Tumor Biol.

(2013) 34:1547–1552
DOI 10.1007/s13277-013-0683-5

RESEARCH ARTICLE

A support vector machine model for predicting non-sentinel

Received: 1 November 2012 / Accepted: 28 January 2013 / Published online: 10 February 2013
# International Society of Oncology and BioMarkers (ISOBM) 2013

Abstract This study aimed to investigate the accuracy and Keywords Breast cancer . SLNB . Prediction . Data model
feasibility of support vector machine (SVM) modeling in
predicting non-sentinel lymph node (NSLN) status in patients
with SLN-positive breast cancer. Clinicopathological data Introduction
were collected from 201 cases with sentinel lymph node
biopsy breast cancer and included patient age, tumor size, Breast cancer is the most common solid tumor in women
histological type and grade, vascular invasion, estrogen re- globally [1]. While the incidence of breast cancer in
ceptor status, progesterone receptor status, CerbB2 status, Western countries has stabilized and even decreased, the
size and number of positive SLNs, number of negative incidence rates in Asia, including China, have been in-
SLNs, and positive SLN membrane invasion. Feature vector creasing rapidly [2–4]. Several large, randomized prospec-
selection was based on a combination of statistical filtration tive trials have shown that sentinel lymph node biopsy
and model-dependent screening. The arbitrary combination (SLNB) substantially reduces the morbidity associated
with the smallest p value for SVM input was selected, the with axillary surgery compared with formal axillary lymph
predicative results of the model were evaluated by a 10-fold node dissection (ALND) [5–11]. Moreover, the National
cross validation, and a training model was established. Using Surgical Adjuvant Breast and Bowel Project B-32 trial has
SLN-positive patients as a double-blind test set, 85 patients demonstrated that when the sentinel node reveals no evi-
were input into the model to analyze its sensitivity and dence of metastatic disease, then no further ALND is
specificity. The combination with the highest cross- required [12]. Recently, the results of the American
validation accuracy was selected for the SVM model and College of Surgeons Oncology Group Z0011 trial have
consisted of the following: the number and size of positive challenged the notion that all patients with metastases to
SLNs, the number of negative SLNs, and the membrane the sentinel node require ALND. The results of this trial
invasion of positive SLNs. The training accuracy of the suggest that in selected sentinel node-positive patients,
model established with the four variables was 92 %, and ALND can be potentially avoided [13]. However, a large
its cross-validation veracity was 87.6 %. The accuracy of number of studies show that the lymph node metastases in
an 85-patient double-blind test of the SVM model was approximately 40–70 % of SLN-positive breast cancer
91.8 %. In conclusion, this SVM model is an accurate and patients are confined to the SLN [14–17]. ALND in this
feasible method for the prediction of NSLN status in cohort of patients would not, therefore, clear residual
SLN-positive breast cancer and is conducive to guide metastatic lesions in the axilla or provide further staging
clinical treatment. information. Furthermore, ALND triples the risk of post-
operative complications (e.g., upper limb lymphedema,
sensory, and functional disturbance), affects the quality
X. Ding (*) : S. Xie : J. Chen : W. Mo : H. Yang
of life of the patient, and greatly increases medical
Department of Breast Surgery, Zhejiang Cancer Hospital,
Hangzhou 310022, China expenses, thereby deeming it unadvisable unless neces-
e-mail: dingxwhz@yeah.net sary. It is obvious that an accurate assessment of non-
1548 Tumor Biol. (2013) 34:1547–1552

