Healthcare Data Analytics

Vaibhav P. Vasani

Assistant Professor
Department of Computer Engineering
K. J. Somaiya College of Engineering
Somaiya Vidyavihar University

10/22/2023 [email protected] 1
 • Sensors measure physical attributes of the world and produce
signals, i.e., time series consisting of ordered measurements of the
form (timestamps, data elements).
• For example, in intensive care, respiration rates are estimated from
measurements of the chest impedance of the patient.
• The resulting time series signals are consumed either by a human or
by other sensors and computing systems.
• For instance, the output of the chest impedance sensor can be
consumed by an apnea detection system to produce a signal
measuring apnea episodes.
• The data elements produced by sensors range from simple scalar
(numerical or categorical) values, to complex data structures.
• Examples of simple data elements include measures such as hourly
average of temperature in a given geographical location, output by a
temperature sensor.
• Examples of more complex data elements include summaries of vital
signs and alerts measured by a patient monitor sensor in a medical
institution.

10/22/2023 2
 Taxonomy of Sensors Used in Medical
Informatics
• Physiological sensors: These sensors measure patient
vital signs or physiological statistics.
• They were first used to measure vitals on astronauts
before appearing in medical institutions, at the
bedside in the 1960s.
• Today, physiological sensors are also available outside
medical institutions, even on pervasive devices (e.g.,
iPhone heart rate monitor applications that make use
of smartphone cameras

10/22/2023 3
 • Wearable activity sensors: These sensors measure
attributes of gross user activity, different from
narrowly focused vital sign sensors.
• Good examples are accelerometers used for gait
monitoring. Shoe manufacturers like Nike have
enabled many of their running shoes with sensors
capable of tracking walking or jogging activities.
• Most smartphones are also equipped with
accelerometers and several wellness management
applications leverage these sensors.

10/22/2023 4
 • Human sensors: Humans play an integral role in the
sensing process. For instance, physicians introduce
important events that relate to the patient health status
during examinations.
• Lab technicians follow rigorous processes to provide
blood content information. Self-reporting (i.e., patients
monitoring their health parameters) is also used in the
management of chronic illnesses like diabetes.
• More recently, with the emergence of social media and
pervasive computing, people use mechanisms like Web
searches and Twitter to generate reports on important
health-related events.

10/22/2023 5
 The sensor data mining process.

10/22/2023 6
 • Contextual sensors:
o These sensors are embedded in the environment around the
user to measure different contextual properties.
o Examples include motion detection sensors, audio and
video sensors, temperature sensors, weather sensors, etc.

10/22/2023 7
 Challenges in Mining Medical
Informatics Sensor Data
• As with standard data mining procedures, healthcare
mining is typically performed in five stages:

10/22/2023 8
 • 1. Data Acquisition: This includes operations involved
in collecting data from external sensor data sources.
• 2. Data Preprocessing: This includes operations
applied to the data to prepare it for further analysis.
Typical preprocessing operations include data
cleaning to filter out noisy data elements, data
interpolation to cope with missing values, data
normalization to cope with heterogeneous sources,
temporal alignment, and data formatting.
• 3. Data Transformation: The includes operations for
representing the data appropriately and selecting
specific features from this representation. This stage is
often called feature extraction and selection.

10/22/2023 9
 • 4. Modeling: This stage, also called mining applies knowledge
discovery algorithms to identify patterns in the data. Modeling
problems can be classified into six broad categories: (1) anomaly
detection to identify statistically deviant data, (2) association rules to
find dependencies and correlations in the data, (3) clustering models
to group data elements according to various notions of similarity, (4)
classification models to group data elements into predefined
classes, (5) regression models to fit mathematical functions to data,
and (6) summarization
models to summarize or compress data into interesting pieces of
information.

• 5. Evaluation: This stage includes operations for evaluation and

10/22/2023 10
 Sensor Data Mining Analytical
Challenges at Each Stage of the Data
Mining Process

10/22/2023 11
 Challenges in Healthcare Data
Analysis
• Despite several standardization efforts, medical sensor manufacturers tend to design
proprietary data models and protocols to externalize sensed signals.
• In healthcare, standard bodies like HL7 and the Continua Health Alliance address data
modeling issues while several IEEE standard protocols address device interoperability issues.
• However, there is a lack of incentives for sensor data manufacturers to adhere to these
standards. With this lack of adherence to standards, mining medical sensor data across
multiple data sources involves several nontrivial engineering challenges, and the design of
custom solutions specific to each sensor data mining application.
• Another key challenge in the acquisition process is related to the protection of user privacy. In
the United States, the Health Insurance Portability and Accountability Act (HIPAA) defines
regulations on access to health data.
• By law, data mining applications that leverage this data must comply with these regulations.
Data de-identification and de-anonymization techniques are often required to comply with
HIPAA.
• Privacy preserving data mining techniques may also be used to extract information from
sensor data while preserving the anonymity of the data.

