Evaluation of 28 Biomarkers to Diagnose Endometriosis
Evaluation of 28 Biomarkers to Diagnose Endometriosis
The present study is an important step in the development and validation of a blood test for endometriosis with a high sensitivity and acceptable specificity in a group of patients with subfertility and/or pain with a negative preoperative US. Multivariate analysis based on two models of four biomarkers (annexin V, VEGF, CA-125, glycodelin; annexin V, VEGF, CA-125 and sICAM-1) during the menstrual phase enabled the diagnosis of endometriosis undetectable by US with a high sensitivity (81–90%) at acceptable specificity (63–81%) in an independent training and test data set. This diagnostic performance was better than the diagnostic performance of any single biomarker in our study.
The novelty of our study is based on the design based on QUADAS guidelines (Whiting et al., 2003; May et al., 2010), the large and well-defined patient population (n = 353), the large number of evaluated plasma biomarkers (n = 28), the advanced multivariate statistical approach to select biomarkers, to classify them by the multivariate logistic regression and LS-SVM, and to validate the models developed in the training set in an independent test set, and the subanalysis of samples from women with negative preoperative US.
The strength of our study is that our design is in accordance with the QUADAS guidelines (Whiting et al., 2003; May et al., 2010) with respect to control group selection and cycle phase correction. In line with these guidelines, sample collection had been performed at a consistent phase of the cycle and results presented were corrected for cycle phases, as recommended (Whiting et al., 2003; May et al., 2010). For instance, three out of nine papers investigating IL-6 failed to adjust for the phase of the menstrual cycle, despite evidence that levels are known to change throughout the cycle (Angstwurm et al., 1997; May et al., 2010).
The choice of an appropriate control group is crucial and depends on the aim of the diagnostic test. In line with the QUADAS guidelines we selected our controls from women with symptoms consistent with endometriosis (such as infertility and/or pelvic pain) but without laparoscopic evidence of endometriosis based on laparoscopic data, obtained by an experienced endometriosis surgeon. Our controls have different pelvic pathology such as non-endometriosis adhesions, myoma, parasalpingeal cyst and hydrosalpinx (Table II).
Interestingly, inflammatory biomarkers such as IL-1beta, IFN-gamma, TNF-alpha and IL-6 were slightly higher in the control group than in the endometriosis group in the training data set. These data were not consistent and not confirmed in the test data set (Table III and Table IV), which included a comparable (P = 0.15) proportion of patients with non-endometriotic adhesions (9/41, 22%) as the training data set (28/80, 35%; Table II). Nevertheless, this observation could be partially explained by the possibility that controls with non-endometriotic pelvic pathology like adhesions and hydrosalpinx had increased plasma concentrations of inflammatory cytokines. As inflammatory cytokines were not included in the final diagnostic model (Table VI), we speculate that they are not relevant in the discrimination of patients with endometriosis from women with non-endometriotic pelvic pathology due to the similar inflammatory pathways which cause endometriotic and non-endometriotic pelvic pathology and results in pelvic pain and infertility. Moreover, at present, there is no consensus regarding the value of inflammatory factors as biomarkers of endometriosis. Comparable serum IL-6 (Kalu et al., 2007; Socolov et al., 2011), TNF-alpha and IL-1 (Kalu et al., 2007; Othman et al., 2008; Socolov et al., 2011) levels were previously reported in women with and without endometriosis. However, other investigators reported elevated peripheral levels of IL-6 (Bedaiwy et al., 2002; Othman et al., 2008), TNF-alpha (Bedaiwy et al., 2002; Xavier et al., 2006); IFN-gamma (Othman et al., 2008) in endometriosis patients compared with controls. These discrepancies could also be partially explained by the differences in study design (different inclusion criteria and different cycle phases), preanalytical variability in the cytokines levels (due to different types of collected samples (serum versus plasma); different clotting times, different conditions of centrifugation) which could influence the study results. For example, measurable concentrations of inflammatory markers are higher in the serum than in simultaneously collected plasma, due to the release of inflammatory markers during the coagulation process in the serum (Skogstrand et al., 2008). Moreover, biological variability due to functional single nucleotide polymorphisms known to influence protein levels of corresponding circulated proteins can also partially explain the discrepancy in study results, since genetic variants in IL-6 and IFN-gamma genes may influence circulating levels of corresponding proteins (Talar-Wojnarowska et al., 2009; Vallinoto et al., 2010).