sentinel lymph node (NSLN) status in patients with SLN- Table 1 Clinicopathological features of modeling cases
positive breast cancer is of great clinical significance. NSLN−, n (%) NSLN+, n (%)
For this purpose, models have been developed to predict Patient characteristics and
NSLN status in patients with SLN-positive breast cancer. tumor pathology 91 cases 110 cases
There are several commonly used assessment models.
Median age (years) 46 (29–78) 49 (25–69)
However, each of these models has limitations [18–20].
ER status
Support vector machines (SVM), developed by Vapnik and
Positive 25 (27.5) 39 (35.5)
Burges, are considered to be superior over traditional linear
Negative 58 (63.7) 65 (59.1)
approaches due to their capability to represent non-linear
Unknown 8 (8.8) 6 (5.5)
features within data [21]. Therefore, we established and
PR status
validated a discriminative SVM model, based on previous
Positive 33 (36.3) 48 (43.6)
SLNB cases in our hospital, to predict the NSLN status of
Negative 49 (53.8) 56 (50.9)
SLN-positive patients.
Unknown 8 (8.8) 6 (5.5)
CerbB2 status
Positive 17 (18.7) 30 (27.3)
Patients and methods
Negative 34 (37.4) 37 (33.6)
Unknown 40 (44.0) 43 (39.1)
From 2003 to 2010, 925 samples of non-selective SLNB
Vascular invasion
breast cancer were collected at our hospital, 201 of them met
+ 40 (44.0) 95 (86.4)
our modeling standards. The inclusion criteria included: no
− 51 (56.0) 15 (13.6)
preoperative systematic treatment, such as neoadjuvant che-
Tumor size
motherapy, was received; a successful SLNB surgery was
T1 47 (51.6) 31 (28.2)
performed; SLNs were positive; and an ALND was com-
T2 41 (45.1) 76 (69.1)
pleted. The exclusion criteria included the following: the
T3 3 (3.3) 3 (2.7)
reception of preoperative systematic treatment such as neo-
Multifocality
adjuvant chemotherapy, negative SLNs, or the lack of an
+ 1 (1.1) 2 (1.8)
ALND. Two-milliliter methylene blue was injected subcuta-
− 90 (98.9) 108 (98.2)
neously into the periareolar area for lymphatic mapping.
Histological type and grade
SLNB procedure was started 5 min after injection, and blue-
IDC grade 1 5 (5.5) 4 (3.6)
dyed lymph nodes with definite blue-dyed afferent drainage
IDC grade 2 59 (64.8) 88 (80.0)
lymphatic vessels were considered sentinel nodes. Additional
IDC grade 3 7 (7.7) 7 (6.4)
palpable-enlarged lymph nodes surrounding the blue- dyed
ILC 5 (5.5) 4 (3.6)
node or nodes, if existing, were also taken out. SLNB was
Mixed type 15 (16.5) 7 (6.4)
aborted once no blue-dyed nodes were identified 20 min
Size of positive SLNs
after injection. Results of SLNB were based on the final
<0.5 cm 43 (47.3) 3 (2.7)
pathological report originating from HE-stained sections
0.5–1.0 cm 44 (48.4) 79 (71.8)
irrespective of whether frozen sections performed or not.
1.0–1.5 cm 4 (4.4) 27 (24.5)
Clinicopathological data of the 201 patients were collected >1.5 cm 0 1 (0.9)
and included the patient’s age, tumor size, histological Number of positive SLNs
type, histological grade, vascular invasion, estrogen recep- 1 68 (74.7) 51 (46.4)
tor (ER) status, progesterone receptor (PR) status, CerbB2 2 19 (20.9) 37 (33.6)
status, size of positive SLNs, number of positive SLNs, 3 4 (4.4) 12 (10.9)
multifocality, number of negative SLNs, and positive SLN >3 0 10 (9.1)
membrane invasion. Positive SLN membrane invasion
We established a discriminative model according to + 34 (37.4) 97 (88.2)
NSLN status (positive or negative) using an SVM method. − 57 (62.6) 13 (11.8)
We employed the leave-one-out cross-validation method to Number of positive SLNs
evaluate the discrimination results of the model. SVM adop- 0 11 (12.1) 68 (61.8)
ted a radial base kernel, gamma was set at 0.6, and penalty 1 19 (20.9) 37 (33.6)
functions (C) were set at 100. Feature vector selection was 2 36 (39.6) 3 (2.7)
based on a combination of statistical filtration and model- 3 14 (15.4) 1 (0.9)
dependent screening. The Wilcoxon rank sum test was per- >3 11 (12.1) 1 (0.9)
formed on each peak of the mass-to-charge ratio, the
Tumor Biol. (2013) 34:1547–1552 1549

