Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Data-driven assessment of the human ovarian reserve

Data-driven assessment of the human ovarian reserve Abstract Human ovarian physiology is still poorly understood, with the factors and mechanisms that control initiation of follicular recruitment and loss remaining particularly unclear. Conventional hypothesis-led studies provide new data, results and insights, but datasets from individual studies are often small, allowing only limited interpretation. Great power is afforded by the aggregation of data from multiple studies into single datasets. In this paper, we describe how modern computational analysis of these datasets provides important new insights into ovarian function and has generated hypotheses that are testable in the laboratory. Specifically, we can hypothesize that age is the most important factor for variations in individual ovarian non-growing follicle (NGF) populations, that anti-Müllerian hormone (AMH) levels generally rise and fall in childhood years before peaking in the mid-twenties, and that there are strong correlations between AMH levels and both NGF populations and rates of recruitment towards maturation, for age ranges before and after peak AMH levels. ovary, fertility, AMH, oocyte, computational analysis Introduction The ovarian reserve may be defined as the remaining pool of non-growing follicles (NGFs) in the ovary at a given age. NGFs are formed in large numbers in the fetal ovary in humans with peak population occurring at 20–22 weeks gestation (Baker, 1963; Wallace and Kelsey, 2010; Mamsen et al., 2011). The population decreases from this peak, largely due to atresia following follicular recruitment towards maturation. The post-menopausal ovary contains fewer than 1000 NGFs, an insufficient number to support recruitment to ovulation. The study of ovarian function throughout life is important for biological, medical and sociological reasons. Human ovarian physiology is still poorly understood, with the factors and mechanisms that initiate and control follicular recruitment and loss through apoptosis remaining unclear. In medical terms, a more complete understanding of ovarian biology is of fundamental importance in assisted conception, and in the assessment of those young woman with cancer and other conditions who may benefit from fertility preservation. In modern western societies, there is an increasing trend towards women starting a family later in life when ovarian reserve is reduced and the menopause is approaching. A better and more complete understanding of how to estimate ovarian reserve for the individual woman would help inform her choices about the timing of having a family. Delayed child bearing in women who have a reduced ovarian reserve is now a major burden on reproductive services and may result in significant emotional and psychological distress if treatment is unsuccessful. There is a good reason for the paucity of knowledge of human ovarian reserve throughout life: direct longitudinal assessment is currently impossible, and is likely to remain impossible for the foreseeable future. No in vivo technique for counting NGFs exists. All studies involving the estimation of NGF populations for ovaries at various chronological ages have analysed tissue post-mortem or post-oophorectomy (Block, 1952, 1953; Baker, 1963; Richardson et al., 1987; Bendsen et al., 2006; Forabosco and Sforza, 2007; Hansen et al., 2008). These estimates are therefore cross-sectional snapshots of populations at known ages, providing no opportunity to assess rates of NGF loss in the subjects, and hence making highly tentative any attempt to describe population trajectories at different ages and stages of reproductive life. A partial solution to this fundamental problem is to consider indirect indicators of ovarian reserve. Endocrine factors, such as anti-Müllerian hormone (AMH) and FSH, are measurable by available serum assays. Physical factors, such as ovarian volume (OV) and antral follicle counts, are measurable by transvaginal sonography. These measurements are inexpensive, non-destructive and repeatable, allowing the possibility of large-scale longitudinal studies. But the key research question remains: how are these measurable indirect indicators related to NGF populations for individuals of known age? Subsidiary questions include the relationships (if any) between indirect factors and NGF recruitment towards maturation, and which combinations of indirect factors (if any) best predict ovarian reserve for which age ranges. The classical approach to this type of research question is to (i) generate a hypothesis, (ii) design an experiment, (iii) collect data and (iv) analyse and discuss. Examples include (Hehenkamp et al., 2006) and (Brett et al., 2009) which considered variations in AMH across menstrual cycles and precision of OV measurement by 2D and 3D ultrasound, respectively. Another common approach is to (i) recruit a cohort of subjects, (ii) collect data, (iii) report and discuss reference ranges, correlations and centiles. High-quality exemplars include (Hansen et al., 2008) and (Hagen et al., 2010), in which NGF populations and serum AMH levels were obtained post-oophorectomy and from juvenile and adolescent volunteers, respectively. Both types