Annals of Surgical Oncology Cite Track
HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
10.1245/ASO.2006.03.019
Annals of Surgical Oncology 13:1113-1122 (2006)
© 2006 Society of Surgical Oncology
This Article
Right arrow Abstract Freely available
Right arrow Full Text (PDF)
Right arrow Alert me when this article is cited
Right arrow Alert me if a correction is posted
Services
Right arrow Similar articles in this journal
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Mocellin, S.
Right arrow Articles by Rossi, C. R.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Mocellin, S.
Right arrow Articles by Rossi, C. R.

Original Article

Support Vector Machine Learning Model for the Prediction of Sentinel Node Status in Patients With Cutaneous Melanoma

Simone Mocellin, MD, PhD1,2, Alessandro Ambrosi, PhD1, Maria Cristina Montesco, MD3, Mirto Foletto, MD1, Giorgio Zavagno, MD1, Donato Nitti, MD1, Mario Lise, MD1 and Carlo Riccardo Rossi, MD1

1 Clinica Chirurgica 2, Dipartimento di Scienze Oncologiche e Chirurgiche, Universitàdi Padova, Via Giustiniani, 2, 35128 Padova, Italy
2 Istituto Oncologico Veneto, Padova, Italy
3 Pathology Section, Department of Oncological and Surgical Sciences, University of Padova, Padova, Italy

Correspondence: Address correspondence and reprint requests to: Simone Mocellin, MD, PhD; E-mail: mocellins{at}hotmail.com.


   ABSTRACT
TOP
 ABSTRACT
INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 REFERENCES
 
Background: Currently, approximately 80% of melanoma patients undergoing sentinel node biopsy (SNB) have negative sentinel lymph nodes (SLNs), and no prediction system is reliable enough to be implemented in the clinical setting to reduce the number of SNB procedures. In this study, the predictive power of support vector machine (SVM)-based statistical analysis was tested.

Methods: The clinical records of 246 patients who underwent SNB at our institution were used for this analysis. The following clinicopathologic variables were considered: the patient’s age and sex and the tumor’s histological subtype, Breslow thickness, Clark level, ulceration, mitotic index, lymphocyte infiltration, regression, angiolymphatic invasion, microsatellitosis, and growth phase. The results of SVM-based prediction of SLN status were compared with those achieved with logistic regression.

Results: The SLN positivity rate was 22% (52 of 234). When the accuracy was ≥80%, the negative predictive value, positive predictive value, specificity, and sensitivity were 98%, 54%, 94%, and 77% and 82%, 41%, 69%, and 93% by using SVM and logistic regression, respectively. Moreover, SVM and logistic regression were associated with a diagnostic error and an SNB percentage reduction of (1) 1% and 60% and (2) 15% and 73%, respectively.

Conclusions: The results from this pilot study suggest that SVM-based prediction of SLN status might be evaluated as a prognostic method to avoid the SNB procedure in 60% of patients currently eligible, with a very low error rate. If validated in larger series, this strategy would lead to obvious advantages in terms of both patient quality of life and costs for the health care system.

Key Words: Melanoma • Sentinel node biopsy • Support vector machine • Sentinel node status • Prognostic factors • Predictive factors


   INTRODUCTION
TOP
 ABSTRACT
 INTRODUCTION
MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 REFERENCES
 
Since the initial cornerstone experience published by Morton et al.1 in 1992, sentinel node biopsy (SNB) has become the most widely used procedure to determine regional lymph node status in patients with cutaneous melanoma, and its application is rapidly gaining the favor of surgeons who deal with other malignancies, particularly breast carcinoma.2

Presently, sentinel lymph node (SLN) status is the single most significant prognostic factor for patients with American Joint Committee on Cancer stage I and II melanoma.3 Moreover, by identifying patients with subclinical lymph node metastases, surgeons can submit patients to therapeutic lymphadenectomy earlier in their natural history while sparing patients with negative SLNs the morbidity associated with radical lymph node dissection. Finally, oncologists can administer potentially useful adjuvant treatments when the disease burden is still limited.4,5

Despite these considerations, the other side of the coin does exist and is multifaceted. A false-negative (FN) rate of 15% has been attributed to the SNB procedure (with standard pathologic evaluation of SLNs), although molecular detection of SLN micrometastasis by means of reverse transcriptase-polymerase chain reaction might significantly reduce this misclassification rate.6 Currently, 18% to 22% of patients undergoing SNB will have a positive result at histological analysis, whereas in the large majority of cases (approximately 80%) this surgical procedure does not change the therapeutic decision-making process. This fact should be evaluated in the light of other considerations.

