A Rule-based Method Applied to the Imbalanced Classification of

Radiation Toxicity

Juan L. Domínguez-Olmedo

, Jacinto Mata

, Victoria Pachón

and Jose L. Lopez Guerra

Escuela Técnica Superior de Ingeniería, University of Huelva, Huelva, Spain

Department of Radiation Oncology, University Hospital Virgen del Rocío, Sevilla, Spain

Keywords: Imbalanced Data Classification, Rules Discovery, Prostate Cancer.

Abstract: This paper describes a rule-based classifier (DEQAR-C), which is set up by the combination of selected

rules after a two-phase process. In the first phase, the rules are generated and sorted for each class, and then

a selection is performed to obtain a final list of rules. A real imbalanced dataset regarding the toxicity during

and after radiation therapy for prostate cancer has been employed in a comparison with other predictive

methods (rule-based, artificial neural networks, trees, Bayesian and logistic regression). DEQAR-C

produced excellent results in an evaluation regarding several performance measures (accuracy, Matthews

correlation coefficient, sensitivity, specificity, precision, recall and F-measure) and by using cross-

validation. Therefore, it was employed to obtain a predictive model using the full data. The resultant model

is easily interpretable, combining three rules with two variables, and suggesting conditions that are mostly

confirmed by the medical literature.

1 INTRODUCTION

Prostate cancer (PC) is the most commonly

diagnosed cancer affecting men, and the third

leading cause of death in men in Europe (Ferlay et

al., 2013). The American Cancer Society estimated

that more than 200,000 men are diagnosed in the

United States with 30,000 deaths (American Cancer

Society, 2014). Although there is an improvement in

tumor control rates using radiation dose escalation,

PC radiotherapy is limited by the proximity of

surrounding normal tissues and because of the

observed dose-effect association with toxicity. It is

essential to understand the true complications

associated with doses delivered to normal anatomy,

to ensure the delivery of a sufficient dose with

minimal complications. The use of intensity-

modulated and image-guided radiation therapy can

decrease acute toxicity in PC patients (Valeriani et

al., 2013); (Morimoto et al., 2014).

Within the field of artificial intelligence and,

more specifically in machine learning, one of the

methods employed to extract knowledge from data is

the use of association rules. Association rule mining

is a technique whose purpose is to extract strong and

interesting relationships between patterns in a set of

data. An association rule takes the form A → C,

where A (the antecedent) and C (the consequent)

express a condition (or a conjunction of conditions)

on variables of the dataset (Agrawal et al., 1993);

(Rudin et al., 2013). The measures support and

confidence are used to assess the quality and

importance of the association rules. The support

measure evaluates the number of cases in which

both the antecedent and the consequent of the rule

hold. The confidence measure is the ratio between

the support of the rule and the number of cases in

which the antecedent holds. In order to filter the

usual huge number of rules generated, the values

minsup (minimum support) and minconf (minimum

confidence) are the thresholds that a rule has to

satisfy to be considered of interest.

Subgroup discovery is a type of descriptive

induction whose objective is to generate models

based on rules using a predictive perspective. It

emerged as the task of discovering properties of a

population by obtaining simple (but significant)

rules, using only one variable in the consequent: the

class or target variable (Wrobel, 1997); (Gamberger

et al., 2003); (Domínguez-Olmedo et al., 2015).

And also, numerous techniques have been

proposed for classification problems. In this kind of

Domínguez-Olmedo, J., Mata, J., Pachón, V. and Lopez-Guerra, J.

A Rule-based Method Applied to the Imbalanced Classiﬁcation of Radiation Toxicity.