of study are crucial for an incremental increase in our understanding of ovarian reserve. They are the current gold standard for scientific endeavour in this field. Another approach is to use the new genetics including micro-array analysis. This methodology has not addressed NGF numbers in individual women thus far, but has provided both detailed analysis of the involvement of key genes in determining premature ovarian insufficiency (Matzuk and Lamb, 2008), and, more recently, large genome-wide association studies addressing the genetic determinants of age of puberty and menopause (Ong et al., 2009; He et al., 2010). Individual studies will typically have the advantages of standardized data acquisition and cohort selection, but the disadvantages of a restricted cohort age range and non-standard data analysis techniques. In this paper, we report initial results for, and discuss the potential of, an alternative scientific approach: the systematic aggregation and analysis of data from several existing studies. We make no claim of originality for this approach: researchers have successfully aggregated data from distinct NGF population studies to produce the first plausible models of NGF decline from birth (Faddy et al., 1992; Faddy and Gosden, 1996). Our aim is to extend this technique, making full use of recent advances in data identification, data retrieval and data analysis. New developments In recent years the biomedical research environment has changed markedly. There has been an exponential increase in the amount of digitized data archived in searchable repositories. Once extracted, the data can be re-analysed using more modern techniques (often using more computationally intensive methods than those available at time of original publication), and datasets from multiple studies can be combined for analysis as a single dataset. The increase in data availability has been matched by enhanced computational capability. Processing power has roughly doubled every 2 years since 1958, and this trend is expected to continue at least until 2015. Many of the algorithms used to model biomedical data are more than 20 years old, but improved processing power means that not only is it possible to analyse larger and more complex datasets, but also that the traditional method of fitting to a theoretically derived model can be replaced by fitting several hundred models and then ranking these by goodness-of-fit. Moreover, increases in computational power allow us to run sophisticated model validation algorithms in those cases where validation by an external dataset is not practical. These algorithms typically involve the repeated generation of models using some of the data, followed by measurement of model error for data that was held back. This has the advantages that small datasets need not be split into test and training sets (thereby compromising predictive power), and we can guard against overfitting by analysis of the trade-off between goodness-of-fit and generalization error. Taken together, these developments motivate and enable systematic and controlled data aggregation and univariate analysis. Notably this approach has clear similarities with that of individualized patient data meta-analyses—the major strength of which is the ability to adjust for confounders in multivariate analyses, however, the lack of provision of the original raw data can bias the included studies and conclusions. Consequently, the structure of a data-led research project might now be (i) perform a structured search for all publications relating to an indicator of ovarian reserve, (ii) import the publications and perform an iterated search based on citation information, (iii) extract raw data from selected publications and combine into a single dataset, (iv) apply knowledge-discovery techniques on the data to produce internally validated models and correlations, (v) report and discuss. Application to the study of the NGF population In an example of the use of this methodology, eight studies were identified that estimated NGF populations for limited age ranges (Table I). Each study used a variation on the standard method for NGF estimation devised by Block in the 1950s: manual stereological counts of NGFs appearing in a small subset of ovarian tissue are integrated to obtain an estimate of the population for the entire ovary. Since the data were collected in a similar manner, they were collated into a single dataset (n= 325) with age range −0.6 through 51.0 years (median 31.0 years). Imaginary but biologically accurate datapoints were added that forced the population at conception to be zero, and non-linear peak models were fitted to the data. After a validation process involving fitting the same models to randomly selected subsets of the data, the best model in terms of goodness-of-fit was reported (Fig. 1; Wallace and Kelsey, 2010). This model is the first that attempts to describe in quantitative terms the dynamics of ovarian reserve for all ages up to menopause. Confidence intervals (CIs) for the model were obtained, together with prediction limits that define normal NGF population ranges. Table I The eight quantitative histological studies forming the combined dataset for the NGF model. Study  Statistics   Number  First author  Year  No. ovaries  