Table 2 Clinicopathological features of the double-blind test cases arbitrary combination with the smallest p value for SVM
NSLN− NSLN+ input was selected, the predicative results of the model were
Patient characteristics and tumor pathology 43 cases 42 cases evaluated by a 10-fold cross-validation, and the combination
with the highest Youden index as predicted by the SVM was
Median age (years) 48 (27–71) 47 (29–73) chosen as the basis for establishment of the training model.
ER status To make an accurate assessment of the model, we re-
Positive 15 (34.9) 15 (35.7) enrolled 85 patients in accordance with the modeling stand-
Negative 28 (65.1) 26 (61.9) ards. These patients were input into the model as a double-
Unknown 0 1 (2.4) blind test set to evaluate and calculate the sensitivity and
PR status specificity of the model.
Positive 20 (46.5) 13 (31.0)
Negative 23 (53.5) 28 (66.7)
Unknown 0 1 (2.4) Results
CerbB2 status
Positive 11 (25.6) 8 (19.0) Clinicopathological data of the 201 patients evaluated
Negative 19 (44.2) 19 (45.2) for modeling
Unknown 13 (30.2) 15 (35.7)
Vascular invasion The clinicopathological data of the 201 patients with SLN-
+ 26 (60.5) 36 (85.7) positive breast cancer used for modeling are shown in
− 17 (39.5) 6 (14.3) Table 1. The lesion was confined to SLNs (NSLN−) in 91
Tumor size cases (45.3 %, 91 of 201), the median age was 46, ER status
T1 17 (39.5) 9 (21.4) was unknown in eight cases (8.8 %, eight of 91), and CerbB2
T2 23 (53.5) 28 (66.7) status was unknown (including cases of CerbB2 IHC++ and
T3 3 (7.0) 5 (11.9) untested CerbB2 status) in 40 cases (44.0 %, 40 of 91).
Multifocality NSLN+ cases numbered 110 (54.7 %, 110 of 201), the
+ 0 2 (4.8) median age was 49, ER status was unknown in six cases
− 43 (100.0) 40 (95.2) (5.5 %, six of 110), and CerbB2 status was unknown in 43
Histological type and grade cases (39.1 %, 43 of 110).
IDC grade 1 3 (7.0) 1 (2.4)
IDC grade 2 17 (39.5) 22 (52.4) Clinicopathological features of cases for the double-blind test
IDC grade 3 11 (25.6) 11 (26.2)
ILC 2 (4.7) 5 (11.9) Clinicopathological data of the 85 patients used for the
Mixed type 10 (23.3) 3 (7.1)
double-blind test is shown in Table 2. For 43 cases
Size of positive SLNs
(50.6 %, 43 of 85), the lesion was confined to the SLNs
<0.5 cm 34 (79.1) 3 (7.1)
0.5–1.0 cm 8 (18.6) 36 (85.7)
Table 3 Variables selected for the SVM model
1.0–1.5 cm 1 (2.3) 3 (7.1)
Number of positive SLNs Variables p value
1 32 (74.4) 24 (57.1)
Age 0.1440008312
2 10 (23.3) 12 (28.6)
Tumor size 0.0015053351
3 1 (2.3) 4 (9.5)
Histological type 0.1163538893
>3 0 2 (4.8)
Histological grade 0.810639
Positive SLN membrane invasion
ER status 0.809162
+ 32 (74.4) 39 (92.9)
PR status 0.598087
− 11 (25.6) 3 (7.1)
Vascular invasion 0.0000091261
Number of negative SLNs
Size of positive SLNs 0
0 8 (18.6) 35 (83.3)
Number of positive SLNs 0.0000085344
1 15 (34.9) 6 (14.3)
Positive SLN membrane invasion 0
2 9 (20.9) 1 (2.3)
Multifocality 0.6959953426
3 8 (18.6) 0
Number of negative SLNs 0
>3 3 (7.0) 0
CerbB2 status 0.238588
1550 Tumor Biol. (2013) 34:1547–1552

Fig. 3 Training of the SVM model

Fig. 1 The training accuracy of the model and the accuracy of cross-
validation
These six variables were combined arbitrarily, and a SVM
training model was constructed for cross validation. The
(NSLN−), the median age was 43, and CerbB2 status was combination with the highest accuracy through a cross-
unknown in 13 cases (30.2 %, 13 of 85). Positive NSLN validation test for modeling included the number of positive
(NSLN+) was expressed in 42 cases (49.4 %, 42 of 85), the SLNs, the size of positive SLNs, the number of negative
median age was 42, ER status was unknown in one case SLNs, and positive SLN membrane invasion. The training
(2.4 %, one of 42), and CerbB2 status was unknown in 15 accuracy of the model constructed with these four variables
cases (35.7 %, 15 of 42). was 92 %, and the accuracy of cross-validation was 87.6 %
(Fig. 1).
Variable selection for use with the SVM model The histogram depicts the 13 variables in order from
smallest p value to greatest. The four variables with the
Thirteen variables including patient age, tumor size, histo- smallest p values were selected for the model. Variables 1–
logical type, histological grade, vascular invasion, ER sta- 13 correspond to the size of the positive SLNs, the number
tus, PR status, CerbB2 status, positive SLN size and of negative SLNs, positive SLN membrane invasion, the
number, negative SLN number, multifocality, and positive number of positive SLNs, vascular invasion, tumor size,
SLN membrane invasion were screened with the SVM. The histological type, age, CerbB2, PR, multifocality, ER, and
following six variables had p values of statistical signifi- histological grade, respectively.
cance: no vascular invasion, the size of the positive SLNs,
the number of positive SLNs, the number of negative SLNs,
positive SLN membrane invasion, and tumor size (Table 3).

