Using few characteristic parameters to enhance the differentiation of prostate cancer from benign pros tatic hyperplasia: a single-center… [601405]

For Peer Review Only

Using few characteristic parameters to enhance the
differentiation of prostate cancer from benign pros tatic
hyperplasia: a single-center retrospective study in China

Journal: IEEE/ACM Transactions on Computational Biology and Bioinformatics
Manuscript ID TCBB-2017-06-0235
Manuscript Type: Regular Paper
Keywords: Aided diagnostic model, decision tree, prostate can cer, prostate-specific
antigen density

Transactions on Computational Biology and Bioinformatics

For Peer Review OnlyIEEE TRANSACTIONS ON JOURNAL NAME, MANUS CRIPT ID 1
Using few characteristic parameters to
enhance the differentiation of prostate cancer
from benign prostatic hyperplasia: a single-
center retrospective study in China
Yi-Yan Zhang, Qin Li, Yi Xin* and Wei-Qi Lv
Abstract — The incidence of prostate cancer annually increases. Prostate cancer has become a vital disease for men in China.
We conducted a cross -sectional study of 392 eligible patients from 710 men with prostate cancer or benign pros tatic
hyperplasia between 2000 and 2003. For total prostate- specific antigen, age, three diameters of prostate, prostate volume and
prostate- specific antigen density 7 indices, the analysis of variance and t test were used to analyze the difference between
groups. Then the decision tree with pruning was established by prostate- specific antigen density, age, transversal diameter and
anteroposterior diameter of prostate. Thus, the risk of prostate cancer was predicted. According to the established decision tr ee
model, prostate- specific antigen density is the most important factor affecting the occurrence of prostate cancer. A benign
prostatic hyperplasia patient with larger transverse diameter and anteroposterior diameter of prostate is prone to suffer from
prostate cancer, with prostate- specific antigen density ranging from 0.32 to 0.5ng/L². With no additional index was introduced,
the detection rate of prostate cancer was 80.0 %. In the meantime, the specificity was enhanced to 80.3% .
Index Terms— Aided diagnostic model , decision tree, prostate cancer , prostate- specific antigen density
——————————  ——————————
1 INTRODUCTION
S the most commonly diagnosed malignancy among
men, prostate cancer (PCa) has become the main
public health concern among non-communicable diseases
in the world. In western countries, PCa is the main cause
of male malignant tumors and the second leading cause
of death after lung cancer [1]. Although the incidence of PCa in China was significantly lower than that of the
United States and Europe, the number of cases increases,
and the onset age tends to be younger due to the aging
population, change of diet, and physical examination or
medical screening [2]. PCa has become an important
threat to the health of men in China. Simil ar to other ma-
lignancies, early prevention, diagnosis, and treatment are
essential to reduce the metastasis and improve the pro g-
nosis of patients with PCa. Discovering tumor markers
with high specificity and sensitivity is the key to early
treatment and de tection. The current main examination
methods for PCa are digital rectal examination, serum
prostate -specific antigen (PSA), and transrectal ultr a-
sound [3].
PSA is a single chain glycoprotein secreted by the pro s-
tate gland bubble and duct epithelial cells; it has a mole-
cular weight of about 34,000 and is the main tool for PCa screening. This protein only exists in prostate tissues, and its level increases in different degrees in the serum of pa-
tients with PCa [4]. PSA plays an important role in the
diagnosis, staging, therapeutic monitoring, and prognosis of PCa because this protein was successfully separated
from the prostate tissue. PCA also advanced the diagnosis
of prostate cancer. In 1986, the U.S. Food and Drug A d-
ministration recommended the applicatio n of serum PSA
for the diagnosis of PCa. Although PSA is the most im-portant tumor marker in PCa, it still has some limitations. When the serum PSA value rang es from 4 μg/L to 10 μ
g/L, the specificity of the diagnosis of PCa is low[5], both PCa and prostatic hyperplasia are likely to occur, and the
biopsy positive rate is low [6]. The level of serum PSA can
be increased by many factors, such as “false positive ”
and “false negative” that usually appear in clinical biopsy,
benign prostatic hyperplasia (BPH), inf ection, and chronic
prostatitis [7]; thus, the confidence on PSA decreases. In addition, the development of a wide range of PCa scree n-
ing based on PSA also leads to over -diagnosis. Therefore,
new tumor markers or effective diagnostic indicators are urgentl y needed for the accurate screening of PCa and
prediction of the prognosis of patients with PCa.
The PSA density (PSAD) was popularized in the mid –
1990s [8] , [9]. Some studies confirmed the diagnostic us e-
fulness of PSAD in the detection of PCa [10], [11], [ 12].
The research of Porcaro et al. showed a direct association between PCa and lower prostate volume index [13]. M o-
schini et al. confirmed that the small size of the prostate
might increase the risk of recurrence after radical prost a-
tectomy in patients w ith moderate risk, whereas this ch a-
racteristic is not correlated with the recurrence in low and high risk groups [14]. Giri et al. found that several single
nucleotide polymorphisms are associated with the size of the prostate gland, and the mechanisms related to loci,
xxxx-xxxx/0x/$xx.00 © 200x IEEE Published by the IEEE Computer Society ————————————————
• The authors are with the School of Life Science , Beijing Institute of Tec h-
nology , Beijing 100081, China. E-mail: zhangyiyan0307 @163.com, {liqin,
ameko }@bit.edu.cn, lwq_voice@126.com .
A Page 1 of 12 Transactions on Computational Biology and Bioinformatics
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only2 IEEE TRANSACTIONS ON JOURNAL NAME, MANUS CRIPT ID
prostate volume, and prostate hyperplasia should be i n-
vestigated [15].
Big data analysis and data mining recently attracted
public attention, particularly the wide application of data
mining in the field of health care. A large number of pa-
tient's medical data were stored with the application and
promotion of hospital information system (HIS). The ab-
undant information inherent in the stored data can be
