Добавил:
ivanov666
Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз:
Предмет:
Файл:Intelligent data analysis in medicine. Study aid
.pdf
11
to paper documents or scanned images for further document processing. Health care
providers collect health biomarkers along with other structured indicators.
Regardless of the source, in order to learn, data must be converted into
information. For example, the blood glucose data of a person with diabetes and its input
into an app is much more meaningful when the blood glucose targets are known to the
system, so it can be understood whether the data entered is within the recommended
target or not.
In the 21
st
century almost every action leaves a transaction trail in one form or
another. Modern mobile devices are more than 30 thousand times more productive than
NASA computers from the era of the Apollo mission, which brought man to the Moon.
Computers are not only smaller, but also significantly more powerful than ever.
As organizations benefit from big data sources, there appeared a shift toward
embedding data-derived insights into every aspect of the user experience—from
product or service purchases to in-app user interactions. The value of data increases
when it is taken in its raw form and transformed into knowledge that changes practice,
determined by the specifics of the project and context. For example, value may result
from faster identification of gaps in adherence and evidence-based patient care. Based
on intelligent analysis, more effective exchange of data and ideas can be organized
within a hospital or other medical organization in order to reduce the number of
incorrect diagnoses.
It's unlikely that AI agents will ever completely replace doctors and nurses, but
machine learning and AI are changing and complementing the healthcare industry.
Machine learning improves the quality of diagnostics, makes it possible to more
accurately predict the course of the disease and opens up new prospects in personalized
medicine. Imagine a situation where a patient comes to the doctor with heart pain. After
listening to the symptoms, he enters them into his computer, which uses the highest
quality evidence to effectively diagnose and even prescribe further treatment. Or, for
example, a patient has an MRI scan and an intelligent computer system helps the
radiologist detect any abnormalities that would be too small for the human eye to see.
Or an electronic wristwatch continuously collected blood pressure and pulse data while
continuously monitoring blood glucose levels in real time. Finally, medical records and
family history are evaluated by a computer system that suggests treatment pathways
tailored to the individual.
Beyond privacy and data governance, the implications of what we can gain from
combining different data pools are exciting. Technologies that track data to predict

12
disease outbreaks are already proliferating. This is often done using real-time data
sources such as social media, as well as historical information from the Internet and
other sources. Malaria outbreaks were predicted using artificial neural networks that
analyzed data including rainfall, temperature, number of cases and various other data.
The possible combination of genomic knowledge with machine learning algorithms
will provide the opportunity to learn about disease risk, improve pharmacogenetics,
and provide more effective treatment pathways for patients (Fig. 2).
Fig. 2. Stages of building intelligent systems in medical practice
Digital therapy from Diabetes Digital Media, the Low Carb Program, helps
people with type 2 diabetes and prediabetes put their disease into remission. The
application provides personalized training and integrated health tracking, studying the
progress of each individual user and the community as a whole. At the end of the year,
most participants who complete the program reduce their dependence on medications
while saving money. The use of machine learning in pre-drug discovery has potential
for a variety of applications, from initial screening of drug compounds to predicting
success based on biological factors. This includes research and development
technologies such as next generation sequencing. Medicines do not have to be
pharmacological in appearance. The use of digital solutions and the aggregation of realworld patient data are providing solutions for diseases that were once considered
chronic and progressive. For example, the Low Carb Program app, which is used by
more than 300 thousand people with type 2 diabetes, leads to remission of the disease
in 26 % of patients who complete the program after 1 year.
Readmission to hospital is a serious problem in healthcare. Doctors as well as
healthcare organizations are trying their best to keep patients healthy, especially when
they return home after treatment in the hospital. Organizations like NextIT have
developed digital health coaches that are similar to a virtual customer service
representative on an e-commerce site. The assistant asks questions about the patient's
medications and reminds him to take the necessary medications, asks about the
symptoms of his condition and transmits relevant information to the doctor in real time.

13
For AI and machine learning to be fully adopted and integrated into the
healthcare system, several important key issues need to be addressed.
There is a huge mismatch between stakeholder expectations and the outcome of
AI and machine learning. Sharing ideas, methodologies, and assessments is critical to
driving the innovation needed to advance AI and machine learning in healthcare. Data
manipulation, including data exchange and integration, is fundamental to the transition
of the healthcare system to the implementation of the logic of personalized medicine.
Building data science teams focused on data science is key to a successful healthcare
strategy. Increasing the value of data analytics’ results for both patient and clinician
requires the data itself and, therefore, data scientists.
There are many obstacles that still need to be overcome. Currently, data is
fragmented and difficult to integrate. Patients store data on their phones, watches and
other devices, and doctors routinely collect biomarkers and demographic data. These
data are never combined based on the patient's previous treatment experience. There is
also no infrastructure to meaningfully and reliably parse and analyze this large data set.
In addition, electronic medical records are being slowly introduced and greater
preference is being given to their paper counterparts.
At the same time, organizations face security and government compliance
challenges, especially around managing patient data and making it available at all
times. Moreover, many healthcare organizations use outdated versions of software,
which may be more vulnerable to cyber attacks.
Along with the topic of secure data storage comes the topic of data management.
Medical data is personal and difficult to access. It is widely believed that a significant
proportion of patients will not want to share their data due to privacy concerns. A
serious problem in machine learning is the ability to justify the result. As AI becomes
more embedded in everyday life — integrated into our homes, work, and roads — it is
imperative that we ask ourselves how and why intelligent systems do what they do and
what they base their results on. In machine learning, self-learning systems create their
own inductive bias based on previous experience. As such, the behavior of a system
can be highly dependent on the data environment in which it operates. It's only when
algorithms are used in real-world environments that people discover built-in biases,
which are often reinforced by working in the real world. Transparency is not the only
criterion; It is necessary to ensure impartial decision-making in order to fully trust the
capabilities of AI systems. What gives people confidence is the ability to see through