First, no general agreement has been reached regarding the effect of SNB on patient survival, either as a direct effect of the early lymphadenectomy or as a consequence of the subsequent adjuvant treatment;7–12 accordingly, most investigators agree that SNB should be considered a diagnostic/prognostic (not therapeutic) procedure until the results from randomized trials designed to address this issue are available.13–15 Second, SNB, which generally requires a minimally invasive surgical operation, is associated with a small but definite risk of morbidity (5%–12%), which is mainly represented by hematoma, seroma, lymphatic leakage, wound infection, nerve damage (greatest in the case of head and neck melanomas), and allergic reaction to the dyes used.16–20 Third, SNB, which requires the intervention of three specialists (a surgeon, a nuclear medicine physician, and a pathologist) is an expensive procedure with an estimated average cost of US$8,000 to $14,000.21,22 Finally, a question has been recently raised on whether the SNB procedure might itself cause an increased occurrence of in-transit melanoma metastases.23,24

Overall, it can be stated that the availability of a prognostic tool for the identification of patients at higher or lower risk of harboring subclinical SLN metastasis would be of great help in the management of patients with cutaneous melanoma. Unfortunately, the high rate of FN prediction that characterizes the statistical models so far used makes impossible their clinical implementation for the selection of patients who are candidates for SNB.

Here we report our findings regarding the application of a statistical learning method, i.e., support vector machine (SVM), to analyze the clinicopathologic variables of 246 patients who underwent SLN biopsy for clinical stage I or II cutaneous melanoma. The primary aim of this study was to assess whether such a novel statistical approach might better predict SLN status as compared with logistic regression, a statistical method traditionally used for the analysis of data with a binary outcome (e.g., SLN positivity/negativity).


   MATERIALS AND METHODS
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
RESULTS
 DISCUSSION
 REFERENCES
 
Patient Demographics and Tumor Characteristics
From May 1999 through June 2004, 246 consecutive patients underwent SLN biopsy for clinical stage I or II cutaneous melanoma at our institution. Inclusion criteria were the following: (1) the patient had to sign a fully informed consent form, and (2) there had to be a pathologic diagnosis of cutaneous melanoma ≥1.0 mm in Breslow thickness or <1.0 mm with at least one adverse histopathologic feature (ulceration or significant regression) or Clark level IV–V. Patients were grouped by the presence or absence of metastatic melanoma in at least one SLN: therefore, cases of one or more positive SLNs were considered positive.

Sentinel Node Biopsy
The SNB procedure is described in detail elsewhere.25,26 Briefly, the day before SNB, lymphoscintigraphy was performed by using the LFOW gamma camera (Siemens, IL). A mean volume of .19 mL (range, .11–.28 mL) of 99mTc-labeled human albumin (Nanocoll Amersham Sorin, Milano, Italy) was intradermally injected into two perilesional sites. A final dose of 102.7 MBq (range, 96–155 MBq) was administered. Early and late images were obtained at 20 and 120 minutes, respectively. The skin site corresponding to the highest emission point was marked. On the day of surgery, 20 minutes before skin incision, .7 to 2.5 mL (mean, 1 mL) of Patent Blue dye (Patent V; Jacopo Monico, Mestre, Italy) was intradermally injected around the scar. A handheld Care Wise gamma camera (Morgan Hill, CA) was used to measure background and SLN radioactivity. After surgical skin incision, the SLN was identified as a blue-stained and/or hot node (defined as the hottest node and any other node with at least 5-fold the background radioactivity). The removed SLNs were sent to the pathologist in buffered formalin at 10%. Definitive wide excision of the primary cutaneous melanoma was performed, if required, after the SNB procedure.

Pathologic Evaluation
The primary tumor and SLN were examined by the same pathologist (M.C.M.) with expertise in melanoma. The following histological features of the primary tumor were considered: histological subtype (superficial spreading melanoma, nodular melanoma, acrolentiginous melanoma, or other) Breslow thickness (millimeters), Clark level (II, III, IV, or V), number of mitoses per square millimeter (≤5 or >5/ mm2), angiolymphatic invasion (present or absent), lymphocyte infiltration (absent, brisk, or nonbrisk), regression (present or absent), microsatellitosis (present or absent), ulceration (present or absent), and growth phase (vertical or radial).

Each formalin-fixed SLN was cut in half in correspondence of the hilum and then subdivided in fragments of 1- to 2-mm thick, including the entire node. All the fragments were embedded in paraffin to be then stained with hematoxylin and eosin and for immunohistochemistry; from each inclusion, 10 sections were prepared for histological and immunohis-tochemical (S-100 protein and HMB-45) examination.

