Predicting Breast Cancer Survivability

Using Data Mining Techniques

Abdelghani Bellaachia, Erhan Guven

Department of Computer Science

The George Washington University
Washington DC 20052
{bell, eguven}@gwu.edu

Abstract relationship of the association. Data driven statistical

research is becoming a common complement to many
In this paper we present an analysis of the prediction scientific areas like medicine and biotechnology. This
of survivability rate of breast cancer patients using trend is becoming more and more visible as in the
data mining techniques. The data used is the SEER studies of Houston et al. [5] and Cios et al. [6].
Public-Use Data. The preprocessed data set consists In this paper, we present data mining techniques
of 151,886 records, which have all the available 16 to predict the survivability rate of breast cancer
fields from the SEER database. We have investigated patients. In our study, we have used the SEER data
three data mining techniques: the Nave Bayes, the and have introduced a pre-classification approach
back-propagated neural network, and the C4.5 that take into account three variables: Survival Time
decision tree algorithms. Several experiments were Recode (STR), Vital Status Recode (VSR), and
conducted using these algorithms. The achieved Cause of Death (COD).
prediction performances are comparable to existing This paper is organized as follows. The next
techniques. However, we found out that C4.5 section reviews related work. Section 3 gives the
algorithm has a much better performance than the methodology used to conduct the prediction analysis.
other two techniques. Experimental results are presented in Section 4.
Conclusion and future work are given in the last
Keywords: Breast cancer survivability, data mining, section.
SEER, Weka.
2. Related Work
1. Introduction
A literature survey showed that there have been
Today, in the United States, approximately one in several studies on the survivability prediction
eight women over their lifetime has a risk of problem using statistical approaches and artificial
developing breast cancer. An analysis of the most neural networks. However, we could only find a few
recent data has shown that the survival rate is 88% studies related to medical diagnosis and survivability
after 5 years of diagnosis and 80% after 10 years of using data mining approaches like decision trees [7,
diagnosis [1]. 8, 9].
The discovery of the survival rate or In this work, we took the study of Delen et al.
survivability of a certain disease is possible by [9] as the starting point of our research. In his study,
extracting the knowledge from the data related to that Delen et al. preprocessed the SEER data (period of
disease. One of these data sources is SEER [2] 1973-2000 with 433,272 records named as breast.txt)
(Surveillance Epidemiology and End Results), which for breast cancer to remove redundancies and missing
is a unique, reliable and essential resource for information. The resulting data set had 202,932
investigating the different aspects of cancer. The records, which then pre-classified into two groups of
SEER database combines patient-level information survived (93,273) and not survived (109,659)
on cancer site, tumor pathology, stage, and cause of depending on the Survival Time Recode (STR) field.
death [3, 4]. The survived class is all records that have a value
The characteristics of a population can be greater than or equal 60 months in the STR field and
observed to establish the factors associated with a the not survived class represent the remaining
specific outcome. Observational studies, such as records. After this step, the data mining algorithms
statistical learning and data mining, can establish the are applied on these data sets to predict the dependent
association of the variables to the outcome, but they field from 16 predictor fields. The results of
do not always establish the cause-and-effect predicting the survivability were in the range of 93%