Min. age  Max. age  Median age  1  Bendsen  2006  11  −0.6  −0.6  −0.6  2  Baker  1963  11  −0.6  7.0  −0.2  3  Forabasco  2007  15  −0.5  0.5  −0.3  4  Block  1953  19  −0.2  0.0  0.0  5  Hansen  2008  122  0.1  51.0  38.0  6  Block  1951  86  6.0  44.0  28.0  7  Gougeon  1987  52  25.0  46.0  39.5  8  Richardson  1987  9  45.0  51.0  46.0  Overall      325  −0.6  51.0  32.0  Study  Statistics   Number  First author  Year  No. ovaries  Min. age  Max. age  Median age  1  Bendsen  2006  11  −0.6  −0.6  −0.6  2  Baker  1963  11  −0.6  7.0  −0.2  3  Forabasco  2007  15  −0.5  0.5  −0.3  4  Block  1953  19  −0.2  0.0  0.0  5  Hansen  2008  122  0.1  51.0  38.0  6  Block  1951  86  6.0  44.0  28.0  7  Gougeon  1987  52  25.0  46.0  39.5  8  Richardson  1987  9  45.0  51.0  46.0  Overall      325  −0.6  51.0  32.0  Reproduced from Wallace and Kelsey (2010). All ages are in years; the year column refers to the year of publication. Data taken from doi:10.1371/journal.pone.0008772.t001. View Large Figure 1 View largeDownload slide The Wallace–Kelsey model of NGF populations from conception to menopause. The best model for the establishment of the NGF population after conception, and the subsequent decline until age at menopause is described by an Asymmetric Double-Gaussian Cumulative (ADC) model with parameters amplitude = 5.56 (95% CI 5.38–5.74), centre = 25.6 (95% CI 24.9–26.4), width = 52.7 (95% CI 51.1–54.2), growth shape = 0.074 (95% CI 0.062–0.085) and decline shape = 24.5 (95% CI 20.4–28.6). Our model has correlation coefficient r2 = 0.81, fit standard error = 0.46 and F-value = 364. The figure shows the dataset (n= 325), the model, the 95% prediction limits of the model and the 95% CI for the model. The horizontal axis denotes age in months up to birth at age zero, and age in years from birth to 51 years. Reproduced from Wallace and Kelsey (2010); doi:10.1371/journal.pone.0008772. Figure 1 View largeDownload slide The Wallace–Kelsey model of NGF populations from conception to menopause. The best model for the establishment of the NGF population after conception, and the subsequent decline until age at menopause is described by an Asymmetric Double-Gaussian Cumulative (ADC) model with parameters amplitude = 5.56 (95% CI 5.38–5.74), centre = 25.6 (95% CI 24.9–26.4), width = 52.7 (95% CI 51.1–54.2), growth shape = 0.074 (95% CI 0.062–0.085) and decline shape = 24.5 (95% CI 20.4–28.6). Our model has correlation coefficient r2 = 0.81, fit standard error = 0.46 and F-value = 364. The figure shows the dataset (n= 325), the model, the 95% prediction limits of the model and the 95% CI for the model. The horizontal axis denotes age in months up to birth at age zero, and age in years from birth to 51 years. Reproduced from Wallace and Kelsey (2010); doi:10.1371/journal.pone.0008772. This model can then be used to assess NGF recruitment, i.e. the rate of loss of NGFs from the pool—bearing in mind that this does not distinguish between NGF activation and loss for other reasons, the degree and mechanisms of which are debated (Gougeon, 1996; Tingen et al., 2009). Analysis of the model showed that NGF recruitment peaks at 14.2 years in the average case. If it is assumed that NGF population dynamics are similar for the general population (i.e. that the shape of the curve is similar for individuals having different peak values and ages at menopause) then NGF recruitment peaks at an average age 13–15 years in normal women (Fig. 2; Wallace and Kelsey, 2010). This is a research hypothesis derived from the detailed examination of aggregated data, and is potentially testable by reproductive biologists in the laboratory. Figure 2 View largeDownload slide NGF model from conception to 25 years. The best model for the establishment of the NGF population after conception, and the subsequent decline until 25 years of age is described by an ADC model with parameters amplitude = 5.79 (95% CI 5.03–6.55), centre = 28.0 (95% CI 15.8–40.2), width = 57.4 (95% CI 33.1–81.8), growth shape = 0.074 (95% CI 0.067–0.081) and decline shape = 34.3 (95% CI −4.2 to 72.8). This model has correlation coefficient = 0.95, fit standard error = 0.29 and F-value = 585. This figure shows the dataset (n= 126), the model, the 95% prediction limits of the model and the 95% CI for the model. The horizontal axis denotes age in months up to birth at age zero, and age in years from birth to 25 years. Reproduced from Wallace and Kelsey (2010); doi:10.1371/journal.pone.0008772. Figure 2 View largeDownload slide NGF model from conception to 25 years. The best model for the establishment of the NGF population after conception, and the subsequent decline until 25 years of age is described by an ADC model with parameters amplitude = 5.79 (95% CI 5.03–6.55), centre = 28.0 (95% CI 15.8–40.2), width = 57.4 (95% CI 33.1–81.8), growth shape = 0.074 (95% CI 0.067–0.081) and decline shape = 34.3 (95% CI −4.2 to 72.8). This model has correlation coefficient = 0.95, fit standard error = 0.29 and F-value = 585. This figure shows the dataset (n= 126), the model, the 95% prediction limits of the model and the 95% CI for the model. The horizontal axis denotes age in months up to birth at age zero, and age in years from birth to 25 years. Reproduced from Wallace and Kelsey (2010); doi:10.1371/journal.pone.0008772. A