Добавил:
ivanov666
Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз:
Предмет:
Файл:Intelligent data analysis in medicine. Study aid
.pdf
21
CHAPTER 3. MACHINE LEARNING IN MEDICINE
Machine learning systems are far from uncommon in healthcare. For example,
there is software that can be considered as an autonomous entity, it learns from millions
of patient records and further acts as an expert system that helps in making a diagnosis.
There are many sources and resources in the field of AI and machine learning.
Knowledge is acquired by the agent through sensors or environmental data.
Knowledge can be used to decide how to act; if stored, they can be used to calculate
previous states of the system, as well as to determine how actions might affect the
environment. Machine learning is data-driven. Thus, the advent of ubiquitous data and
pervasive computing is well suited to the development of systems with real-world
applications.
Training data
Training data is the data used by a learning algorithm to explore possible
hypotheses and infer patterns. An example would be some sample x, including its
output from the objective function.
In 1959 machine learning was defined as “a field of study that gives computers
the ability to learn without being explicitly programmed.” Machine learning was born
from pattern recognition and the theory that computers can learn without being
programmed to perform specific tasks. As a result, learning is data-driven and
intelligence is acquired through the ability to make effective decisions based on the
nature of the learning signal or feedback. The usefulness of these decisions is assessed
in relation to the goal. Machine learning focuses on developing algorithms that adapt
to the presentation of new data and new patterns. Machine learning illustrates the
principles of data mining, but can also infer correlations and learn from them to apply
them to new algorithms. The goal is to mimic a person's ability to learn through
experience and complete a given task with little or no external (human) assistance.
As with human learning, machine learning consists of many approaches. At its
core, basic rules and patterns are remembered. We learn by retrieving information
(reading, listening, learning something new). And thirdly, we learn by example.

22
For example, if we were taught the squares of numbers by showing a set of numbers
Y = {1, 4, 9, 16, ...} and demonstrating that Y = nn, then we would typically understand
the concept of square numbers without explicit teaching every square number.
Typical software structure
Traditional software development and machine learning have the common goal
of solving a problem or set of problems. The approach to problem solving is what
distinguishes the two paradigms. Traditional software development, or programming,
refers to the task of computer science or automation a task so that a function or
program, given an input, produces correct output. In other words, it is necessary to
write a function f such that, given input data x, the output is y = f(x). This is done using
logic, typically if-else statements, while loops, and logical operators.
Machine learning differs from conventional programming in that instead of
being given instructions about a function f, the computer is given input x and output y
(a training set) and is expected to determine or predict function f. While traditional
programs were written by humans to solve a problem, machine learning programs learn
through reasoning to solve a problem from examples, rules, and information.
Many classifiers in machine learning help make predictions. Machine learning
programs can also learn to generalize from data and help with uncertainty problems
through the use of statistics and probability-based methods. Models can learn from
previous calculations or experience to produce reliable, repeatable solutions and
results.
An approach to classifying machine learning methods
Tasks for machine learning algorithms are generally classified as:
– supervised learning (also known as inductive learning);
– training without a teacher;
– semi-supervised learning;
– reinforcement learning.
In general, the classification of machine learning methods can be presented as
shown in Fig. 4.

23
Fig. 4. Typical classification of machine learning methods
Tutored training
In supervised learning, algorithms are presented with examples of inputs and
subsequently required outputs in the form of training data, with the goal of identifying
general rules that map inputs to outputs. The input data is called training data and has
a known result (output) associated with it. The training data determines how the
algorithm is developed. The model is created through a learning process in which it
makes predictions and is adjusted if the resulting predictions are incorrect.
Training continues until the model reaches the desired level of accuracy on the
training data set. Supervised algorithms include, for example, logistic regression.
Supervised machine learning algorithms can apply what has been learned in the past to
new data, using examples of labeled data to predict future events. The learning
algorithm can also compare the output with the correct data to find any errors and
change the model accordingly. Supervised learning tasks can be used in the following
forms:
– classification: to predict an outcome based on a training data set where the
output variable is represented as separate categories. Models are built by inputting
training data in the form of pre-labeled data. Classification methods determine decision
boundaries and include, for example, support vector machine (SVM), naive Bayes
classifier, Gaussian Bayes, k-nearest neighbors (kNN), and logistic regression. An
example of classification is a diagnosis (healthy or sick) based on a set of symptoms;
– regression. In general, a regression problem is very similar to a classification
problem. The only difference between classification and regression is that regression