DOI: 10.5220/0006586401470155

In Proceedings of the 11th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2018) - Volume 5: HEALTHINF, pages 147-155

ISBN: 978-989-758-281-3

147

task, a predictive model (classifier) tries to predict,

with some certainty, the objective variable (of

categorical type). Some examples of predictive

methods are artificial neural networks, decision

trees, logistic regression and Bayesian networks,

among others (Hastie et al., 2009); (Golhar et al.,

2017); (Cortés et al., 2015); (Liu et al., 1998). In

binary classification tasks with imbalanced data, as

in the case at hand, most algorithms are not usually

capable of obtaining good results for the minority

class and, therefore, the overall classification

performance does not reach adequate values (Sun et

al., 2009); (Rastgoo et al., 2016). The technique

proposed in this paper has achieved an improved

precision in both classes, thanks in part to an

alternating selection of rules for each class.

The rest of the paper is organized as follows.

Section 2 gives a description of the methods

employed in this work. The experimental setup is

presented in Section 3. Section 4 describes the

experimental results and discussion. And the last

section presents the conclusions.

2 METHODS EMPLOYED

2.1 Description of DEQAR-C

DEQAR-C is a rule-based classifier that works by

using a list of selected rules and a default class, both

of them obtained during the training process. Figure

1 illustrates this training process, which is composed

of two phases.

In the first phase, the "rule generation phase",

DEQAR-C generates rules from the training dataset.

This generation of rules is based on a method

developed to extract knowledge in the form of

association rules (Domínguez-Olmedo et al., 2011);

(Domínguez-Olmedo et al., 2012); (Domínguez-

Olmedo and Mata, 2016). It employs a deterministic

approach to generate rules without a previous

discretization of the numerical variables. Instead of

discretizing, what may result in suboptimal results

(Grosskreutz and Ruping, 2009), the process uses a

dynamic generation of conditions. DEQAR-C

obtains an ordered list of rules (called ranking) for

each possible value of the class variable, storing

separately the best rules in each class according to

their values of confidence and support.

In the second phase, the "rule selection phase", a

selection of the rules from all the rankings is done,

by starting in the ranking of the rule with the highest

confidence-support value and alternating iteratively

between these rankings to select rules from them.

The parameter maxrules determines the maximum

number of rules that will form the classifier. Figure

2 presents an example of this rule generation for a

dataset with two classes, and also shows a possible

selection of rules using a value of 3 for maxrules.

Figure 1: Training process in DEQAR-C.

The detail of the final selection process is shown

in Algorithm 1, which takes as input the set of cases

in the training dataset, the rankings of rules and the

parameter maxrules. After starting in the ranking

with the best rule (step 3), the process continues

selecting rules from the different rankings, but only

those rules covering some case not covered by a rule

previously selected (step 8). In the case of a binary

classification, the process would alternate in the

selection of rules for the two possible classes (if

there were still rules not processed in both rankings).

The procedure stops when all the rules have been

processed or the number of selected rules reaches

maxrules (step 17). At the end, the default class will

be the one having the greatest number of cases not

covered by any of the selected rules.

HEALTHINF 2018 - 11th International Conference on Health Informatics

148

Figure 2: Example of the generation and selection of rules.

Algorithm 1: Rules Selection in DEQAR-C.

Input: training dataset, rankings of rules, maxrules

Output: final list of rules, default class

1: T = set of cases in the training dataset

2: nRules = 0

3: r = ranking with the best rule according

to the highest confidence-support value

4: stop = FALSE

5: while NOT stop do

6: if ranking r has more rules to process

then

7: R = next rule in ranking r

8: if rule R covers at least one case

in T then

9: add R to the final list of rules

10: nRules = nRules + 1

11: T = T - cases covered by rule R

12: r = next ranking (alternate)

13: end if

14: else

15: r = next ranking (alternate)

16: end if

17: if all the rules were processed OR

nRules = maxrules then

18: stop = TRUE

19: end if

20: end while

21: return the list of selected rules and

the default class

To classify a new case, DEQAR-C will search

the list of rules for the first one where the case

matches its antecedent, assigning the class of that

rule. If no rule is found, the default class is assigned.

Algorithm 2 shows this process.