mined by using data mining methods and can be useful
for doctors during treatment and even i n the scientific
field.
The prevalence of PCa in China is different from that
in Europe and America. This paper studied several cha-racteristic parameters of PCa and BPH patients who were
admitted to the Department of Urology in a large three –
grade hospital in China. This paper aims to explore and
mine the relevant factors of PCa based on the medical
data, provide theoretical support for early prevention and
screening, and improve the prognosis of patients with PCa. By modeling a decision tree, the diagnost ic aided
model of PCa was established to analyze the laboratory data of related factors within PCa and BPH patients, mine the association between PCa and related factors, and pr e-
dict the occurrence of PCa.
2 MATERIALS AND METHODS
2.1 Data Sources
The research data source is the database of prostate dis-
ease patients during 2000 –2003, which is collected from
the Urology Department of Chinese PLA General Hospit-al and is provided by the Clinical Data Center. The dat a-
base contains all outpatient and inpatie nt cases from 1
January 2000 to 31 December 2003. Data items include patient ID, gender, age, date of birth, total PSA (tPSA),
superoinferior diameter (S- I diameter), anteroposterior
diameter (A -P diameter), transversal diameter of prostate,
volume of pros tate, admitting diagnosis, and discharge
diagnosis.
2.2 Inclusion Criteria and Data Extraction
The database includes 710 cases of prostate disease in the
urological department from January 2000 to December
2003. The inclusion criteria are set as follows: ( 1) patients
with PCa (including acinar adenocarcinoma, tubular ad e-
nocarcinoma, urothelial carcinoma, squamous cell carc i-
noma, and adenosquamous carcinoma) or BPH as co n-
firmed by the pathological diagnosis; (2) records of labo r-
atory examination results that contain S- I diameter, A -P
diameter, transversal diameter of prostate measured by B –
mode ultrasound; (3) initial tPSA examination records of
each patient (to reduce the interference during treatment
as caused by the regulation index); and (4) patients olde r
than 40 years of age at the time of examination. Accor d-
ing to the inclusion criteria, 392 patients are eligible and constitute the research dataset. The participants were
aged 44 –97 years with an average of 73.19 ± 7.41 years old.
2.3 Statistical Analys es
The mean ± standard deviation and distribution of me a-surement data, such as tPSA, age, S -I diameter, A -P di-
ameter, and transversal diameter of prostate, were calc u-
lated for eligible patient records. Shapiro -Wilk test was
used to analyze whether each meas urement follows a
normal distribution. The analysis of variance and t test
were used to analyze the difference between groups. PV
was calculated according to the general formula of pro s-
tate volume (PV) [13]:
V=0.52×S-I diameter ×A-P diameter ×transversal diameter (1)
PSAD was obtained from the ratio of tPSA and PV.
The corresponding distribution test and statistics of PV
and PSAD were also conducted.
R version 3.2.2 was used to analyze data in the study,
and the p values less than 0.01 were considered signif i-
cant.
2.4 Decision Tree
The decision tree was originated from artificial intell i-
gence machine learning technology to explore the inh e-
rent law of the data and predict the unknown data. Dec i-
sion trees are widely used in many fields because of their
excellent data analysis and ease of understanding. As one of the representative supervised learning algorithms, the
decision tree model is a simple and easy -to-use non-
parametric classifier. In a decision tree, any priori as-
sumption on the data is no t required, and the results are
easy to explain. In addition, the decision tree has other
advantages, such as fast calculation speed, easy -to-
interpret results, strong robustness, and insensitivity to noise and missing values.
The induction method of decision tree follows a un i-
fied recursive mode. First, the root node represents the current given dataset. Starting from the root node, the remaining attributes are detected in turn at each node.
According to the information theoretical criteria, attribute
selection is performed, and the data set is divided into some smaller subsets. The subdivision stops when the
purity of each subset is sufficiently high or the initial stop
condition is satisfied [16]. The over -fitting might lead to
the loss of universality; thus, the model cannot be used to
predict unknown data. The pruning step is necessary to
avoid over -fitting of the training data. The commonly
used pruning techniques are pre -pruning and post –
pruning; the former is mainly used to limit the full
growth of t he decision tree, and the latter performs pru n-
ing after the decision tree is fully grown [17]. In this study,
the pathological diagnosis, namely, suffering from PCa or
BPH, was considered as the target variable, which is r e-
garded as the root node in the ab ove principle. As
attribute variables, the information gain for tPSA, S- I di-
ameter, A -P diameter, transversal diameter of prostate,
PV, and PSAD are compared at the next node, and the
variable with maximum gain is selected as the split vari a-
ble of this nod e. In this step, the remaining variables are
continuously compared and selected until the preset stop condition is reached. Finally, each leaf node contains only
one type of diagnosis (PCa or BPH).
The variable importance can be evaluated according to
the average reduction Gini index of the split node in the
decision tree generation. Higher reduction in Gini index Page 2 of 12 Transactions on Computational Biology and Bioinformatics
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review OnlyAUTHOR ET AL.: TITL E 3
indicates the greater importance of this variable. This
study evaluates the importance of each variable by using
the Gini index reduction of variables. The top positions in
the ranking are viewed as the relevant factors of PCa.
3 RESULTS
3.1 Statistical Results
The examination records of 392 eligible prostate disease patients from January 2000 to December 2003 were i n-
cluded in this study. According to the preliminary statis-
tics, 150 subjects with PCa (38.27% of the total subjects) and 242 cases of BPH (61.73%) were included. The distr i-
bution of tPSA in patients with PCa and BPH is shown in
Figure 1.

