The 5th IEEE International Conference on E-Health and Bioengineering - EHB 2015

Grigore T. Popa University of Medicine and Pharmacy, Iaúi, Romania, November 19-21, 2015

Benefits of Using Artificial Intelligence in Medical

Predictions
Adriana ALBU1, Loredana STANCIU2
Affiliation 1: Automation and Applied Informatics Department, Politehnica University Timiúoara,
Timiúoara, România, adriana.albu@aut.upt.ro
Affiliation 2: Automation and Applied Informatics Department, Politehnica University Timiúoara,
Timiúoara, România, loredana.stanciu@aut.upt.ro

Abstract − Medicine is a domain where predictions are very
important. This paper encourages the use of artificial
intelligence (AI) as technological support for decision-making in
medical life. Not because the human intelligence would not be
enough, but because the mechanisms of artificial intelligence
have several benefits that are suitable for this area. These
positive features come as a result of combining two strong
qualities in the same systems: the precision of mathematics and
the power of current technologies. It would be a real waste to
ignore them.
Keywords − decision-making, inference, Bayes theorem,
prediction.
they bring to medical decision-making. There are also
discussed some weak points of each mechanism; it is
necessary to underline them in order to determine the most
suitable AI method for a specific task. The mechanisms that
are described in this paper are applied to decisional systems
developed for liver’s diseases in order to emphasize the
features of these mechanisms [6]. Some of these features
(advantages or disadvantages) can be found in the domain’s
literature and they are evidenced by the decisional systems
presented here. Some others have been noticed once these
systems were developed.

INTRODUCTION
ARTIFICIAL INTELLIGENCE
First of all, a decisional system is not intended to replace
Artificial intelligence domain contains mechanisms that
the physicians (however, it would be impossible). It only
behave in a way considered intelligent if a human would act
makes suggestions. The final medical decision belongs to
so [2]. The work in AI field started after the Second World
humans. But some automatic systems are useful for two
War; therefore, it is a new field beside other domains of
important reasons:
science, an area that can still be exploited by the researches.
⎯ to avoid simple, usual things that are time
The mechanisms of this field can be successfully used to
consuming and overload the medical stuff with a
solve problems connected to a large variety of applications
lot of activities that could be easily performed by
that need to have an intelligent behavior. This behavior can
a machine (for instance, to verify if the medical
be related to the capacity of perception and learning, to the
features of a patient recommend that person for a
ability to make decisions (e.g. in a game, when a car must be
certain job, or to predict when is the optimal time
driven, when a medical diagnosis or a treatment scheme must
to supply the store for a specific medicine);
be set) or to other aspects that involve intelligent reasoning.
⎯ to point out medical conditions which can be
The intelligent reasoning is a feature of human beings. The
hardly detected by humans; a suggestion or an
artificial intelligence comes with a rationalist approach. This
alert can make the difference between life and
doesn’t mean that the human approach is irrational [3], but
death (for instance, to detect some cancerous cells
there are a lot of other mechanisms that should be involved.
in an image, or to predict the risk of a genetic
These mechanisms are part of mathematics (including logic
disease).
and probabilities) and engineering and they are using the
A medical decision-making system can benefit of some
power of computer science. Therefore, this lead to the main
important advantages of technology: it is not perturbed by
benefit of artificial intelligence: it creates tools that should
causes that are specific to human beings (stress, fatigue,
function better than humans.
reduced attention), it has a superior speed, it is efficient, it can
However, even if technology provides power and speed,
be repeated, it can quickly store huge amounts of data being
features that exceed the human ability to perform calculation,
able to make complex connections between them. These
the technology is far away from the human judgment. For this
systems can select or can generate important data, very useful
reason, humans are still the best standard for the intelligent
for the physicians.
systems [1].
This paper presents several mechanisms from the large
domain of artificial intelligence, emphasizing the benefits

