3.

ARTIFICIAL INTELLIGENCE 75

misses, demonstrate this in practice. Layered with intelligent algorithms, predic-

tive data advance to the use of prescriptive data, which allow nurses to ensure
more precisely what to do about events beforehand to administer the most
appropriate, evidence-based practices and standards for better outcomes. The
subsequent foresight that prescriptive data impart empowers nurses to move be-
yond hindsight and provides insight to use information, knowledge, and wisdom
along the path of data complexity. The advancement along the spectrum from
simple to more intricate data enables enhanced data analysis from less to more
difficult to promote randomized testing, predictive modeling through intelligent
algorithm use, and forecasting to optimize care and processes (HIMSS, 2019).

DECISION-MAKING USING PREDICTIVE DATA: INPATIENT LENGTH OF STAY

To highlight the value of using basic to complex data to identify the best inter-
ventions that lead to quality outcomes, consider the example of using prediction
All rights reserved. May not be reproduced in any form without permission from the publisher, except fair uses permitted under U.S. or applicable copyright law.

for the appropriate patient length of stay (LOS) in the inpatient setting upon
patient admission. When a patient enters the hospital, if nurses use descriptive and
diagnostic data alone to determine LOS, it will likely be an uninformed decision,
forcing heuristics to merely guess how long a patient may or should stay, based
on perhaps only an admitting diagnosis code estimating what might be appropri-
ate or reasonable. With the hindsight that this creates, looking backward is the
sole method for nurses to analyze these data through reports and dashboards
to show trends, including how long a patient stayed—for too long or not long
enough. The sources and actions that led to the outcomes prompt the need to
drill down in standard or ad hoc reports if the patient has a delayed discharge,
or hospital readmission resulting from an early discharge.
In contrast, predictive analytics can determine inpatient LOS using automa-
tion and smart machine algorithms to determine the appropriate or reasonable
LOS based on data-driven patterns from multiple Big Data sources that alert
nurses of what to expect as the most beneficial time frame from the first contact
with the patient. Nurses can then be prescriptive through the patient’s hospital
journey. With foresight, applying evidence-based interventions more quickly is
possible, including discharging the patient if he or she has been in the hospital
for too long, to decrease the risk of hospital-acquired conditions and increase
patient satisfaction, and identifying if the patient is not staying long enough, to
prevent a hospital readmission or emergency department visit. With insight and
foresight, nurses can anticipate the best LOS to plan and take necessary actions
earlier, allowing more time for precision treatments and therapies that can re-
duce them. Knowledge, gained through prediction, enhances nurses’ ability to
provide high-quality care planning coordination and determine breakdowns in
hospital flow, such as long radiology wait times or poor bed management, and
make adjustments as needed, including canceling a discharge or readjusting staff-
ing to improve care for all hospital patients throughout the continuum of care.
2021. Springer Publishing Company.

Machine Learning

ML, a branch of AI (Iriondo, 2019), is the study of computer algorithms and

programs said to improve automatically through experience concerning a class
of tasks and performance measures (Mitchell, 1997). Specifically, with predictive

EBSCO Publishing: eBook Collection (EBSCOhost) printed on 3/9/2025 12:20:24 PM UTC via EWHA WOMANS UNIV 2359857; Whende M. Carroll, MSN, RN-BC; Emerging
Technologies for Nurses : Implications for Practice Account:ehost.

Carroll46496_PTR_CH03_57-94_01-06-20.indd 75 1/7/2020 2:56:25 PM

 76 EMERGING TECHNOLOGIES FOR NURSES: IMPLICATIONS FOR PRACTICE

analytics, ML is a set of algorithms (models) that can take a set of data inputs
(variables and features) and return a prediction (Caffo, Leek, & Peng, 2016). ML
differs from traditional statistical analysis for prediction, in that statisticians cast
confidence more often on data modeling, and that ML methods focus more on
the predictive models’ accuracy (Breiman, 2001). This distinction is vital as nurses
may wonder why statistics are not used in clinical and administrative settings for
intelligent prediction. Statistics underscore inference, the process of using data
analysis designed to deduce the characteristics or properties of a population; in
contrast, ML smart machines automatically learn over time through experience,
emphasize predictions, evaluate results via prediction performance, and serve to
highlight algorithm performance and robustness of outcomes (Caffo et al., 2016).
A simple analogy for ML is that it is the equivalent of teaching a system the rules
of a game and getting the program to practice it at elementary and intermediate
levels. After preparation and further training, the system can, in real time, play
at advanced levels (Bari et al., 2017).
ML is well suited for using complex healthcare data, data in various forms
collected from multiple sources, and data that are voluminous for prediction.
While data mining, the process used to extract usable data from a broader set
of any raw data (The Economic Times, n.d.), can unearth previously unknown
connections in data, ML categorizes the new and upcoming unknowns, learns
from them based on its previous processing of the data, and gets better at incor-
porating them into known data (see Figure 3.3). Both techniques lead to richer
insight and improved understanding of the data (Bari et al., 2017). There are
four techniques of ML, and each has its own approach based on the business
problem that needs to be solved, as well as the amount, type, and volume of the
data (Hurwitz & Kirsch, 2018).

