal.org 1239
 Big Data in the ICU

Patient information Contextual Diagnoses Interventions Cellular/Molecular Imaging Natural language Physiologic

Demographics Encounter Billing diagnoses Medications Routine laboratory Radiology Clinician notes Clinician-charted
information
Social Physician-curated Procedures “Omics” data Pathology Diagnostic reports Device-generated
ICU & hospital
diagnoses
Family operations Organ support Patient Video recording Patient narratives Monitors
Data-driven
Financial Environmental phenotypes Other therapies Microbial Photographs Speech recording Wearables

Generally available in the electronic health record (EHR) Not available or limited information in the EHR

Figure 1 – Some of the major sources of big data in the ICU. The term “omics” refers to the data derived from modern molecular techniques
(eg, genomics, transcriptomics, proteomics, metabolomics, microbiomics). EHR ¼ electronic health record.

increasingly evident that intensivists should be familiar These and other pertinent definitions in data science are
with the promise and perils of these approaches. The presented in Table 1.
present article reviews the definitions, types of
algorithms, applications, challenges, and future of big
data and data science in critical care. The Ecosystem of Data Science in Health Care
The data revolution in health care would not be possible
if it were not for several key developments, including: (1)
Definitions in Data Science the data science movement that has transformed other
Big data can be defined as digital data that are generated industries; (2) the extraordinary growth in
in high volume and high variety and that accumulate at computational power; (3) the availability of open source
high velocity, resulting in datasets too large for tools and low-cost equipment to perform advanced
traditional data-processing systems.2,9 In practice, big analyses; and (4) the increasing availability of
data in health care depend on both the breadth and the educational resources and advanced degrees in data
depth of the data being captured. For example, science and related fields.12 Open source programming
administrative health-care datasets with few data and scripting languages, such as R and Python, have
elements per patient record (low depth) are usually extensive libraries of statistical packages and machine
considered big data problems when they contain learning algorithms that are relatively easy to use by
millions of records (wide breadth). Conversely, when researchers with some training and have greatly
applying next-generation sequencing and other “-omics” democratized the access to data science techniques.
approaches (high depth), just a few dozen patients can
Educational resources, such as graduate programs in
become a big data problem (narrow breadth).10
data science, are also increasingly available. Notable
Data science can be defined as “the set of fundamental among these are massive open online courses, which
principles that support and guide the principled extraction have led to the increased popularity of data science and
of information and knowledge from data.”9 A closely the applications of machine learning techniques to real-
related term is data mining, which is the actual extraction world problems.13 There are now a multitude of courses
of knowledge from data via machine learning algorithms from prestigious institutions that cover data science and
that incorporate data science principles. Machine learning machine learning, and many of them are free. These
is the field of study that focuses on how computers learn courses provide the interested researcher with
from data and the development of algorithms that make educational opportunities from world-class experts at
this learning possible.11 Finally, another important the touch of a button. Other websites contain user-
concept in data science is domain expertise, which in created code to run machine learning algorithms from
health care can be defined as the understanding of real- scratch (eg, https://github.com/) or host data science
world clinical problems and the realities of patient care competitions for participants around the world (eg,
that help frame and contextualize the application of data https://www.kaggle.com/). This rich online environment
science to health-care problems.8,9,11 is ideal for data scientists to learn and grow, with the