10/22/2023 12
 • Acquisition Challenges In a clinical setting such as the ICU,
includes different types of physiological sensors (e.g., ECG
sensors, SpO2, Temperature sensors), contextual sensors (e.g.,
RFID sensors linked with care providers, video and cameras)
and human sensors (e.g., care-provider notes, entries in the
Electronic Medical Records).
• More recently, with the emergence of wearable devices,
and network connectivity, additional information is provided
(even in nonclinical settings) by activity sensors (e.g., wearable
devices such as cell phones) and completely nontraditional
sources of information (e.g., community discussions in
healthcare-related sites, aggregated views of user
searches, etc.).

10/22/2023 13
 • Acquiring and integrating this data is nontrivial because of the inherent
heterogeneity and lack of standards and protocols. Physiological sensor
manufacturers have mostly designed proprietary data models and protocols
to externalize sensed signals, despite the efforts of standard bodies like HL7
and the Continua Health Alliance to address data modeling issues, and IEEE
standard protocols to address device interoperability issues. Additionally,
there is little standardization or interoperability studies of contextual and
activity sensors, and data from healthcare providers is captured poorly, often
requiring manual entry and transcription – all making the data acquisition
task extremely complex.
• This has led to the emergence of data aggregators in ICU and EHRs for
general clinical settings, however these aggregator solutions operate only on
a narrow set of sources, and often do not interoperate with each other.
• Hence, mining medical sensor data across multiple data sources has
involved several nontrivial engineering challenges, and the design of
custom solutions specific to each sensor data mining application.

10/22/2023 14
 • These acquisition challenges are compounded by the need to
provide privacy protection for this
often very sensitive personal information.
• This includes conforming with regulations such as the Health
Insurance Portability and Accountability Act (HIPAA) act and
providing appropriate controls
with mechanisms for authentication, authorization,
anonymization, and data de-identification.
• This also requires the design of privacy preserving data mining
and analysis techniques.
• There are also several open, unresolved questions related to the
privacy protection of data generated using
nontraditional, contextual, and activity sensors.

10/22/2023 15
 Preprocessing Challenges
• Data in the real world is inherently noisy. The preprocessing stage needs to address this
problem with sophisticated data filtering, sampling, interpolation, and summarization
techniques (such as sketches, descriptive statistics to minimize the effects of noise.
• The preprocessing also needs to account for the heterogeneity of data, and the lack of
standards adoption by medical sensor manufacturers. Indeed, data generated in different
formats needs to be syntactically aligned and synchronized before any analysis can take place.
• Sensors report data with timestamps based on their internal clocks. Given that clocks across
sensors are often not synchronized, aligning the data across sensors can be quite challenging.
In addition, sensors may report data at different rates.
• For instance while the ECG signal is generated at several 100s of Hz, the EMR may only be
updated hourly. Aligning these datasets requires a careful design of strategies. These
preprocessing techniques need to handle different types of structured data such as transactions,
numeric measurements, and completely unstructured data such as text and images, often
jointly.
• It is critical that the preprocessing of these sources retains the appropriate correlation
structures across sources, so that meaningful and subtle
indicators of patient health can be detected.

10/22/2023 16
 • Furthermore, a semantic normalization is often required to
cope with differences in the sensing process.
• As an illustration, a daily reported heart rate measure may
correspond to a daily average heart rate in some cases,
while in other cases it may represent a heart rate average
measured every morning when the subject wakes up.
• Comparing these values in a data mining application can
yield incorrect conclusions, especially if they are not
semantically distinguished. All of these issues make
the preprocessing task very complex