Fig. 2 The training accuracy of the model, the cross-validation verac-

Training and test results of the model was 91.8 %. These data demonstrate that this model is stable
and reliable.
The training accuracy of the model established with the four Nevertheless, because all of the data for modeling were
variables was 92 %, the cross-validation veracity was obtained from a single tumor treatment center, this method
87.6 %, and the accuracy of the double-blind test was will require a larger number of cases, different groups of
91.8 %. Training of the SVM model on 201 patients people, and multiple treatment centers for verification. If the
revealed a sensitivity of 90.0 %, a specificity of 94.6 %, final results show that the model is reliable and accurate, the
and an accuracy of 92 %, as shown in Figs. 2 and 3. A SVM model will be used to predict SLN status in SLN-
double-blind test on the 85 follow-up cases revealed a positive breast cancer cases and to carry out a study aimed at
sensitivity of 97.7 %, a specificity of 85.7 %, and an accu- determining an alternative to ALND as the next step in the
racy of 91.8 %, as shown in Fig. 4. By leave-one-out cross- treatment of early breast cancer. SLNB, as a substitute for
validation, the sensitivity of the model was 84.4 %, the ALND, will be carried out on these patients on the premise
specificity was 91.0 %, and the accuracy was 87.6 %. that locoregional control and long-term survival rate is not
affected, thus reducing surgical complications, improving
the quality of life of the patients, shortening the duration
Discussion of the operation, decreasing treatment expenses and medical
costs, and protecting social productivity. This model there-
Currently, many ways can be used to classify patterns. fore has the potential to be of great value to theoretical
Traditional ways include the Bayesian method, distance research and in clinical applications, providing huge eco-
discrimination method, and Fisher discrimination method. nomic and social benefits.
SVM can successfully address regression problems (time
series analysis), pattern recognition (classification, discrim- Conflicts of interest None
inant analysis), and many other issues. The main advantage
of the SVM method is that it seeks to achieve the optimum
answer under the condition of existing information on lim-
ited samples, not just an optimum value when the number of References
samples tends toward infinity [21].
In recent years, four models have been used for assessing 1. Nakhlis F, Golshan M. Bevacizumab: where do we go from here in
the NSLN status in SLN-positive breast cancer, including breast cancer? Transl Cancer Res. 2012;1(1):55–6.
the BCN of MSKCC, Tenon model, Stanford model, and 2. Chen W, Zheng R, Zhang S, Zhao P, Li G, Wu L, et al. Report of
incidence and mortality in China cancer registries, 2009. Chin J
Cambridge model. The results show that the BCN of
Cancer Res. 2013;25(1):10–21.
MSKCC is the most reliable. Different models contain dif- 3. Zhang BN, Cao XC, Chen JY, Chen J, Fu L, Hu XC, et al.
ferent variables, most often in the aspects of the primary Guidelines on the diagnosis and treatment of breast cancer (2011
tumor and SLN, such as primary tumor size, grade, ER and edition). Gland Surg. 2012;1(1):39–61.
4. Rashid OM, Takabe K. The evolution of the role of surgery in the
CerbB2 expression, vascular invasion, the size of positive
management of breast cancer lung metastasis. J Thorac Dis. 2012;4
SLNs, the number of positive SLNs, positive SLN mem- (4):420–4.
brane extension, and positive SLN testing method (HE or 5. Veronesi U, Paganelli G, Viale G, Luini A, Zurrida S, Galimberti
IHC). The variables that have predictive values are primary V, et al. A randomized comparison of sentinel-node biopsy with
routine axillary dissection in breast cancer. N Engl J Med.
tumor size, the size and number of positive SLNs, and
2003;349(6):546–53.
vascular invasion. 6. Mansel RE, Fallowfield L, Kissin M, Goyal A, Newcombe RG,
In this project, using the SVM method, we isolated six Dixon JM, et al. Randomized multicenter trial of sentinel node
variables with statistical significance, including the number biopsy versus standard axillary treatment in operable breast cancer:
the ALMANAC Trial. J Natl Cancer Inst. 2006;98(9):599–609.
of SLNs, vascular invasion, the size of positive SLNs, the
7. Fan F. Sentinel lymph node micrometastases and isolated tumor cells
number of negative SLNs, positive SLN membrane inva- in breast cancer: an evolving field. Gland Surg. 2012;1(1):5–6.
sion, and tumor size. These six variables were combined 8. Purushotham AD, Upponi S, Klevesath MB, Bobrow L, Millar K,
arbitrarily, and a SVM training model was constructed for Myles JP, et al. Morbidity after sentinel lymph node biopsy in
primary breast cancer: results from a randomized controlled trial. J
cross-validation. The combination with the highest accuracy
Clin Oncol. 2005;23(19):4312–21.
according to the cross-validation test for modeling consisted 9. Bernet L, Cano R. Metastatic sentinel node and axillary lympha-
of the number of positive SLNs, the size of positive SLNs, denectomy revisited. Gland Surg. 2012;1(1):7–8.
the number of negative SLNs, and positive SLN membrane 10. McLaughlin SA, Wright MJ, Morris KT, Giron GL, Sampson MR,
Brockway JP, et al. Prevalence of lymphedema in women with
invasion. The training accuracy of the model constructed
breast cancer 5 years after sentinel lymph node biopsy or axillary
with the four variables was 92 %, the accuracy of cross- dissection: objective measurements. J Clin Oncol. 2008;26
validation was 87.6 %, and the accuracy of double-blind test (32):5213–9.
1552 Tumor Biol. (2013) 34:1547–1552

