Cluster analysis reveals a homogeneous subgroup of PCOS women with metabolic disturbance associated with adverse reproductive outcomes

Introduction

Polycystic ovarian syndrome (PCOS) is one of the most common reproductive endocrinological diseases affecting 6–20% women of reproductive age depending on the population studied and diagnostic criteria applied.[1] PCOS is a heterogeneous and complex disease with three key clinical manifestations: ovulatory dysfunction (OD), hyperandrogenism (HA), and polycystic ovarian morphology (PCOM).[2] PCOS could lead to anovulatory infertility, obesity, insulin resistance (IR), type II diabetes mellitus (T2DM), metabolic syndrome (MS), etc., which have always been challenging for early diagnosis and clinical management.[3]

The gold standard for diagnosing PCOS has been debating for decades. The international evidence-based guideline[4] recommended the application of the 2003 Rotterdam criteria of the European Society for Human Reproduction and Embryology/American Society for Reproductive Medicine (ESHRE/ASRM)[2] that require at least two of the three characteristics (OD, clinical and/or biochemical HA, and PCOM). The 1990 National Institutes of Health (NIH) criteria[5] require both the presence of HA and OD, but do not necessarily require PCOM. The 2006 Androgen Excess and PCOS (AE-PCOS) Society criteria[6] emphasized the presence of HA, and OD is also required in the form of oligo-/anovulation and/or PCOM.

In 2012, NIH defined four different clinical phenotypic subgroups of Rotterdam criteria that aimed to provide a more convenient approach for research and clinical practice: (1) Full-blown PCOS consists of OD + HA + PCOM; (2) Non-PCOM consists of HA + OD; (3) Ovulatory PCOS consists of PCOM + HA; and (4) Non-hyperandrogenic PCOS consists of PCOM + OD.[7] However, the PCOS phenotypic subgroups have been questioned by many researchers. Some suggested that there is no evidence for the metabolic variances between HA + OD and HA + OD + PCOM.[8] Recently, elevated serum anti-Müllerian hormone (AMH) levels have been recognized as a substitute for PCOM for its strong association with the number of follicles.[9] However, some reported that insulin resistance (IR) parameters and their relation to AMH did not differ between the PCOS phenotypic subgroups.[10] Also, it remains unclear which subgroups of PCOS have the best reproductive outcomes.

Cluster analysis is a well-established statistical method to aggregate individuals with similar characteristics into clusters. Cluster analysis has been widely applied in many fields including asthma, obesity, T2DM, cardiovascular diseases, etc.[11–13] Recently, one study adopted cluster analysis to investigate clinical characteristics and found similar genetic architecture inside one certain subtype, thus connecting phenotypic subtyping and genetic background.[14] However, the reproductive outcomes of each PCOS cluster are still unclear. Therefore, in this retrospective study, using cluster analysis, we attempted to identify the cluster of PCOS that is associated with the best or worst reproductive outcomes during in vitro fertilization (IVF) treatment and clarify the features of that cluster.

Methods Ethical approval

This study was approved by the Ethical Review Board of West China Second University Hospital, Sichuan University (No. 2021-033). Written informed consent was waived since there was no intervention for the participants.

Study population

Infertile PCOS patients who underwent their first cycle of IVF in the Reproductive Center, Department of Obstetrics and Gynecology, West China Second University Hospital, Sichuan University from January 2019 to December 2021 were included in this retrospective study. Patients were diagnosed with PCOS under the Rotterdam criteria.[2] For patients who underwent more than one controlled ovarian stimulation (COS) cycles in our reproductive center, only the information regarding the first cycle was included. And we only enrolled patients using fixed GnRH-antagonist protocol for COS to reduce the heterogeneity. Patients with any known endocrine or systemic diseases associated with infertility, such as Cushing syndrome, thyroid diseases, functional hypothalamic amenorrhea (FHA), androgen-secreting ovarian tumors or adrenal tumors, hyperprolactinemia, or premature ovarian insufficiency (POI), were excluded. Besides, male factor infertility, couples with abnormal chromosome karyotype not including chromosome polymorphisms, patients with a history of recurrent pregnancy loss or gynecological surgery were also excluded.