14
the black box and understand the cause-and-effect relationships behind the conclusions
of computers.
The first AI systems were developed in Prolog and Lisp. Today, most machine
learning systems are written in Python due to the fact that many of the mathematical
foundations of machine learning are available in the form of a huge number of updated
and improved libraries. However, machine learning algorithms can be developed in
most programming languages, including C, C++, C#, Java, etc.
The potential applications of machine learning in medicine are vast and
extremely exciting. Intelligent systems can help determine the risk of developing
cancer or build a course of treatment based on bio-markers in real time. With it also
comes enormous responsibility and moral questions. We don't yet fully understand
what can be learned from health data. As a result, the ethics of using intelligent systems
in medicine is one of the fundamental topics for consideration.
Questions for self-control
1. What is artificial intelligence?
2. What do the basic concepts of artificial intelligence include?
3. What are the main problems of creating and implementing intelligent systems
in medical practice?
4. Define the term “machine learning.”
5. Can an intelligent computer system completely replace a medical specialist?
Justify your answer.
6. To diagnose, treat and monitor what diseases are intelligent systems used
today?
7. What is the problem with storing medical data securely?
8. What programming languages are used to develop medical intelligent
systems?

15
CHAPTER 2. MEDICAL DATA AND THEIR FEATURES
Data and their structure
Data are forms of information representation that information systems and their
users deal with.
Data can take various forms — symbolic, textual, numeric, graphic and audio.
At their core, in machine learning theory, they represent a set of values of qualitative
or quantitative variables. In order for data to become information, it must be
interpreted. Information is data that is organized or classified in some way.
Information itself is processed data on the basis of which certain decisions can
be made and further actions can be planned. Today, healthcare is experiencing an
information revolution thanks to two significant points:
1) the need to contain the ever-increasing costs of material support creates new
incentives and reimbursement structures;
2) digital health empowers consumers through the introduction of more and more
digital devices.
Clinical trends are changing. The emergence of social networks and immediate
access to information via the Internet, participation in medical communities increase
the level of patient awareness about their health. Patients generate data almost
continuously and have the right to manage it, including transferring it between various
medical services. At the same time, medicine strives for personalized treatment based
on an evidence-based approach to decision making are developed. Both patients and
healthcare providers express a desire to receive all available clinical data in order to
make better evidence-based treatment decisions. Research shows that one in three
patients would like to share their data with their doctor, but only one in five patients
have the opportunity to do so.
Combining individual data sets into more meaningful samples also provides
more reliable evidence, since features in small subsamples are quite infrequent and do
not immediately appear in statistical analysis of the data. For example, pooling patients
with different diseases may more reliably represent trends in patient health status for
both groups than separate samples of patients. The use of big data in modern healthcare
offers the prospect of significant improvements in many areas. As data mining-based
medicine advances, it becomes increasingly evidence-based and personalized.

16
Data classification
There are two classes of data — structured and unstructured. Structured data
usually refers to something stored in a database, i.e. in some structure corresponding
to a model or diagram. Many organizations are familiar with this form of data and may
already be using it effectively. Most organizations use a simplified form of data
description for storage, such as spreadsheets and electronic processors. In medical
centers, data is most often stored in a structured form. Readings from medical sensors,
smartphones, smart watches and various devices are usually in the form of structured
data, be it blood glucose readings, steps taken, calories burned, heart rate or blood
pressure.
Structured data is similar in nature to machine language. Structured data, well
organized in its format, facilitates information retrieval operations. For this purpose
data is usually stored in a relational database.
All other methods of data storage are unstructured. Unstructured data does not
have a predefined model or scheme. In addition, they do not have an identifiable
structure, and this creates problems for querying and retrieving information. Email,
text messages, social media posts are good examples of unstructured data. Data volume
is expected to grow 800 % over the next few years, and 80 % of this data will be in the
unstructured form. The problem that unstructured data presents is volume. The lack of
structure makes compilation and interpretation a labor-intensive task in terms of time
and computational resources. Any medical organization will benefit from using an
unstructured data analysis mechanism.
Structured data is generally easier to analyze than unstructured data due to the
raw and unorganized form of unstructured data. However, analyzed on unstructured,
informal data sets is being improved by using data science and machine learning
techniques, including natural language processing, to help understand and classify raw
data. It is not always possible to transform unstructured data into a structured model.
For example, an email communication or notification contains information such as time
sent, subject, and sender as uniform fields. However, the content of the message cannot
be easily analyzed and classified. Organizations that develop various types of spam
filters to deal with such problems.