1240 Contemporary Reviews in Critical Care Medicine [ 154#5 CHEST NOVEMBER 2018 ]
TABLE 1 ] Definitions of Common Terms in Data Science
Term Definition
Big data Digital data that are generated in high volume and high variety and that accumulate at high
velocity, resulting in datasets too large for traditional data-processing systems
Data science The set of fundamental principles that support and guide the principled extraction of information
and knowledge from data
Data mining The extraction of knowledge from data via machine learning algorithms that incorporate data
science principles
Domain expertise The understanding of real-world problems in a given domain (eg, critical care medicine) that helps
frame and contextualize the application of data science to solve these problems
Machine learning The field of study that focuses on how computers learn from data and the development of
algorithms that make this learning possible
Features The data elements, also known as independent variables, used to train a model. Features can be
simple transformations of the raw data (eg, average heart rate in the last 24 h) or complex
transformation such as the ones performed by neural networks (see Table 2)
Outcomes The data elements, also known as dependent variables, represent the target for training in a
supervised learning model. Outcomes can be categorical (eg, yes/no) or continuous (eg, length
of hospital stay). Categorical binary outcomes are the most common in medicine (eg, died or
alive by 28 days). Binary outcomes are typically represented as a Boolean logic (ie, true/false or
1/0) but can also be represented using fuzzy logic (ie, a range of probabilities, or degrees of
truth, between 0 and 1)
Supervised learning Algorithms that are used to uncover the relationship between a set of features and one or more
known outcomes
Unsupervised learning Algorithms that are used to uncover naturally occurring patterns or groupings in the data, without
targeting a specific outcome
Model training The process through which machine learning algorithms develop a model of the data by learning
the relationships between features and, in supervised learning, between features and
outcomes. This is also referred to as model derivation or data fitting
Model validation The process of measuring how well a model fits new, independent data. For example, evaluating
the performance of a supervised model at predicting an outcome in new data. This approach is
also referred to as model testing.
Predictive model A model generally trained to predict the likelihood of a condition, event, or response. The US Food
and Drug Administration specifically considers predictive strategies as those geared toward
identifying groups of patients more likely to respond to an intervention
Prognostic model A model specifically trained to predict the likelihood of a condition-related endpoint or outcome
such as mortality. In general, the goal is to estimate a prognosis given a set of baseline features,
regardless of what ultimately leads to the outcome
Overfitting The phenomenon that occurs when an algorithm learns from idiosyncrasies in the training data,
usually referred to as noise. Noisy data are data that are randomly present in the training
dataset but do not represent the generalizable truth (usually referred to as signal) that explains
the relationships between the features and the outcomes. Overfitting will generally lead to poor
performance of the model in an independent validation dataset
Digitization The conversion of something analog or physical (eg, paper documents, printed images) into a
digital format (ie, bits or 1s and 0s)
Digitalization The wide adoption of digital technologies by an organization to leverage their digitized data with
the goal of improving operations and performance. The adoption of electronic health records
and other digital technologies (eg, picture archiving and communication systems for medical
images, pharmacy management systems, billing systems) are examples of digitalization in
health care
Data curation The process of integrating data from different sources, structuring it, authenticating it, and
annotating it to ensure its quality, add value, and facilitate its use and reuse
Structured data Data (usually discrete or numeric) that are easy to search, summarize, sort, and quantify.
Examples include vital signs (eg, heart rate) or laboratory test results (eg, CBC)
Unstructured data Data that do not conform to a prespecified structure, such as a written narrative, images, video, or
audio. Unstructured data are generally harder to search, sort, and quantify. Examples include
clinician notes, pathology slides, and radiology images

chestjournal.org 1241
field of medicine benefiting from advances shared A
around the world freely through the Internet. Patients Clinical features Outcomes

Pt. 1 Survived
Types of Algorithms in Data Science Pt. 2 Not survived
Machine learning algorithms are generally divided into Pt. 3 Survived
two categories: supervised and unsupervised.11 Semi- ...
LEARN
supervised algorithms represent a hybrid of the two but
have been used less often in health-care problems. PREDICT
Finally, deep learning algorithms defy this classification,
New Pt. ?
even though they derive from artificial neural network
algorithms, which are generally classified as supervised
algorithms. The most defining characteristic of deep B
Heterogeneous
learning is their focus on learning data representations population Cluster 1:
• High Oxygenation Index
(or features) that can then be used in supervised, • Low Glasgow Coma Scale
• High mortality risk
unsupervised, or semi-supervised problems. The
following discussions review these types of algorithms in
more detail (as well as in Fig 2 and Table 2). Cluster 2:
• On inotropes
• Acute kidney injury
Supervised Learning Algorithms • High mortality risk

Supervised learning algorithms are used to uncover the

relationship between variables of interest and one or more Cluster 3:
• No comorbidities
target outcomes.11,14 For supervised problems, the target • Normal respiratory rate
• Low mortality risk
outcome(s) must be known. For example, if researchers
want to know whether a set of clinical features (eg, vital
signs, laboratory tests) can predict ICU mortality, they C
Hidden Layers
could apply a supervised learning algorithm to a dataset in
which each patient record contains the set of clinical Input Layer Output Layer
features of interest and a label specifying their outcome
Consolidation
(“survived” or “not survived” in this case) (Table 1).
Examples of supervised learning algorithms include Effusion

regression-based methods (eg, linear and logistic Atelectasis

Chest Findings
regression, lasso, elastic net), tree-based methods (eg, X-ray
classification and regression trees, random forest, gradient Shapes,
Densities
boosted trees), k-nearest neighbor, artificial neural Structures, Lungs, Heart,
networks, and support vector machines (Table 2).15 Background Bones, Fluid