Basic clinical information of each individual was extracted from the hospital record, including age, height, weight, body mass index (BMI), type of infertility, OD, HA, PCOM, baseline serum testosterone (T), dehydroepiandrosterone sulfate (DHEAS), androstadienone (A), sex hormone-binding globulin (SHBG), estradiol (E2), progesterone (P), follicular stimulation hormone (FSH), luteinizing hormone (LH), AMH, fasting insulin (Ins0), and fasting glucose (Glu0) levels, were also collected. HA was evaluated by the presence of hirsutism, alopecia, or hyperandrogenemia. Serum sex hormone levels and PCOM were detected on Days 2–4 of the menstrual cycle. Homeostatic model assessment of insulin resistance (HOMA-IR) was calculated using the following formula: HOMA-IR = (Glu0 × Ins0)/22.5.

Phenotypic clustering

Phenotypic clustering was performed based on combinations of several adjusted continuous variables: BMI, T, DHEAS, Ins0, Glu0, HOMA-IR, SHBG, LH, FSH, LH/FSH, and AMH. The methods for phenotypic clustering were previously detailed described by Dapas et al.[14] Briefly, continuous variables were loge(ln)-normalized and adjusted for age and assay method using a linear regression model, after which an inverse normal transformation was performed. An unsupervised, agglomerative, hierarchical clustering was used to cluster the normalized trait residuals.[15] The stability of clustering was evaluated by the mean Jaccard coefficient from a repeated non-parametric bootstrap resampling (n = 1000) of the dissimilarity matrix and a mean Jaccard coefficient more than 0.6 indicated high similarity between the sample sets.[16]

COS and fresh embryo transfer

All of the patients enrolled in this study underwent fixed gonadotropin-releasing hormone (GnRH)-antagonist protocols. Briefly, exogenous gonadotropin (Gn) (Gonal-F, 100–375 IU, q.d., Merck Serono, Germany) was administered starting from Days 2 to 3 of the menstrual cycle. GnRH-antagonist (Cetrotide, 0.25 mg, q.d., Merck Serono, Germany) was daily administrated after 5–7 days of exogenous Gn application and the administration of GnRH-antagonist continued until the trigger day (TD). Triggering ovulation was performed using urine human chorionic gonadotropin (hCG) (Ovidrel, 8000–10,000 IU, Merck Serono, Germany) when at least two leading follicles reached a mean diameter of 18 mm, or more than three follicles reached a mean diameter of 17 mm. Information regarding initial Gn dosage, Gn stimulation time, total Gn dosage, E2, P, LH, endometrial thickness, number of oocytes with a diameter more than 14 mm on TD, and number of retrieved oocytes were extracted. Transvaginal oocytes retrieval was performed 36–38 h after triggering. IVF or intracytoplasmic sperm injection (ICSI) was performed according to previous fertilization history and semen parameters.

In the present study, data including mature oocyte (MII) rate, two pronuclei (PN) rate, IVF fertilization rate, ICSI fertilization rate, good-quality day 3 (D3) embryo rate, available D3 embryo rate, blastocyst formation rate, good-quality blastocyst rate, available blastocyst rate, fresh embryo transfer (ET) cancelation rate, clinical pregnancy rate after fresh ET, and ovarian hyperstimulation syndrome (OHSS) rate were collected. Embryo or blastocyst was conventionally evaluated for its morphology to assess its quality.[17] Normal fertilization was defined by the presence of two-pronuclear zygotes (2PN). Good-quality D3 embryo was defined by presented 2PN when fertilization, had 6–10 blastomeres, no more than 20% fragmentation, and no multinucleated blastomeres. Good-quality blastocyst was defined by the inner cell mass/trophectoderm score of AA, AB+, AB-, B+A, B-A, B+B+, or BB. Fresh ET cancelation and freeze-all strategy were performed if the patients showed higher P level or at high risk of severe OHSS. Serum hCG test was routinely performed 14 days after fresh ET and transvaginal ultrasound (TVS) was conducted 28 days after fresh ET. Clinical pregnancy was defined as at least one gestational sac observed by the TVS.

Statistical analyses

Numeric variables were presented as mean ± standard deviation (SD) if they were normally distributed otherwise were presented as median and interquartile range (IQR). Categorical variables were expressed as frequency and percentage. To compare means between three or more groups, a one-way analysis of variance (ANOVA) test was performed and the least-significant differences (LSD) test was adopted as the post hoc test. Kruskal–Wallis H tests were adopted to compare medians between different groups. For the comparison between the category variables in different groups, Chi-squared test and/or Fisher's exact test was used and Bonferroni correction was performed. For all analyses, a two-sided P-value ≤0.05 was considered statistically significant. Logistic regression analyses were conducted to compare the fresh ET cancelation rate and clinical pregnancy rate after fresh ET among the three PCOS clusters. Crude and adjusted odds ratio (OR) and 95% corresponding confidential interval (CI) were calculated. All statistical analyses were conducted via SPSS version 22.0 (IBM, Armonk, NY, USA). Phenotypic clustering was performed using R package "fpc" (version 2.2–9), "stats" (version 4.0.2), and "clusterSim" (version 0.49–2). Heatmaps were drawn by R package "gplots" (version 3.1.1) and PCA plots were drawn using R package "factoextra" (version 1.0.7) and "FactoMineR" (version 2.4). Chord Diagram of the relations between old and new PCOS classifications was conducted using R package "circlize" (version 0.4.13).

Results PCOS clustering

A total of 236 PCOS patients were enrolled in this study and only 232 patients were included in the clustering because four patients did not have enough clinical information for clustering. As mentioned before, five combinations of serval traits (BMI, T, DHEAS, Ins0, Glu0, HOMA-IR, SHBG, LH, FSH, LH/FSH, and AMH) were used to perform the phenotypic clustering. When deciding which traits should be included in the clustering models, we found that the clustering model using FSH and LH could better separate the three subgroups than the model using LH/FSH, and similar results were obtained regarding Glu0, Ins0, and HOMA-IR [Table 1 and Supplementary Table 1, https://links.lww.com/CM9/B645]. Therefore, we adopted the model based on BMI, T, DHEAS, Ins0, Glu0, SHBG, LH, FSH, and AMH as the final clustering model [Table 1] and details of other clustering were displayed in Supplementary Tables 1 and 2 and Supplementary Figures 1–3, https://links.lww.com/CM9/B645. PCOS patients were classified into three groups after clustering which revealed three different phenotypic subtypes: (1) reproductive group: 113/232 (48.7%) of the PCOS patients were classified as reproductive group which was characterized by high T, SHBG, FSH, LH, and AMH level; (2) metabolic group: 72/232 (31.0%) of the patients were identified as metabolic group which was characterized by high BMI, Glu0, and Ins0 level; (3) balanced group: 47/232 (20.3%) of the patients were assigned to balanced group which was characterized by low level of both reproductive and metabolic parameters except for SHBG (i.e., low FSH, LH, AMH, BMI, Glu0, and Ins0 levels) [Table 1 and Figure 1]. The core traits that could separate the reproductive and metabolic clusters were AMH, T, SHBG, LH, and FSH of the reproductive traits and BMI, Glu0, and Ins0 of the metabolic traits [Figure 2]. The reproductive and metabolic clusters were relatively stable compared to the balanced cluster with a mean bootstrapped Jaccard coefficient of 0.59, 0.59, and 0.44, respectively.

Table 1 - Baseline characteristics of PCOS patients in three clusters and different phenotypic subgroups of Rotterdam criteria. Characteristics Clustering Phenotypic subgroups of Rotterdam criteria

Reproductive

(n = 113)

Metabolic

(n = 72)

Balanced

(n = 47)

Statistical value P-value

Full-blown

(n = 95)

Non-PCOM

(n = 39)

Ovulatory

(n = 37)

Non-hyperandrogenic

(n = 65)

Statistical value P-value Age (years) 29
(27–31) 29
(27–31) 29
(26–33) 0.615a 0.735 29
(26–31) 29
(27–32)|| 31
(28–33)¶,** 29
(27–31) 26.776a <0.001 Height (cm) 158(155–160) 159(155–161) 160
(158–163) 2.153a 0.341 158
(156–163) 159
(157–160) 158
(155–160) 158
(155–160) 0.632a 0.889 Weight (kg) 54
(50–60) 63(58–70)*,† 53
(51–58) 68.435a <0.001 59
(53–62) 58
(52–61) 53
(51–60) 55
(51–62) 2.262a 0.520 BMI (kg/m2) 21.6(20.3–23.6) 25.6
(23.7–27.7)*,† 21.1
(19.4–22.1) 84.454a <0.001 23.1
(21.0–25.3) 23.0
(20.8–24.4) 21.5
(19.7–23.7) 21.6
(20.3–25.6) 2.065a 0.559 E2 (pg/mL) 47.4(39.5–55.8) 39.1
(31.4–53.2) 38.0
(29.2–45.1)‡ 7.674a 0.022 43.6
(35.8–55.3) 41.2
(32.7–59.6) 45.4
(39.2–52.1) 42.6
(29.9–56.1) 1.184a 0.757 P (ng/mL) 0.5
(0.4–0.6) 0.6
(0.4–0.7) 0.6
(0.4–0.7) 3.704a 0.157 0.5
(0.4–0.7) 0.6
(0.4–0.8) 0.4
(0.3–0.6) 0.4
(0.3–0.7) 5.100a 0.165 T (ng/mL) 0.5
(0.3–0.6) 0.4
(0.3–0.5)* 0.3
(0.2–0.4)‡ 18.115a <0.001 0.4
(0.3–0.5) 0.4
(0.3–0.5) 0.5
(0.3–0.7) 0.4
(0.2–0.5) 5.898a 0.117 DHEAS (μg/dL) 198(149–264) 224
(157–308) 239
(191–263)‡ 7.192a 0.027 223
(171–288) 241
(166–271) 200
(139–228) 198
(113–259)‡‡ 10.232a 0.017 A (ng/mL) 3.7
(3.1–4.4) 3.5
(2.9–4.5) 3.6
(2.6–4.3) 2.004a 0.367 4.1
(3.5–4.9) 3.8
(3.4–5.0) 4.0
(3.6–4.7)** 2.7
(1.9–3.1)††,‡‡ 104.757a <0.001 SHBG (nmol/L) 49.2(33.2–67.0) 21.8
(16.7–30.2)*,† 52.1
(36.7–63.2) 79.854a <0.001 35.7
(23.3–50.2) 32.7
(21.3–56.1) 47.2
(38.4–65.2) 46.3
(26.7–64.7) 3.171a 0.366 FAI 3.1
(2.0–5.3) 6.6
(4.0–9.6)*,† 2.8
(1.3–4.8) 36.526a <0.001 5.2
(2.8–7.8) 4.1
(2.1–7.5) 4.3
(2.0–5.9) 2.5
(1.6–4.2)†† 18.376a <0.001 FSH (IU/L) 6.6(6.0–8.4) 6.8
(4.9–7.7) 5.7
(4.2–6.8)‡ 15.895a <0.001 7.0
(6.0–8.0) 6.6
(4.9–7.7) 6.9
(6.1–7.8) 5.9
(4.9–7.4)†† 15.162a 0.002 LH (IU/L) 11.7(8.1–16.1) 7.6
(3.7–12.3)*,† 4.6
(3.2–6.1)‡ 72.066a <0.001 9.6
(5.7–14.4) 7.8
(5.7–14.7) 7.9
(4.7–11.4) 6.5
(3.6–10.4) 5.764a 0.124 LH/FSH 1.7(1.2–2.3) 1.2
(0.8–1.8)*,† 0.8
(0.5–1.1)‡ 49.316a <0.001 1.4
(0.9–2.0) 1.6
(0.8–2.1) 1.1
(0.7–1.5) 1.2
(0.7–1.8) 2.594a 0.459 HOMA-IR 1.9(1.2–2.9) 4.0
(3.1–5.4)*,† 1.6
(1.2–2.1) 74.587a <0.001 2.5
(1.6–3.8) 3.1
(1.7–4.2) 1.5
(1.1–3) 2.2
(1.2–3.1) 6.905a 0.075 Glu0 (mmol/L) 4.9(4.7–5.1) 5.2
(4.9–5.5)*,† 4.9
(4.7–5.1) 25.546a <0.001 5.0
(4.8–5.3) 5.1
(4.8–5.3) 4.9
(4.5–5.2) 5.0
(4.7–5.2) 2.288a 0.515 Ins0 (μIU/mL) 8.6 (5.7–13.0) 17.6
(13.0–24.0)*,† 7.1
(5.3–9.2) 79.339a <0.001 11.6
(7.1–16.9) 14.5
(8.5–18.1) 6.9
(5.7–13.0) 9.2
(5.4–14.0) 5.320a 0.150 AMH (ng/mL) 12.7 (9.1–16.6) 4.7
(3.6–8.1)*,† 7.2
(5.3–10.9)‡ 82.838a <0.001 10.0
(6.2–14.6) 5.2
(3.8–12.1)§ 7.1
(4.8–14.6) 8.2
(5.4–13.1) 9.741a 0.021 PCOM (%) 87.61(99/113) 77.78
(56/72) 80.85
(38/47) 3.271b 0.195 – – – – – – OD (%) 79.65(90/113) 86.11
(62/72) 87.23
(41/47) 2.004b 0.367 – – – – – – HA (%) 74.34(84/113) 70.83
(51/72) 70.21
(33/47) 0.413b 0.813 – – – – – –

Data are presented as median (interquartile range) or (%) n/N. aKruskal–Wallis H test. bχ² test. *P <0.05 between Reproductive Group and Metabolic Group. †P <0.05 between Metabolic Group and Balanced Group C. ‡P <0.05 between Reproductive Group and Balanced Group. §P <0.0125 between Full-blown Group and Non-PCOM Group. ||P <0.0125 between Non-PCOM Group and Ovulatory Group. ¶P <0.0125 between Full-blown Group and Ovulatory Group. **P <0.0125 between Ovulatory Group and Non-hyperandrogenic Group. ††P <0.0125 between Full-blown Group and Non-hyperandrogenic Group. ‡‡P <0.0125 between Non-PCOM Group and Non-hyperandrogenic Group. A: Androstadienone; AMH: Anti-Müllerian hormone; BMI: Body mass index; DHEAS: Dehydroepiandrosterone sulfate; E2: Estradiol; FAI: Free androgen index; FSH: Follicle-stimulating hormone; Glu0: Fasting glucose; HA: Hyperandrogenism (hirsutism, alopecia, hyperandrogenemia); HOMA-IR: Homeostatic Model Assessment of Insulin Resistance; Ins0: Fasting insulin; LH: Luteinizing hormone; OD: Ovulatory dysfunction; P: Progesterone; PCOM: Polycystic ovarian morphology; PCOS: Polycystic ovarian syndrome; SHBG: Sex hormone-binding globulin; T: Testosterone; –: Not applicable.

Figure 1:

Heatmap of hierarchical clustering of PCOS cohort. Hierarchical clustering of PCOS patients according to adjusted continuous traits revealed three distinct phenotypic subtypes, "reproductive", "metabolic", and "balanced". The colors of heatmap correspond to trait z-scores, in which red refers to high values and blue refers to low values. AMH: Anti-Müllerian hormone; BMI: Body mass index; DHEAS: Dehydroepiandrosterone sulfate; FSH: Follicle-stimulating hormone; Glu0: Fasting glucose; Ins0: Fasting insulin; LH: Luteinizing hormone; PCOS: Polycystic ovarian syndrome; SHBG: Sex hormone-binding globulin; T: Testosterone.

Figure 2:

PCA plot of continuous traits used for clustering of PCOS cohort. Metabolic, reproductive, and balanced PCOS clusters are shown as 95% concentration ellipses, assuming bivariate normal distributions. The direction of trait and relative magnitude correlations with the PCs were shown with black arrows. AMH: Anti-Müllerian hormone; BMI: Body mass index; DHEAS: Dehydroepiandrosterone sulfate; FSH: Follicle-stimulating hormone; Glu0: Fasting glucose; Ins0: Fasting insulin; LH: Luteinizing hormone; PCA: Principal component analysis; PC1: Principal component 1; PCOS: Polycystic ovarian syndrome; SHBG: Sex hormone-binding globulin; T: Testosterone.

Table 1 and Figure 3 display the prevalence of three key PCOS characteristics (i.e., PCOM, OD, and HA) in each cluster and the proportion of each cluster in different PCOS phenotypic subgroups of Rotterdam criteria. Our results suggested that the reproductive cluster had a higher prevalence of PCOM (87.61% [99/113] vs. 77.78% [56/72] vs. 80.85% [38/47], P = 0.195) and lower prevalence of OD (79.65% [90/113] vs. 86.11% [62/72] vs. 87.23% [41/47], P = 0.367) compared to the other two clusters, but the differences were not statistically significant. About half of the enrolled PCOS patients in the present study were full-blown phenotype (40.25%) which remained the largest proportion in each cluster, 41.59% in reproductive cluster, 36.11% in metabolic cluster, and 38.30% in balanced cluster. Non-hyperandrogenic phenotype was the second most common phenotype in each cluster with 26.55%, 29.17%, and 29.79% of the patients in reproductive, metabolic, and balanced clusters, respectively. 12.39%, 22.22%, and 19.15% of the patients in reproductive, metabolic, and balanced clusters were non-PCOM phenotype, respectively. Lastly, 19.47%, 12.50%, and 12.77% of the patients in reproductive, metabolic, and balanced clusters were ovulatory phenotype, respectively.

Figure 3:

Chord diagram of the relations between three clusters and different phenotypic subgroups of Rotterdam criteria. This figure illustrates the proportions of the typical presentations identified by clustering analysis (upper) by different phenotypic subgroups of Rotterdam criteria (lower). For the sake of readability, the results were presented separately for PCOS patients with (A) full-blown, (B) non-PCOM, (C) ovulatory, and (D) non-hyperandrogenic. HA: Hyperandrogenism; OD: Ovulatory dysfunction; PCOM: Polycystic ovarian morphology; PCOS: Polycystic ovarian syndrome.

COS parameters and reproductive outcomes between different clusters of PCOS

We first compared the COS parameters and reproductive outcomes of different PCOS phenotypic subgroups of Rotterdam criteria. Our results showed that only MII rate, IVF fertility rate, good quality D3 embryo rate, and good quality blastocyst rate were different among the four subgroups, and other outcomes, including fresh ET cancelation rate, clinical pregnancy rate after fresh ET, OHSS rate, remained insignificant [Table 2].

Table 2 - COS parameters and reproductive outcomes of PCOS patients in three clusters and different phenotypic subgroups of Rotterdam criteria. Characteristics Clustering Phenotypic subgroups of Rotterdam criteria