10/22/2023 17
 Transformation Challenges
• Data transformation involves taking the normalized and cleaned input data and converting it to
a representation such that attributes or features relevant to the mining process can be extracted.
• This may include applying different types of linear (e.g., Fourier Transform, Wavelet Transform) and
nonlinear transformations to numeric data, converting unstructured data such as text and images
into numeric representations (e.g., using a bag of words representations, or extracting color, shape,
and texture properties), and applying dimensionality reduction and de-correlation techniques (e.g.,
Principal Component Analysis), and finally summarizing the result with a set of representative
features that can then be used for analysis and modeling.
• The choice of the appropriate transformations and representations for the features is heavily
dependent on the task that needs to be performed.
• For instance a different set of features may be required for an anomaly detection task, as opposed to a
clustering or classification task.
• Additionally, the choice of appropriate features requires understanding of the healthcare problem
at hand (e.g., the underlying physiology of the patient) and often requires inputs from domain experts.
• For instance, in neurological intensive care environments, spectral decomposition techniques
for feature extraction have been defined, in conjunction with domain experts, to aid the interpretation
of electroencephalograms (EEG) signals for brain activity monitoring and diagnosis of conditions
such as seizures

10/22/2023 18
 • In addition to such signals, human sensing adds different types of
unstructured data that need to be effectively integrated.
• This includes textual reports from examinations (by physicians or nurses)
that need to be transformed into relevant features, and aligned with the rest
of the physiological measurements.
• These inputs are important to the data mining process as they provide expert
data, personalized to the patients. However, these inputs can be biased by
physician experiences, or other diagnosis and prognosis techniques they
use.
• Capturing some of these aspects during the mining process is extremely
challenging.
• Finally, there is a lot of external domain knowledge in open source
repositories, medical journals, and patient guidelines that can be relevant to
patient care, and features should be placed in context of this knowledge for
appropriate interpretation.

10/22/2023 19
 Modeling Challenges
• There are several challenges that need to be overcome in the modeling stage of the data mining
process for medical sensor data. First of all, the time series nature of the data often requires
the application of sequential mining algorithms that are often more complex than conventional
machine learning techniques (e.g., standard supervised and unsupervised learning approaches).
• Nonstationarities in time series data necessitate the use of modeling techniques that can
capture the dynamic nature of the state of the underlying processes that generate the data.
Known techniques for such problems, including discrete state estimation approaches (e.g.,
dynamic Bayesian networks and hidden Markov models) and continuous state estimation
approaches (e.g., Kalman filters or recurrent neural networks) have been used only in limited
settings.
• Another challenge arises due to the inherent distributed nature of these applications. In many
cases, communication and computational costs, as well as sharing restrictions for patient
privacy prevent the aggregation of the data in a central repository.
• As a result, the modeling stage needs to use complex distributed mining algorithms. In remote
settings, there is limited control on the data acquisition at the sensor. Sensors may be
disconnected for privacy reasons or for resource management reasons (e.g., power constraints),
thereby affecting the data available for analysis.
• Modeling in these conditions may also require the distribution of analytic approaches between
the central repository and the sensors. Optimizing the modeling process becomes a challenging
distributed data mining problem that has received only limited attention in the data mining
community.

10/22/2023 20
 • Modeling in healthcare mining is also hindered by the ability to obtain ground truth on the
data.
• Labels are often imprecise and noisy in the medical setting. For instance, a supervised learning
approach for the early detection of a chronic disease requires well-labeled training data.
However, domain experts do not always know exactly when a disease has started to manifest
itself in a body, and can only approximate this time. Additionally, there are instances of
misdiagnosis that can lead to incorrect or noisy labels that can degrade the quality of any
predictive models.
• In clinical settings, physicians do not have the luxury of being able to try different treatment
options on their patients for exploration purposes. As a result, historical data sets used in the
mining process tend to be quite sparse and include natural biases driven by the way care was
delivered to the patient.
• Standard approaches are not well-equipped to cope with this bias in the data, especially as it is
hard to quantify precisely. Furthermore, most studies in medical informatics are retrospective.
• Well-done prospective studies are hard to do, and are often done on small populations, limiting
the statistical significance of any derived results.

10/22/2023 21
 • Evaluation and Interpretation Challenges
• Data mining results consist of models and predictions that need to be
interpreted by domain experts.
• Many modeling techniques produce models that are not easily interpretable.
• For example, the weights of a neural network may be difficult to grasp for a
domain expert. But for such a model to be adopted for clinical use, it needs
to be validated with existing medical knowledge.
• It becomes imperative to track provenance metadata describing the process
used to derive any results from data mining to help domain expert interpret
these results.
• Furthermore, the provenance of the data sets, and analysis decisions used
during the modeling are also required by the experts to evaluate the validity
of the results.
• This imposes several additional requirements on the selected models and
analysis.