Statistical Analysis
We assessed the ability of two models, logistic regression and SVM, to determine the association between the clinicopathologic independent variables and the SLN status (the binary dependent variable was positive or negative SLN). No variable selection was made (e.g., by a stepwise procedure), because our intent was to compare the prediction power of SVM and logistic regression by using in both cases the same available information. All statistical analyses and graphics were computed in the R environment (version 2.0.1).

Performance Assessing
The ability of both models (SVM and logistic regression) to correctly classify subjects was measured by means of leave-one-out cross-validation. In brief, we considered n samples of size n – 1, in which for each subject i, i = 1...n is held out. In each step, the model is trained on the basis of n – 1 subjects, and its capability of correctly predicting the classes is valued on the held-out subject i, who is counted as true positive (TP), true negative (TN), false positive (FP), or FN. On the basis of these predicted classifications, we estimated accuracy [(TP + TN)/(TP + TN + FP + FN)], negative predictive value [NPV; TN/(TN + FN)], positive predictive value [PPV; TP/ (TP + FP)], sensitivity [TN/(TN + FP)], and specificity [TP/(TP + FN)].

Logistic Model
Logistic models extend classic linear models to accommodate for the binary response variable.27 In this study, the event of interest was defined as the presence of one or more positive SLNs.

The linear combinations of factors derived in the logistic regression model generate a risk score that is converted into a probability of a positive SLN by using the equation


Formula

where Pis the probability of SLN positivity and RS indicates the risk score and is a linear combination of the chosen covariates. A decision rule is constructed to categorize the patient’s SLN as predicted to be positive if the probability is greater than a given cutoff and is predicted to be negative otherwise.

For each possible cutoff value, the sensitivity of the rule is the fraction of correct calls among the TPs, and the specificity is the fraction of correct calls among the TNs. Because the cutoff is varied, the sensitivity and specificity change and are plotted in the receiver operator characteristic curve, which is commonly adopted to evaluate diagnostic procedures. We considered the rule P > .5, which is well motivated and widely applied, although it is not the only possible choice: the cutoff can be moved upward or downward depending on the necessity and the specific problem.

A logistic regression model was first fit for each prognostic variable to test the correlation with the SLN status (univariate analysis). Then, a multivariate analysis was performed that considered all variables: this model was used both for describing the correlation of the variables with the SLN status and for predicting the SLN status. Different generalized logistic additive models were fit to take into account any possible nonlinear pattern of the numerical variables (age and tumor thickness) by means of penalized regression splines. Results were cross-validated by means of the leave-one-out method.

Support Vector Machine
The basic idea behind the SVM is to construct an optimal separating hyperplane between two classes (e.g., cases/controls and positive/negative).28–30 In the two-dimensional case (i.e., if we consider only two explanatory variables), a hyperplane is a straight line: therefore, we are searching for the straight line (or hyperplane in more dimensions) that optimally separates the two classes, where optimally means that the distances of the hyperplane to the closest points of either class are maximized (support vectors). Hence, only these data are necessary to build the classification rule, whereas the rest of the training cases do not contribute to the expression of the dividing hyper-plane (Fig. 1Go). This reveals a very important aspect of SVM, which is one of the reasons for its effectiveness as a classifier. The hyperplane is determined only by a relatively small number of cases that are close to the opposite group: the data points that are far away have no influence on the results, because it is clear as to which group they belong.


Figure 1
View larger version (8K):
[in this window]
[in a new window]
  		FIG. 1. Graphical representation of support vector machine (SVM) classification. A hyperplane (dashed line) is determined that optimally separates the two groups of samples. SVM classifiers are determined only by subjects who are close to the interface between the two populations (called support vectors) and not by other subjects who are further away. This allows fine-tuning of the classifier to the subtle details of the interface between the two groups.

 
Ultimately, the classifier inherently focuses on the subtleties of the morphological differences between the two groups and not on gross differences that are not difficult to detect, and it is therefore more effective. This is possible when we consider the linear case, i.e., when the dependent variables are separable by a linear hyperplane. In real situations, however, it is more likely that two groups overlap to some degree and are better separated by a nonlinear hyperplane (e.g., a polynomial of degree d). This difficulty is tackled by the SVM model in two ways: first, some points are allowed to appear on the "wrong" side of the hyperplane, thus introducing a parameter C, which represents the trade-off between minimizing the error and the maximizing the margin; second, nonlinear functions of the original set of variables are introduced into the model. To construct such a nonlinear hyperplane, the data are mapped into a higher-dimensional space by using Kernel functions K(•). Of note, this statistical approach is meant to classify cases into one of two categories, but it does not provide any estimate of membership probability (the likelihood that a case classified as positive or negative is really positive or negative, respectively) for each given case. This intrinsic property of SVM differentiates it from logistic regression, in which a cutoff must be chosen by the investigator to classify each calculated probability as a positive or negative case.