second finding based on selective analysis of the aggregated data is that age is the dominant factor for differences in individual NGF populations for ages up to 25 (Fig. 3; Wallace and Kelsey, 2010). The coefficient of determination for this subset of the data is 0.95, indicating that 95% of individual variation is due to age alone. For ages from conception to menopause, this falls to 81%, despite the dataset being larger and thus expected to be more robust, indicating that (i) lifestyle and genetic factors are more important later in life, and (ii) age remains the dominant factor that explains population variations. Figure 3 View largeDownload slide Rates of NGF recruitment towards maturation. Each sub-figure describes the absolute number of NGFs recruited per month, for ages from birth to 55 years, based on population decline predicted by the ADC model. The red curve denotes recruitment for individuals whose decline is in line with the average age at menopause; maximum recruitment of 880 follicles per month occurs at 14 years 2 months. The green curve denotes recruitment for individuals whose decline is in line with early age at menopause (the lower 95% prediction limit of the model); maximum recruitment of 104 follicles per month occurs at 14 years 2 months. The yellow curve denotes recruitment in line with late age at menopause (the upper 95% prediction limit of the model); maximum recruitment of 7520 follicles per month occurs at 14 years 2 months. Reproduced from Wallace and Kelsey (2010); doi:10.1371/journal.pone.0008772. Figure 3 View largeDownload slide Rates of NGF recruitment towards maturation. Each sub-figure describes the absolute number of NGFs recruited per month, for ages from birth to 55 years, based on population decline predicted by the ADC model. The red curve denotes recruitment for individuals whose decline is in line with the average age at menopause; maximum recruitment of 880 follicles per month occurs at 14 years 2 months. The green curve denotes recruitment for individuals whose decline is in line with early age at menopause (the lower 95% prediction limit of the model); maximum recruitment of 104 follicles per month occurs at 14 years 2 months. The yellow curve denotes recruitment in line with late age at menopause (the upper 95% prediction limit of the model); maximum recruitment of 7520 follicles per month occurs at 14 years 2 months. Reproduced from Wallace and Kelsey (2010); doi:10.1371/journal.pone.0008772. Application to the study of serum AMH in healthy females In a recent aggregation study we found 25 studies reporting normative levels of serum AMH for cohorts with limited age ranges. We aggregated the data extracted from these studies with our own data, giving a total dataset of 12 241 AMH measurements for ages from −0.3 to 68 years. Many data points were associated with women attending infertility clinics (for whom normal ovarian reserve indicators could not be assumed) or with individuals with chronic disease, such as lupus or cancer. After careful exclusion of these data, a single dataset of healthy women (n= 3260, −0.3 through 54.3 years, median 28.3 years) was obtained from 20 separate sources (Table II; Kelsey et al., 2011). After fixing AMH levels to zero at conception, a range of non-linear models were fitted to this data, with a cross-validation stage to guard against reporting a model that provides a good fit to the data but which is unlikely to generalize well to unseen data. The best model, in terms of validated goodness-of-fit, is the first quantitative model for AMH levels for the normal population from conception to post-menopausal ages (Fig. 4; Kelsey et al., 2011). As for the NGF population model, the AMH model both supports existing hypotheses and identifies new hypotheses. Consistent with previous reports (Hagen et al., 2010) our model shows that serum AMH falls shortly after birth in girls, with concentrations only increasing again after about 2 years of age. This feature is in line with evidence of a mini-puberty seen in neonatal girls (Chellakooty et al., 2003; Lee, 2003). Secondly, our model demonstrated a non-linear decline from age 24.5, with 99% correlation to the decline reported for large infertility cohorts (n= 25 435; Nelson et al., 2011a, b). Table II Serum AMH data summary. Ref.  First author  Data  Assay  n  Average age  Age range  Detection limits  Intra CV  Inter CV  [35]  Soto  Graph  IBC  58  30.3  ±8.7  0.10  5.3  8.7  [38]  Guibourdenche  Graph  IBC  192  NS  −0.3 to 1.0  0.30  5.3  8.7  [39]  Hudecova  Graph  IBC  64  46.3  ±6.4  0.70  12.3  12.3  [40]  Mulders  Graph  IBC  82  29.9  20–36  NS  5.0  8.0  [41]  Pastor  Graph  IBC  42  NS  18–50  0.10  5.3  7.8  [42]  Piltonen  Graph  IBC  44  31.6  21–44  NS  5.1  6.6  [20]  van Rooilj  Graph  IBC  162  NS  25–46  0.05  5.0  8.0  [43]  Laven  Graph  IBC  41  NS  20–36  0.05  5.0  8.0  [19]  de Vet  Graph  IBC  82  29.0  ±4.0  0.05  5.0  8.0  [44]  Knauf  Graph  IBC  83  34.2  ±3.4  0.03  11.0  11.0  [45]  Lee  Graph  IBC  225  NS  0–51  0.5  9.0  15.0  [36]  La Marca  Graph  IBC  24  44.0  ±2.8  0.24  5.0  8.0  [29]  Hagen  Graph  IBC  891  NS  0–68  0.03  7.8  11.6  [46]  van Beek  Graph  DSL  82  29.0  20–35  NS  5.0  15.0  [47]  Sanders  Graph  DSL  43  24.1  0–51  0.01  NS  11.4  [34]  van Disseldorp  Graph  DSL  144  37.9  25–46  0.03  11.0  11.0  [48]  Tehrani  Graph  DSL  267  27.1  16–44  0.01  5.2  9.1  [49]  Dorgan  Graph  DSL  204  44.7  33–55  0.06  8.0  8.0  [30]  Ahmed  Raw  DSL  128  8.5  1–17  0.50  8.0  8.0  [25]  Nelson  Raw  DSL  441  36.1  22–48  0.03  3.4  8.6    Total IBC      1990    −0.3 to 68          Total DSL      1309    0–55          Total n      3299    −0.3 to 68          Censored total n      3260    −0.3 to 54.3        Ref.  