Figure 2 – A-C, Types of machine learning algorithms applicable to

Unsupervised Learning Algorithms critical care. A, Supervised learning algorithms can be used, for example,
to uncover the relationship between patient clinical features (eg, labora-
Unsupervised learning algorithms are used to uncover tory tests and vital signs) and mortality to predict the outcome in future
naturally occurring patterns or groupings in the data, cases. B, Unsupervised learning algorithms can be used to uncover
naturally occurring groupings or clusters of patients based on their clinical
without targeting a specific outcome.11 The most characteristics, without targeting a specific outcome. C, Deep learning
compelling use case of unsupervised learning in health algorithms can be used, for example, to extract meaningful features from
care is in precision medicine, in which the goal is to imaging data (eg, chest radiograph) to represent information in an
increasingly higher order of hierarchical complexity and be able to make
uncover subsets of patients who share similar clinical or predictions, such as the presence of pathologic findings.
molecular characteristics and are, in theory, more likely
to respond to targeted therapies directed at their shared include clustering algorithms (eg, hierarchical clustering,
underlying pathobiology.16,17 For example, an k-means clustering), latent class analysis, and principal
unsupervised learning algorithm may be used to uncover component analysis (Table 2).11,14,19
subgroups of patients with sepsis who have distinct
molecular and clinical characteristics and will respond Deep Learning Algorithms
differently to specific drugs, such as corticosteroids.18 Deep learning algorithms are designed to extract
Some examples of unsupervised learning algorithms meaningful features from the data to represent

1242 Contemporary Reviews in Critical Care Medicine [ 154#5 CHEST NOVEMBER 2018 ]
TABLE 2 ] Examples of Algorithms Use in Data Science
Algorithm Class Examples Description
Classic regression Linear regression, Linear regression is a supervised learning algorithm that models the
logistic regression relationship between one or more features and a continuous outcome
by fitting a regression line that minimizes the sum of all the residuals,
which are the distances between each feature in the training data and
the line being fitted to model them. Logistic regression is a
generalization of the linear model that uses the logistic function to
estimate the probability of a binary outcome. To do this, the fitted
sigmoid-shaped curve of the logistic function maps the feature values
into a probability between 0 and 1
Regularized Lasso, ridge An extension of the classic regression algorithms in which a penalty is
regression regression, elastic imposed to the fitted model to reduce its complexity and decrease the
net risk of overfitting (see Table 1).
Tree-based Classification and A class of supervised learning algorithm based on decision trees.
regression trees, Decision trees are a sequence of “if-then-else” splits that are derived
random forest, by iteratively separating the data into groups based on the
gradient boosted relationship of the features with the outcome. Random forest and
trees gradient boosted trees are example of ensemble tree models.
Ensemble models combine the output of many trained models to
estimate an outcome
Support vector Linear, polynomial, A class of supervised learning algorithms that represents the data in a
machines radial basis kernel multidimensional feature space and then fits a “hyperplane” that best
separates the data based on the outcomes of interest
K-nearest neighbor K-nearest neighbor A type of supervised learning algorithm that represents data in
multidimensional feature space and uses local information about
observations closest to a new example to predict the outcome for that
example
Bayesian Naive Bayes, Bayesian A class of supervised learning algorithms that use Bayes’ theorem of
network conditional probability, which is the probability that something will
happen given that something else has already occurred. In general,
Bayesian algorithms work by iteratively updating the probability of an
outcome (or posterior belief) given new data
Neural network Artificial neural A class of nonlinear algorithms built using layers of nodes that extract
network, deep features from the data and perform combinations that best represent
neural network the underlying structure, usually to predict an outcome. Neural
networks can be shallow (eg, a perceptron with two layers) or deep
(multiple layers), which form the basis for the field of deep learning
Dimensionality Principal component A class of unsupervised learning algorithms that exploit the inherent
reduction analysis, linear structure in the data to describe data using less information. Principal
algorithms discriminant components, for example, summarize a large set of correlated
analysis features into a smaller number of representative features
Latent class analysis Latent class analysis A type of unsupervised learning algorithm that identifies unseen
subgroups, or latent classes, in the data. Class membership is
unknown for each example so the probability of class membership is
indirectly estimated by measuring the patterns in the data
Cluster analysis K-means, hierarchical A class of unsupervised learning algorithm that uses the inherent
cluster analysis structures in the data to best organize the data into subgroups of
maximum commonality based on some distance measure between
features

information in an increasingly higher order of complex features. After deeper layers of nodes have
hierarchical complexity in the form of stacked layers of perceived increasingly more complex features in an
nodes (or “neurons”).20 For example, if the input is a unsupervised way, they can then be used to perform
photo of several people, the first layer of nodes might specific tasks, such as matching the faces in the photo to
simply extract straight lines, curves, and color hues. certain specific people with known features. In medical
Deeper layers may combine some of those lines, curves, applications, deep learning has been used, for example,
and hues to represent eyes, noses, ears, and other more to detect diabetic retinopathy on funduscopic images,21