978-1-4673-7545-0/15/$31.00 ©2015 IEEE

Authorized licensed use limited to: M/s Shanti Education Society. Downloaded on November 06,2023 at 14:50:07 UTC from IEEE Xplore. Restrictions apply.
 THE MECHANISMS
The artificial intelligence domain provides a large variety
of solutions for problems that require reasoning. This paper
presents two approaches suitable for medical predictions: a
knowledge-based system that involves logic and a system
based on probabilistic reasoning.
A. The Knowledge-Based Solution
This approach comes to implement the human process of
reasoning applied to a representation of knowledge. The most
important part of this type of decisional systems is the
knowledge base. It consists of a set of sentences that are
related, but not identical, to the sentences of a natural
language [3]. Here comes the first advantage of knowledge-
based systems: knowledge representation using a language
that is very close to the natural one makes them easy to Fig. 1. Structure of the knowledge-based system
understand and easy to create.
Therefore, this mechanism represents the things that an representation – Fig. 1). The nodes (vertices) of the graph
automated decisional system should know in order to contain knowledge pieces or logical operators. An arc (edge)
generate a solution to the problem it was developed for. The connects a pair of nodes and carries a truth value.
example presented here comes from medical diagnosis [6]. Unfortunately, a problem can be expected here: the graph
This area is very suitable for the use of automated prediction becomes too complex if the task that must be solved implies
systems; it needs a suggestion regarding a conclusion (this many interconnected premises (many data with complex
case, a diagnosis) starting from well-defined prior states. The connections). Another weak point is that sometimes it is
initial information constitutes the system’s premises. These difficult to express the logic that should be used by the
are connected through logical operators, creating the system’s system to draw its conclusions. It is also possible to encounter
rules. one more problem: the information content could be damaged
The aim of the knowledge-based system presented here is when the natural knowledge is transformed into rules [5].
to diagnose three different types of hepatitis: B, B+D and C. However, this type of decisional system still has the main
The premises follow the values of some laboratory tests that advantage that it is easy to be implemented for a knowledge
analyze three markers involved into these diseases: AgHBs, base with simple rules. The results it offers are very clear;
anti-VHD and anti-VHC. These markers can be positive or using logic, it decides whether the output is true or false. And
negative; therefore, they have a binary representation and the it has also a great chance to be accepted by those who are
premises have a truth value (true or false). It thus fulfills an involved in the medical field because its mathematical
important requirement of such systems: that premises must be support is closely related to the natural language, offering a
standardized [4]. high level of trust to the medical stuff who use it.
In order to create the rules for this system, the following A discussion around a mechanism from artificial
logical connectives have been used [3]: intelligence domain is not completed without presenting the
⎯ ¬ – not (negation) – unary operator; accuracy of the system described. This is a measure of the
⎯ ∧ – and (conjunction) – binary operator; results’ precision. The simple system, which was presented
⎯ ∨ – or (disjunction) – binary operator; here, always produces a correct result; it is obvious, for
⎯ Ÿ – implies (implication or conditional) – it needs instance, that if the marker AgHBs is positive, then the
some premises and produces a conclusion; it is hepatitis B virus is present. But if the graph becomes too
also known as rule or if-then statement. complex, then unexpected results can be encountered.
The following rules have been created for the decisional
B. The Probabilistic Reasoning
system:
When the information that should be represented by the
AgHBs ∧ ¬ anti-VHD Ÿ Hepatitis B
decisional system is too complex to use a knowledge-based
AgHBs ∧ anti-VHD Ÿ Hepatitis B+D
solution, the probabilistic approach can be a choice for the
¬ AgHBs ∧ anti-VHC ŸHepatitis C
problem. For example, once a patient is diagnosed with
These rules can be written in sentences closer to the natural
hepatitis B, his medical investigations continue. The problem
language this way:
that rises is connected to the evolution of the patient’s
if AgHBs and not anti-VHD then Hepatitis B
medical condition. There are three evolutionary types of this
if AgHBs and anti-VHD then Hepatitis B+D
disease (usual, with relapses and with de-compensations) and
if not AgHBs and anti-VHC then Hepatitis C
six severity levels (easy, medium, serious, prolonged,
In fact, the knowledge base has an oriented graph
cholestasis and comatose). It would be very useful to have
structure [4], and the system becomes clearer (having a visual