As regards the selection of the optimal SVM model, we considered SVM with a radial basis Kernel function:


Formula

Hence, given the training sample, the model depends only on the two parameters C and {gamma}, the latter being the dispersion parameter. In the space of these parameters, we searched for the model that maximizes the NPV; this best fit the purpose of our research, as discussed previously). One constraint (accuracy ≥80%) was introduced to avoid abnormal model fitting and to preserve clinical utility.

Adopting a standard leave-one-out training/testing scheme, the SVM was trained separately on training sets made up of all cases but one and then tested on the single held-out case. During the training phase, the SVM takes as input all independent variables from n – 1 subjects and labels each case as belonging to either either one class (positive SLN) or the other (negative SLN). Each case is treated as a point in a high-dimensional feature space, where the number of independent variables (predictors) determines the dimensionality of the space. Then, the SVM learning model identifies a hyperplane in this space that best separates the positive and negative training examples. During the testing phase, the trained SVM is used to make predictions about a test sample’s membership in the class. The leave-one-out strategy allowed us to collect unbiased measurements of the ability of the SVM to classify all cases (n = 234).


   RESULTS
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
DISCUSSION
 REFERENCES
 
Sentinel Node Biopsy
Of 246 consecutive patients identified in our prospective database, 234 patients had an identified and evaluated SLN and all available data, and they form the basis of this article. There were 129 men and 105 women, and the median age was 53 years (range, 16–82 years). Primary melanoma characteristics are listed in Table 1Go. Thirty-nine patients (17%) underwent simultaneous SLN biopsies from two or three different nodal basins (primary lesion location with more than one primary nodal basin, such as the trunk), and these cases were considered positive if at least one biopsy in one or more basins identified a positive SLN. The mean number of identified SLNs per patient was 2.0 (range, 1–6), and 1 SLN was removed in 105 patients (45%).


View this table:
[in this window]
[in a new window]
  		TABLE 1. Primary melanoma characteristics
 
One or more histologically or immunohistochemically positive lymph nodes were found in 52 of 234 patients (SLN positivity rate, 22%). SNB-related morbidity occurred in 24 patients (morbidity rate, 10%) and included lymphatic leakage (n = 16), seroma (n = 12), hematoma (n = 2), and wound infection (n = 3).

Logistic Regression
At univariate analysis, Breslow thickness, Clark level, ulceration, mitotic index, lymphocyte infiltration, vascular invasion, and microsatellitosis significantly (P < .01) correlated with SLN status (Table 2Go), whereas patient age and sex, tumor growth phase, and histological regression did not have any significant effect on the SNB outcome. At multivariate analysis, only Breslow thickness, mitotic index, and microsatellitosis were independently associated with SLN status (Table 2Go).


View this table:
[in this window]
[in a new window]
  		TABLE 2. Logistic regression: univariate and multivariate analysis of prognostic factors in 234 patients who underwent sentinel node biopsy for cutaneous melanoma
 
The accuracy of the models built with multivariate logistic regression to predict SLN status was represented by using a receiver operating characteristic curve (Fig. 2Go). When the accuracy was 80%, the best NPV-oriented results of SLN status prediction were the following: sensitivity, 69%; specificity, 93%; PPV, 41 %; and NPV, 82% (Table 3Go).


Figure 2
View larger version (5K):
[in this window]
[in a new window]
  		FIG. 2. Logistic regression–based prediction model: negative predictive value (NPV) and accuracy varying with the cutoff of 0 < C < 1. For comparison purposes, the best combination of NPV and accuracy obtained with the support vector machine is indicated by an asterisk.

 

View this table:
[in this window]
[in a new window]
  		TABLE 3. Prediction of sentinel lymph node status in 234 patients with cutaneous melanoma who underwent sentinel node biopsy (SNB): results from multivariate logistic regression and support vector machine (SVM)-based analysis
 
We then calculated the diagnostic error rate (percentage of patients who would not be submitted to SNB even though they had SLN metastasis) and the SNB percentage reduction (the percentage of patients who would not be submitted to SNB and who had no SLN metastasis) of the test: these were 15% and 73%, respectively (Table 3Go). This means that out of 100 patients who were candidates for SNB, the logistic regression–based prediction system would spare 73 patients with negative SLNs the surgical procedure and would lead to an erroneous decision (not to undergo SNB) in 15 cases with positive SLNs (Fig. 3Go). The generalized regression additive model neither yielded significantly better results (accuracy, 79%; NPV, 81%) nor provided any evidence of a nonlinear influence of numerical variables in predicting the SLN status (age, P = .845; tumor thickness, P = .682).