chestjournal.org 1243
detect cancer in skin photographs,22 or predict clinical colleagues18 used a combination of a classification and
outcomes by using EHR data.23,24 regression tree-based biomarker risk model and gene
expression profiles in pediatric patients with sepsis to
Data Science Applications in Critical Care identify a subgroup of patients who were more likely to
benefit from corticosteroids.
Predictive and Prognostic Models
The most common applications of data science to Clustering and Phenotyping
critical care problems are predictive and prognostic Unsupervised learning algorithms in critical care have
models using supervised learning algorithms. Although mainly been used to uncover naturally occurring
identical from a modeling perspective, predictive and subgroups or clusters of patients who share similar
prognostic models can be distinguished semantically by clinical and/or molecular characteristics. These clusters
the fact that predictive models are generally trained to are oftentimes called phenotypes, subphenotypes, or
predict the likelihood of a condition, event, or response, subtypes, although there is still little consensus on the
whereas prognostic models are specifically trained to terminology.16 For example, Calfee and colleagues32
predict the likelihood of a condition-related endpoint or applied latent class analysis and identified two
outcome, such as mortality (Table 1).16,25 This subphenotypes of ARDS using clinical and cytokine data
distinction, however, is not always clear in the literature from two randomized controlled trials of ARDS. The
and, depending on the use case, might be irrelevant. subphenotypes identified had distinct differences in
inflammatory profiles, response to ventilator strategies,
One of the oldest and best-known prognostic models to
and clinical outcomes. Knox and colleagues33 used self-
estimate risk of mortality in ICU patients is the Acute
organizing maps and k-means clustering to identify four
Physiology and Chronic Health Evaluation score, which
distinct clusters of patients with sepsis-associated
was first developed in the 1980s by Knaus and
multiple organ dysfunction syndrome that were
colleagues26,27 using logistic regression. Since then, many
independently associated with outcomes after adjusting
other groups have developed predictive and prognostic
for severity of illness. Luo and colleagues34 analyzed
models using larger, more granular datasets and applying
multiple physiological variable trends of patients in the
modern machine learning methods. For example,
MIMIC dataset and applied nonnegative matrix
Churpek and colleagues28 developed a logistic regression
factorization to group-related trends, which were shown
model in a dataset of > 250,000 hospital admissions that
to effectively predict 30-day mortality while maintaining
accurately estimated the risk for ICU transfer, cardiac
model interpretability. Finally, Vranas and colleagues35
arrest, or death in ward patients. In a follow-up study, the
applied clustering analysis to discover and validate six
same group showed that more modern machine learning
clinically recognizable subgroups of ICU patients who
methods, such as random forests and gradient boosted
differed significantly in all baseline characteristics and
machines, could more accurately predict clinical
clinical trajectories despite sharing common diagnoses.
deterioration compared with classic logistic regression.15
In another example, Joshi and Szolovits29 used 54 clinical Applications With Nontraditional Data Types
variable time series to predict 30-day mortality in ICU
Natural language processing: Much of the data used in
patients in the publicly available Multiparameter
critical care studies, such as vital signs or laboratory test
Intelligent Monitoring in Intensive Care (MIMIC)
results, are structured data that can be easily entered into a
dataset. They clustered the physiological measurements
relational database or spreadsheet and be sorted and
into organ-specific patient states and achieved a state-of-
summarized. However, there is a significant amount of
the-art 30-day mortality prediction area under the
clinical information contained in the form of
receiver-operating characteristic curve (AUC) of 0.91.
unstructured clinical narratives (eg, progress notes,
Predictive models aimed at identifying patients with discharge summaries, nursing notes, diagnostic
specific conditions or those more likely to respond to a reports).36 Methods for analyzing narrative data,
specific therapy are more commonly used in the field of generally known as natural language processing (NLP),
oncology, with multiple examples of biomarker-based are designed to extract features from texts that can then be
models used to diagnose particular subtypes of cancer that used in task-specific algorithms for different purposes (eg,
respond to targeted therapy.4,16 However, there are some prognostic modeling). Lehman and colleagues37 used
examples in the critical care literature, particularly in clinical data and unstructured progress notes from the
sepsis and septic shock.30,31 For example, Wong and first 24 h of ICU admissions to estimate the risk of