10/22/2023 22
 • Generic Systems Challenges
• Beyond analytical challenges, sensor data mining also comes with a set of
systems challenges
that apply to medical informatics applications.
• The mining of sensor data typically requires more than conventional data
management (database or data warehousing) technologies for the following
reasons:
o The temporal aspect of the data produced by sensors sometimes generate large
amounts of
data that can overwhelm a relational database system.
o For example, a large population monitoring solution requiring the real-time analysis
of physiological readings, activity sensor readings and social media interactions,
cannot be supported with relational database technologies alone.
o Sensor mining applications often have real-time requirements. A conventional store-
thenanalyze paradigm leveraging relational database technologies may not be
appropriate for such time-sensitive applications.
o The unstructured nature of some of the data produced by sensors coupled with the
real-time
requirements imposes requirements on the programming and analysis models used
by developers of sensor data mining applications.

10/22/2023 23
 • Hence, sensor mining in healthcare requires the use of
emerging stream processing system technology in
conjunction with database and data warehousing
technologies.
• Stream processing systems are designed to cope with
large amounts of real-time data, and their
programming models are geared towards the analysis
of structured and unstructured sensor data. They are
also time sensitive and analyze data within small
latency bounds.

10/22/2023 24
 • Figure 4.2 presents an extended architecture for
sensor data mining that illustrates this integration.
• The rationale behind this architecture is to use a
stream processing system for the real-time analysis of
sensor data, including the preprocessing and
transformation stages of the analytical data mining
process.

10/22/2023 25
 A generic architecture for sensor data mining systems.

10/22/2023 26
 • The sensor data acquisition is performed by a layer of software
that interfaces with sensors and feeds into the stream
processing system.
• The results of the transformation stage may be persisted in a
data warehouse for offline modeling with machine learning
techniques.
• The resulting models may be interpreted by analysts and
redeployed on the stream processing platform for real-time
scoring.
• In some cases, online learning algorithms may be implemented
on the stream processing system.
• This integration of stream processing with data warehousing
technologies creates a powerful architecture that addresses the
system challenges outlined above.

10/22/2023 27
 Sensor Data Mining Applications
• Intensive Care Data Mining
In 2003, it has been reported that intensivists have to handle over 200 variables, some of them
being temporal, on a per patient basis to provide care. Anecdotal evidence tells us that this
number has increased significantly since 2003 with the emergence of more and more sensing
devices in critical care.
• Today, critically ill patients are often attached to large numbers of body sensors
connected to sophisticated monitoring devices producing these large volumes of physiological
data.
• These data streams originate from medical devices that include electrocardiogram, pulse
oximetry, electroencephalogram, and ventilators, resulting in several kilobits of data each
second. While these monitoring systems aim at improving situational awareness to provide
better patient care with increased staff productivity, they clearly have introduced a data
explosion problem.
• In fact, the vast majority of data collected by these monitoring systems in Intensive Care Units
(ICUs) is transient.

10/22/2023 28
 • In talking with medical professionals, we learned that the typical practice in
ICUs is for a nurse to eyeball representative readings and record summaries
of these readings in the patient record once every 30–60 minutes.
• The rest of the data remains on the device for 72-96 hours (depending on
the device’s memory capacity) before it times out and is lost forever.
• Hospitals are simply not equipped with the right tools to cope with most of
the data collected on their patients, prompting many to state that medical
institutions are data rich but information poor.
• The potential of data mining in this area has been recognized by many.
Several efforts are underway to develop systems and analytics able for the
modeling of patient states and the early detection of complications.
• In general, early detection of complications can lead to earlier interventions
or prophylactic strategies to improve patient outcomes.
• Early detection rests on the ability to extract subtle yet clinically meaningful
correlations that are often buried within several multimodal data streams
and static patient information, spanning long periods of time.