Figure 3
View larger version (28K):
[in this window]
[in a new window]
  		FIG. 3. The graphics represent the diagnostic error rate (proportion of patients who would not be submitted to sentinel node biopsy [SNB] although they had sentinel lymph node [SLN] metastasis) and the SNB percentage reduction (proportion of patients who would not undergo SNB and had no SLN metastasis) for logistic regression–based (top) and the support vector machine–based (bottom) prediction models.

 
Support Vector Machine
The NPV of the models built with SVM as a function of the parameters C and {gamma} to predict SLN status is illustrated in Fig. 4Go. When the accuracy was 81%, the best NPV-oriented results of SLN status prediction by SVM were the following: sensitivity, 94%; specificity, 77%; PPV, 54%; and NPV, 98% (Table 3Go). This model was also characterized by a diagnostic error rate and an SNB percentage reduction of 1% and 60%, respectively (Table 3Go). This implies that of 100 patients who were candidates for SNB, the SVM-based prediction system would spare 60 patients with negative SLNs the surgical procedure and would lead to an erroneous decision (not to undergo SNB) in 1 case with a positive SLN (Fig. 3Go).


Figure 4
View larger version (20K):
[in this window]
[in a new window]
  		FIG. 4. Support vector machine–based prediction of sentinel lymph node (SLN) status in patients (n = 234) with cutaneous melanoma. To identify patients who might avoid sentinel node biopsy (SNB), 11 clinicopathologic variables were used to identify subjects with a high probability of having a negative SLN. For such a prediction model to be implemented in the clinical setting, the negative predictive value (NPV) must be close to 100%, because any value <100% would mean that 100% – NPV% of patients would not undergo the SNB procedure although they had a positive SLN. Considering the z-axis (NPV), the maximum (arrow) identifies the combination of parameters C and {gamma}of the optimal model for our data set (NPV = 98%).

 

   DISCUSSION
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
REFERENCES
 
The SNB is one of the most significant advances in melanoma management over the past decade and allows for selective application of radical lymph node dissection only to node-positive patients. As such, it maximizes the likelihood that a survival benefit will be realized with this approach. However, approximately 80% of patients currently undergoing SNB have no metastasis in their SLN, and this implies that the selection criteria currently adopted are largely inefficient. This is even more relevant considering that some issues (morbidity rates and costs) and controversies (effect on patient survival) regarding SNB do exist. Therefore, any prediction system reliable enough to spare SLN-negative subjects this surgical procedure would be of great clinical benefit, both in terms of patients quality of life and costs for the health care system.

Numerous studies have reported clinicopathologic factors that correlate with the SLN status.11,31–40 The single factor that uniformly correlates with SLN status is the Breslow thickness of the primary lesion: as the Breslow depth increases, the probability of a positive SLN increases. Several reports have studied other primary melanoma pathologic features associated with SLN status. Microsatellitosis, vascular invasion, higher mitotic rate, vertical growth phase, age <60 years, Clark level greater than III, and ulceration have been reported to be significantly associated with a positive SLN, although with con-flicting results. Most studies focus on the correlation between single prognostic factors and SLN status: a {chi}2 statistic is generally used to show differences between frequencies of SLN positivity in different subgroups, and sometimes a PPV is calculated. Unfortunately, none of these parameters can predict the SLN status with a statistical reliability of clinical value, either when they are used singularly or when a probabilistic model based on multivariate logistic regression is fitted to the data.41

SVM is a supervised statistical method that has received much attention in the medical field since the broad diffusion of DNA microarray technology42,43: however, to our knowledge, this is the first time that SVM has been applied for the analysis of clinicopathologic predictors of SLN status. This novel statistical approach is based on complex mathematical models aimed at recognizing patterns of behavior of the relationships between independent variables (i.e., clinicopathologic factors) and a binary dependent variable (i.e., SLN status).

In our series, which closely mirrors the experience published worldwide both in terms of patient/tumor characteristics and SNB outcome, the results of the univariate logistic regression analysis showed that Breslow thickness, mitotic rate, and microsatellitosis were correlated with the likelihood of SLN positivity. Then, to additionally exploit the prognostic potential of all available clinicopathologic variables, we applied multivariate logistic regression and SVM to build models predictive of SLN status.