1244 Contemporary Reviews in Critical Care Medicine [ 154#5 CHEST NOVEMBER 2018 ]
in-hospital mortality. They inferred topic models from Challenges and Pitfalls of Data Science in
progress notes and achieved an AUC of 0.82, which was Critical Care
superior to severity of illness scores based only on Like most emerging technologies, the products of data
structured clinical variables. Ghassemi and colleagues38 science research in critical care will undoubtedly go
further investigated the prognostic power of topics as through a series of hype and disillusionment cycles before
features from the first 24 h and achieved an AUC of 0.85 becoming accepted, proven assets in the study and care of
for in-hospital mortality when combining text and critically ill patients. One of the first challenges that data
structured data. Weissman and colleagues39 applied NLP science faces in critical care is that, despite the increasing
to analyze discharge documents of ARDS survivors and number of studies and publications in the field, thus far
found that ARDS itself is rarely mentioned in those there have been few examples of data science projects
documents, as opposed to more frequent mentions of that have resulted in successful implementation of data-
“mechanical ventilation” and “ICU stay.” Conversely, driven systems in the ICU.11 This lack of exposure in the
their NLP-based document classifier reported clinical setting inevitably results in a degree of mistrust by
100% accuracy for ARDS identification, suggesting that clinicians in these data-driven systems.47,48 Although
NLP can be used to effectively identify patients with clinicians are happy to use similar systems to browse
certain types of conditions. their smart televisions, shop online, or interact with
Physiological waveform analysis: Physiological social media apps, they are wary of the idea of sharing
waveform data from bedside monitors and wearable clinical decision-making responsibilities with machine
devices are increasingly being used in data science learning algorithms, particularly if they view them as
studies in critical care. Many institutions collect and “black boxes.”48 It is likely that only the implementation
store physiological monitor data, such as of well-designed, interpretable, and effective data-driven
electrocardiography, photoplethysmography, impedance systems in the ICU will make clinicians start to gain trust
pneumography, invasive arterial manometry, end-tidal in them. Furthermore, the implementation of these data-
capnography, and electroencephalography. The publicly driven systems must be performed under the rigorous
available MIMIC databases contain physiological auspices of well-controlled experimental studies,
waveform data for ICU patients at Beth Israel Deaconess including (but not limited to) simulation testing,
Medical Center, which has facilitated the development of preintervention and postintervention studies, and
the state-of-the-art waveform analysis in the field.40 For randomized controlled trials. The medical informatics
example, researchers have used waveform data to literature has good examples of using scientifically
estimate cardiac output data using pulse contour rigorous approaches to the implementation and testing of
analysis techniques,41 detect hypovolemia using digital solutions such as clinical decision support tools
photoplethysmography data,42 and predict and can serve as a model to follow.49-51
hyperlactatemia using combined physiologicaldata.43
Clinicians and researchers appraising a data-driven
Image analysis: The advancement in the field of deep system and the literature that supports it must be aware of
learning, which is particularly useful for image analysis, common pitfalls that can raise concerns about its value.
has resulted in a rapid increase in the number of studies in The effectiveness of a data-driven system goes beyond a
this area in the last few years.44 However, none of the measure of performance, such as an AUC or a P value. To
current published studies has tested the usefulness of be effective, a data-driven system must produce
automated image analysis in an ICU setting. The rapid actionable outputs for the right patients, at the right time.
growth of this field, however, will undoubtedly result in For example, the output can be predictive information
many uses applicable to critical care situations. Perhaps that can help a clinician decide the most effective
most pertinent to critical care clinicians is the treatment for a particular patient as soon as a diagnosis is
advancement in techniques to detect pulmonary made. Furthermore, when evaluating the clinical
pathology in chest radiographs,45,46 as well as normal and implementation of such a system, it is important to know
abnormal findings in brain and abdominal imaging.44 whether it has been tested in an experimental setting and
These techniques could be particularly helpful in ICUs whether it has shown a meaningful impact in a population
with limited availability of specialists who can accurately similar to the one for which it is being considered.
interpret radiographic images in a timely fashion, but
their effectiveness and safety should first be thoroughly Unfortunately, bad data science abounds. We must
tested before any clinical implementation is considered. make a collective effort to ensure that only good data