24
is the result of a given sample, where the output variable is in the form of real values.
Examples may include height, body temperature, weight, and other characteristics.
Linear regression, polynomial regression, ensembles, decision trees, and neural
networks are examples of regression models;
– forecasting is a method of constructing future data based on past and present
data. Also known as time series forecasting;
– ensemble is a type of supervised learning that combines several different
machine learning models to predict an outcome on a new sample.
Unsupervised learning
When users point to systems that can learn on their own, they are actually
referring to unsupervised learning. In unsupervised learning, the learning algorithm
does not receive a label for the data, leaving the algorithm itself to find the existing
structure in the input. Since the data is not initially labeled, there is no assessment of
the accuracy of the structure produced by the algorithm. Both the original classification
and the labels may be missing from the data. As a result, the model learns through
interpretation, i.e. through searching for hidden structures and logical conclusions in
the input data. This can be done through rule extraction, data redundancy reduction, or
through data reorganization.
Unsupervised learning includes clustering, dimensionality reduction, and
association rule learning. The algorithm may never find the correct result, but instead
simulate the underlying structure of the data.
There are three main types of unsupervised learning problems:
1) association — determining the probability of coincidence of elements in a
collection. Widely used in marketing as well as healthcare. For example, the percentage
probability of developing any form of cancer in the presence of other diseases.
Association is similar to classification, however, any attribute can be predicted to be
continuous in the association, whereas classification is binary;
2) clustering — grouping elements in such a way that elements in one cluster
are more similar to each other than to elements from another cluster (Fig. 5);
3) dimensionality reduction. It can be achieved by selecting and extracting
individual features. Dimensionality reduction mathematically re-transforms the data.
Feature extraction transforms data from high-dimensional space to low-dimensional

25
space. This may include reducing data set variables while maintaining data integrity and
ensuring the most important information is presented. Feature extraction techniques can
be used to reduce dimensionality. Typically, a new feature set is created from the original
feature set. An example would be combining all of a patient's clinical test results into a
health risk score based on results that have demonstrated an impact on mortality.
Reducing the number of dimensions simplifies visualization, especially in two or three
dimensions, and also reduces the time and space required for storage. It uses hidden
Markov models, principal component analysis, singular value decomposition, neural
networks and deep learning. Deep learning applies the architecture of deep neural
networks, which are types of machine learning algorithms.
Fig. 5. Data clustering (3 clusters)
Semi-supervised learning
Semi-supervised learning is a hybrid method in which the input is a set of both
labeled and unlabeled data. The model may also include an output. The model also
learns structures to organize data and make predictions. Examples of problems are
classification and regression.

26
Reinforcement learning
This is where systems interact with a dynamic environment in which an agent
must perform a specific task. Reinforcement learning combines machine learning,
behavioral psychology, ethics, and information theory. The algorithm is given
feedback regarding rewards and punishments as it solves the problem. Reinforcement
learning allows an agent to select the next best action based on its current state, and
learning behavior that maximizes reward. The optimal action (or optimal strategy) is
usually learned through trial, error, and feedback. This allows the algorithm to
determine ideal behavior in context.
Reinforcement learning is commonly used in robotics, for example a robot
vacuum cleaner learns to avoid collisions by receiving negative feedback by bumping
into tables, chairs, etc. At the same time, today reinforcement learning is combined
with computer vision methods. Reinforcement learning differs from standard
supervised learning in a way that correct examples are not always presented and
suboptimal solutions are not explicitly corrected. The focus is on real-time
performance. Similar to how people learn, reinforcement learning involves rational
behavior towards learning from experience and as such helps to understand how people
learn from experience to make good decisions. Reinforcement learning is used in
autonomous vehicles. In the real world, the agent needs to take into account aspects of
the environment, including current speed, road obstacles, surrounding traffic, road
information, and operator controls. The agent learns to act or develop rules that
correspond to a particular state of the environment. This leads to new ethical dilemmas
that AI and machine learning create for agents with reward-oriented programming.
Data collection
Data mining, also known as knowledge discovery in databases, is defined as the
process of mining data to extract new information (including patterns and
relationships) using complex algorithms. It can be applied to various aspects of life,
medicine and industry and allows you to:
1) predicti patterns based on trends and behavior;
2) forecast based on probable results;
3) analyze large data sets (especially unstructured ones);
4) cluster by identifying previously unknown facts.

27
Data mining uses the power of machine learning, statistics, and software
algorithms to form the basis of analytical data mining techniques.
Data mining typically uses large amounts of data and can also produce simpler
descriptive analytics. For example, in healthcare, data mining applications can be used
to identify conditions or medical procedures that frequently occur together, e.g.
correlate with each other. Data mining also allows you to make predictions by
identifying relationships in data. This could, for example, help insurance companies
identify patterns of risky behavior among customers.
Machine learning can be used for data mining. Data mining will inform, what
decisions are made, such as which learning models are best suited to solve the current
problem. However, it also contains methods beyond those included in machine learning
(see Fig. 3). In data mining, the goal is to discover unknown knowledge, while the
effectiveness of machine learning is assessed in terms of reproducing already known
knowledge. For example, if you have a dataset of a patient's blood pressure, you can
perform anomaly detection, which is considered a data mining task to identify
previously unknown patterns. To solve this problem, machine learning methods can be
used, for example, the k-means algorithm for cluster analysis. Data mining is not the
same as, for example, natural language processing. Natural language processing can be
used as a technique used in data mining to help the user understand text.
Parametric and non-parametric algorithms
Algorithms can be parametric or non-parametric in form. Algorithms that can be
simplified to a known final form are called parametric, while non-parametric
algorithms learn a functional form from the training data. In other words, when using
nonparametric algorithms, the complexity of the model increases with the amount of
training data.
Parametric models have a fixed structure or set of parameters. As a result, they
perform faster than nonparametric models. However, nonparametric models can
achieve higher accuracy if they are provided with sufficient training resources in terms
of training data and time. Linear models such as linear regression and logistic
regression are parametric, whereas k-nearest neighbors, neural networks, and decision
trees are nonparametric.