Both logistic regression and SVM-based prediction analysis showed that, using the currently available predictors, the best PPVs (54% and 59%, respectively; data not shown) were too low to be of clinical value, because their prediction power is similar to that of a coin toss. These findings strengthen the idea that no current prediction system can reliably identify patients at higher risk of SLN positivity so that only these subjects undergo SNB. However, our results suggested another way to tackle this problem: the identification of patients with a high probability of harboring no metastasis in their SLNs. To this aim, an NPV (the probability that a patient who tests negative has no metastasis in the SLN) close to 100% is required for a test to be safely implemented in clinical practice, because any value <100% means that 100% – NPV% (diagnostic error; Table 3Go) of patients who test negative would not undergo SNB even though they had a metastatic SLN (FN cases). Consequently, we chose the logistic regression and SVM models that maximized the NPV (NPV-oriented model). Moreover, to be of practical usefulness, the ideal prognostic test should significantly reduce the number of SNBs presently performed: thus, from the statistical viewpoint, such a test should be characterized by a high sensitivity (the probability that a patient without SLN metastasis tests negative), which indicates how many SNBs would be correctly spared out of 100 procedures performed by using the current eligibility criteria (SNB percentage reduction; Table 3Go).

In the light of these considerations, the results were still unsatisfactory with logistic regression: in fact, the NPV (82%) was unacceptably low, because it would lead to a diagnostic error rate (15%) unsuitable for clinical purposes. By contrast, the data obtained with SVM were definitely more encouraging, because the NPV (98%), sensitivity (94%), diagnostic error rate (1%), and SNB percentage reduction (60%) seem favorable enough to deserve further investigation. In fact, if these data were confirmed in larger series, SVM-based prediction might drastically change the current management of patients who are candidates for SNB because a significant proportion (60%) of patients who currently undergo SNB and have a high probability of having a negative SLN would correctly avoid the psychological/physical discomfort as well as the potential complications of an operation; consequently, the SNB costs presently supported by the health care system would significantly decrease.

As regards the issue of the low but definite risk (1%) of incorrectly submitting patients to observation (instead of SNB) with the described SVM model, two considerations can be made. First, it has been repeatedly reported that ultrasound scanning combined with fine-needle aspiration is a reliable method for the early detection of lymph node metastasis, at least when disease deposits are >2 mm but still clinically silent;25,26,44–47 in addition, recent technological advances (e.g., proton magnetic resonance spectroscopy of fine-needle aspirate material) might further improve the reliability of such a strategy.48 Second, a potential development of the SVM-based prediction of SLN status is represented by the addition of molecular markers of tumor aggressiveness (e.g., expression of metastasis-related genes/proteins by primary melanoma) to the set of independent variables. This might significantly improve the performance of this test, thus further increasing the reliability of its clinical implementation. To validate this hypothesis, a large multicentric prospective study is under way on behalf of the Italian Melanoma Intergroup.


   ACKNOWLEDGMENTS
 
Supported by a grant from the Regione Veneto.

Received for publication March 3, 2005. Accepted for publication January 12, 2006.