Algorithm 2: Classification process in DEQAR-C.

Input: list of rules, default class, a new case to classify

Output: predicted class

1: r = 1

2: matched = FALSE

3: while NOT matched AND r ≤ number of

rules do

4: if the case matches the antecedent of

rule r then

5: predictedClass = class of rule r

6: matched = TRUE

7: end if

8: r = r + 1

9: end while

10: if NOT matched then

11: predictedClass = default class

12: end if

13: return predictedClass

2.2 Classifiers Used in the Comparison

Several predictive methods have been employed in a

comparison with DEQAR-C. Methods based on

rules (ZeroR, PART), artificial neural networks

(MultilayerPerceptron), trees (J48, RandomForest),

Bayes (BayesNet, NaiveBayes) or logistic regression

(Logistic) have been used. Some of their

characteristics are shown below:

 ZeroR. It is a classification method that only

relies on the target variable (class), simply

predicting the majority class. It can be useful to

determine a baseline performance.

 PART. It generates a decision list by using a

separate-and-conquer strategy (Frank and

Witten, 1998).

 MultilayerPerceptron. A classifier that uses an

artificial neural network with backpropagation.

The nodes in this network are all sigmoid

(Rumelhart et al., 1986).

 J48. It uses a pruned or unpruned C4.5 decision

tree (Quinlan, 1993). A decision tree builds a

classification model in the form of a tree

structure.

 RandomForest. It constructs a forest of random

trees, an ensemble learning method for

classification, regression and other tasks

(Breiman, 2001).

 BayesNet. It employs a Bayes network, a

probabilistic graphical model that represents a

set of random variables and their conditional

dependencies (Pearl, 1985).

 NaiveBayes. It is based on Bayes theorem with

independence assumptions between predictors.

Despite its simplicity, it often outperforms more

sophisticated classification methods (John and

Langley, 1995).

A Rule-based Method Applied to the Imbalanced Classiﬁcation of Radiation Toxicity

149

Table 1: Variables and units of the dataset.

Variable

Units/Values

Age

years

Indication treatment

Post-prostatectomy , Primary Prostate Cancer, Recurrence

Radiation technique

Tomotherapy, RapidArc

Gleason score

2..10

T stage

T1 , T1b , T1c , T2 , T2a , T2b , T2c , T3 , T3a , T3b , T4

Diagnosis PSA

ng/mL

Risk

Low, Intermediate, High

ADT

No ADT, Short Term, Long Term

Radiation time

days

Planning tumor volume

Prostate radiation dose

Fractionation

Pelvic treatment

Yes, No

Bladder volume

Bladder mean dose

Bladder median dose

GU acute toxicity

+, -

PSA: prostate specific antigen

ADT: androgen deprivation therapy

 Logistic. It builds a multinomial logistic

regression model with a ridge estimator (Le-

Cessie and van Houwelingen, 1992).

3 EXPERIMENTAL SETUP

3.1 Dataset Description

In this work, a dataset about the toxicity effects

during and after treatment of PC (Lopez et al., 2015)

has been used. This dataset includes the clinical (i.e.

age), pathological (i.e. Gleason score, T score), and

therapeutic (i.e. radiation dose, fractionation, whole

pelvic lymph node irradiation, radiation technique)

information as well as the out-come (acute

genitourinary [GU] toxicity) of 162 PC patients

treated with arc radiation therapy from June 2006

through May 2012 at two institutions from different

nationalities (Europe and Latin-America).

The names of the 17 selected variables in the

dataset are shown in Table 1. The numerical

variables are 10 and the class variable is binary ('+'

for a toxicity grade ≥ 2, '-' for a toxicity grade < 2),

with a distribution for class '+' of 23.5% of the cases.

Therefore, it is an imbalanced dataset with a 3.3:1

ratio of negative/positive cases.

Ethical Considerations. All identifiable

information about the patients was adequately

removed from the da-ta to preserve anonymity.