First author  Data  Assay  n  Average age  Age range  Detection limits  Intra CV  Inter CV  [35]  Soto  Graph  IBC  58  30.3  ±8.7  0.10  5.3  8.7  [38]  Guibourdenche  Graph  IBC  192  NS  −0.3 to 1.0  0.30  5.3  8.7  [39]  Hudecova  Graph  IBC  64  46.3  ±6.4  0.70  12.3  12.3  [40]  Mulders  Graph  IBC  82  29.9  20–36  NS  5.0  8.0  [41]  Pastor  Graph  IBC  42  NS  18–50  0.10  5.3  7.8  [42]  Piltonen  Graph  IBC  44  31.6  21–44  NS  5.1  6.6  [20]  van Rooilj  Graph  IBC  162  NS  25–46  0.05  5.0  8.0  [43]  Laven  Graph  IBC  41  NS  20–36  0.05  5.0  8.0  [19]  de Vet  Graph  IBC  82  29.0  ±4.0  0.05  5.0  8.0  [44]  Knauf  Graph  IBC  83  34.2  ±3.4  0.03  11.0  11.0  [45]  Lee  Graph  IBC  225  NS  0–51  0.5  9.0  15.0  [36]  La Marca  Graph  IBC  24  44.0  ±2.8  0.24  5.0  8.0  [29]  Hagen  Graph  IBC  891  NS  0–68  0.03  7.8  11.6  [46]  van Beek  Graph  DSL  82  29.0  20–35  NS  5.0  15.0  [47]  Sanders  Graph  DSL  43  24.1  0–51  0.01  NS  11.4  [34]  van Disseldorp  Graph  DSL  144  37.9  25–46  0.03  11.0  11.0  [48]  Tehrani  Graph  DSL  267  27.1  16–44  0.01  5.2  9.1  [49]  Dorgan  Graph  DSL  204  44.7  33–55  0.06  8.0  8.0  [30]  Ahmed  Raw  DSL  128  8.5  1–17  0.50  8.0  8.0  [25]  Nelson  Raw  DSL  441  36.1  22–48  0.03  3.4  8.6    Total IBC      1990    −0.3 to 68          Total DSL      1309    0–55          Total n      3299    −0.3 to 68          Censored total n      3260    −0.3 to 54.3        Reproduced from Kelsey et al. (2011). The references relate to the bibliography section of the original paper. Age information is given as median and range, or as mean ± SD depending on which form was reported in the referenced study. Detection limits are given in nanogram/millilitre. Intra- and inter-assay coefficients of variation (CV) are percentages. NS denotes not stated. For longitudinal studies we report the average age of participants at first measurement. The censored total excludes any values >54.3 years (i.e. one standard deviation above the average age at menopause). Data taken from doi:10.1371/journal.pone.0022024.t001. View Large Figure 4 View largeDownload slide A validated model of serum AMH from conception to menopause. The red line is the model that best fits the 3260 datapoints shown as triangles. The coefficient of determination, r2, is 0.34, indicating that 34% of variation in serum AMH concentrations is due to age alone. Peak serum AMH is at 24.5 years. Reproduced from Kelsey et al. (2011); doi:10.1371/journal.pone.0022024. Figure 4 View largeDownload slide A validated model of serum AMH from conception to menopause. The red line is the model that best fits the 3260 datapoints shown as triangles. The coefficient of determination, r2, is 0.34, indicating that 34% of variation in serum AMH concentrations is due to age alone. Peak serum AMH is at 24.5 years. Reproduced from Kelsey et al. (2011); doi:10.1371/journal.pone.0022024. We have performed two analyses to compare NGF and AMH models, one from birth to the average age of peak serum AMH concentration (24.5 years), and a second from this age until average age at menopause (51 years). We found that from birth to peak AMH, NGF population is strongly negatively correlated with AMH (r = −0.93), since NGF populations are falling while AMH concentrations are rising in general (Fig. 5). Over the same age range AMH is positively but less closely correlated with rate of NGF recruitment (r = 0.52). Close relationships were also found between AMH concentrations and NGF populations from peak AMH until the menopause. From 24.5 to 51 years both NGF population and NGF recruitment correlate well and positively with AMH (r = 0.83 and r = 0.88, respectively). Figure 5 View largeDownload slide Comparison of serum AMH concentrations with NGF population and with NGF recruitment. The red line is the log-unadjusted validated AMH model (Kelsey et al., 2011), peaking at 24.5 years. The blue line denotes the decline in NGF population (Wallace and Kelsey, 2010), with peak population at 18–22 weeks gestation. The green line denotes the numbers of NGFs recruited towards maturation population (Wallace and Kelsey, 2010), with peak numbers lost at age 14.2 years on average. Each quantity has been normalized so that the peak occurs at 100%. Correlation coefficients (r) are given for AMH concentrations against the other two curves for birth to 24.5 years and for 24.5–51 years. Figure 5 View largeDownload slide Comparison of serum AMH concentrations with NGF population and with NGF recruitment. The red line is the log-unadjusted validated AMH model (Kelsey et al., 2011), peaking at 24.5 years. The blue line denotes the decline in NGF population (Wallace and Kelsey, 2010), with peak population at 18–22 weeks gestation. The green line denotes the numbers of NGFs recruited towards