Authorized licensed use limited to: M/s Shanti Education Society. Downloaded on November 06,2023 at 14:50:07 UTC from IEEE Xplore. Restrictions apply.
some predictions regarding the evolution of the patient during
the treatment, but unfortunately the influence of premises
(symptoms and laboratory tests results) on the patient’s
evolution is difficult to be represented in clear rules as those
that are required by a knowledge-based system.
The solution could be provided by statistical methods.
Thus, in a database will be stored all the standardized
premises for as many patients as possible. Then, for a new
patient, the database is processed by statistical methods in
order to calculate the probability of each evolutionary type
and severity level of hepatitis B. Therefore, the solution
offered by this method is more accurate compared to the
previous one (based on logic), which establishes whether a
statement is true or false. The statistical approach associates a
probability to each possible output [6].
One of the most suitable methods used in decisional
systems that implement probabilistic reasoning is Bayes’
theorem. It is a formula with conditional probabilities and (1)
presents it [4].
p ( S | Dk ) ⋅ p( Dk )
p( Dk | S ) =
p( S )
This theorem describes the connection between two events
Dk and S, Dk being a diagnosis (among many possible
diseases) from available evidence S, which here is a vector
that stores the patient’s symptoms and laboratory test results
in a standardized form (binary values). Some of the patient’s
features have a binary form from the beginning: sex (M/F),
living conditions (town/village), had contact with other
infected persons (yes/no), made transfusions during the last
twelve months, made injected treatments during the last
months and a lot of symptoms or subjective signs (anorexia,
nausea, queasiness, asthenia, fever, hyperchrome urine,
tegument jaundice, arthralgia, myalgia, skin eruption). Other
features must be processed in order to decide whether they
belong to a particular range or not:
⎯ Age, a feature that was divided into four intervals
(< 20 years old / [20, 30) / [30, 45) / > 45).
Therefore, a patient that is 34 years old will have
the following binary values stored in the fields
that belong to this feature: 0, 0, 1, 0.
⎯ Starting time was split into three ranges (”7 days /
[8, 21] / • 22). If, for instance, the starting time
was 21 days, then the values connected to this
feature are: 0, 1, 0.
⎯ Jaundice, the color of the skin, was also divided
into three intervals ([0, 1] / 2 / [3, 6]) associated to
different intensities of color.
⎯ Liver is measured in centimeters and has two
possibilities ([0, 2) / •2).
⎯ Bilirubin is a laboratory test which was divided
into three ranges (<1 / [1, 10) / •10).
⎯ Tymol has two alternatives ([0, 4) / •4).
⎯ TGP has two intervals ([0, 400) / •400).
All these ranges are established by a physician, not by the
system’s developer, because they have a strong connection to
Authorized licensed use limited to: M/s Shanti Education Society. Downloaded on November 06,2023 at 14:50:07 UTC from IEEE Xplore. Restrictions apply.
the patient’s medical status and to his evolution. If other
features must be added to the database, a physician should
analyze them in order to decide which will be their
standardized form.
As (1) shows, Bayes’ theorem expresses a probability
based on three other probabilities. This choice could be
considered a disadvantage, but it is useful in practice, because
the three probabilities can be easily calculated, as (2) to (6)
demonstrate. This is because the probability in the causal
direction (p(S|Dk)) is more evident than the probability in the
diagnosis direction (p(Dk|S)) [3].
The probability p(Dk) is very simple to be calculated (and
(2) proves this). It is the frequency of a disease in the
statistical population Ω (in fact the number of records r that
have the diagnosis Dk equal to 1).
cardinal{r ∈ Ω | Dk ( x) = 1}
p ( Dk ) = (2)
cardinal Ω
The probability p(S|Dk) is calculated assuming the
conditional independence of all symptoms for a disease Dk.
This is an important restriction that Bayes’ theorem imposes
(1)
and sometimes it is difficult to fulfill it. In (3), which
calculates this probability, n is the number of symptoms
stored as binary values in the input vector S (34 for this case),
and σi is a symptom.
n n
p(σ i , Dk )
p( S | Dk ) = ∏ p(σ i | Dk ) = ∏ p ( Dk )
(3)
i =1 i =1
The probability p(σi,Dk) is called compound probability
and means that both events σi and Dk happen at the same time
(a patient with the symptom σi also has the disease Dk). This
probability, calculated by (4), is the frequency of σi and Dk in
the statistical population Ω (number of records r that have
these two parameters equal to 1).
cardinal{r ∈ Ω | Dk (r ) = 1,σ i (r ) = 1}
p(σ i , Dk ) = (4)
cardinal Ω
The last probability, p(S), is calculated considering that the
diseases are mutually exclusive (only one disease is present at
a moment) and that it is sure that one disease is present. This
restriction is not a problem for the example presented here,
because it is obvious that hepatitis B cannot evolve in two
different directions for the same patient at the same time, it
cannot have two different severity levels, and also it is sure
that the disease will have an evolutionary type and a severity
level.
Equation (1) is used again (in (5)) to calculate, for the same