   REFERENCES
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 REFERENCES
 

  1. Morton DL, Wen DR, Wong JH, et al. Technical details of intraoperative lymphatic mapping for early stage melanoma. Arch Surg 1992; 127:392–9.[Abstract]
  2. Leong SP. Paradigm of metastasis for melanoma and breast cancer based on the sentinel lymph node experience. Ann Surg Oncol 2004; 11:192S–7S.[Abstract/Free Full Text]
  3. Tsao H, Atkins MB, Sober AJ. Management of cutaneous melanoma. N Engl J Med 2004; 351:998–1012.[Free Full Text]
  4. Morton DL. Lymphatic mapping and sentinel lymphadenectomy for melanoma: past, present, and future. Ann Surg Oncol 2001; 8:22S–28S.[Medline]
  5. McMasters KM, Reintgen DS, Ross MI, et al. Sentinel lymph node biopsy for melanoma: controversy despite widespread agreement. J Clin Oncol 2001; 19:2851–5.[Abstract/Free Full Text]
  6. Takeuchi H, Morton DL, Kuo C, et al. Prognostic significance of molecular upstaging of paraffin-embedded sentinel lymph nodes in melanoma patients. J Clin Oncol 2004; 22:2671–80.[Abstract/Free Full Text]
  7. Mohrle M, Schippert W, Rassner G, et al. Is sentinel lymph node biopsy of therapeutic relevance for melanoma? Dermatology 2004; 209:5–13.[CrossRef][Medline]
  8. Fife K, Thompson JF. Lymph-node metastases in patients with melanoma: what is the optimum management?. Lancet Oncol 2001; 2:614–21.[CrossRef][Medline]
  9. Caraco C, Celentano E, Lastoria S, et al. Sentinel lymph node biopsy does not change melanoma-specific survival among patients with Breslow thickness greater than four millimeters. Ann Surg Oncol 2004; 11:198–202S.
  10. Roka F, Kittler H, Cauzig P, et al. Sentinel node status in melanoma patients is not predictive for overall survival upon multivariate analysis. Br J Cancer 2005.
  11. Vuylsteke RJ, van Leeuwen PA, Statius Muller MG, et al. Clinical outcome of stage I/II melanoma patients after selective sentinel lymph node dissection: long-term follow-up results. J Clin Oncol 2003; 21:1057–65.[Abstract/Free Full Text]
  12. Shen J, Wallace AM, Bouvet M. The role of sentinel lymph node biopsy for melanoma. Semin Oncol 2002; 29:341–52.[CrossRef][Medline]
  13. McMasters KM. What good is sentinel lymph node biopsy for melanoma if it does not improve survival? Ann Surg Oncol 2004; 11:810–2.[Free Full Text]
  14. Doting MH, Hoekstra HJ, Plukker JT, et al. Is sentinel node biopsy beneficial in melanoma patients? A report on 200 patients with cutaneous melanoma. Eur J Surg Oncol 2002; 28:673–8.[CrossRef][Medline]
  15. Thompson JF, Stretch JR, Uren RF, et al. Sentinel node biopsy for melanoma: where have we been and where are we going?. Ann Surg Oncol 2004; 11:147S–51S.[Abstract/Free Full Text]
  16. Wrightson WR, Wong SL, Edwards MJ, et al. Complications associated with sentinel lymph node biopsy for melanoma. Ann Surg Oncol 2003; 10:676–80.[Abstract/Free Full Text]
  17. Hettiaratchy SP, Kang N, O’Toole G, et al. Sentinel lymph node biopsy in malignant melanoma: a series of 100 consecutive patients. Br J Plast Surg 2000; 53:559–62.[CrossRef][Medline]
  18. Cimmino VM, Brown AC, Szocik JF, et al. Allergic reactions to isosulfan blue during sentinel node biopsy—a common event. Surgery 2001; 130:439–42.[CrossRef][Medline]
  19. Roaten JB, Pearlman N, Gonzalez R, et al. Identifying risk factors for complications following sentinel lymph node biopsy for melanoma. Arch Surg 2005; 140:85–9.[Abstract/Free Full Text]
  20. Wasserberg N, Tulchinsky H, Schachter J, et al. Sentinel-lymph-node biopsy (SLNB) for melanoma is not complication-free. Eur J Surg Oncol 2004; 30:851–6.[CrossRef][Medline]
  21. Brobeil A, Cruse CW, Messina JL, et al. Cost analysis of sentinel lymph node biopsy as an alternative to elective lymph node dissection in patients with malignant melanoma. Surg Oncol Clin North Am 1999; 8:435–45viii.[Medline]
  22. Agnese DM, Abdessalam SF, Burak WE Jr, et al. Cost-effectiveness of sentinel lymph node biopsy in thin melanomas. Surgery 2003; 134:542–7discussion 547–8.[CrossRef][Medline]
  23. Estourgie SH, Nieweg OE, Kroon BB. High incidence of intransit metastases after sentinel node biopsy in patients with melanoma. Br J Surg 2004; 91:1370–1.[CrossRef][Medline]
  24. Thomas JM, Clark MA. Selective lymphadenectomy in sentinel node-positive patients may increase the risk of local/intransit recurrence in malignant melanoma. Eur J Surg Oncol 2004; 30:686–91.[CrossRef][Medline]
  25. Rossi CR, Mocellin S, Scagnet B, et al. The role of preoperative ultrasound scan in detecting lymph node metastasis before sentinel node biopsy in melanoma patients. J Surg Oncol 2003; 83:80–4.[CrossRef][Medline]
  26. Rossi CR, Scagnet B, Vecchiato A, et al. Sentinel node biopsy and ultrasound scanning in cutaneous melanoma: clinical and technical considerations. Eur J Cancer 2000; 36:895–900.[CrossRef][Medline]
  27. Hastie T, Tibshirani R. Generalized Additive Models. Chapman & Hall, 1990.
