978-1-4244-9352-4/11/$26.00 ©2011 IEEE 1689
2011 4th International Conference on Biomedical Engineering and Informatics (BMEI)
Comparison of Bayesian Network and Binary
Logistic Regression Methods for Prediction of
Prostate Cancer
Selen Bozkurt
Akdeniz University, Faculty of Medicine Department of
Biostatistics and Medical Informatics,
Antalya, Turkey
Asli Uyar
Akdeniz University, Faculty of Medicine Department of
Biostatistics and Medical Informatics,
Antalya, Turkey
Kemal Hakan Gulkesen
Akdeniz University, Faculty of Medicine Department of Biostatistics and Medical Informatics,
Antalya, Turkey
Abstract—Prostate cancer is one of the most common cancers in
men. Luckily, Serum PSA level, age, digital rectal examination
(DRE), and clinical symptoms are helpful for early detection of
this tumor. The aim of this study was to examine and compare
the methods used for improving the diagnostic accuracy of serum
PSA in Turkey, a country with low incidence of prostate cancer.
The predictors used for early detection of prostatic carcinoma
were identified by both Logistic Regression and Bayesian
networks. The results of the methods were compared in terms of
predicting performance and advantages
Keywords-component; Prostate Cancer, Bayesian Networks,
Logistic Regression
I. INTRODUCTION (HEADING 1)
Prostate cancer is one of the most common cancers in men
[1-2]. In the United States, approximately 32,000 men have
died of the prostate tumor as a part of the 217,730 men who
were detected with that disease in 2010 [2]. Luckily, there are
some predictors used for early detection of prostatic carcinoma.
Serum PSA level, age, digital rectal examination (DRE), and
clinical symptoms are helpful for early detection of this tumor
[3-4].
When a patient is suspected to have a prostate tumor, a
biopsy from the prostate is advised by the physician.
Sometimes, because of the presence of strong indicators such
as a very high serum PSA level, the decision for biopsy is easy.
However, when the findings are in the grey zone, physicians
and patients have to make a choice between the risk of missing
an early detection of a tumor and the risk of an unnecessary
biopsy [5]. There are several studies [3-4, 6] trying to establish
methods to improve the sensitivity and specificity of different
examinations in these grey zone cases. Generally, a PSA level
between 4-10 ng/ml is accepted as having a 70% sensitivity
and a 70% specificity [5, 7]. Since 1989, several concepts to
further improve the diagnostic accuracy of PSA have been
developed with the aim of avoiding unnecessary biopsies.
Likewise, the aim of this study was to examine and
compare the methods used for improving the diagnostic
accuracy of serum PSA in Turkey, a country with low
incidence of prostate cancer. The predictors used for early
detection of prostatic carcinoma were identified by both
Logistic Regression and Bayesian networks. The results of the
methods were compared in terms of predicting performance
and advantages.
II. METHOD
A. Study Population
All the transrectal ultrasound (TRUS)-guided prostatic
biopsy cases who were admitted to Akdeniz University
Hospital, Department of Urology, between January 2000-April
2007 was retrospectively evaluated. TRUS-guided biopsy
could be performed only in Akdeniz University in Antalya
district which has a population around 1,800,000. Original
dataset included medical records of 1453 patients, samples
including missing variables were excluded from the study and
the remaining 983 cases, whose serum PSA level ranged
between 0.05 and 1000, have been analyzed. In patients with
multiple biopsies only the first one was included in the study.
B. Dataset characteristics
Akdeniz University Hospital Information System (HIS) and
the patients’ medical records of the Department of Urology
were used as data sources. HIS contains the demographic
information about all patients, laboratory data of the last eight
years, and pathology reports.
C. Statistical Analysis
Logistic Regression Analysis