Fig. 1. Distribution of tPSA in patients with P Ca and BPH.
From Figure1, the tPSA level in patients with BPH is
mostly below 40μg/L, whereas that for75% of patients
with PCa is higher than the average level for BPH pa-
tients. The levels of tPSA were different in patients with
PCa and BPH. The range of serum tPSA in patients with
PCa is larger than that in BPH patients, and some ove r-
lapping areas are observed. Based on the Shapiro -Wilk
test, the tPSA values of PCa and BPH patients do not fo l-
low a normal distribution; the kurtosis values are
6.910729 and 11.8886, and the skewness values are
2.461218 and 2.883435, respectively. According to the analysis of variance, the serum tPSA levels of patients
with PCa and BPH are significantly different; the F -value
is 61.96, and the p- value is 3.5 10
-14.
3.2 Decision Tree Model Establishment
In this study, under -sampling and stratified random
sampling methods are combined to sample the examina-
tion records of 150 patients with PCa and 242 cases of
BPH. A total of 120 patients are randomly selected from
each of the two groups to comprise the training data. These sampling methods are helpful to enhance the acc u-
racy and stability of classification prediction, thereby e n-
suring the full study of the current data set.
According to the existing examination indices and pa-
thological diagnosis, a decision tree is established to learn
the PCa classification. The decision tree prediction model
and the corres ponding rule set are obtained. The decision
tree prediction model is shown in Figure 2, and the rule
set is listed in Table 1.