1
accuracy. After a careful analysis of the breast cancer We have used the Weka toolkit to experiment
data used in [9], we have noticed that the number of with these three data mining algorithms [12]. The
not survived patients used does not match the Weka is an ensemble of tools for data classification,
number of not alive (field VSR) patients in the first regression, clustering, association rules, and
60 months of survival time. As a matter of fact, the visualization. The toolkit is developed in Java and is
number of not survived patients is expected to be an open source software issued under the GNU
around 20% based on the breast cancer survival General Public License [10].
statistics of 80% [1]. In our discussion with the Preprocessing the input data set for a knowledge
authors of [9], we found out that the pre-classification discovery goal using a data mining approach usually
process was not accurate in determining the records consumes the biggest portion of the effort devoted in
of the not survived class. They did not take into the entire work [10]. We have developed a set of
consideration neither the Vital Status Recode (VSR), tools to extract and cleanup the raw SEER data.
nor the Cause of Death (COD). They assume that all A simple analysis shows that the SEER data has
patients are dead with cancer, which is not always missing information in the fields of Extent of Disease
true. (EOD) and Site Specific Surgery (SSS) fields for
In our study, we have used a newer version of almost half of the records. Most of the missing
SEER database (period of 1973-2002 with 482,052 information is in the records, which are gathered
records) and, unlike [9], we have included two other prior to 1988. Since we wanted to use all the
fields in the pre-classification process: available fields in the SEER database, we removed
Survival Time Recode (STR), these records from the test data set. These records
Vital Status Recode (VSR), have Coding System for EOD coded as 4. The SSS
Cause of Death (COD) field usage has changed after 1998. Instead of the
The next section presents our pre-classification regular field, the information is split in five other
process. fields. A mapping scheme from new SSS to old SSS
is developed to fill the missing SSS fields. After this
3. Methodology step, the records with missing information are
removed from the data set.
In this paper, we have investigated three data mining The EOD field is composed of five fields
techniques: the Nave Bayes, the back-propagated including the EOD code. These fields (size of tumor,
neural network, and the C4.5 decision tree number of positive nodes, number of nodes, and
algorithms. In this paper, we used these algorithms to number of primaries) contain missing information
predict the survivability rate of SEER breast cancer coded such as 999, 99 or 9 representing the
data set. We selected these three classification unknown information. Please note that, the statistics
techniques to find the most suitable one for in Table 1 do not contain fields with unknown
predicting cancer survivability rate. values. The table also shows the fields used in our
The Nave Bayes technique depends on the analysis.
famous Bayesian approach following a simple, clear
and fast classifier [10]. It has been called Nave due Nominal variable name Number of
distinct values
to the fact that it assumes mutually independent
attributes. In practice, this is almost never true but is Race 19
achievable by preprocessing the data to remove the Marital status 6
Primary site code 9
dependent categories [10]. This method has been
Histologic type 48
used in many areas to represent, utilize, and learn the Behavior code 2
probabilistic knowledge and significant results have Grade 5
been achieved in machine learning [10]. Extension of tumor 23
Lymph node involvement 10
The second technique uses artificial neural
Site specific surgery code 19
networks. In this study, a multi-layer network with Radiation 9
back-propagation (also known as a multi-layer Stage of cancer 5
perceptron) [10] is used.
The third technique is the C4.5 decision-tree Numeric variable name Mean Std. Dev. Range
generating algorithm [11]. C4.5 is based on the ID3 Age 58 13 10-110
algorithm. Tumor size 20 16 0-200
No of positive nodes 1.5 3.7 0-50
It has been shown that the last two techniques
Number of nodes 15 6.8 0-95
have better performance [7, 8, 9]. Therefore we have Number of primaries 1.25 0.5 1-8
included them in our analysis. Table 1: Survivability Attributes

2
 As stated in the previous section, we have Figure 1 shows the ranked survivability
adopted a different approach in the pre-classification attributes of our data as calculated by the Weka
process. Unlike [9], we have included three fields: toolkit. It clearly shows that Extension of Tumor has
STR, VSR, and COD. The STR field ranges from 0 a higher rank than the Tumor Size.
to 180 months in the SEER database. The pre- 0.25

classification process is outlined as follows. Ranked Survivability Attributes

0.2

// Setting the survivability dependent variable for 60

0.15
// months threshold
if STR 60 months and VSR is alive then 0.1
the record is pre-classified as survived
else if STR < 60 months and COD is breast cancer, then 0.05
the record is pre-classified as not survived
else 0
Ignore the record

Extention of tumor (EOD)

Tumor size (EOD)

Number of primaries
Primary site
Lymph node involv (EOD)

Number of nodes (EOD)

Grade
Site Specific Surgery

Radiation
Age

Behavior code
No of pos nodes (EOD)
Stage of cancer

Marital status
Hostologic type

Race
end if

In the above approach, the ignored records

correspond to those patients that have an STR less
than 60 months and are still alive, or those patients
Figure 1: Ranked Survivability Attributes
that have an STR less than 60 months but the cause
of their death is not breast cancer.
We will use the performance metrics of
Table 2 and Table 3 show the classes of our pre-
accuracy, precision and recall to compare the three
classification process and the approach used in [9],
techniques. In order to have a fair measure of the
respectively.
performance of the classifier; we used a cross-
Class No of instances Percentage
validation with 10 folds. In its most elementary form,
cross-validation consists of dividing the data into k
0: not survived 35,148 23.2 subgroups. Each subgroup is predicted via the
1: survived 116,738 76.8
Total 151,886 100
classification rule constructed from the remaining
(k-1) subgroups, and the estimated error rate is the
Table 2: Proposed Survivability Class Instances
average error rate from these k subgroups. In this
Class No of instances Percentage
way, the error rate is estimated in an unbiased way.
The final classifier rule is calculated from the entire
0: not survived 162,381 58.3 data set. After running the classifier 10 times with 10
1: survived 116,282 41.7
Total 278,663 100
folds, we obtain the metrics of precision, recall,
accuracy Ai and the Cross Validation Accuracy
Table 3: Survivability Class Instances based on the
Previous Work (study [9])
(CVA) to represent a classifier performance:
CVA = (1/10)Ai i = 1, 2, , 10
After the preprocessing step, a common analysis
would be determining the effect of the attributes on Ai = # records correctly classified / total # records
the prediction, or attribute selection. We used the
information gain measure [10] to rank the attributes The Weka toolkit can calculate all these
due to the fact that it is a common method and the performance metrics after running a specified k-fold
C4.5 decision tree technique utilizes this measure. cross-validation.
Information gain (IG) is measured as the amount of
the entropy (H) difference when an attribute 4. Experimental Results
contributes the additional information about the class.
The following is the information gain and the entropy In this study, the accuracy of three data mining
before and after observing the attribute Xi for the techniques is compared. The goal is to have high
class C: accuracy, besides high precision and recall metrics.
Although these metrics are used more often in the
field of information retrieval, here we have
H(C) = p(c)logp(c) , cC
considered them as they are related to the other
H(C|Xi) = p(x) p(c|x)logp(c|x) , xXi, cC existing metrics such as specificity and sensitivity.
IGi = H(C) H(C|Xi) These metrics can be derived from the confusion