10/22/2023 29
 • Systems for Data Mining in Intensive Care Modern patient monitors have evolved
into complex system that not only measure physiological signals but also produce
alerts when the physiological state of the patient appears to be out of
range.
• State-of-the-art patient monitors allow physicians to program thresholds defining
normality
ranges for physiological systems. For example, one can program a patient monitor to
produce an audible alert if the oxygen saturation level of the blood is below 85%.
The values of these thresholds
are typically obtained from general guidelines or from data mining processes. Such
simple alerting schemes are well known to produce very large numbers of false
alarms.
• In, it is reported that more than 92% of alarms generated in an ICU are of no
consequence.
• Furthermore, there are many complex physiological patterns of interest to
physicians that cannot be represented by a set of thresholds on sensor data streams.
Several research initiatives are addressing this problem with the design of platforms
facilitating analysis beyond the simple thresholding capabilities of
existing patient monitoring systems.

10/22/2023 30
 • One example is BioStream a system that performs
real-time processing and analysis of physiological
streams on a general purpose streaming infrastructure.
• The authors use ECG data along with temperature,
oxygen saturation, blood pressure, and glucose levels
as inputs into patient- specific analytic applications.
• The system supports a different processing graph (for
analysis) per patient, where the graph can be
composed of system supplied operators (functions)
and user implemented operators.

10/22/2023 31
 State-of-the-Art Analytics for Intensive
Care Sensor Data Mining
• State-of-the-art analytics and mining approaches for in-hospital sensor data
monitoring (Figure 4.3) tend to generate innovations on data preprocessing
and transformation.
• Modeling is typically done with well-known families of machine learning
techniques such as classification, clustering, and dynamic system modeling
with sequential learning.
• These analytical techniques often attempt to derive features from
physiological time series to model the inflammatory response of the body,
as it is known to be highly correlated with early sign of complications in
general.
• The inflammatory response is a reaction from the body to different harmful
stimuli such as pathogens, various irritants, or even damaged cells.
• Hence, accurate modeling of it enables a wide range of early detection
applications in intensive care. In particular, devastating complications such
as sepsis are known to produce an inflammatory response well before the
appearance of clinical symptoms

10/22/2023 32
 • Sensor Data Mining in Operating Rooms
Data mining applications that relate to operating rooms tend to focus on the analysis
of Electronic Medical Record data where most sensor data inputs are filtered and
summarized.
• For example, in, EMR data is used to improve the efficiency of operating rooms, in
terms of scheduling (start times, turnover times) and utilization.
• In, knowledge management and data mining techniques are used to improve
orthopedic operating room processes, yielding more effective decision making.
A few researchers have reported applications directly mining physiological sensor
data produced by operating room monitoring systems.
• Exceptions are presented in where the authors correlate EEG signals with cerebral
blood flow measurements for patients undergoing carotid endarterectomy.
• This finding is quite valuable as it proves that EEG signals can be used to monitor
complex mechanisms including cerebral blood flow for this patient population.
• In, machine learning techniques are proposed for the closed-loop control of
anesthesia procedures.
• In, the authors present a prototype of a context-aware system able to analyze patient
data streams collected in an operating room during surgical procedures, to detect
medically significant events, and apply them in specific EMR systems.

10/22/2023 33
 • General Mining of Clinical Sensor Data
• Recent stimuli from the federal government and the increased ease of adoption of
electronic health records system has led to widespread use of EHR in clinical
practice.
• Large providers such as EPIC and McKesson have essentially unified the elements
of data entry by having common platforms, although the use of free text and
contextual rather than templated data is more common.
• EHR systems are a unique healthcare sensor, since real-time data and vast data
troves are similar to other sensors, yet the relatively unstructured data makes it
difficult to view this as a typical sensor.
• They typically contained structured and unstructured comprising of all the key
administrative clinical data relevant to patients, demographics, progress notes,
problems, medications, vital signs, past medical history, immunizations, laboratory
data, diverse test results, and radiology reports.

10/22/2023 34
 • Furthermore, there are no widely accepted standards for the
representation of all these data points stored in EHR systems.
Several code systems (e.g., ICD-9, ICD-10, CPT-4, SNOWMED-
CT) and interoperability standards (e.g., HL7, HIE) are in use by
many systems but there are no overarching standards that EHR
vendors are adhering to.
• Despite this lack of global standardization that is hindering the
realization of very large-scale data mining, many researchers are
spending considerable efforts to analyze these data sets to improve
healthcare in general.
• Mining of such sensors have been undertaken by various groups. The
use of EHR mining todetect delays in cancer diagnosis, hospital-
acquired complications, and the ability for groups to
develop high throughput phenotyping to identify patient cohorts
based on modular and consistent resources has been reported

10/22/2023 35
10/22/2023 36