Fig. 2. Decision treeprediction model of prostate patients .
TABLE 1
Decision rules generated by the decision tree model

PSAD: prostate -specific antigen density, A- P diameter: anteroposterior
diameter .

Number Decision rule
1 [Class=BPH cover=64 (27%) prob=0.88]
PSAD< 0.3218
PSAD< 0.1709
2 [Class=BPH cover=10 (4%) prob=0.80]
PSAD>=0.3218
PSAD< 0.7684
Transversal diameter >=5.45
3 [Class=BPH cover=37 (15%) prob=0.76]
PSAD< 0.3218
PSAD>=0.1709
Age< 78.5
4 [Class=BPH cover=9 (4%) prob=0.67]
PSAD>=0.3218
PSAD< 0.7684
Transversal diameter < 5.45
PSAD>=0.6433
5 [Class=BPH cover=11 (5%) prob=0.64]
PSAD>=0.3218
PSAD< 0.7684
Transversal diameter < 5.45
PSAD< 0.6433
PSAD< 0.4993
A-P diameter < 3.15
6 [Class=PCa cover=15 (6%) prob=0.27]
PSAD< 0.3218
PSAD>=0.1709
Age>=78.5
7 [Class= PCa cover=18 (8%) prob=0.22]
PSAD>=0.3218
PSAD< 0.7684
Transversal diameter < 5.45
PSAD< 0.6433
PSAD< 0.4993
A-P diameter >=3.15
8 [Class= PCa cover=60 (25%) prob=0.10]
PSAD>=0.3218
PSAD>=0.7684
9 [Class= PCa cover=16 (7%) prob=0.06]
PSAD>=0.3218
PSAD< 0.7684
Transversal diameter < 5.45
PSAD< 0.6433
PSAD>=0.4993 Page 3 of 12 Transactions on Computational Biology and Bioinformatics
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only4 IEEE TRANSACTIONS ON JOURNAL NAME, MANUS CRIPT ID

The following rules are obtained from the rightmost
branch of the decision tree in Figure2: if the PSAD value
is not “> = 0.32” nor “> = 0.17” (that is, when the patient has a PSAD value of less than 0.17ng/L²), then the patient
suffers from BPH. This rule is consistent with the rule1 in
the rule set.
3.3 Relevant Factor Analyses
The importance of each input variable can be dete r-
mined based on the change of the Gini index of each var i-
able during decision tree construction. As shown in T a-
ble2, the Gini index of PSAD has the greatest change. Compared with other indices, PSAD is the most impo r-
tant factor affecting the occurrence of PCa. The second
important factor is tPSA. A larg e gap is found between
PSAD, tPSA, and other indices.
TABLE 2
Importance evaluation of various indices

PSAD: prostate -specific antigen density, tPSA: total prostate -specific ant i-
gen, S -I diameter: superoinferior diameter, PV: prostate volume, A- P di-
amete r: anteroposterior diameter.

3.4 Model Evaluation
Except for the training set subjects, the test set includes
the remaining 30 PCa patients and 122 BPH patients. The
decision tree prediction model is executed on the test set
to validate the generality of the prediction effect.
In this paper, we use the sensitivity, specificity, accur a-
cy, and Youden index (Y) to evaluate the model predic-tion results. Sensitivity is the probability that the PCa pa-tients are correctly classified by the prediction model.
Specificity is the possibility of correctly predicting the
patients with BPH based on the prediction model. Acc u-
racy refers to the percentage of all results that are correc t-
ly predicted. Y is calculated by Y = sensitivity + specific i-
ty-1. All the aforementioned evaluate indices are ind e-
pendent of disease prevalence. To evaluate the effectiv e-
ness of a model, we need to synthetically consider the sensitivity, specificity, and several other indices. The evaluation indices of the PCa general prediction model
are sho wn in Table3.
TABLE 3
Evaluation indices of the prediction model

Y: Youden index
Table3 shows that for the training and test sets, the
specificity values of the decision tree prediction model are 87.5% and 80.3%, respectively, which are greater than
80%. The sensitivity values of the decision tree prediction model are 78.3% and 80.0% respectively, that is, the PCa
screening rate of the model is about 80%. The accuracy
values of the training and test sets are 82.9% and 80.3% respectively. Therefore, th e model has high generality
and is not over -fitted.
4 DISCUSSION
The statistical analysis results show that the distribution of tPSA is different in PCa and BPH patients. The range of
serum tPSA in patients with PCa is greater than that in
BPH patients, co nfirming the rationality of using PSA as a
screening tool for PCa in prior research and other studies.
According to the decision tree prediction model estab-
lished in this study, PSAD has a greater correlation with PCa than PSA. Therefore, PSAD is statistic ally analyzed
to perform the post hoc test, and the results are shown in Figure3.