3
matrix and can be easily converted to true-positive data in the EOD field from the old EOD fields prior
(TP) and false-positive (FP) metrics [10]. to 1988. This might increase the performance as the
The experimental results of our approach as size of the data set will increase considerably.
presented in Table 4. Finally, we would like to try survival time
prediction of certain cancer data such as respiratory
Classification Accuracy Class Precision Recall cancer where the survivability is seriously low. We
Technique (%)
think of discretizing the survival time in terms of one
Nave Bayes 84.5 0 0.70 0.57 year and then classifying using the aforementioned
1 0.88 0.93 data mining algorithms.
Artificial 86.5 0 0.83 0.52
Neural Net 1 0.87 0.97
0 0.80 0.56 References
C4.5 86.7
1 0.88 0.96
[1] American Cancer Society. Breast Cancer Facts
Table 4: Combined Results (our study) & Figures 2005-2006. Atlanta: American
Cancer Society, Inc. (http://www.cancer.org/).
Classification Accuracy Class Precision Recall
Technique (%) [2] Surveillance, Epidemiology, and End Results
(SEER) Program (www.seer.cancer.gov)
C4.5 81.3 0 0.86 0.81 Public-Use Data (1973-2002), National Cancer
1 0.76 0.81
Institute, DCCPS, Surveillance Research
Table 5: Results for C4.5 (dataset as in Table 3) Program, Cancer Statistics Branch, released
April 2005, based on the November 2004
As can be seen in Table 4, neural net and submission.
decision tree have comparable performances. [3] Cox DR. Analysis of survival data. London:
Table 5 shows the experimental results using Chapman & Hall; 1984.
the pre-classification approach used in [9] and the [4] Benjamin F. Hankey, et. al. The Surveillance,
same dataset used in our approach. The results clearly Epidemiology, and End Results Program: A
show that the classification rate (81%) is much lower National Resource. Cancer Epidemiology
than the classification rate of our approach (~87%). Biomarkers & Prevention 1999; 8:1117-1121.
It may be worth noting that the computation [5] Houston, Andrea L. and Chen, et. al.. Medical
times of the algorithms Nave Bayes, neural net and Data Mining on the Internet: Research on a
C4.5 (on an AMD Athlon 64 4000+ machine) were in Cancer Information System. Artificial
the ranges of 1 minute, 12 hours and 1 hour, Intelligence Review 1999; 13:437-466.
respectively. [6] Cios KJ, Moore GW. Uniqueness of medical
These obtained results in this work differ from data mining. Artificial Intelligence in Medicine
the study of Delen et al. [9] due to the facts that we 2002; 26:1-24.
used a newer database (2000 vs. 2002), a different [7] Zhou ZH, Jiang Y. Medical diagnosis with C4.5
pre-classification (109,659 and 93,273 vs. 35,148 and Rule preceded by artificial neural network
116,738) and different toolkits (industrial grade tools ensemble. IEEE Trans Inf Technol Biomed.
vs. Weka). 2003 Mar; 7(1):37-42.
[8] Lundin M, Lundin J, Burke HB, Toikkanen S,
5. Conclusions and Future Work Pylkkanen L, Joensuu H. Artificial neural
networks applied to survival prediction in breast
This paper has outlined, discussed and resolved the cancer. Oncology 1999; 57:281-6.
issues, algorithms, and techniques for the problem of [9] Delen D, Walker G, Kadam A. Predicting breast
breast cancer survivability prediction in SEER cancer survivability: a comparison of three data
database. Unlike the pre-classification process used mining methods. Artificial Intelligence in
in [9], our approach takes into consideration, besides Medicine. 2005 Jun; 34(2):113-27.
the Survival Time Recode (STR), the Vital Status [10] Ian H. Witten and Eibe Frank. Data Mining:
Recode (VSR) and Cause of Death (COD). The Practical machine learning tools and techniques,
experimental results show that our approach 2nd Edition. San Fransisco:Morgan Kaufmann;
outperforms the approach used in [9]. 2005.
This study clearly shows that the preliminary [11] J. R. Quinlan, C4.5: Programs for Machine
results are promising for the application of the data Learning. San Mateo, CA:Morgan Kaufmann;
mining methods into the survivability prediction 1993.
problem in medical databases. [12] Weka: Data Mining Software in Java,
Our analysis does not include records with http://www.cs.waikato.ac.nz/ml/weka/
missing data; future work will include the missing