3.2 Evaluation Criteria

In a binary classification problem, such as the one

we are presenting, we can denote with TP (True

Positive) the number of positive cases correctly

classified, with TN (True Negative) the number of

negative cases correctly classified, with FN (False

Negative) the number of positive cases incorrectly

classified, and with FP (False Positive) the number

of negative cases incorrectly classified.

The following evaluation measures were

employed in the comparison: accuracy, Matthews

correlation co-efficient, the average value of

sensitivity and specificity, precision, recall and F-

measure. A description of these measures is

presented below:

 Accuracy: the proportion of true results (both

true positives and true negatives) among the

total number of cases examined.

Accuracy =

FP+FN+TN+TP

TN+TP

(1)

 MCC: Matthews correlation coefficient, which

measures the quality of binary classifications

(Matthews, 1975).

MCC =

FN)+(TNFP)+(TNFN)+(TPFP)+(TP

FNFP-TN TP





(2)

HEALTHINF 2018 - 11th International Conference on Health Informatics

150

 Average value of sensitivity and specificity:

sensitivity is the proportion of positives cases

that are correctly identified as such, and

specificity is the proportion of negatives cases

that are correctly identified as such.

avg (Se, Sp) =

5.0















FP+TN

FN+TP

(3)

 Precision: analogous to positive predictive value

(PPV).

precision =

FP+TP

(4)

 Recall: analogous to sensitivity.

recall =

FN+TP

(5)

 F-measure: the harmonic mean of precision and

recall.

F-measure =

2





ecallrprecision

ecallr precision

(6)

4 EXPERIMENTAL RESULTS

4.1 Results of the Comparison

The classifiers previously mentioned were evaluated

in the task about the prediction of toxicity effects in

the radiotherapy treatment of PC. The evaluation

measures were calculated by using stratified 10-fold

cross-validation. Cross-validation reduces the

variance of the estimates and improves the

estimation of the generalization performance. In k-

fold cross-validation, the original data is partitioned

into k equal size subsets. Then, a single subset is

retained as the validation data and the remaining k-1

subsets are used as training data. The process is

repeated k times, with each of the k subsets used

exactly once as the validation data (Arlot and

Celisse, 2010). At the end, the final validation result

is calculated from all the partial results.

The machine learning software Weka (Frank et

al., 2016) was used to run the classifiers ZeroR,

PART, MultilayerPerceptron, J48, RandomForest,

BayesNet, NaiveBayes and Logistic. For a fair

comparison, the final values of the parameters used

in all the classifiers were the ones that yielded the

best results after testing several combinations of

values (grid search).

The results for accuracy, MCC and average

value of sensitivity and specificity are displayed in

Table 2 and Figure 3. As can be seen, DEQAR-C

obtained excellent results, which seems to support

the proposed selection of high confidence-support

rules for each class, not only to obtain high values of

general accuracy but also to get a satisfactory

prediction for both classes. The imbalance in the

dataset (38 positive cases and 124 negative cases)

adds more difficulty to the classification task. The

results of DEQAR-C were the best regarding these

three evaluation measures. The classifiers

NaiveBayes and MultilayerPerceptron also obtained

good results, but the difference for MCC, in

comparison with DEQAR-C, is important. Matthews

correlation coefficient is generally regarded as being

one of the best measures to describe the confusion

matrix of true and false positives and negatives by a

single number, especially suitable to the case of

imbalanced data learning (Powers, 2011).

Table 2: Results for accuracy, MCC and average(Se, Sp).