The data of our study confirmed the hypothesis (Robin et al., 2009; May et al., 2010) that a panel of biomarkers can improve the sensitivity and specificity of diagnostic test compared with the diagnostic performance of any single biomarker. Our panel of four biomarkers (annexin V, VEGF, CA-125, sICAM-1 or glycodelin) had a better diagnostic performance than any single biomarker in our study. So far, only a limited number of studies have focused on the prediction of endometriosis based on a panel of markers, and these studies were limited by univariate analysis (Somigliana et al., 2004; Agic et al., 2008). Our multivariate statistical approach allowed us to model the relationship between diagnostic categories and all biomarkers simultaneously, taking into account the correlation that may exist between those biomarkers, while univariate analysis only deals with the relationship between one predictor and diagnostic category.
In the present study the two selected panels of four biomarkers (annexin V, VEGF, CA-125, sICAM-1/glycodelin) performed robustly by using two fundamentally different classifiers (multivariate logistic regression and LS-SVM) and could predict US-negative endometriosis with a sensitivity of 81–90% and a specificity of 63–81% (Table VI). Indeed, both methods (multivariate logistic regression and LS-SVM) are widely used in biomarker studies (Robin et al., 2009) and none of these methods is clearly superior when compared with the other (Robin et al., 2009). However, the multivariate logistic regression is more sensitive to feature selection (Pochet and Suykens, 2006; Mihalyi et al., 2010) since it tends to build a classification model that fits patients from a training set optimally (Pochet and Suykens, 2006), but is not always possible to make good predictions for novel patients from an independent test set, a problem defined as overfitting (Pochet and Suykens, 2006). In contrast, the LS-SVM is less sensitive to feature selection and effect of outliers, preventing the model from overfitting the training data (Pochet and Suykens, 2006; Mihalyi et al., 2010). In addition, it has an internal mechanism for modeling non-linearity, giving rise to increased robustness and therefore good performance in an independent test data set (Pochet and Suykens, 2006). Our data show that it is possible to overcome the problem of data overfitting by using the biomarkers selection method, as described in the methodology section. Indeed, the diagnostic performance of the models based on the training set was confirmed on the independent test set by using both classifiers (multivariate logistic regression and LS-SVM).
The relevance of the selected diagnostic panel (annexin V, VEGF, CA-125 and sICAM-1/glycodelin) is confirmed by the fact that the selected biomarkers are involved in apoptosis, angiogenesis, adhesion and tumorogenesis, which are highly related to the pathogenesis of endometriosis.
Annexin V, a marker of apoptosis, has been recently reported by our group to be a promising semi-invasive biomarker for diagnosis of minimal–mild endometriosis (Kyama et al., 2011). Indeed, alterations in the regulation of apoptosis in eutopic and ectopic endometrium from women with endometriosis could contribute to the survival of endometrial cells into the peritoneal cavity and development of endometriosis (reviewed by Taniguchi et al., 2011).
Glycodelin is an endometrium-derived protein with known angiogenic, immunosuppressive and contraceptive effects, which could contribute to the development of endometriosis and endometriosis-related infertility (reviewed by Seppälä et al., 2009). VEGF is one of the main stimuli for angiogenesis and increased vessel permeability, which contributes to the development of endometriotic lesions (Taylor et al., 2002, Becker and D'Amato, 2007). sICAM-1 is one of the major adhesion molecules which inhibits natural killer cell-mediated cytotoxicity (Becker et al., 1991), resulting in defective immune surveillance and is involved in the implantation and development of endometriotic lesions (Wu and Ho, 2003).
CA-125 is the most extensively investigated and used peripheral biomarker of endometriosis (Gupta et al., 2006). CA-125 is produced by endometrial and mesothelial cells and exudes into circulation via the endothelial lining of capillaries in response to inflammation (Bischof, 1993; Zeillemaker et al., 1994; reviewed by Gupta et al., 2006). However, CA-125 levels in the peripheral blood lack diagnostic power as a single biomarker of endometriosis (Mol et al., 1998; Kennedy et al., 2005).