Fig. 3. Distribution of PSAD in patients with PCa and BPH.
As seen from Figure 3, the PSAD in the two groups of
BPH and PCa patients shows evident differences and re l-
atively less coincidence parts than PSA. PSAD is more conducive to the discrimination of PCa. Horiguchi et al.
claimed that PSAD is the most valuable predictor of PSA
related indices [18]. Considering the similar results, their proposed hypothesis serves as a theoretical support for
the current study. Matsuyama et al. reported that when
the PSAD value higher than 0.19ng/L² is used as the cr i-
terion for the diagnosis of PCa, the sensitivity and spec i-
ficity of diagnosis are 58% and 79%, respectively. If the 0.20ng/L² is set as a diagnostic threshold, then the sensi-
tivity and specificity are 82.6% and 74.6%, respectively
[19]. The decision tree model constructed in this study
simultaneously considers both the PSAD and the known information such as the S- I diamet er, A -P diameter, and Indices Variation of Gini index
PSAD 44.863056
tPSA 23.508395
S-Idiameter 15.794838
PV 14.945968
Transversal diameter 13.614238
A-Pdiameter 12.480382
Age 6.715725

Data set Sensitivity Specificity Accuracy Y
Training set 0.783 0.875 0.829 0.658
Test set 0.800 0.803 0.803 0.603

Page 4 of 12 Transactions on Computational Biology and Bioinformatics
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review OnlyAUTHOR ET AL.: TITL E 5
transversal diameter of prostate. The sensitivity and spe-
cificity are 80.0% and 80.3%, respectively, when no other
examination indices are introduced. Thus, the specificity
is improved, and the high detection rate of PCa is ensured. The waste of time and resources caused by too many in-
spection items can be minimized, and the clinical pat h-
way can be shortened.
To our best knowledge, this study is the first to report
the relationship of transverse diameter and A- P diameter
of prostate to PCa. PV is an important parameter in d e-
tecting PCa, as reported by the reviews of the literature
stating that smaller PV sizes are associated with PCa [20],
[21]. However, our results indicate that a BPH patient
with larger transverse diameter and A -P dia meter of
prostate is prone to suffer from PCa, with PSAD ranging from 0.32 to 0.5 ng/L².
Considering the limitations of the data, the prostate
disease dataset only included PCa and BPH patients but the normal population is not covered. In addition, this research is limited to the cases with tPSA and PV values
from 2000 to 2003; other new examination items added in recent years are not included. In some reports, the free PSA and the ratio of free PSA and tPSA significantly in-
fluence the detection of PCa. Fur thermore, the diagnosis
and treatment data of the same patient at different time
can be covered, and a longitudinal study and mining can
be conducted when the amount of data is sufficient. This
factor is conducive for the accurate identification of the
factors associated with prostate -related diseases and the
precision prediction of risk degree.
5 CONCLUSION
The clinical application of PSA still has limitations as an
important tool for PCa screening. In this paper, a decision
tree prediction model is established to aid PCa diagnosis
and proved that PSAD has stronger distinguishing ability, especially in combination with age, tranversal diameter
and anteroposterior diameter of prostate 3 indices. This
work can help provide better diagnosis and treatment mode for PCa screening, diagnosis, prognosis, and fo l-
low-up and provide support for doctors and experts to
make the right decision. Considering the limitation of the data, only the age, tPSA, PV, and PSAD are studied in
this paper. Some other indicators will be further studied
and explored to improve the sensitivity and accuracy of
the PCa screening model.