Classifier

accuracy

MCC

avg(Se,Sp)

ZeroR

0.765

0.000

0.500

PART

0.710

0.185

0.592

MultilayerPerceptron

0.710

0.199

0.601

J48

0.698

0.118

0.556

RandomForest

0.765

0.161

0.546

BayesNet

0.710

0.185

0.592

NaiveBayes

0.698

0.210

0.611

Logistic

0.716

0.132

0.559

DEQAR-C

0.772

0.358

0.677

The results for the measures associated with a

particular class (precision, recall and F-measure) are

shown in Tables 3 and 4. DEQAR-C did not obtain

the best F-measure result for negative toxicity (the

majority class); but its result was close to the best,

and obtained the best precision. Regarding the

positive toxicity, DEQAR-C obtained the best

precision, recall and F-measure; the classifier

NaiveBayes was the second best. As can be seen, the

F-measure for this minority class was not very high

in all the classifiers, and only DEQAR-C surpassed

the value 0.5.

Table 3: Results for precision, recall and F-measure

(toxicity '+').

Classifier

Precision

Recall

F-Measure

ZeroR

0.000

PART

0.378

0.368

0.373

MultilayerPerceptron

0.385

0.395

0.390

J48

0.333

0.289

0.310

RandomForest

0.500

0.132

0.208

BayesNet

0.378

0.368

0.373

NaiveBayes

0.378

0.447

0.410

Logistic

0.357

0.263

0.303

DEQAR-C

0.514

0.500

0.507

A Rule-based Method Applied to the Imbalanced Classiﬁcation of Radiation Toxicity

151

Figure 3: Results for accuracy, MCC and average(Se, Sp).

Table 3: Results for precision, recall and F-measure

(toxicity '+').

Classifier

Precision

Recall

F-Measure

ZeroR

0.000

PART

0.378

0.368

0.373

MultilayerPerceptron

0.385

0.395

0.390

J48

0.333

0.289

0.310

RandomForest

0.500

0.132

0.208

BayesNet

0.378

0.368

0.373

NaiveBayes

0.378

0.447

0.410

Logistic

0.357

0.263

0.303

DEQAR-C

0.514

0.500

0.507

Table 4: Results for precision, recall and F-measure

(toxicity '-').

Classifier

Precision

Recall

F-Measure

ZeroR

0.765

1.000

0.867

PART

0.808

0.815

0.811

MultilayerPerceptron

0.813

0.806

0.810

J48

0.791

0.823

0.806

RandomForest

0.783

0.960

0.862

BayesNet

0.808

0.815

0.811

NaiveBayes

0.821

0.774

0.797

Logistic

0.791

0.855

0.822

DEQAR-C

0.848

0.855

0.851

The results for the measures associated with a

particular class (precision, recall and F-measure) are

shown in Tables 3 and 4. DEQAR-C did not obtain

the best F-measure result for negative toxicity (the

majority class); but its result was close to the best,

and obtained the best precision. Regarding the

positive toxicity, DEQAR-C obtained the best

precision, recall and F-measure; the classifier

NaiveBayes was the second best. As can be seen, the

F-measure for this minority class was not very high

in all the classifiers, and only DEQAR-C surpassed

the value 0.5.

4.2 Prediction Model

After testing and comparing the described classifiers

by stratified cross-validation, the full dataset was

used to obtain a prediction model for the GU

toxicity. DEQAR-C was executed with the same

parameters that achieved the best results in cross-

validation (minsens = 0.7, delta = 0.05, maxAttr = 2,

maxrules = 3). The parameters minsens, maxAttr and

delta are used in the rules generation phase,

controlling the search for rules and the conditions

for the numerical variables (Domínguez et al.,

2015). After this execution, three rules were selected

(see Table 5) and the default class was set to '-'.

As can be seen, the rules are simple with two

variables, because of the constraint due to the

parameter maxAttr (maximum number of variables

in the antecedent).

The combination of these three rules achieves a

covering of 76% of the cases, and from the 39

remaining cases, 28 are negative ones.

The simplicity of the obtained classifier also

makes it more interpretable. It can be easily

analyzed to discover the conditions most likely to be

of influence in the toxicity effects, in contrast with

the greater complexity of other models such as

artificial neural networks or Random Forest. As an

example, Figure 4 shows some of the 24 rules

HEALTHINF 2018 - 11th International Conference on Health Informatics

152

Table 5: Rules obtained in the final classifier.