chestjournal.org 1245
science evolves into data-driven systems that can be complex nonlinear interactions between variables, they
safely tested and used in critically ill patients. The ease of never sleep, and they can multitask effortlessly.
access to large amounts of data and computing power However, clinical thinking and medical decision-making
can lead to data mining “fishing expeditions” that can are not reproducible by current technologies.10 The
result in low-quality research.3 Poorly framed clinical qualitative aspect of clinical decision-making—the so-
problems, bad data, or debatable methods will result in called “art of medicine”—is impossible to model
flawed data science, and it can create more problems quantitatively. Furthermore, many factors influencing
than it solves.3,6,48 Using epidemiologic best practices to clinical decision-making, including clinical, societal, and
analyze retrospective data, including thoughtfully personal factors, are not necessarily reflected in the
adjusting for confounding variables, is just as important digital records, and thus any output from a data-driven
in large datasets as it is in smaller ones. In addition, a system will need to be first evaluated, interpreted, and
model may fit well only on the training data but enriched by clinicians before any action is taken.
generalize poorly to other data, a phenomenon known as However, to achieve a successful partnership between
overfitting (Table 1). Overfitting may occur when clinicians and computers, we must first improve the
algorithms learn from idiosyncrasies, or noise, in the skills of bedside clinicians at interpreting and using the
training data; techniques such as cross-validation and output from these data-driven systems.48
regularization can be used to mitigate this problem.52,53
Finally, another challenge faced by data science teams in
Poorly implemented digital technologies can harm critical care is balancing the need for data openness and
patients,54 and only a rigorous approach to their reproducibility with the demand for data privacy and
evaluation and implementation can mitigate this risk. security. The open data science movement calls for
Partnerships between data scientists, clinical domain transparent and reproducible research with seamless
experts, medical informaticians, and implementation data-sharing across institutions. Indeed, a recent study
science specialists will result in more effective and safer showed an alarming lack of reproducibility in data
data-driven systems. Clinicians with data science skills, science studies using the same ICU data, which suggests
clinical research expertise, and an intimate knowledge of that algorithms, study procedures, computer code, and
the clinical realities in the ICU can help data science even datasets should be openly available to ensure
teams capture the right data, address the right clinical reproducibility.58 However, this data openness must not
problems, and produce the right actionable knowledge.10 result in poor data governance, lack of data security, or
Furthermore, clinician input can help minimize the loss of confidentiality, all of which are necessary to
number of unnecessary alerts or prompts these systems perform ethical research and maintain public trust.59
might produce, thereby reducing the risk of alert fatigue,
which is another common problem among front-line
providers working with novel digital technologies.55 The Future of Big Data and Data Science in
Critical Care
Another common concern among clinicians is the
We imagine a future in critical care in which data-driven
perceived loss of autonomy in the face of increasingly
systems and clinicians work hand-in-hand. Large
more sophisticated computational systems. This concern
quantities of clinical, physiologic, and “omics” data are
exists despite the fact that clinicians will readily
analyzed by computational systems and are served to the
acknowledge that the complexity of medicine nowadays
bedside clinicians in the form of manageable,
far exceeds the capacity of the unaided human mind and
interpretable, and actionable knowledge that augment
that perhaps these novel computational systems can help
the clinician’s decision-making capacity. Predictive
manage some of this complexity.10 To put it in
models perform diagnostic and therapeutic
perspective, humans typically make decisions using
recommendations, while clinicians contextualize these
fewer than six data points, because anything more than
recommendations and coordinate their implementation.
that becomes cognitively too expensive.56 However, an
False alerts are kept to a minimum and systems are
ICU patient can generate thousands of data points in a
continuously improved through a collaborative and
single day, and when you add fatigue, interruptions, and
scientifically rigorous approach.
the clinicians’ own cognitive biases, it is not surprising
that many clinical decisions end up being suboptimal.57 Data science can be transformative. There is a real
Conversely, computers can sift seamlessly through tens opportunity that this scenario will become a reality in
of thousands of data points, they can easily analyze the near future, but there is still a lot of work ahead of