We realize that a diagnostic test may do more harm than good, e.g. by subjecting patients to unnecessary or even potentially harmful procedures (Evers and Van Steirteghem, 2009) since the benefits of treating women with asymptomatic endometriosis is unclear (May et al., 2010). Therefore, we do not recommend to develop or use a blood test for screening purpose in asymptomatic women. However, up to 45% of subfertile women with a regular cycle whose partner has normal sperm quality, with or without pelvic pain, and with normal clinical examination and a normal pelvic US may have endometriosis (Meuleman et al., 2009). A blood test could identify those most likely to have endometriosis or other pelvic conditions and likely to benefit from surgical therapy for both subfertility and pain (Kennedy et al., 2005; D'Hooghe et al., 2006). In our study, the biomarker panels allowed to rule in these women with a high sensitivity (81–90%) and acceptable specificity (63–81%) and distinguish them from women without endometriosis who had symptoms similar to those with endometriosis (subfertility and/or pain), which is in line with published recommendations (May et al., 2010).
In our future work we plan to develop a computer application based on the selected best predictive models (Table VI) that would be freely available for clinical use. The two models, based on the incorporation of measured plasma levels of selected biomarkers (model 1: annexin V, VEGF, CA-125, glycodelin; model 2: annexin V, VEGF, CA-125, sICAM-1) during menstrual cycle phase will be used to identify high-risk groups of patients with a predicted probability of developing endometriosis based on the developed threshold. If the value of predicted probability is greater than the given threshold we conclude that the endometriosis status is positive while in the reverse case we conclude a negative status. Further prospective study is required to validate these models in clinical setting.
Our study is marked by the following limitations.
Firstly, the best diagnostic model was based on the analysis of plasma samples obtained during the menstrual phase. This is not surprising, since it is well known that plasma CA-125 levels in women (O'Shaughnessy et al., 1993) and baboons (Falconer et al., 2005) with endometriosis are higher during the menstrual phase than during other phases of the cycle. In practice, this could be a limitation since the blood sampling has to be limited to the menstrual cycle phase only.
Secondly, stress factors directly before surgery might have affected plasma biomarker levels, as blood was taken just prior to anaesthesia (as previously described by Mihalyi et al., 2010). More research is needed to validate our diagnostic models in plasma samples obtained in an outpatient clinic. However, in our study, the priority was to ensure that the blood sample was taken at the time of surgery in order to have a direct temporal comparison between laparoscopic diagnosis and staging of endometriosis disease and the plasma levels of the biomarkers studied.
Thirdly, we did not evaluate possible diurnal variability in biomarkers levels, as previously observed for serum IL-6 levels (Arvidson et al., 1994). From a practical perspective, we were looking for a robust biomarker panel not depending on diurnal variability, as suggested by the validation of our model in an independent test data set.
Fourthly, the selection of control group was based only on the laparoscopically exclusion of endometriosis by an experienced endometriosis surgeon without histological evaluation, which could be a limitation, especially for the patients with non-endometriotic adhesions. Indeed, it is difficult to rule out that women with a normal pelvis or with non-endometriotic adhesions may have microscopic endometriosis, and that laparoscopic absence of endometriosis may be a temporary phenomenon. However, since intraperitoneal adhesions are accepted as aetiologic factors for infertility (Hammoud et al., 2004), inclusion of patients with non-endometriosis adhesions based on laparoscopy data in the control group can be justified in a biomarker study for endometriosis.
In conclusion, multivariate analysis of four biomarkers (annexin V, VEGF, CA-125 and sICAM-1/or glycodelin) in plasma samples obtained during menstruation enabled the diagnosis of endometriosis undetectable by US with a sensitivity of 81–90% and a specificity of 63–81% in the independent training- and test data set. The next research step is to predict the presence of endometriosis with a high sensitivity, using the models presented in this study, in an independent set of patients with infertility and/or pain without US evidence of endometriosis scheduled for surgery, and to compare the predicted with the actual presence of endometriosis. Although our current study is an important step in the development of a blood test with a high sensitivity for the diagnosis of endometriosis in subfertile patients with a normal gynaecological US, new system biology approaches, i.e. proteomics, are needed to identify novel and specific biomarkers of endometriosis to further increase the sensitivity and specificity of a blood test for endometriosis.