Antecedent

GU acute toxicity

Technique = Tomotherapy AND Planning tumor volume ≤ 218.62

Prostate radiation dose ≥ 70.02 AND Bladder volume ≥ 63.67

Technique = Tomotherapy AND Fractionation ≤ 2.52

obtained after executing the classifier PART (some

of them with five conditions).

Figure 4: Rules obtained by PART classifier.

4.3 Application to Clinical Practice

As can be seen by analyzing the rules in Table 5,

five of the sixteen independent variables in the

dataset are employed in the model: Technique,

Planning tumor volume, Prostate radiation dose,

Bladder volume and Fractionation. They can be

considered of great relevance in this model for GU

toxicity, because the predictions mainly depend on

their values. Also, some relative importance could

be established between them, because there exists an

ordering in the rules and the search for a match

follows this order.

These variables and their associated values could

be seen as risk factors for GU toxicity. These risk

factors are mostly confirmed by the literature

(Acevedo-Henao et al., 2014); (Ahmed et al., 2013);

(Aizer et al., 2011); (Lopez et al., 2013), which may

corroborate the value of the method employed.

Better stratification of patients based on their

own expected tumor and normal tissue factors will

enable therapy to be highly tailored. Prostate cancer

patients with low-risk toxicity (e.g., men treated

with Tomotherapy and having a lower prostate

volume) might be able to receive a more intense

treatment. Additionally, we can better define the

individual patient subgroups that benefit from

specific components of radiation therapy.

5 CONCLUSIONS

In this work we have presented the application of

several predictive methods to data regarding the

toxicity of radiation therapy for prostate cancer. This

dataset exhibits some imbalance in the classes, with

a 3.3:1 ratio of negative/positive cases.

A rule-based classifier (DEQAR-C) was

described, which works without discretizing the

numerical variables and by selecting a subset of the

best rules extracted for each class. This method was

compared to other classifiers by using cross-

validation with several evaluation measures.

DEQAR-C produced outstanding results in this

classification task, with higher prediction

performance in both classes than the rest of

classifiers. Therefore, it was employed to obtain a

predictive model using the full data. The simplicity

of the model (three rules with two variables) also

makes it more interpretable, which may be useful in

obtaining knowledge from medical data and

subsequently applying it into the clinical practice.

As future work, it would be interesting to test the

proposed approach in another real classification

problem or simultaneously with more datasets.

ACKNOWLEDGMENTS

The authors would like to thank Dr. Ignacio

Azinovic for his support with the employed data.

A Rule-based Method Applied to the Imbalanced Classiﬁcation of Radiation Toxicity

153

The research presented in this paper was partially

funded by the Regional Government of Andalusia

(Junta de Andalucía) under grant number TIC-7629

and Spanish Ministry of Education and Science

(Grant Number: TIN2009-14057-C03-03).

REFERENCES

Ferlay, J., Steliarova-Foucher, E., Lortet-Tieulent, J.,

Rosso, S., Coebergh, J.W.W., Comber, H., Forman, D.

and Bray, F., 2013. Cancer incidence and mortality

patterns in Europe: estimates for 40 countries in 2012.

European Journal of Cancer 49(6), 1374-1403.

American Cancer Society, September 2014.

http://www.cancer.org/cancer/

prostatecancer/detailedguide/prostate-cancer-key-

statistics.

Valeriani, M., Carnevale, A., Osti, M.F., de Sanctis, V.,

Agolli, L. and Enrici, R.M., 2013. Hypofractionated

intensity-modulated simultaneous integrated boost and

image-guided radiotherapy in the treatment of high-

risk prostate cancer patients: a preliminary report on

acute toxicity. Tumori 99(4), 474-479.