  28. Vapnik V. Statistical Learning Theory. Wiley, 1998.
  29. Hastie T, Tibshirani R, Friedman J. The Elements of Statistical Learning. Springer, 2001.
  30. Byvatov E, Schneider G. Support vector machine applications in bioinformatics. Appl Bioinformatics 2003; 2:67–77.[Medline]
  31. Wagner JD, Gordon MS, Chuang TY, et al. Predicting sentinel and residual lymph node basin disease after sentinel lymph node biopsy for melanoma. Cancer 2000; 89:453–62.[CrossRef][Medline]
  32. Gershenwald JE, Thompson W, Mansfield PF, et al. Multiinstitutional melanoma lymphatic mapping experience: the prognostic value of sentinel lymph node status in 612 stage I or II melanoma patients. J Clin Oncol 1999; 17:976–83.[Abstract/Free Full Text]
  33. Cascinelli N, Belli F, Santinami M, et al. Sentinel lymph node biopsy in cutaneous melanoma: the WHO Melanoma Program experience. Ann Surg Oncol 2000; 7:469–74.[Abstract]
  34. Essner R, Conforti A, Kelley MC, et al. Efficacy of lymphatic mapping, sentinel lymphadenectomy, and selective complete lymph node dissection as a therapeutic procedure for early-stage melanoma. Ann Surg Oncol 1999; 6:442–9.[Abstract]
  35. Chao C, Martin RC II, Ross MI, et al. Correlation between prognostic factors and increasing age in melanoma. Ann Surg Oncol 2004; 11:259–64.[Abstract/Free Full Text]
  36. McMasters KM, Wong SL, Edwards MJ, et al. Factors that predict the presence of sentinel lymph node metastasis in patients with melanoma. Surgery 2001; 130:151–6.[CrossRef][Medline]
  37. Nguyen CL, McClay EF, Cole DJ, et al. Melanoma thickness and histology predict sentinel lymph node status. Am J Surg 2001; 181:8–11.[CrossRef][Medline]
  38. Mraz-Gernhard S, Sagebiel RW, Kashani-Sabet M, et al. Prediction of sentinel lymph node micrometastasis by histological features in primary cutaneous malignant melanoma. Arch Dermatol 1998; 134:983–7.[Abstract/Free Full Text]
  39. Cuellar FA, Vilalta A, Rull R, et al. Small cell melanoma and ulceration as predictors of positive sentinel lymph node in malignant melanoma patients. Melanoma Res 2004; 14:277–82.[CrossRef][Medline]
  40. Thompson JF. The Sydney Melanoma Unit experience of sentinel lymphadenectomy for melanoma. Ann Surg Oncol 2001; 8:44S–47S.[Medline]
  41. Sondak VK, Taylor JM, Sabel MS, et al. Mitotic rate and younger age are predictors of sentinel lymph node positivity: lessons learned from the generation of a probabilistic model. Ann Surg Oncol 2004; 11:247–58.[Abstract/Free Full Text]
  42. Furey TS, Cristianini N, Duffy N, et al. Support vector machine classification and validation of cancer tissue samples using microarray expression data. Bioinformatics 2000; 16:906–14.[Abstract/Free Full Text]
  43. Mocellin S, Provenzano M, Rossi CR, et al. DNA array-based gene profiling: from surgical specimen to the molecular portrait of cancer. Ann Surg 2005; 241:16–26.[Medline]
  44. Machet L, Nemeth-Normand F, Giraudeau B, et al. Is ultrasound lymph node examination superior to clinical examination in melanoma follow-up? A monocentre cohort study of 373 patients. Br J Dermatol 2005; 152:66–70.[CrossRef][Medline]
  45. Hocevar M, Bracko M, Pogacnik A, et al. The role of preoperative ultrasonography in reducing the number of sentinel lymph node procedures in melanoma. Melanoma Res 2004; 14:533–6.[CrossRef][Medline]
  46. Bafounta ML, Beauchet A, Chagnon S, et al. Ultrasonography or palpation for detection of melanoma nodal invasion: a meta-analysis. Lancet Oncol 2004; 5:673–80.[CrossRef][Medline]
  47. Schmid-Wendtner MH, Dill-Muller D, Baumert J, et al. Lymph node metastases in patients with cutaneous melanoma: improvements in diagnosis by signal-enhanced color Doppler sonography. Melanoma Res 2004; 14:269–76.[CrossRef][Medline]
  48. Lean CL, Bourne R, Thompson JF, et al. Rapid detection of metastatic melanoma in lymph nodes using proton magnetic resonance spectroscopy of fine needle aspiration biopsy specimens. Melanoma Res 2003; 13:259–61.[CrossRef][Medline]



This article has been cited by other articles:


Home page
 Ann. Surg. Oncol.Home page
C. R. Rossi, G. L. De Salvo, E. Bonandini, S. Mocellin, M. Foletto, S. Pasquali, P. Pilati, M. Lise, D. Nitti, E. Rizzo, et al.
Factors Predictive of Nonsentinel Lymph Node Involvement and Clinical Outcome in Melanoma Patients With Metastatic Sentinel Lymph Node
Ann. Surg. Oncol., April 1, 2008; 15(4): 1202 - 1210.
[Abstract] [Full Text] [PDF]

This Article
Right arrow Abstract Freely available
Right arrow Full Text (PDF)
Right arrow Alert me when this article is cited
Right arrow Alert me if a correction is posted
Services
Right arrow Similar articles in this journal
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Mocellin, S.
Right arrow Articles by Rossi, C. R.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Mocellin, S.
Right arrow Articles by Rossi, C. R.

HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS