PCOSt: A non-invasive and cost-effective screening tool for polycystic ovary syndrome

Dear Editor, Polycystic ovary syndrome (PCOS) affects more than 1 in 10 women worldwide.1Stener-Victorin E. Padmanabhan V. Walters K.A. et al.Animal models to understand the etiology and pathophysiology of polycystic ovary syndrome.Endocr. Rev. 2020; 41bnaa010Crossref PubMed Scopus (129) Google Scholar Despite its high prevalence, PCOS and its accompanying morbidities are likely underdiagnosed.2Teede H.J. Misso M.L. Costello M.F. et al.Recommendations from the international evidence-based guideline for the assessment and management of polycystic ovary syndrome.Fertil. Steril. 2018; 110: 364-379Abstract Full Text Full Text PDF PubMed Scopus (563) Google Scholar To induce a PCOS-like phenotype in animals, testosterone and dihydrotestosterone are usually used, indicating that androgen excess is a major contributor to PCOS.3Padmanabhan V. Veiga-Lopez A. Animal models of the polycystic ovary syndrome phenotype.Steroids. 2013; 78: 734-740Crossref PubMed Scopus (99) Google Scholar Anti-Müllerian hormone (AMH) is an upstream regulator of androgen and is well correlated with the androgens.4Zhang X. Xu H. Feng G. et al.Sensitive HPLC-DMS/MS/MS method coupled with dispersive magnetic solid phase extraction followed by in situ derivatization for the simultaneous determination of multiplexing androgens and 17-hydroxyprogesterone in human serum and its application to patients with polycystic ovarian syndrome.Clin. Chim. Acta. 2023; 538: 221-230PubMed Google Scholar Some are advocating that AMH should be considered as a diagnostic marker for PCOS. Previously, we conducted an analysis of 11,720 ovarian stimulation cycles to determine if certain independent variables could be used to predict PCOS. These variables included age, body mass index (BMI), upper limit of menstrual cycle length (UML), antral follicle counts, serum AMH levels, basal follicle-stimulating hormone, basal luteinizing hormone, basal estradiol, basal testosterone, and basal androstenedione (A4) levels. The final four predictive variables included in the original PCOS model, denoted PCOS-4, were serum AMH, UML, BMI, and immune-based detection of serum A4.5Xu H. Feng G. Alpadi K. et al.A model for predicting polycystic ovary syndrome using serum AMH, menstrual cycle length, body mass index and serum androstenedione in Chinese reproductive aged population: a retrospective cohort study.Front. Endocrinol. 2022; 13: 821368Crossref PubMed Scopus (3) Google Scholar In the PCOS-4 model, AMH was the most important predicting variable, contributing 41.2% to the prediction, whereas immune-based A4 contributed only 4.3%. Testosterone was not included in the PCOS-4 model.5Xu H. Feng G. Alpadi K. et al.A model for predicting polycystic ovary syndrome using serum AMH, menstrual cycle length, body mass index and serum androstenedione in Chinese reproductive aged population: a retrospective cohort study.Front. Endocrinol. 2022; 13: 821368Crossref PubMed Scopus (3) Google Scholar In another study, we showed that AMH correlated well with mass spectrometry (MS)-based detection of testosterone and A4 levels in human serum, with Pearson correlation coefficients (r) of 0.57 and 0.51, respectively.4Zhang X. Xu H. Feng G. et al.Sensitive HPLC-DMS/MS/MS method coupled with dispersive magnetic solid phase extraction followed by in situ derivatization for the simultaneous determination of multiplexing androgens and 17-hydroxyprogesterone in human serum and its application to patients with polycystic ovarian syndrome.Clin. Chim. Acta. 2023; 538: 221-230PubMed Google Scholar Nevertheless, the correlation between the A4 level measured in the MS platform and the immune-based chemiluminescence platform was weak in women, with a r2 of 0.285,4Zhang X. Xu H. Feng G. et al.Sensitive HPLC-DMS/MS/MS method coupled with dispersive magnetic solid phase extraction followed by in situ derivatization for the simultaneous determination of multiplexing androgens and 