11. Nelson V, Rademaker A, Kaklamani V. Paradigm of polyendocrine axilla: a study to evaluate the need for complete axillary lymph
therapy in endocrine responsive breast cancer: the role of fulves- node dissection. Arch Surg. 2001;136(6):688–92.
trant. Chin Clin Oncol. 2013;2(1):10. 17. Van Zee KJ, Manasseh DM, Bevilacqua JL, Boolbol SK, Fey JV,
12. Ashikaga T, Krag DN, Land SR, Julian TB, Anderson SJ, Brown Tan LK, et al. A nomogram for predicting the likelihood of
AM, et al. Morbidity results from the NSABP B-32 trial comparing additional nodal metastases in breast cancer patients with a posi-
sentinel lymph node dissection versus axillary dissection. J Surg tive sentinel node biopsy. Ann Surg Oncol. 2003;10(10):1140–51.
Oncol. 2010;102(2):111–8. 18. Barranger E, Coutant C, Flahault A, Delpech Y, Darai E, Uzan S.
13. Shah-Khan M, Boughey JC. Evolution of axillary nodal staging in An axilla scoring system to predict non-sentinel lymph node status
breast cancer: clinical implications of the ACOSOG Z0011 Trial. in breast cancer patients with sentinel lymph node involvement.
Cancer Control. 2012;19(4):267–76. Breast Cancer Res Treat. 2005;91(2):113–9.
14. Giuliano AE, Haigh PI, Brennan MB, Hansen NM, Kelley MC, 19. Pal A, Provenzano E, Duffy SW, Pinder SE, Purushotham AD. A
Ye W, et al. Prospective observational study of sentinel lympha- model for predicting non-sentinel lymph node metastatic disease
denectomy without further axillary dissection in patients with when the sentinel lymph node is positive. Br J Surg. 2008;95
sentinel node-negative breast cancer. J Clin Oncol. 2000;18 (3):302–9.
(13):2553–9. 20. Kohrt HE, Olshen RA, Bermas HR, Goodson WH, Wood DJ, Henry
15. Turner RR, Ollila DW, Krasne DL, Giuliano AE. Histopathologic S, et al. Bay Area SLN Study. New models and online calculator for
validation of the sentinel lymph node hypothesis for breast carci- predicting non-sentinel lymph node status in sentinel lymph node
noma. Ann Surg. 1997;226(3):271–6. positive breast cancer patients. BMC Cancer. 2008;8:66.
16. Kamath VJ, Giuliano R, Dauway EL, Cantor A, Berman C, Ku 21. Sattlecker M, Bessant C, Smith J, Stone N. Investigation of sup-
NN, et al. Characteristics of the sentinel lymph node in breast port vector machines and Raman spectroscopy for lymph node
cancer predict further involvement of higher-echelon nodes in the diagnostics. Analyst. 2010;135(5):895–901.