Discussion
The present study is an important step in the development and validation of a blood test for endometriosis with a high sensitivity and acceptable specificity in a group of patients with subfertility and/or pain with a negative preoperative US. Multivariate analysis based on two models of four biomarkers (annexin V, VEGF, CA-125, glycodelin; annexin V, VEGF, CA-125 and sICAM-1) during the menstrual phase enabled the diagnosis of endometriosis undetectable by US with a high sensitivity (81–90%) at acceptable specificity (63–81%) in an independent training and test data set. This diagnostic performance was better than the diagnostic performance of any single biomarker in our study.
The novelty of our study is based on the design based on QUADAS guidelines (Whiting et al., 2003; May et al., 2010), the large and well-defined patient population (n = 353), the large number of evaluated plasma biomarkers (n = 28), the advanced multivariate statistical approach to select biomarkers, to classify them by the multivariate logistic regression and LS-SVM, and to validate the models developed in the training set in an independent test set, and the subanalysis of samples from women with negative preoperative US.
The strength of our study is that our design is in accordance with the QUADAS guidelines (Whiting et al., 2003; May et al., 2010) with respect to control group selection and cycle phase correction. In line with these guidelines, sample collection had been performed at a consistent phase of the cycle and results presented were corrected for cycle phases, as recommended (Whiting et al., 2003; May et al., 2010). For instance, three out of nine papers investigating IL-6 failed to adjust for the phase of the menstrual cycle, despite evidence that levels are known to change throughout the cycle (Angstwurm et al., 1997; May et al., 2010).
The choice of an appropriate control group is crucial and depends on the aim of the diagnostic test. In line with the QUADAS guidelines we selected our controls from women with symptoms consistent with endometriosis (such as infertility and/or pelvic pain) but without laparoscopic evidence of endometriosis based on laparoscopic data, obtained by an experienced endometriosis surgeon. Our controls have different pelvic pathology such as non-endometriosis adhesions, myoma, parasalpingeal cyst and hydrosalpinx (Table II).
Interestingly, inflammatory biomarkers such as IL-1beta, IFN-gamma, TNF-alpha and IL-6 were slightly higher in the control group than in the endometriosis group in the training data set. These data were not consistent and not confirmed in the test data set (Table III and Table IV), which included a comparable (P = 0.15) proportion of patients with non-endometriotic adhesions (9/41, 22%) as the training data set (28/80, 35%; Table II). Nevertheless, this observation could be partially explained by the possibility that controls with non-endometriotic pelvic pathology like adhesions and hydrosalpinx had increased plasma concentrations of inflammatory cytokines. As inflammatory cytokines were not included in the final diagnostic model (Table VI), we speculate that they are not relevant in the discrimination of patients with endometriosis from women with non-endometriotic pelvic pathology due to the similar inflammatory pathways which cause endometriotic and non-endometriotic pelvic pathology and results in pelvic pain and infertility. Moreover, at present, there is no consensus regarding the value of inflammatory factors as biomarkers of endometriosis. Comparable serum IL-6 (Kalu et al., 2007; Socolov et al., 2011), TNF-alpha and IL-1 (Kalu et al., 2007; Othman et al., 2008; Socolov et al., 2011) levels were previously reported in women with and without endometriosis. However, other investigators reported elevated peripheral levels of IL-6 (Bedaiwy et al., 2002; Othman et al., 2008), TNF-alpha (Bedaiwy et al., 2002; Xavier et al., 2006); IFN-gamma (Othman et al., 2008) in endometriosis patients compared with controls. These discrepancies could also be partially explained by the differences in study design (different inclusion criteria and different cycle phases), preanalytical variability in the cytokines levels (due to different types of collected samples (serum versus plasma); different clotting times, different conditions of centrifugation) which could influence the study results. For example, measurable concentrations of inflammatory markers are higher in the serum than in simultaneously collected plasma, due to the release of inflammatory markers during the coagulation process in the serum (Skogstrand et al., 2008). Moreover, biological variability due to functional single nucleotide polymorphisms known to influence protein levels of corresponding circulated proteins can also partially explain the discrepancy in study results, since genetic variants in IL-6 and IFN-gamma genes may influence circulating levels of corresponding proteins (Talar-Wojnarowska et al., 2009; Vallinoto et al., 2010).