17
Big Data
The term “big data” is used to describe voluminous, digital, most often
unstructured data, the sources of which are various devices, generators, etc.
Big data is a popular term that defines data sets that are too large to be stored
and processed in a traditional relational database system.
Thus, the term "big data" is somewhat vague: while size is certainly an important
characteristic of big data, scale itself is not defined. There are no arbitrary criteria that
make big data really big. Thanks to big data analytics, we can identify hidden patterns,
unknown correlations, trends, preferences and other information that can help
stakeholders make more informed and correct decisions. Machine learning provides a
set of techniques that can be applied to such data sets (Fig. 3).
Fig. 3. Areas of knowledge used in big data processing
Big data was first described in 2001 and the following characteristics were
introduced:
– volume — the amount of generated and stored data. Big data usually takes up
a significant amount of disk space. The huge size of the data leads to difficulties in the
form of storage, it’s indexing and searching;
– diversity — big data varies in type and nature, which requires effective storage
and analysis, as well as the implementation of systems for processing such data;
– speed — big data is received at speed, which also has a number of
requirements.

18
Patients and doctors are the source of big data. Patients generate data through
clinical examinations, wearable devices, mobile applications. Meanwhile, doctors
leave behind a wealth of clinical data (such as written notes, images), patient histories,
and data generated by various devices.
Data volume
Storage and transmission have always been key issues when implementing big
data in AI. The transition of data centers from on-premises to existing cloud storage
has solved these problems to some extent. Storage costs were significantly reduced,
and flexibility in data placement, collaboration and disaster recovery was provided.
Cloud storage and cloud services refer to distributed computing carried out through
network protocols. Moore's Law predicted not only the capacity of data storage, but
also its cost. The healthcare field has a lot to do with big data. Data are continuously
collected in both digital and other formats and cover a variety of patient behavioral,
epidemiological, environmental and health information. Data is processed, distributed
and generated for use in diagnostic purposes and scientific research. To date, patient
medical records are not available in digital form, but are stored in paper format, which
significantly slows down the implementation of intelligent systems in this area. Many
organizations strive to make information publicly available, but often run into legal
problems.
For continuous monitoring, patients and healthcare organizations are switching
to smartphones, wearable devices such as smart watches and fitness trackers, home
sensors, and intelligent personal systems for analyzing accumulated data. Such
digitally stored data is plentiful, accessible, and considered reliable.
Heterogeneity of data
Another important property of data is diversity or, as it is also called,
the property of heterogeneity. This applies not only to differences in data types, but
also to their sources and uses. Twenty years ago, we stored data in spreadsheets and
databases. Today, data can be in the form of photographs, information from sensors,
encrypted files, etc. This variety of unstructured data creates challenges for data
storage, retrieval, and analysis. This is one of the areas where machine learning can
greatly help people.

19
Data relevance
Similar to the validity of big data, there is also the issue of data relevance, which
concerns whether the data is correct, accurate, and timely for its intended use. In
clinical applications, validity may be considered a priority property of a data set to
ensure that only useful and relevant data are used. The veracity or reliability of data is
absolute, while reliability depends on the context.
Variability
Big data is volatile. The variability property characterizes data with regularly
changing values. Variability is very relevant when analyzing natural data.
Variability is often confused with heterogeneity. For example, a florist might sell
five types of roses. This is diversity. Now, if you go to the florist for two weeks in a
row and buy the same white rose every day, each day it will have a slightly different
shape and scent. This is variability. To conduct proper data mining, algorithms must
be able to understand the context of texts and decipher the exact meaning of a word in
a given specific passage. This remains a challenging task despite advances in natural
language processing techniques.
Metadata
Metadata is data about data, i.e. descriptive data about each instance or
individual piece of data. Metadata provides detailed information about a file,
supporting the ability to discover patterns and trends from the metadata. Metadata
provides information about a file's origin, date, time, and format; it may also include
notes or user comments.
The key to efficient use of resources, especially time, in strategic information
management is to ensure that all data samples are correctly named, tagged, stored and
archived. This facilitates faster linking of data sets and supports a consistent
methodology for managing that data so that individual records are easy to find, retrieve,
and share. The hype around big data has eclipsed metadata. However, with big data
comes big metadata, allowing organizations to generate and use knowledge. Metadata
can be useful in any data-driven project.

20
Questions for self-control
1. What is data?
2. What is information?
3. What is the key difference between data and information?
4. What are the ways to classify data? Give examples.
5. What is big data? What are the characteristics of big data?
6. What is the problem of data heterogeneity?
7. Why is the relevance of data in mining technologies important?
8. How is data variability different from heterogeneity?
9. What is metadata? What is it used for?
Соседние файлы в предмете [НЕСОРТИРОВАННОЕ]