maturation population (Wallace and Kelsey, 2010), with peak numbers lost at age 14.2 years on average. Each quantity has been normalized so that the peak occurs at 100%. Correlation coefficients (r) are given for AMH concentrations against the other two curves for birth to 24.5 years and for 24.5–51 years. Our validated models have thus enabled us to examine the relationship of AMH with the dynamics of the primordial follicle pool. We can demonstrate that circulating AMH concentrations in adult women fall in line with the rate of loss of NGFs. AMH concentrations also broadly rise in line with this rate in children and young adults, but with a difference of about 10 years between the two peaks. The relationship of AMH to the total number of primordial follicles remaining in the pool for future recruitment therefore has two distinct phases, with the direction of association changing in early adulthood. In adult women serum AMH concentrations correlate with the size of the stereologically determined primordial follicle pool (Hansen et al., 2011), despite AMH production being restricted to later stages of follicular development. Similar data have been obtained in rodents (Kevenaar et al., 2006). The fate of the majority of NGFs (i.e. primordial and transitional stages) in the human is unclear. However, there is evidence that the greatest contribution to this is activation of growth, following which almost all follicles undergo atresia with only very few resulting in ovulation, rather than primordial follicles undergoing atresia directly (Gougeon, 1996). Thus the rate of loss of follicles from the non-growing population largely reflects the rate of initiation of follicle growth. It has recently been shown that initial follicular recruitment rises during childhood and adolescence (Wallace and Kelsey, 2010), despite an ongoing decline in the primordial follicle pool (Faddy et al., 1992; Faddy and Gosden, 1996; Hansen et al., 2008; Wallace and Kelsey, 2010). In this study, we have demonstrated that this increasing recruitment during childhood and adolescence is accompanied by a general increase in circulating AMH. In contrast, initial follicular recruitment rates decline in adult life (Wallace and Kelsey, 2010), as do circulating AMH concentrations (Kelsey et al., 2011) in accordance with previous cross-sectional and longitudinal analyses (de Vet et al., 2002; van Disseldorp et al., 2008; La Marca et al., 2005; Nelson et al., 2011a, b). These relationships between AMH and the age dependent changes in rate of follicular recruitment are entirely consistent with AMH being produced from the earliest stages of follicular development (Weenen et al., 2004). Increasing age and lower AMH both independently predict reduced oocyte (i.e. growing follicle) yield following ovulation induction for assisted conception (Nelson et al., 2007; La Marca et al., 2010), which is consistent both with this study and with the mathematical models suggesting declining follicular recruitment with increasing age (Hansen et al., 2008; Wallace and Kelsey, 2010). For these reasons, we speculate that the strong correlations between AMH and NGF dynamics do reflect a causal link, and hence that AMH can be used as an indirect approximation of age-related NGF populations and rates of recruitment. Our data intriguingly indicate a substantial distinction between peak follicle recruitment at puberty and peak AMH 10 years later. It is possible that this is an artefact reflecting the limitations of the data, as both sets are derived from cross-sectional sampling with limited numbers of young adults. However, if confirmed by more robustly designed studies, this may reflect that the organization of the adult follicle hierarchy within the ovary continues to mature after menarche as ovulatory cycles are initiated and become more prevalent over several years (Treloar et al., 1967; Flug et al., 1984). The finding that during childhood AMH tends to rise with time/age while NGF number falls does not mean that individuals with highest AMH will have lowest NGF number, although you might expect them to have lowest NGF recruitment based on the known biology of AMH (Durlinger et al., 2002). Speculatively, this might be associated with a later menopause (Broer et al., 2011) but our study finds the opposite, that there is a positive relationship between AMH and NGF recruitment during childhood. The relationship between AMH and NGF population seems to change in early adulthood. This may reflect the absence of larger AMH-producing follicles in childhood, which will change post-puberty thus causing a rise in AMH independent of NGF number. However, the relationship between AMH and NGF recruitment is positive both before and after the peak. Implications A model obtained by analysis of aggregated data serves as a series of reference points, which generate testable hypotheses. For our NGF study, the hypotheses are (i) that age is the dominating factor for variations in NGF populations [being responsible for about 80% of the variation for all ages from conception to menopause (Fig. 1), and for about 95% for ages up to 25 years (Fig. 2)], and (ii) that, in general, the recruitment of NGF towards maturation peaks at 13–15 years of age (Fig. 3). The hypotheses that arise from our AMH study are (i) that, in general, AMH levels in females fall shortly after birth, rise sharply during childhood before levelling out or even declining in pubertal years, thereafter rising to a peak level in the mid-twenties, and (ii) that about one-third of the variation in AMH levels for the healthy population is due to age alone (Fig. 4). By considering both models together, we can speculate that (i) NGF populations are strongly and negatively correlated with AMH from birth to peak AMH, (ii) for the same age range, NGF recruitment is positively correlated with AMH but much less strongly, and (iii) for ages from peak AMH to menopause both NGF population and NGF recruitment are strongly and positively correlated with AMH, with NGF recruitment having the higher rate of correlation (Fig. 5). Our NGF and AMH models are testable, either directly through histological analysis or indirectly through markers of ovarian reserve. Future in vitro estimates of NGF populations can be matched against the model, either confirming accuracy or identifying systematic bias. Future measurements of serum AMH can also be matched against the model, giving an indication of how well the model predicts unseen data. The results of these tests increase our understanding of the dynamics of human ovarian biology in either case. Evidence for inaccuracies in the models will lead to improved models; supporting evidence will lead to increased confidence in the models, and, in all likelihood, another set of testable hypotheses. In most studies involving the measurement of serum AMH, detection limits and inter- and intra-assay variations are reported as a matter of course, and are broadly similar across studies although the assay methodology also needs to be noted as there remains no recognized international standard. This supports the assumption that an aggregated dataset can be analysed as a single entity. However, the exclusion of data so that the final dataset approximates the general healthy population is not a simple process. Each paper has to be studied in detail to identify biases for inclusion or exclusion of subsets of the data. This process cannot be automated at any level: there is a requirement for intellectual input from experts in the field. This research methodology is limited to data that can be found and retrieved. Articles published in languages other than English, or published in older, more obscure, journals that have not yet been digitized will be unobtainable and/or uninterpretable. Naming conventions can also cause problems: AMH is referred to as MIS (Müllerian inhibiting substance) in many publications. Often data are summarized by descriptive statistics, as opposed to being reported in tabular or chart form. There are well-known techniques for the re-creation of datasets from summary descriptions, but another validation step is needed before such data can be used as a reasonable approximation to the data from the original study. The overall implication is that studies based on aggregated data add to our collective ability to accurately model ovarian reserve using both direct and indirect factors. The models and correlations obtained from these studies provide hypotheses that can be tested by traditional studies. Moreover, the results from data aggregation studies provide the estimates of event sizes needed to determine how many subjects are needed to test the hypotheses to a reasonable level of statistical significance. The resulting scientific process then becomes iterative: data from hypothesis-led studies are combined and analysed, the results lead to further hypothesis-led studies, the data from which can be aggregated anew. Authors' roles All authors contributed extensively to the work presented in this paper. Each author contributed to the data aggregation phase, the assessment of data quality, the modelling and the writing of the manuscript. All authors discussed the results and implications and commented on the manuscript at all stages. Funding This work was supported by UK Engineering & Physical Sciences Research Council grant EP/H004092/1. References Baker TG.  A quantitative and cytological study of germ cells in human ovaries,  Proc R Soc Lond B Biol Sci ,  1963, vol.  158 (pg.  417- 433)  Containing papers of a Biological character. Royal Society (Great Britain) Google Scholar CrossRef Search ADS PubMed  Bendsen E,  Byskov AG,  Andersen CY,  Westergaard LG.  Number of germ cells and somatic cells in human fetal ovaries during the first weeks after sex differentiation,  Hum Reprod ,  2006, vol.  21 (pg.  30- 35) Google Scholar CrossRef Search ADS PubMed  Block E.  Quantitative morphological investigations of the follicular system in women; variations at different ages,  Acta Anat ,  1952, vol.  14 (pg.  108- 123) Google Scholar CrossRef Search ADS PubMed  Block E.  A quantitative morphological investigation of the follicular system in newborn female infants,  Acta Anat ,  1953, vol.  17 (pg.  201- 206) Google Scholar CrossRef Search ADS PubMed  Brett S,  Bee N,  Wallace WHB,  Rajkhowa M,  Kelsey TW.  Individual ovarian volumes obtained from 2-dimensional and 3-dimensional ultrasound lack precision,  Reprod Biomed Online ,  2009, vol.  18 (pg.  348- 351) Google Scholar CrossRef Search ADS PubMed  Broer SL,  Eijkemans MJ,  Scheffer GJ,  van Rooij IA,  de Vet A,  Themmen AP,  Laven JS,  de Jong FH,  Te Velde ER,  Fauser BC, et al.  