ACKNOWLEDGMENT
The authors wish to thank the Clinical Data Center . This
work was supported in part by a grant from the staffs.
REFERENCES
[1] H. T. Lynch, O . Kosoko- Lasaki, S . W. Leslie, M . Rendell, T .
Shaw, C . Snyder , A. V. D’Amico, S . Buxbaum, W . B. Isaacs, S .
Loeb, J. W. Moul and I . Pow ell, “Screening for familial and h e-
reditary prostate cancer: Screening in prostate cancer families ,”
INT J CANCER , vol.138, no.11 , pp.2579– 2591, 2016. [2] D.–W. Ye and Y. Zhu, “Epidemiology of prostate cancer in
China：an overview and clinical implication ,” Chin J Surg , vol.
53, no. 4, pp. 249-252, 2015.
[3] X.-H. Liao , S.-Y. Zhu, H. Yang , Y. He and D. -Y. Liu , “Transre c-
tal Prostatic Biopsy Combined with Serum Prostate Antigen
and Transrectal Ultrasonography in Diagnosis of Prostate Car-
cinoma ,” Chinese J Ul trasound Med , vol. 26, no. 8, pp. 744-747,
2010.
[4] Q.-W. Huang , G.-M. Wang and Y. Wang , “Progress in Research
on Biomarkers f or Prosta te Cancer ,” Chinese Journal of Clinical
Medicine , vol. 20, no. 1 , pp. 99–101, 2013.
[5] Y.-Q. Wu and X.-H. Liu, “The diagnostic value of prostate
health index for prostate cancer ,” Chin J Exp Surg , vol. 32, no. 5,
pp. 1190– 1192, 2015.
[6] X.-Z. Duan and Z.-H. Tao, “The progress of biomarkers in pro s-
tate cancer ,” Chin J Lab Med , vol. 38, no. 10, pp. 657– 660, 2015.
[7] W. Wang , C.-H. Li, H. Liu and B. Yang, “Evaluation of the D i-
agnostic Value of PSA in Prostate Cancer Patients, ” Journal of
China Medical University , vol. 45, no. 1 , pp. 61-65, 2016.
[8] M. C. Benson, I . S. Whang , A. Pantuck , K. Ring , S.A. Kaplan,
C.A. Olsson and W.H. Cooner , “Prostate specific antigen densi-
ty: a means of distinguishing benign prostatic hypertro phy and
prostate cancer ,” J Urol , vol. 147, pp. 815–816, 1992.
[9] M. Ohori, J.K. Dunn and P.T. Scardino , “Is prostate -specific
antigen density more useful than prostate -specific antigen l e-
vels in the diagnosis of prostate cancer? ” Urology , vol. 46, no. 5,
pp. 666–671, 1995.
[10] Y. Arai, H. Maeda, S. Ishitoya, K. Okubo, T . Okada and Y. Aok i,
“Prospective evaluation of prostate specific antigen density and
systematic biopsy for detecting prostate cancer in Japanese p a-
tients with normal rectal examinations and intermediate pro s-
tate specific antigen levels ,” J Uro l, vol. 158, pp. 861– 864, 1997.
[11] M. Bazinet , A.W. Meshref, C. Trudel , S. Aronson, F. Péloquin ,
M. Nachabe , L.R. Bégin and M.M. Elhilali, “Prospective evalu a-
tion of prostate -specific antigen density and systematic biopsies
for early detection of prostatic carci noma ,” Urol ogy, vol. 43, no.
1, pp. 44 –51,1994.
[12] H.-J. Yu and M. -K. Lai, “The usefulness of prostate -specific
antigen (PSA) density in patients with intermediate serum PSA
level in a country with low incidence of prostate cancer ,” Urol-
ogy, vol. 51, pp. 125– 130, 1998.
[13] A.B. Porcaro, G. Novella, A. Molinari, A. Terrin, A. Minja, V . De
Marco, G. Martignoni, M . Brunelli, M .A. Cerruto, P . Curti, S .
Cavalleri and W. Artibani , “Prostate Volume Index and Chro n-
ic Inflammation of the Prostate Type IV with Respect to the
Risk of Prostate Cancer ,” UROL INT , vol. 94, no. 3 , pp. 270- 285,
2015.
[14] M. Moschini, G. Gandaglia, N. Suardi, N. Fossati , V. Cucchiara ,
R. Damiano , F. Cantiello, S .F. Shariat , F. Montorsi and A. Bri-
ganti, “Importance of prostate volume in the stratification of
patients with intermediate -risk prostate cancer ,” INT J UROL ,
vol. 22, no. 6 , pp. 555- 561, 2015.
[15] G. Ayush, L.E. Todd, S.M. Saundra, H. B . Susan and H. F. Jay,
“Genetic Determinants of Metabolism and Benign Prostate E n-
largement: Associations with Prostate Volume ,” Plos One , vol.
10, no. 7 , pp. e01320282015.
[16] J.R. Quinlan, “Induction o f Decision Trees ,” MACH LEARN ,
vol. 1, no. 1, pp. 81–106, 1986.
[17] J.-W. Han, K. Micheline and J. Pei, Data Mining: Concepts and
Techniques, 3rd ed. San Francisco: Morgan Kaufmann Publish-Page 5 of 12 Transactions on Computational Biology and Bioinformatics
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only6 IEEE TRANSACTIONS ON JOURNAL NAME, MANUS CRIPT ID
ers, pp. 344, 2011 .
[18] H. Akio, N. Jun, H. Yutaka, K. Nakagawa, M . Oya, T . Ohigashi,
K. Marumo and M. Murai , “Prediction of extraprostatic cancer
by prostate specific antigen density, endorectal MRI, and bio p-
sy Gleason score in clinically localized prostate cancer ,” Pros-
tate, vol. 56, no. 1 , pp. 23–29, 2003.
[19] M. Hideyasu , B. Yoshikazu , Y. Gen -ichiro , N. Yamamoto and K.
Naito , “Diagnostic value of prostate -specific antigen -related p a-
rameters in discriminating prostate cancer ,” INT J UROL , vol. 7,
no. 11, pp. 409–414, 2000.
[20] M. J. Roobol, F.H. Schröder, J. Hugosson, J. S. Jones, M . W.
Kattan, E . A. Klein, F. Hamdy, D . Neal, J. Donovan, D. J. Pa rekh,
D. Ankerst, G. Bartsch, H . Klocker, W. Hornin ger, A. Benchikh,
G. Salama, A . Villers, S . J. Freedland, D . M. Moreira, A. J. Vic k-
ers, H. Lilja and E . W. Steyerberg , ”Importance of prostate v o-
lume in the European Randomised Study of Screening for Pro s-
tate Cancer (ERSPC) risk calculators: results from the prostate
biopsy collaborativ e group ,” World J Urol , vol. 30, no. 2, pp.
149– 155, 2012.
[21] R. Al-Azab, A. Toi, G. Lockwood, G.S. Kulkarni and N. Flesh-
ner, “Prostate volume is strongest predictor of cancer diagnosis
at transrectal ultrasound -guided prostate biopsy with prostate –
specific antigen values between 2.0 and 9.0 ng/ml ,” Urology ,
vol. 69, no. 1): 103 –107, 2007.