The data of our study confirmed the hypothesis (Robin et al., 2009; May et al., 2010) that a panel of biomarkers can improve the sensitivity and specificity of diagnostic test compared with the diagnostic performance of any single biomarker. Our panel of four biomarkers (annexin V, VEGF, CA-125, sICAM-1 or glycodelin) had a better diagnostic performance than any single biomarker in our study. So far, only a limited number of studies have focused on the prediction of endometriosis based on a panel of markers, and these studies were limited by univariate analysis (Somigliana et al., 2004; Agic et al., 2008). Our multivariate statistical approach allowed us to model the relationship between diagnostic categories and all biomarkers simultaneously, taking into account the correlation that may exist between those biomarkers, while univariate analysis only deals with the relationship between one predictor and diagnostic category.
In the present study the two selected panels of four biomarkers (annexin V, VEGF, CA-125, sICAM-1/glycodelin) performed robustly by using two fundamentally different classifiers (multivariate logistic regression and LS-SVM) and could predict US-negative endometriosis with a sensitivity of 81–90% and a specificity of 63–81% (Table VI). Indeed, both methods (multivariate logistic regression and LS-SVM) are widely used in biomarker studies (Robin et al., 2009) and none of these methods is clearly superior when compared with the other (Robin et al., 2009). However, the multivariate logistic regression is more sensitive to feature selection (Pochet and Suykens, 2006; Mihalyi et al., 2010) since it tends to build a classification model that fits patients from a training set optimally (Pochet and Suykens, 2006), but is not always possible to make good predictions for novel patients from an independent test set, a problem defined as overfitting (Pochet and Suykens, 2006). In contrast, the LS-SVM is less sensitive to feature selection and effect of outliers, preventing the model from overfitting the training data (Pochet and Suykens, 2006; Mihalyi et al., 2010). In addition, it has an internal mechanism for modeling non-linearity, giving rise to increased robustness and therefore good performance in an independent test data set (Pochet and Suykens, 2006). Our data show that it is possible to overcome the problem of data overfitting by using the biomarkers selection method, as described in the methodology section. Indeed, the diagnostic performance of the models based on the training set was confirmed on the independent test set by using both classifiers (multivariate logistic regression and LS-SVM).
The relevance of the selected diagnostic panel (annexin V, VEGF, CA-125 and sICAM-1/glycodelin) is confirmed by the fact that the selected biomarkers are involved in apoptosis, angiogenesis, adhesion and tumorogenesis, which are highly related to the pathogenesis of endometriosis.
Annexin V, a marker of apoptosis, has been recently reported by our group to be a promising semi-invasive biomarker for diagnosis of minimal–mild endometriosis (Kyama et al., 2011). Indeed, alterations in the regulation of apoptosis in eutopic and ectopic endometrium from women with endometriosis could contribute to the survival of endometrial cells into the peritoneal cavity and development of endometriosis (reviewed by Taniguchi et al., 2011).
Glycodelin is an endometrium-derived protein with known angiogenic, immunosuppressive and contraceptive effects, which could contribute to the development of endometriosis and endometriosis-related infertility (reviewed by Seppälä et al., 2009). VEGF is one of the main stimuli for angiogenesis and increased vessel permeability, which contributes to the development of endometriotic lesions (Taylor et al., 2002, Becker and D'Amato, 2007). sICAM-1 is one of the major adhesion molecules which inhibits natural killer cell-mediated cytotoxicity (Becker et al., 1991), resulting in defective immune surveillance and is involved in the implantation and development of endometriotic lesions (Wu and Ho, 2003).
CA-125 is the most extensively investigated and used peripheral biomarker of endometriosis (Gupta et al., 2006). CA-125 is produced by endometrial and mesothelial cells and exudes into circulation via the endothelial lining of capillaries in response to inflammation (Bischof, 1993; Zeillemaker et al., 1994; reviewed by Gupta et al., 2006). However, CA-125 levels in the peripheral blood lack diagnostic power as a single biomarker of endometriosis (Mol et al., 1998; Kennedy et al., 2005).