Supervised
Learning Classification

Develop predictive
model based on both
input and output data

Machine
Learning Regression
Unsupervised
Learning

Group and interpret

data based only on
input data
Clustering

Figure 3.3 Two types of classic machine learning methods. Supervised learning, which trains a model on
known input and output data to predict future outputs, and unsupervised learning, which discovers unknown
patterns or basic structures in input data.
Source: MathWorks. (2019). What is machine learning? How it works—Techniques & applications. Retrieved from https://www.mathworks.com/
discovery/machine-learning.html

EBSCOhost: eBook Collection (EBSCOhost) printed on 3/9/2025 12:20:24 PM UTC via EWHA WOMANS UNIV. All use subject to https://www.ebsco.com/terms-of-use.

Carroll46496_PTR_CH03_57-94_01-06-20.indd 76 1/7/2020 2:56:25 PM

 3. ARTIFICIAL INTELLIGENCE 77

SUPERVISED LEARNING
In supervised learning, machines build prediction models based on evidence
in the presence of uncertainty from both input and output data (Figure 3.4). A
supervised learning algorithm takes a known set of input data and known responses
to the output data and trains a model to generate reasonable predictions for the
answers to new data (MathWorks, 2019). This learning technique is used most
commonly when known data for the output data have labels that define their
meaning (Caffo et al., 2016). To develop predictive models, supervised learning
uses classification and regression techniques. Classification models classify input
data and discrete output for predictions, for example, whether a tumor is benign
or cancerous. Typical applications include medical imaging and speech recogni-
tion, including voice clinical charting or dictation. Classification algorithms data
can be tagged or divided into specific groups or classes. Regression approaches
predict continuous responses, for example, changes in a patient’s temperature or
variations in blood pressure. Scientists use regression techniques when working
with a numeric output variable or if the nature of the response is a real number,
such as values from vital signs, lab results, or the precise time until failure for a
piece of equipment (MathWorks, 2019).
UNSUPERVISED LEARNING
Unsupervised learning discovers hidden patterns or basic structures in data.
It is used to draw inferences from datasets consisting of input data only, with-
out labeled responses. This technique allows the computer to learn how to do
something and use this to determine structure and patterns in data; for example,
trying to uncover unobserved factors in unlabeled data, such as image recognition

Unsupervised
Supervised Learning
Learning

Linear Classification Clustering

Regression

Readmit = Yes
Length of Stay

Number of
Diagnoses

Readmit = No

Number of Age
Diagnostic Tests

Length of Stay Readmits Heart Failure

Subgroups
Figure 3.4 Outputs of common machine learning prediction algorithms using supervised and unsupervised
learning for length of stay, readmits, and heart failure subtypes.
Source: Qureshi, I. (2018, October 26). Healthcare data scientists: 6 steps for utilizing your data analysts. Retrieved from https://www.healthcatalyst
.com/insights/healthcare-data-scientists-steps-leveraging-data-analysts

EBSCOhost: eBook Collection (EBSCOhost) printed on 3/9/2025 12:20:24 PM UTC via EWHA WOMANS UNIV. All use subject to https://www.ebsco.com/terms-of-use.

Carroll46496_PTR_CH03_57-94_01-06-20.indd 77 1/7/2020 2:56:25 PM

 78 EMERGING TECHNOLOGIES FOR NURSES: IMPLICATIONS FOR PRACTICE

processing, where recognition techniques are used for image data segmentation
(Holehouse, n.d.). It is known as “unsupervised” because there is no standard or
benchmark outcome to judge against for the prediction outputs, segmenting data
into groups of features; this may be used before moving the predictive outputs
to a supervised learning approach (Caffo et al., 2016). Clustering is the most
common unsupervised learning technique (Figure 3.4). It is used for exploratory
data analysis to find hidden patterns or groupings in data (MathWorks, 2019).
Applications for cluster analysis include gene sequence analysis and scoring pa-
tient satisfaction results. Also, if an Internet of Things (IoT) health IT vendor
wants to adjust the locations where he or she is placing sensors in a hospital unit
for optimal connectivity and sensing, the vendor may use unsupervised ML to
estimate the number of clusters of patients requiring the sensors for the best
technical outcomes.
REINFORCEMENT LEARNING
Reinforcement learning (RL) is a type of behavioral model of ML technique
that enables an agent, a piece of software in an AI program, to learn in an interac-
tive environment through trial and error by using feedback from its own actions
and experiences (Bhatt, 2018). Both supervised learning and RL use mappings
between inputs and outputs. However, unlike supervised learning where feedback
provided to the software program is a correct set of actions for performing a
task, RL employs rewards and punishment as signals for positive and negative
behavior (Alzantot, 2017). RL is different in terms of goals, compared to unsu-
pervised learning. The goal in unsupervised learning is to relate data points; in
RL, the purpose is to find a suitable action model that would maximize the total
cumulative reward of the agent (see Figure 3.5; Bhatt, 2018). In clinical practice,
for cohorts of patients with chronic disease, clinicians can use RL to determine
real-life human actions to better determine treatment regimes in population
health settings (Liu et al., 2017).
DEEP LEARNING
DL is a subset of ML and is a method that encapsulates cognitive learning, ML,
and functions in the same way, but it has different capabilities. The main difference
between DL and ML is ML models become better progressively but the model
still needs some direction to do so. When an ML model (regression, classifica-
tion, or clustering) returns an incorrect prediction, a scientist needs to fix that
problem purposefully, but in the case of DL, the model will do it autonomously
(Hariharan, 2018). DL is associated with an ML concept known as an artificial
neural network (ANN), which is inspired from neurons in the human brain to

Agent

State Reward Action

Environment

Figure 3.5 Process of reinforcement learning.

EBSCOhost: eBook Collection (EBSCOhost) printed on 3/9/2025 12:20:24 PM UTC via EWHA WOMANS UNIV. All use subject to https://www.ebsco.com/terms-of-use.

Carroll46496_PTR_CH03_57-94_01-06-20.indd 78 1/7/2020 2:56:25 PM

 3. ARTIFICIAL INTELLIGENCE 79

facilitate the modeling of arbitrary functions, requires a massive amount of data,

and is highly flexible when it comes to modeling multiple outputs simultane-
ously (Jain, 2015). Neural networks (NNs), or nodes, were initially motivated by
looking at machines that replicate the brain’s functionality, with the hypothesis
that the brain has a single learning algorithm and it can process and learn from
data from any source (Holehouse, n.d.; see Figure 3.6).
Reasons for the increased popularity of DL today are the immense increase
in chip processing abilities, the significantly increased volume of Big Data used
for ML, and the recent advances in ML (Deng & Yu, 2014). This progress en-
ables DL techniques to exploit complex datasets effectively, through nonlinear
compositional functions, to learn distributed and hierarchical features that are
domain-specific; these are used to describe the data and representations, a set of
classifiers, or the language that a computer understands (Deng & Yu, 2014) to
make effective use of both labeled and unlabeled data (Faggella, 2019). Research
is now focusing on developing data-efficient ML, mainly DL systems that can
learn more efficiently, with the same performance, quicker, and with fewer data.
For cutting-edge nurse-led use cases that support precision healthcare, includ-
ing chatbots, population health management, and genomics, abstraction helps
to make sense of data such as images, sound, and text—an automatic car driving
system is an example of DL (Deng & Yu, 2014).

The ML Process

The foundation of ML is to seek an answer to the question, “How can we build

computer systems that automatically improve with experience, and what are the
fundamental laws that govern all learning processes?” (Mitchell, 2006 as cited in
Faggella, 2019, “How We Arrived at Our Definition”). The method of ML is a
five-step process that involves obtaining input data, cleaning and preparing data,
training data through the chosen algorithm(s), data testing and validating, and
receiving and improving the output. Learning and improvement occur through
the cyclic method of retraining and retesting data from ML outputs (Figure 3.7)

IL HL HL HL OL

Y
Diabetes
N

Feature Extraction
and Classification

Figure 3.6 Deep learning artificial neural networks.

Cognitive intelligence algorithms process clinical data inputs (IN, input layer) through hidden layers of cells (HL) that adjust and connect based on well-
constructed rules in a smart machine to perform a task, and subsequently develop responses (OL, output layer) based on pattern recognition similar
to how a human’s brain responds to external stimuli.

EBSCOhost: eBook Collection (EBSCOhost) printed on 3/9/2025 12:20:24 PM UTC via EWHA WOMANS UNIV. All use subject to https://www.ebsco.com/terms-of-use.