Morimoto, M., Yoshioka, Y., Konishi, K., Isohashi, F.,

Takahashi, Y., Ogata, T., Koizumi, M., Teshima, T.,

Bijl, H.P., van der Schaaf, A., Langendijk, J.A. and

Ogawa, K., 2014. Comparison of acute and subacute

genitourinary and gastrointestinal adverse events of

radiotherapy for prostate cancer using intensity-

modulated radiation therapy, three-dimensional

conformal radiation therapy, permanent implant

brachytherapy and high-dose-rate brachytherapy.

Tumori 100(3), 265-271.

Agrawal, R., Imielinski, T. and Swami, A., 1993. Mining

Association Rules between Sets of Items in Large

Databases. In Proceedings of ACM SIGMOD ICMD,

207-216.

Rudin, C., Letham, B. and Madigan, D., 2013. Learning

theory analysis for association rules and sequential

event prediction. Journal of Machine Learning

Research, 14:3441-3492.

Wrobel, S., 1997. An algorithm for multi-relational

discovery of subgroups. Principles of data mining and

knowledge discovery, 78-87.

Gamberger, D., Lavrac, N. and Krstacic, G., 2003. Active

Subgroup Mining: A Case Study in Coronary Heart

Disease Risk Group Detection. Artificial Intelligence

in Medicine 28, 27–57.

Domínguez-Olmedo, J. L., Mata, J. and Pachón, V., 2015.

Deterministic Extraction of Compact Sets of Rules for

Subgroup Discovery. In Proceedings of Intelligent

Data Engineering and Automated Learning – IDEAL,

138-145.

Hastie, T., Tibshirani, R. and Friedman, J., 2009. The

Elements of Statistical Learning: Data Mining,

Inference, and Prediction. Springer-Verlag.

Golhar, M., Iwahori, Y., K. Bhuyan, M., Funahashi, K.

and Kasugai, K., 2017. A Robust Method for Blood

Vessel Extraction in Endoscopic Images with SVM-

based Scene Classification. In Proceedings of the 6th

International Conference on Pattern Recognition

Applications and Methods: ICPRAM, 148-156.

Cortés, X., Serratosa, F. and Moreno-García, C., 2015. An

Interactive Model for Structural Pattern Recognition

based on the Bayes Classifier. In Proceedings of the

International Conference on Pattern Recognition

Applications and Methods: ICPRAM, 240-247.

Liu, B., Hsu, W. and Ma, Y., 1998. Integrating

classification and association rule mining. In

Proceedings of Fourth International Conference on

Knowledge Discovery and Data Mining.

Sun, Y., Wong, A.K.C., Kamel, M.S., 2009. Classification

of imbalanced data: a review. International Journal of

Pattern Recognition and Artificial Intelligence 23(4),

687–719.

Rastgoo, M., Lemaitre, G., Massich, J., Morel, O.,

Marzani, F., Garcia, R. and Meriaudeau, F.,

2016. Tackling the Problem of Data Imbalancing for

Melanoma Classification . In Proceedings of the 9th

International Joint Conference on Biomedical

Engineering Systems and Technologies:

BIOIMAGING (BIOSTEC).

Domínguez-Olmedo, J. L., Mata, J., Pachón, V. and Maña,

M., 2011. A deterministic approach to association rule

mining without attribute discretization. In Proceedings

of the International Conference on Digital Information

Processing and Communications, 188, 140–150.

Domínguez-Olmedo, J. L., Mata, J., Pachón, V. and Maña,

M., 2012. Rule extraction from medical data without

discretization of numerical attributes. In Proceedings

of the International Conference on Health Informatics

(HEALTHINF), 397-400.

Domínguez-Olmedo, J.L. and Mata, J., 2016. Comparison

of Standard Discretization with a New Method for

Quantitative Association Rules. IEEE Latin America

Transactions 14(4), 1879-1885.