17-hydroxyprogesterone in human serum and its application to patients with polycystic ovarian syndrome.Clin. Chim. Acta. 2023; 538: 221-230PubMed Google Scholar indicating the low precision of the widely used chemiluminescence-based assay for assessing A4. Taking into account the limited contribution of immune-based A4 and its low accuracy, we propose a model for PCOS prediction that omits A4. For this new model, we investigated the same data from 11,720 ovarian stimulation cycles with only three predictors, AMH, UML, and BMI, and called it the PCOS-3. The modeling process was comparable to that used in the PCOS-4 model.5Xu H. Feng G. Alpadi K. et al.A model for predicting polycystic ovary syndrome using serum AMH, menstrual cycle length, body mass index and serum androstenedione in Chinese reproductive aged population: a retrospective cohort study.Front. Endocrinol. 2022; 13: 821368Crossref PubMed Scopus (3) Google Scholar To be specific, the continuous predictors were initially converted into categorical variables with the grouping criteria employed for the PCOS-4 model.5Xu H. Feng G. Alpadi K. et al.A model for predicting polycystic ovary syndrome using serum AMH, menstrual cycle length, body mass index and serum androstenedione in Chinese reproductive aged population: a retrospective cohort study.Front. Endocrinol. 2022; 13: 821368Crossref PubMed Scopus (3) Google Scholar The data were then randomly divided into a training set and a test set in a ratio of 7:3. Model building was performed using the training set, and the performance was assessed using the test set. The newly developed PCOS-3 model was established through 10-fold cross validation in the training set. Finally, we evaluated whether the performance of the simplified PCOS-3 model was similar to that of the PCOS-4 model. The performances of PCOS-3 and PCOS-4 can be compared directly due to the use of the same data (Table 1). A default cutoff of 0.5 was implemented to differentiate between PCOS and non-PCOS; thus, those with a predicted probability of more than 0.5 were considered to be diagnosed with predicted PCOS. However, no significant differences were found between the two models in terms of area under the curve, sensitivity, and specificity, as indicated by the overlapping 95% confidence intervals, indicating that the performances of the two models were comparable. For the purpose of evaluating the calibration of the models, the Brier score was utilized, with 0.067 and 0.066 for the training sets of the PCOS-3 and PCOS-4 models, respectively, and 0.066 and 0.066 for the test sets of the PCOS-3 and PCOS-4 models, respectively. The Brier scores of the two models were similar, suggesting that the calibration of the two models was virtually the same. The net reclassification index (NRI) was utilized to measure the enhancement in the performance of the PCOS-3 model. In the training and test sets, the respective NCIs were −0.022 (−0.035, −0.009) and 0 (−0.023, 0.023). The small negative value in the training set demonstrates that the PCOS-3 model was slightly weaker than the PCOS-4 model. Nonetheless, the performance of both models was comparable in the test set, with the 95% confidence interval including zero. Figure 1A provides an overview of the contribution of each predictor in the two models, including the main effect, which reflects the relative contribution of each variable independently, and the total effect, which takes into account the relative contribution of each variable, both individually and in combination with other variables.Table 1The performance of PCOS-3 and PCOS-4 modelsTraining setValidation setTest setPCOS-3PCOS-4PCOS-3PCOS-4PCOS-3PCOS-4AUC (95% CI)0.850 (0.842, 0.858)0.855 (0.838–0.870)0.851 (0.828, 0.874)0.848 (0.791–0.891)0.841 (0.826, 0.856)0.846 (0.812–0.875)Sensitivity (95% CI)0.339 (0.308, 0.370)0.362 (0.331–0.395)0.384 (0.294, 0.482)0.394 (0.303–0.492)0.383 (0.324, 0.445)0.383 (0.324–0.445)Specificity (95% CI)0.981 (0.978, 0.984)0.981 (0.997–0.983)0.983 (0.972, 0.990)0.985 (0.974–0.991)0.981 (0.974, 0.986)0.981 (0.974–0.986)AUC, area under the curve; CI, confidence interval; PCOS, polycystic ovary syndrome; PCOS-3, PCOS model with three predictors of anti-Müllerian hormone, upper limit of menstrual cycle length, and body mass index; PCOS-4, PCOS model with four predictors of anti-Müllerian hormone, upper limit of menstrual cycle length, body mass index, and androstenedione. Open table in a new tab AUC, area under the curve; CI, confidence interval; PCOS, polycystic ovary syndrome; PCOS-3, PCOS model with three predictors of anti-Müllerian hormone, upper limit of menstrual cycle length, and body mass index; PCOS-4, PCOS model with four predictors of anti-Müllerian hormone, upper limit of menstrual cycle length, body mass index, and androstenedione. Based on the association between the PCOS probability predicted by the PCOS-4 model and the actual incidence of PCOS, we have previously classified the population into three groups: low risk (predicted probability <10%), medium risk (predicted probability 10%–50%), and high risk (predicted probability >50%). Figure 1B illustrates the relationship between the PCOS probability predicted by the PCOS-3 model and the real incidence of PCOS. The results demonstrate that the original three groups still apply to the new PCOS-3 model. Subsequently, we once again divided the population into low-, medium-, and high-risk groups. We employed a default cutoff of 0.5 (i.e., predicted probability ≥50%) with the JMP PRO v.16.0 (Cary, NC, USA) software to identify positive PCOS cases. The numbers of predicted and actual PCOS-positive and -negative cases were then depicted in a Venn diagram, as illustrated in Figure 1C. Upon optimizing the cutoff value to 12.2% based on the Youden index (sensitivity + specificity −1),6Fluss R. Faraggi D. Reiser B. Estimation of the youden index and its associated cutoff point.Biom. J. 2005; 47: 458-472Crossref PubMed Scopus (1513) Google Scholar the number of positive cases increased significantly (Figure 1D). Nevertheless, regardless of the cutoff value, the predicted probability of PCOS remained unchanged. To address the misclassification problem, we divided the population into three distinct groups according to the predicted PCOS probability instead of two. We have further improved the PCOS screening online tool, namely PCOSt, which now contains both PCOS-3 and PCOS-4 models (http://121.43.113.123:8888/). Although the onset of PCOS starts in prepuberty, its symptoms are more pronounced than in adulthood; hence, the diagnostic criteria for adults with PCOS are not suitable for adolescent girls.7Peña A.S. Witchel S.F. Hoeger K.M. et al.Adolescent polycystic ovary syndrome according to the international evidence-based guideline.BMC Med. 2020; 18: 72Crossref PubMed Scopus (101) Google Scholar Nonetheless, the PCOS-3 model divides the population into 100 subgroups and arranges them according to their predicted PCOS likelihood (predicted severity). This model could possibly be adapted for adolescent girls by altering the thresholds of medium-risk and high-risk groups; however, this needs to be confirmed through further validation. We developed the enhanced PCOS-3 model to predict PCOS, incorporating just one serum predictor (AMH) as well as two easily accessible predictors: menstrual cycle length and BMI. By omitting immune-based androgen measurements, the PCOS-3 model eliminates the potential imprecision of such an approach. Consequently, PCOS-3 is non-invasive, cost effective, and suitable for general use, not just in hospitals. However, it should be noted that PCOS-3 is only for initial screening; for those diagnosed with PCOS, further androgen-focused8Azziz R. Carmina E. Chen Z. et al.Polycystic ovary syndrome.Nat. Rev. Dis. Prim. 2016; 2: 16057Crossref PubMed Scopus (792) Google Scholar,9Puvithra T. Pandiyan N. Polycystic ovary syndrome is an epiphenomenon - an opinion.Chettinad Health City Med. J. 2016; 5: 106-107Google Scholar subtype diagnosis and treatment will be required.