vector of symptoms (S), the probability of not having the
particular disease Dk (¬ Dk).
p( S | ¬ Dk ) ⋅ p(¬ Dk )
p(¬ Dk | S ) = (5)
p(S )
It can be noticed that both (1) and (5) contain the term
1/p(S). It is considered a normalization constant [3] and it
ensures that the sum of probabilities is 1. This should be 1,
because it represents all possible cases, the certain event,
which has the probability 1. For the system presented, the
hepatitis B has three evolutionary types; the patient will
evolve to one of these and nothing else is possible. Therefore,
if the term 1/p(S) is ignored in (1) and (5), the obtained partial
probabilities are still in the correct relative proportion. Their
sum is not 1 anymore, but this can be solved dividing each
one by their sum [3]. Thus, this sum can replace p(S) (as (6)
shows).
p( S ) = p( S | Dk ) ⋅ p( Dk ) + p( S | ¬ Dk ) ⋅ p (¬ Dk ) (6)
Equation (7) is the final form of the Bayes’ theorem. It will
be applied for each of the three evolutionary types of hepatitis
B and for each of the six severity levels of this disease. The
results will be numbers in the range [0, 1] (because they are
probabilities). Their sum will be 1 for all the evolutionary
types and also 1 for all the severity levels of hepatitis B.
p( Dk | S ) =
n
p ( Dk ) ⋅ ∏ p(σ
i =1
i | Dk )
(7)
n n
p( Dk ) ⋅ ∏ p(σ
i =1
i | Dk ) + p(¬ Dk ) ⋅ ∏ p(σ
i =1
i | ¬ Dk )
The database that was investigated by this mechanism
contains 165 patients, 15 of them being reserved for the
system’s validation. The accuracy of this decisional system is
73.33%. A physician can decide, according to the medical
domain where the system is used, whether this accuracy is
acceptable or not. If, for instance, this type of decisional
system is used to predict the evolution of a particular
medicine store in a pharmacy, then the accuracy could be
enough. But if the system should predict a diagnosis for a
patient, then its performances must be improved. This can be
done through a couple of ways. First, the statistical
population Ω should be large enough in order to include
uncommon diseases [4]. Second, a better accuracy is obtained
if the number of analyzed inputs (symptoms and laboratory
test results) is increased. These two features of the database
are connected and a compromise is usually necessary. The
reason is that not all the patients need (or have) all the
laboratory tests; therefore, their records are not completed.
There are two choices here: to ignore those patients or to
ignore those missing laboratory tests for all the other patients.
The database cannot contain blank locations.
An important feature of this decisional system is that it
doesn’t provide a clear result, but a probability. Its output
quantifies uncertainty, offering a degree of belief [3]. For
instance, it is not sure that a patient will develop a serious
level of severity regarding hepatitis B, but it is more probably
because the system says 0, 0.2777, 0.7223, 0, 0, 0. Further, it
Authorized licensed use limited to: M/s Shanti Education Society. Downloaded on November 06,2023 at 14:50:07 UTC from IEEE Xplore. Restrictions apply.
is the physician’s responsibility to decide how will be treated
a patient with 72.23% risks to develop a serious form of
hepatitis B.
CONCLUSIONS
This paper described two important mechanisms of
artificial intelligence domain that can be used to develop
decisional systems for medical predictions. There were
emphasized general features of these mechanisms, advantages
and disadvantages of each one.
The benefits provided by these systems are obvious. On
one hand they can simplify the physicians’ work, saving their
time and energy (which otherwise would be wasted on too
many things they have to do). This way, the physicians can
focus on more important activities. On the other hand, an
automated system can detect imperceptible things (things that
are hardly noticed or that are the result of complex
computation and reasoning, things that are not evident or that
are the effect of too many factors involved).
It is difficult to decide where should be the limit between
artificial and human reasoning, but one thing is sure:
automated systems can be of much help with their
suggestions, even though the final decision belongs to the
humans.
ACKNOWLEDGMENT
This work was supported by a grant from the Partnerships
in priority areas – PN II program of the Romanian Ministry of
Education and Research (MEdC) – the Executive Agency for
Higher Education, Research, Development and Innovation
Funding (UEFISCDI), project number PN-II-PT-PCCA-
2013-4-0544.
REFERENCES
[1] J. Beal, P. H. Winston, “Guest Editors' Introduction: The New Frontier
of Human-Level Artificial Intelligence”, Intelligent Systems, IEEE
24.4: 21-23, 2009
[2] J. McCarthy, “What is Artificial Intelligence?”, Computer Science
Department, Stanford University, 2007
[3] S. Russell, P. Norvig, “Artificial Intelligence: A Modern Approach”,
third ed., Pearson Education, Inc., Publishing as Prentice Hall, United
States, 2010, ISBN 978-0-13-207148-2
[4] N. Szirbik, D. Pescaru, “Expert Systems”, Computer Science
Department, Politehnica University Timisoara, 2000
[5] J. M. Zurada, “Introduction to Artificial Neural Systems”, Jaico
Publishing House, 2012, ISBN 978-8172246501
[6] D. Dragulescu, A. Albu, “Medical Predictions System”, Acta
Polytechnica Hungarica, Vol. 4, No. 3, 89-101, 2007