Carroll46496_PTR_CH03_57-94_01-06-20.indd 79 1/7/2020 2:56:25 PM

 80 EMERGING TECHNOLOGIES FOR NURSES: IMPLICATIONS FOR PRACTICE

Learning and improvement

through retraining and retesting

Train Data
Get Input Clean and Through Test and Improve
Data Prepare Data Chosen Validate Data Output
Algorithm

Figure 3.7 The machine learning process.

To begin the ML process, the initial, and arguably the most time-consuming,
step is feature engineering, the process of collecting and manipulating datasets
that will train the model. This step is critical because it determines the dataset
labels, algorithmic logic, and ensures accurate connections during the cyclic pro-
cess (Ng, 2019). Cleaning prepares data by removing irrelevant and redundant
data; further, the methods of feature selection and extraction use only the most
relevant data attributes and transform them into new formats that describe vari-
ants within the data and reduce the amount of information required to develop
the ML model (Zhou, 2018). This stage is where, for example, extraneous and
duplicative patient demographics, vital signs, and lab results from healthcare data
pulled from multiple sources, such as from the EHR, medical claims, and data
sensed through IoT devices, are scrubbed and fixed for dataset completeness.
Training the labeled dataset, once collected and prepared, entails choosing a
model that best fits the use case. Commonly used algorithmic prediction models
include linear and logistic regression, decision trees, random forest, gradient
boosting, and neural networks (NNs), discussed later in this chapter. The size
and dataset quality will determine each, as will the nature of the data, computa-
tional time availability, task urgency, and use case for the model (Le, 2019), and
algorithm determination will generate various data visualizations. It is customary
to trial several algorithms on a dataset to fine-tune them with slightly different
parameters to identify variation in performance. Typically, most of the models
perform similarly; however, the quality and underlying structure of the dataset
for predicting the data outputs is essential, as an algorithm might perform well
on specific datasets and poorly on others, which leads to the crucial need to trial
different algorithms (Mastanduno, 2017).
Algorithm training will fit and tune data, and testing and validation of their
performance is a key step toward precision learning and output improvement.
Test datasets are separated as “unseen” data to evaluate the model. Data are always
first split to determine reliability before training the model and are put aside until
choosing the algorithm (Ng, 2019). Comparing test and training performance
avoids overfitting when a model does not generalize well from training data to
test data; this is the outcome if model performance is high on training dataset
but poor on the test data, in which case it is overfit due to lack of original data
structure and quality (Elite Data Science, n.d.). Improvements of unknown pattern

EBSCOhost: eBook Collection (EBSCOhost) printed on 3/9/2025 12:20:24 PM UTC via EWHA WOMANS UNIV. All use subject to https://www.ebsco.com/terms-of-use.

Carroll46496_PTR_CH03_57-94_01-06-20.indd 80 1/7/2020 2:56:25 PM

 3. ARTIFICIAL INTELLIGENCE 81

recognition and outcomes including predictions materialize with rigorous dataset

training and testing, after which machines begin to learn on their own.
DIFFERENCES BETWEEN ML AND DL
Like ML, DL completes feature extraction for processing and outputs, and the
data are trained, tested, and validated. However, the DL technique uses an ANN
architecture organized into layers composed of interconnected nodes. The nodes
of the network perform a weighted sum of the input data that transfer to an activa-
tion function. Weights become dynamically optimized during the training phase
(Pesapane et al., 2018). In DL, there are three different kinds of layers: the input
layer, which receives input data, and the output layer, which produces the results
of data processing, with hidden layer(s) in between. A deep ANN differs from the
single hidden layer, NNs, or classical NNs, by having a large number of hidden
layers, which characterize the depth of the network (Erickson, Korfiatis, Akkus, &
Kline, 2017). DL avoids the design of a dedicated high-level feature extraction
process from raw sensory data (Topol, 2019) by using a deep NN approach that
automatically unearths and represents complex features as a compilation of
simpler ones (Pesapane et al., 2018; Case Study 3.2).