1246 Contemporary Reviews in Critical Care Medicine [ 154#5 CHEST NOVEMBER 2018 ]
us. Our patients entrust us with their precious data and 19. Luo Y, Ahmad FS, Shah SJ. Tensor factorization for precision
medicine in heart failure with preserved ejection fraction.
we—clinicians, researchers, data scientists, and leaders J Cardiovasc Transl Res. 2017;10(3):305-312.
in critical care—have an obligation to use it in the best 20. Goodfellow I, Bengio Y, Courville A. Deep learning. Adapt Comput
possible way. Mach Le. 2016:1-775.
21. Gulshan V, Peng L, Coram M, et al. Development and validation of a
deep learning algorithm for detection of diabetic retinopathy in
Acknowledgments retinal fundus photographs. JAMA. 2016;316(22):2402-2410.
Financial/nonfinancial disclosures: The authors have reported to 22. Esteva A, Kuprel B, Novoa RA, et al. Dermatologist-level
CHEST the following: M. M. C. has a patent pending for risk classification of skin cancer with deep neural networks. Nature.
stratification algorithms for hospitalized patients; he is also supported 2017;542(7639):115-118.
by a career development award from the National Heart, Lung and
23. Miotto R, Li L, Kidd BA, Dudley JT. Deep patient: an unsupervised
Blood Institute and a research project grant program award from
representation to predict the future of patients from the electronic
the National Institute of General Medical Sciences. None declared health records. Sci Rep. 2016;6:26094.
(L. N. S.-P., Y. L.).
24. Aczon M, Ledbetter D, Ho L, et al. Dynamic mortality risk
predictions in pediatric critical care using recurrent neural networks.
References arXiv preprint arXiv:170106675. 2017 Jan 23.
1. Smith M, Saunders R, Stuckhardt L, McGinnis JM. Best Care at 25. Wong HR. Intensive care medicine in 2050: precision medicine.
Lower Cost: The Path to Continuously Learning Health Care in Intensive Care Med. 2017;43(10):1507.
America. Washington, DC: National Academies Press; 2013.
26. Knaus WA, Zimmerman JE, Wagner DP, Draper EA, Lawrence DE.
2. Bates DW, Saria S, Ohno-Machado L, Shah A, Escobar G. Big data in APACHE-acute physiology and chronic health evaluation: a
health care: using analytics to identify and manage high-risk and physiologically based classification system. Crit Care Med. 1981;9(8):
high-cost patients. Health Aff (Millwood). 2014;33(7):1123-1131. 591-597.
3. Badawi O, Brennan T, Celi LA, et al. Making big data useful for 27. Knaus WA, Draper EA, Wagner DP, Zimmerman JE. APACHE II: a
health care: a summary of the inaugural mit critical data conference. severity of disease classification system. Crit Care Med. 1985;13(10):
JMIR Med Inform. 2014;2(2):e22. 818-829.
4. Iwashyna TJ, Liu V. What’s so different about big data?. A primer 28. Churpek MM, Yuen TC, Winslow C, et al. Multicenter development
for clinicians trained to think epidemiologically. Ann Am Thorac and validation of a risk stratification tool for ward patients. Am J
Soc. 2014;11(7):1130-1135. Respir Crit Care Med. 2014;190(6):649-655.
5. Anthony Celi L, Mark RG, Stone DJ, Montgomery RA. “Big data” in 29. Joshi R, Szolovits P. Prognostic physiology: modeling patient severity
the intensive care unit. Closing the data loop. Am J Respir Critic Care in intensive care units using radial domain folding. Paper presented
Med. 2013;187(11):1157-1160. at: American Medical Informatics Association Annual Symposium
Proceedings; November 3-7, 2012; Chicago, IL.
6. Ghassemi M, Celi LA, Stone DJ. State of the art review: the data
revolution in critical care. Crit Care. 2015;19:118. 30. Henry KE, Hager DN, Pronovost PJ, Saria S. A targeted real-time
early warning score (TREWScore) for septic shock. Sci Translat Med.
7. Buchman TG, Billiar TR, Elster E, et al. Precision medicine for 2015;7(299):299ra122.
critical illness and injury. Crit Care Med. 2016;44(9):1635-1638.
31. Nemati S, Holder A, Razmi F, Stanley MD, Clifford GD,
8. Johnson AE, Ghassemi MM, Nemati S, Niehaus KE, Clifton DA, Buchman TG. An interpretable machine learning model for accurate
Clifford GD. Machine learning and decision support in critical care. prediction of sepsis in the ICU. Critical Care Medicine. 2018;46(4):
Proceedings IEEE. 2016;104(2):444-466. 547-553.
9. Provost F, Fawcett T. Data science and its relationship to big data 32. Calfee CS, Delucchi K, Parsons PE, et al. Subphenotypes in acute
and data-driven decision making. Big Data. 2013;1(1):51-59. respiratory distress syndrome: latent class analysis of data from two
10. Obermeyer Z, Lee T. Lost in thought—the limits of the human mind randomised controlled trials. Lancet Respir Med. 2014;2(8):611-620.
and the future of medicine. N Engl J Med. 2017;377(13):1209. 33. Knox DB, Lanspa MJ, Kuttler KG, Brewer SC, Brown SM.
11. Deo RC. Machine learning in medicine. Circulation. 2015;132(20): Phenotypic clusters within sepsis-associated multiple organ
1920-1930. dysfunction syndrome. Intensive Care Med. 2015;41(5):814-822.
12. Murdoch TB, Detsky AS. The inevitable application of big data to 34. Luo Y, Xin Y, Joshi R, Celi L, Szolovits P. Predicting ICU Mortality
health care. JAMA. 2013;309(13):1351-1352. Risk by Grouping Temporal Trends from a Multivariate Panel of
Physiologic Measurements. Paper presented at: Proceedings of the
13. Hardin J, Hoerl R, Horton NJ, et al. Data science in statistics 30th AAAI Conference on Artificial Intelligence; 2016.
curricula: preparing students to "think with data." Am Stat 2015;
2015;69(4):343-353. 35. Vranas KC, Jopling JK, Sweeney TE, et al. Identifying distinct
subgroups of ICU patients: a machine learning approach. Crit Care
14. James G, Witten D, Hastie T, Tibshirani R. An Introduction to Med. 2017;45(10):1607-1615.
Statistical Learning, Vol. 112. New York, NY: Springer; 2013.
36. Sjoding MW, Liu VX. Can you read me now? Unlocking narrative
15. Churpek MM, Yuen TC, Winslow C, Meltzer DO, Kattan MW, data with natural language processing. Ann Am Thorac Soc.
Edelson DP. Multicenter comparison of machine learning methods 2016;13(9):1443-1445.
and conventional regression for predicting clinical deterioration on
the wards. Crit Care Med. 2016;44(2):368-374. 37. Lehman LW, Saeed M, Long W, Lee J, Mark R. Risk stratification of
ICU patients using topic models inferred from unstructured progress
16. Seymour CW, Gomez H, Chang CH, et al. Precision medicine for notes. Paper presented at: AMIA annual symposium
all? Challenges and opportunities for a precision medicine approach proceedings; 2012.
to critical illness. Crit Care. 2017;21(1):257.
38. Ghassemi M, Naumann T, Doshi-Velez F, et al. Unfolding
17. Mayhew MB, Petersen BK, Sales AP, Greene JD, Liu VX, Wasson TS. physiological state: mortality modelling in intensive care units. Paper
Flexible, cluster-based analysis of the electronic medical record of presented at: Proceedings of the 20th ACM SIGKDD International
sepsis with composite mixture models. J Biomed Inform. 2018;78: Conference on Knowledge Discovery and Data Mining; 2014.
33-42.
39. Weissman GE, Harhay MO, Lugo RM, Fuchs BD, Halpern SD,
18. Wong HR, Atkinson SJ, Cvijanovich NZ, et al. Combining Mikkelsen ME. Natural language processing to assess documentation
prognostic and predictive enrichment strategies to identify children of features of critical illness in discharge documents of acute
with septic shock responsive to corticosteroids. Crit Care Med. respiratory distress syndrome survivors. Ann Am Thoracic Soc.
2016;44(10):e1000-e1003. 2016;13(9):1538-1545.