28
Examples
Categorization is one of the “skills” that the system can learn from known
examples. It can be used for forecasting tasks, such as predicting the risk of a disease
or the need for certain resources. Some problems arise in healthcare settings, such as
storing medical appointments and diagnoses in handwritten form. Poorly organized
data storage can lead to additional errors in models. For classification purposes,
examples are usually given in the form of positive and negative results.
Background information
Background information refers to the knowledge and axioms associated with
machine learning problems. This could include metadata, attributes, or relationships
between concepts. For example, in a hypertension prediction method, the diagnosis
would be related to the likelihood of hypertension based on medical indications and
data. This is the information that is used in the concepts learned to achieve
categorization.
Data errors
It is always important to keep in mind that real-world errors may occur, such as
misclassifications, missing data, incorrect reference information, and duplicated data.
Errors can occur, for example, when digitizing paper documents and machine reading
written text. It's important to remember that a model is only as good as the data it uses.
It's worth taking the time to review and clean your data to minimize human error.
Data preparation
Machine learning algorithms learn from pre-trained labeled data. It is extremely
important to provide the model with reliable and correct training data. Data should be
prepared in a format that is easy to use. In a real-world scenario, this is usually the
interpretation of the data that will be used to model the problem and export the data.
The data must then be processed to ensure correct formatting, removal of erroneous
data, and correction of any missing data.

29
The data set size may be larger than required, so an additional subsample is often
created. Data preprocessing is necessary to obtain accurate and reliable data. Accurate,
reliable data is the key to obtaining reliable and valid results. First, we need to select
the data. When first extracted they are unprocessed. You may need to get your data in
order or need to clean it up, which includes deleting or correcting incorrect data.
Selecting Attributes
Attribute selection is also called variable selection or feature selection.
Essentially, this is filtering and refers to the selection of a subset of the original set of
examples that is most suitable for modeling and prediction. Feature selection differs
from dimensionality reduction as dimensionality reduction techniques create new
combinations of attributes. Feature selection includes and excludes attributes rather than
creating new ones. This is where domain knowledge is an important advantage.
Understanding the context in which a task fits allows you to better define the feature set.
Data Conversion
In most industries, especially healthcare, many situations require the use of
anonymized data, so some features and functions may be removed from the set entirely.
You may need to change the formatting; for example, you may need to export data to
a flat file rather than a traditional relational database. You may only want to work with
a subset of the sample data.
Digitized text, such as handwritten text, may contain errors, so you need to check
your data sets for errors, inaccuracies, and inconsistencies. This is a very common
problem. The data may also need to be transformed. This is usually determined by the
algorithm you are using and the data available. You can also apply data scaling.
Choosing a teaching method
A common question for any machine learning newbie is “Which algorithm
should I use?” The selection of the algorithm depends on several factors, including the
size, quality and nature of the data, the timing of the task and available resources, and
the motivation for using the data. The learning method is also called a decision
algorithm because each machine learning approach represents data differently. It is
almost impossible to predict which approach will work best with the data.

30
Tools for applying teaching methods
Machine learning tasks are usually performed in various programming
languages, mainly R, Python, Scilab and SQL. Java and C are also widely used to solve
this kind of problem. Let's look at the most popular machine learning tools today:
1. R is commonly used for statistical analysis. It allows you to interpret and
explore data using statistical methods and graphs, and also contains a wide range of
machine learning algorithms.
2. Python is a language well suited for machine learning. Extensions such as
NumPy and SciPy are especially useful for machine learning and data analytics;
3. Scilab is the language that many data mining professionals start with. It is
useful for rapid prototyping as it contains a large repository of machine learning
algorithms;
4. SQL is a language used to manage data stored in a relational database
management system.
Questions for self-control
1. What is training data? What are they used for?
2. What are the requirements for training data?
3. What classifications of machine learning methods exist?
4. What determines the choice of a machine learning method to solve a specific
problem?
5. What tools are used in machine learning methods?
6. What is data conversion and what is it used for?
7. Give the main differences between parametric and nonparametric algorithms.
8. What steps does data preprocessing include?
Соседние файлы в предмете [НЕСОРТИРОВАННОЕ]