CASE STUDY 3.2

PIECES TECHNOLOGIES

USING AI TO IMPROVE HOSPITAL PATIENT LENGTH OF STAY,

READMISSIONS, AND DETERIORATION

Pieces Technologies, a health information technology (IT) spinoff of Dallas-based

Parkland Center for Clinical Innovation, has developed an AI cloud-based software
solution called DS (Decision Sciences) that uses language semantics processing and
predictive modeling techniques to process both patients’ clinical and social deter-
minants data. The intelligent technology enables nurses to follow patients through
their care journey and provide decision support across the care continuum (Cohen,
2018a), tackling high-risk social and clinical issues that can impact adverse outcomes,
particularly during the hospital stay (Pieces Tech, 2018b).
Pieces DS is a cloud-based, back-end AI technology that predicts patients at risk
for adverse outcomes and excess hospital utilization. The solution applies advanced
analytics, using ML in conjunction with clinical human augmentation, as well as chart
reviews by Pieces’ clinicians, to ensure the highest accuracy in the predictive models.
In addition to processing vast amounts of structured data, the DS language process-
ing technology reads and interprets unstructured data such as free text progress and
dictated notes for actionable insights. With this assistive intelligence, DS discovers
critical patterns within the data, which enables the activation of precision, expedited
interventions—the right care, the team member at the right time—for enhanced and
efficient decision-making to maximize operational and clinical effectiveness, without
overalerts improving human factors in practice (Pieces Tech, 2018b). DS continuously

(continued )

EBSCOhost: eBook Collection (EBSCOhost) printed on 3/9/2025 12:20:24 PM UTC via EWHA WOMANS UNIV. All use subject to https://www.ebsco.com/terms-of-use.

Carroll46496_PTR_CH03_57-94_01-06-20.indd 81 1/7/2020 2:56:25 PM

 82 EMERGING TECHNOLOGIES FOR NURSES: IMPLICATIONS FOR PRACTICE

CASE STUDY 3.2 (continued)

monitors patient data for risk of readmission, overutilization of ED services, exces-
sive LOS, and social determinants that may affect patient quality outcomes (Pieces
Tech, 2018b). The platform’s Clinician-in-the-Loop™ prediction and communica-
tion capabilities are evidence-based, clinically developed, and DS’s configured ML
algorithms that maintain, monitor, and continually improve in clinical practice, with
demonstrated 80% to 95% positive predictive value (PPV), or precision; 80% to 95%
sensitivity (true positive rate); and five times fewer false-positive results upon model
testing and validation (Pieces Tech, 2018b).

PREDICTION FOR LENGTH OF STAY AND READMISSIONS

Despite efforts to prevent avoidable hospitalizations, research shows that 3.3 million
patients are readmitted within 30 days of discharge, resulting in $17B of preventable
costs annually (Pieces Tech, 2018b). Reducing readmissions and avoidable hospital-
izations can have an immediate, positive impact on a hospital’s bottom line. While
multidisciplinary teams work to find solutions to this issue, many still lack actionable
insights needed for precise results, particularly recognizing the importance of social
determinants and their impact on avoidable readmissions. To improve outcomes, as an
example, a major academic, integrated delivery network licensed a Pieces DS machine
learning platform to sit atop their electronic health records (EHRs) and clinical data
warehouses to predict and monitor the likelihood of costly clinical events including
readmissions and excess LOS to identify at-risk patients. With the solution in place
at several of their hospitals, the organization realized reductions in readmissions and
LOS excess; the ability to discern where readmission efforts should be focused; more
than 95% PPV identifying barriers to discharge, including pending tests, delayed
provider consultant feedback, and unclear follow-up activities and post-discharge
placement (Amarasingham, 2019); and more than 850 nursing review hours saved
per year, per hospital (Pieces Tech, 2018b). Overall, the DS clinical decision support
tools assist nurses in creating a holistic and personalized discharge plan, making smart
recommendations in the patient’s EHR (Miliard, 2018).

EARLY WARNING SYSTEMS FOR PATIENT DETERIORATION AND SEPSIS

With the use of Pieces intelligent technologies, The Parkland Center for Clinical
Innovation, Pieces’ nonprofit research affiliate, realizes the benefits of AI’s ability to
analyze massive volumes of data and alert clinicians up to 48 hours in advance of a
patient requiring a rapid response team (RRT) for clinical deterioration (Miliard, 2018).
While some hospitals use automated early warning systems (EWSs) from structured
data and RRTs, these tools are often limited (Pieces Tech, 2018a). To improve response
time and the opportunity window for care teams to intervene on patients at risk of
deterioration, a study was conducted with a client hospital to evaluate the effectiveness
of the machine learning, predictive solution. Researchers surveyed 4,600 encounters

(continued )

EBSCOhost: eBook Collection (EBSCOhost) printed on 3/9/2025 12:20:24 PM UTC via EWHA WOMANS UNIV. All use subject to https://www.ebsco.com/terms-of-use.

Carroll46496_PTR_CH03_57-94_01-06-20.indd 82 1/7/2020 2:56:25 PM