chestjournal.org 1247
40. Saeed M, Villarroel M, Reisner AT, et al. Multiparameter Intelligent 50. Awdishu L, Coates CR, Lyddane A, et al. The impact of real-time
Monitoring in Intensive Care II (MIMIC-II): a public-access alerting on appropriate prescribing in kidney disease: a cluster
intensive care unit database. Crit Care Med. 2011;39(5):952. randomized controlled trial. J Am Med Inform Assoc. 2016;23(3):
41. Sun JX, Reisner AT, Saeed M, Heldt T, Mark RG. The cardiac output 609-616.
from blood pressure algorithms trial. Crit Care Med. 2009;37(1):72. 51. Bright TJ, Wong A, Dhurjati R, et al. Effect of clinical decision-
42. Roederer A, Weimer J, DiMartino J, Gutsche J, Lee I. Robust support systems: a systematic review. Ann Intern Med. 2012;157(1):
monitoring of hypovolemia in intensive care patients using 29-43.
photoplethysmogram signals. Paper presented at: Engineering in 52. Hinton GE, Srivastava N, Krizhevsky A, Sutskever I,
Medicine and Biology Society (EMBC), 2015 37th Annual Salakhutdinov RR. Improving neural networks by preventing
International Conference of the IEEE; 2015. co-adaptation of feature detectors. arXiv preprint arXiv:
43. Dunitz M, Verghese G, Heldt T. Predicting hyperlactatemia in the 12070580:2012.
MIMIC II database. Paper presented at: Engineering in Medicine 53. Bishop CM. Pattern recognition and Machine Learning. New York,
and Biology Society (EMBC), 2015 37th Annual International NY: Springer ScienceþBusiness Media LLC; 2006.
Conference of the IEEE; 2015.
54. Han YY, Carcillo JA, Venkataraman ST, et al. Unexpected
44. Litjens G, Kooi T, Bejnordi BE, et al. A survey on deep learning in increased mortality after implementation of a commercially sold
medical image analysis. Medical Image Analysis. 2017;42:60-88. computerized physician order entry system. Pediatrics.
45. Gonzalez G, Ash SY, Vegas-Sanchez-Ferrero G, et al. Disease staging 2005;116(6):1506-1512.
and prognosis in smokers using deep learning in chest computed
tomography. Am J Respir Crit Care Med. 2018;197(2):193-203. 55. Kizzier-Carnahan V, Artis KA, Mohan V, Gold JA. Frequency of
passive EHR alerts in the ICU: another form of alert fatigue
46. Bar Y, Diamant I, Wolf L, Greenspan H. Deep learning with non- [published online ahead of print June 22, 2016]? J Patient Saf.
medical training used for chest pathology identification. Paper https://doi.org/10.1097/PTS.0000000000000270.
presented at: Proc. SPIE2015.
56. Kahneman D. Thinking, Fast and Slow. New York, NY: Farrar, Staus
47. Tangri N, Kent DM. Toward a modern era in clinical prediction: the and Giroux; 2011.
TRIPOD statement for reporting prediction models. Am J Kidney
Dis. 2015;65(4):530-533. 57. Neuraz A, Guerin C, Payet C, et al. Patient mortality is associated
with staff resources and workload in the ICU: a multicenter
48. Verghese A, Shah NH, Harrington RA. What this computer needs is observational study. Crit Care Med. 2015;43(8):1587-1594.
a physician: humanism and artificial intelligence. JAMA.
2018;319(1):19-20. 58. Johnson AE, Pollard TJ, Mark RG. Reproducibility in critical care: a
mortality prediction case study. Paper presented at: Machine
49. Pickering BW, Dong Y, Ahmed A, et al. The implementation of
Learning for Healthcare Conference; 2017.
clinician designed, human-centered electronic medical record viewer
in the intensive care unit: a pilot step-wedge cluster randomized trial. 59. Raghupathi W, Raghupathi V. Big data analytics in healthcare:
Int J Med Inform. 2015;84(5):299-307. promise and potential. Health Inf Sci Syst. 2014;2:3.

1248 Contemporary Reviews in Critical Care Medicine [ 154#5 CHEST NOVEMBER 2018 ]