Yi-Yan Zhang received the BE degree from
Hangzhou Dianzi University, Hangzhou, Chi-
na. She is currently working toward the PhD
degree in the Department of Biomedical E n-
gineering, Beijing Institute of Technology . Her
research focuses on data mining, machine
learning and healthcare data analysis . She
has published about 3 journal and conference
papers.

Qin Li received the PhD degree from Beijing
Institute of Technology , Beijing , China. Sh e is
a professor in the Department of Biomedical
Engineering, Beijing Institute of Technology .
Her research focuses on data mining, health-
care data analysis and biophotonics . She has
published more than 10 journal papers.

Yi Xin received the PhD degree from Beijing
Institute of Technology , Beijing , China. Sh e is
currently working as a senior lecturer in the
Department of Biomedical Engineering, Bei-
jing Institute of Technology . Her research
focuses on data mining, healthcare data
analysis and signal processing. She has
published more than 10 journal papers.

Wei-Qi Lv received the BE degree from Bei-
jing Institute of Technology , Beijing , China.
He is currently working toward the Master
degree in the Department of Biomedical E n-
gineering, Beijing Institute of Technology . His
research focuses on data mining and health-
care data analysis .

Page 6 of 12 Transactions on Computational Biology and Bioinformatics
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only

122x155mm (600 x 600 DPI)

Page 7 of 12 Transactions on Computational Biology and Bioinformatics
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only

68x70mm (600 x 600 DPI)

Page 8 of 12 Transactions on Computational Biology and Bioinformatics
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only

146x185mm (600 x 600 DPI)

Page 9 of 12 Transactions on Computational Biology and Bioinformatics
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only

297x420mm (300 x 300 DPI)

Page 10 of 12 Transactions on Computational Biology and Bioinformatics
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only

297x420mm (300 x 300 DPI)

Page 11 of 12 Transactions on Computational Biology and Bioinformatics
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only

297x420mm (300 x 300 DPI)

Page 12 of 12 Transactions on Computational Biology and Bioinformatics
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60