We realize that a diagnostic test may do more harm than good, e.g. by subjecting patients to unnecessary or even potentially harmful procedures (Evers and Van Steirteghem, 2009) since the benefits of treating women with asymptomatic endometriosis is unclear (May et al., 2010). Therefore, we do not recommend to develop or use a blood test for screening purpose in asymptomatic women. However, up to 45% of subfertile women with a regular cycle whose partner has normal sperm quality, with or without pelvic pain, and with normal clinical examination and a normal pelvic US may have endometriosis (Meuleman et al., 2009). A blood test could identify those most likely to have endometriosis or other pelvic conditions and likely to benefit from surgical therapy for both subfertility and pain (Kennedy et al., 2005; D'Hooghe et al., 2006). In our study, the biomarker panels allowed to rule in these women with a high sensitivity (81–90%) and acceptable specificity (63–81%) and distinguish them from women without endometriosis who had symptoms similar to those with endometriosis (subfertility and/or pain), which is in line with published recommendations (May et al., 2010).
In our future work we plan to develop a computer application based on the selected best predictive models (Table VI) that would be freely available for clinical use. The two models, based on the incorporation of measured plasma levels of selected biomarkers (model 1: annexin V, VEGF, CA-125, glycodelin; model 2: annexin V, VEGF, CA-125, sICAM-1) during menstrual cycle phase will be used to identify high-risk groups of patients with a predicted probability of developing endometriosis based on the developed threshold. If the value of predicted probability is greater than the given threshold we conclude that the endometriosis status is positive while in the reverse case we conclude a negative status. Further prospective study is required to validate these models in clinical setting.
Our study is marked by the following limitations.
Firstly, the best diagnostic model was based on the analysis of plasma samples obtained during the menstrual phase. This is not surprising, since it is well known that plasma CA-125 levels in women (O'Shaughnessy et al., 1993) and baboons (Falconer et al., 2005) with endometriosis are higher during the menstrual phase than during other phases of the cycle. In practice, this could be a limitation since the blood sampling has to be limited to the menstrual cycle phase only.
Secondly, stress factors directly before surgery might have affected plasma biomarker levels, as blood was taken just prior to anaesthesia (as previously described by Mihalyi et al., 2010). More research is needed to validate our diagnostic models in plasma samples obtained in an outpatient clinic. However, in our study, the priority was to ensure that the blood sample was taken at the time of surgery in order to have a direct temporal comparison between laparoscopic diagnosis and staging of endometriosis disease and the plasma levels of the biomarkers studied.
Thirdly, we did not evaluate possible diurnal variability in biomarkers levels, as previously observed for serum IL-6 levels (Arvidson et al., 1994). From a practical perspective, we were looking for a robust biomarker panel not depending on diurnal variability, as suggested by the validation of our model in an independent test data set.
Fourthly, the selection of control group was based only on the laparoscopically exclusion of endometriosis by an experienced endometriosis surgeon without histological evaluation, which could be a limitation, especially for the patients with non-endometriotic adhesions. Indeed, it is difficult to rule out that women with a normal pelvis or with non-endometriotic adhesions may have microscopic endometriosis, and that laparoscopic absence of endometriosis may be a temporary phenomenon. However, since intraperitoneal adhesions are accepted as aetiologic factors for infertility (Hammoud et al., 2004), inclusion of patients with non-endometriosis adhesions based on laparoscopy data in the control group can be justified in a biomarker study for endometriosis.
In conclusion, multivariate analysis of four biomarkers (annexin V, VEGF, CA-125 and sICAM-1/or glycodelin) in plasma samples obtained during menstruation enabled the diagnosis of endometriosis undetectable by US with a sensitivity of 81–90% and a specificity of 63–81% in the independent training- and test data set. The next research step is to predict the presence of endometriosis with a high sensitivity, using the models presented in this study, in an independent set of patients with infertility and/or pain without US evidence of endometriosis scheduled for surgery, and to compare the predicted with the actual presence of endometriosis. Although our current study is an important step in the development of a blood test with a high sensitivity for the diagnosis of endometriosis in subfertile patients with a normal gynaecological US, new system biology approaches, i.e. proteomics, are needed to identify novel and specific biomarkers of endometriosis to further increase the sensitivity and specificity of a blood test for endometriosis.