Grosskreutz, H. and Ruping, S., 2009. On Subgroup

Discovery in Numerical Domains. Data Mining and

Knowledge Discovery 19(2), 210–226.

Frank, E. and Witten, I. H., 1998. Generating Accurate

Rule Sets without Global Optimization. In

Proceedings of the Fifteenth International Conference

on Machine Learning, 144-151.

Rumelhart, D. E., Hinton, G. E. and Williams, R. J., 1986.

Parallel distributed processing: explorations in the

microstructure of cognition. MIT Press Cambridge,

318-362.

Quinlan, R., 1993. C4.5: Programs for Machine Learning.

Morgan Kaufmann Publishers, San Mateo.

Breiman, L., 2001. Random Forests. Machine Learning

45(1), 5-32.

Pearl, J., 1985. Bayesian Networks: A Model of Self-

Activated Memory for Evidential Reasoning. UCLA

Computer Science Department.

John, G. H. and Langley, P., 1995. Estimating Continuous

Distributions in Bayesian Classifiers. In Proceedings

of the Eleventh Conference on Uncertainty in Artificial

Intelligence, 338-345.

HEALTHINF 2018 - 11th International Conference on Health Informatics

154

Le-Cessie, S. and van Houwelingen, J. C., 1992. Ridge

Estimators in Logistic Regression. Applied Statistics

41(1), 191-201.

Lopez Guerra, J. L., Matute R, Puebla F, Sánchez-Reyes

A, Pontes B, Rubio C, Nepomuceno I, Acevedo C, Isa

N, Lengua R, Praena-Fernandez JM, del Campo ER,

Ortiz M. J and Azinovic I., 2015. Ethnic difference in

risk of toxicity in prostate cancer patients treated with

dynamic arc radiation therapy. Tumori 101(4), 461-8.

Matthews, B. W., 1975. Comparison of the predicted and

observed secondary structure of T4 phage lysozyme.

Biochimica et Biophysica Acta (BBA)-Protein

Structure 405(2), 442–451.

Arlot, S. and Celisse, A., 2010. A survey of cross-

validation procedures for model selection. Statistics

Surveys, 40-79.

Frank, E., Hall, M. A. and Witten, I. H., 2016. Data

Mining: Practical Machine Learning Tools and

Techniques. Morgan Kaufmann.

Powers, D. M. W., 2011. Evaluation: from precision,

recall and F-measure to ROC, informedness,

markedness and correlation. International Journal of

Machine Learning Technology 2(1), 37-63.

Acevedo-Henao, C. M., Lopez, J. L., Matute, R. and

Azinovic, I., 2014. Image-guided radiation therapy

based on helical tomotherapy in prostate cancer:

minimizing toxicity. Oncology Research and

Treatment 37(6), 324-30.

Ahmed, A. A., Egleston, B., Alcantara, P., Li, L., Pollack,

A., Horwitz, E.M. and Buyyounouski, M. K., 2013.

A novel method for predicting late genitourinary

toxicity after prostate radiation therapy and the need

for age-based risk-adapted dose constraints.

International Journal of Radiation Oncology, Biology,

Physics, 86(4), 709-15.

Aizer, A. A., Anderson, N. S., Oh, S. C., Yu, J. B.,

McKeon, A. M., Decker, R. H. and Peschel, R. E.,

2011. The impact of pretreatment prostate volume on

severe acute genitourinary toxicity in prostate cancer

patients treated with intensity-modulated radiation

therapy. International Journal of Radiation Oncology,

Biology, Physics, 79(2), 379-84.

Lopez, J. L., Isa, N., Matute, R., Russo, M., Puebla, F. and

Miran, K. M., 2013. Hypofractionated helical

tomotherapy using 2.5-2.6 Gy daily fractions for

localized prostate cancer. Clinical and Translational

Oncology 15(4), 271-7.

A Rule-based Method Applied to the Imbalanced Classiﬁcation of Radiation Toxicity

155