Anti-müllerian hormone predicts menopause: a long-term follow-up study in normoovulatory women,  J Clin Endocrinol Metab ,  2011, vol.  96 (pg.  2532- 2539) Google Scholar CrossRef Search ADS PubMed  Chellakooty M,  Schmidt IM,  Haavisto AM,  Boisen KA,  Damgaard IN,  Mau C,  Petersen JH,  Juul A,  Skakkebaek NE,  Main KM.  Inhibin A, inhibin B, follicle-stimulating hormone, luteinizing hormone, estradiol, and sex hormone-binding globulin levels in 473 healthy infant girls,  J Clin Endocrinol Metab ,  2003, vol.  88 (pg.  3515- 3520) Google Scholar CrossRef Search ADS PubMed  de Vet A,  Laven JS,  de Jong FH,  Themmen AP,  Fauser BC.  Antimüllerian hormone serum levels: a putative marker for ovarian aging,  Fertil Steril ,  2002, vol.  77 (pg.  357- 362) Google Scholar CrossRef Search ADS PubMed  Durlinger AL,  Gruijters MJ,  Kramer P,  Karels B,  Ingraham HA,  Nachtigal MW,  Uilenbroek JT,  Grootegoed JA,  Themmen AP.  Anti-Müllerian hormone inhibits initiation of primordial follicle growth in the mouse ovary,  Endocrinology ,  2002, vol.  143 (pg.  1076- 1084) Google Scholar PubMed  Faddy MJ,  Gosden RG.  A model conforming the decline in follicle numbers to the age of menopause in women,  Hum Reprod ,  1996, vol.  11 (pg.  1484- 1486) Google Scholar CrossRef Search ADS PubMed  Faddy MJ,  Gosden RG,  Gougeon A,  Richardson SJ,  Nelson JF.  Accelerated disappearance of ovarian follicles in mid-life: implications for forecasting menopause,  Hum Reprod ,  1992, vol.  7 (pg.  1342- 1346) Google Scholar PubMed  Flug D,  Largo RH,  Prader A.  Menstrual patterns in adolescent Swiss girls: a longitudinal study,  Ann Hum Biol ,  1984, vol.  11 (pg.  495- 508) Google Scholar CrossRef Search ADS PubMed  Forabosco A,  Sforza C.  Establishment of ovarian reserve: a quantitative morphometric study of the developing human ovary,  Fertil Steril ,  2007, vol.  88 (pg.  675- 683) Google Scholar CrossRef Search ADS PubMed  Gougeon A.  Regulation of ovarian follicular development in primates: facts and hypotheses,  Endocr Rev ,  1996, vol.  17 (pg.  121- 155) Google Scholar CrossRef Search ADS PubMed  Hagen CP,  Aksglaede L,  Sørensen K,  Main KM,  Boas M,  Cleemann L,  Holm K,  Gravholt CH,  Andersson AM,  Pedersen AT, et al.  Serum Levels of Anti-Müllerian Hormone as a Marker of Ovarian Function in 926 Healthy Females from Birth to Adulthood and in 172 Turner Syndrome Patients,  J Clin Endocrinol Metab ,  2010, vol.  95 (pg.  1- 8) Google Scholar CrossRef Search ADS PubMed  Hansen KR,  Knowlton NS,  Thyer AC,  Charleston JS,  Soules MR,  Klein NA.  A new model of reproductive aging: the decline in ovarian non-growing follicle number from birth to menopause,  Hum Reprod ,  2008, vol.  23 (pg.  699- 708) Google Scholar CrossRef Search ADS PubMed  Hansen KR,  Hodnett GM,  Knowlton N,  Craig LB.  Correlation of ovarian reserve tests with histologically determined primordial follicle number,  Fertil Steril ,  2011, vol.  95 (pg.  170- 175) Google Scholar CrossRef Search ADS PubMed  He C,  Kraft P,  Chasman DI,  Buring JE,  Chen C,  Hankinson SE,  Paré G,  Chanock S,  Ridker PM,  Hunter DJ.  A large-scale candidate gene association study of age at menarche and age at natural menopause,  Hum Genet ,  2010, vol.  128 (pg.  515- 527) Google Scholar CrossRef Search ADS PubMed  Hehenkamp WJ,  Looman CW,  Themmen AP,  de Jong FH,  Te Velde ER,  Broekmans FJ.  Anti-Müllerian hormone levels in the spontaneous menstrual cycle do not show substantial fluctuation,  J Clin Endocrinol Metab ,  2006, vol.  91 (pg.  4057- 4063) Google Scholar CrossRef Search ADS PubMed  Kelsey TW,  Wright P,  Nelson SM,  Anderson RA,  Wallace WH.  A validated model of serum anti-Müllerian hormone from conception to menopause,  PLoS ONE ,  2011, vol.  6 pg.  e22024  Google Scholar CrossRef Search ADS PubMed  Kevenaar ME,  Meerasahib MF,  Kramer P,  van de Lang-Born BM,  de Jong FH,  Groome NP,  Themmen AP,  Visser JA.  Serum anti-müllerian hormone levels reflect the size of the primordial follicle pool in mice,  Endocrinology ,  2006, vol.  147 (pg.  3228- 3234) Google Scholar CrossRef Search ADS PubMed  La Marca A,  De Leo V,  Giulini S,  Orvieto R,  Malmusi S,  Giannella L,  Volpe A.  Anti-Müllerian hormone in premenopausal women and after spontaneous or surgically induced menopause,  J Soc Gynecol Investig ,  2005, vol.  12 (pg.  545- 548) Google Scholar CrossRef Search ADS PubMed  La Marca A,  Sighinolfi G,  Radi D,  Argento C,  Baraldi E,  Artenisio AC,  Stabile G,  Volpe A.  Anti-Müllerian hormone (AMH) as a predictive marker in assisted reproductive technology (ART),  Hum Reprod Update ,  2010, vol.  16 (pg.  113- 130) Google Scholar CrossRef Search ADS PubMed  Lee MM.  Reproductive hormones in infant girls—a harbinger of adult reproductive function?,  J Clin Endocrinol Metab ,  2003, vol.  88 (pg.  3513- 3514) Google Scholar CrossRef Search ADS PubMed  Mamsen LS,  Lutterodt MC,  Andersen EW,  Byskov AG,  Andersen CY.  Germ cell numbers in human embryonic and fetal gonads during the first two trimesters of pregnancy: analysis of six published studies,  Hum Reprod ,  2011, vol.  26 (pg.  2140- 2145) Google Scholar CrossRef Search ADS PubMed  Matzuk MM,  Lamb DJ.  The biology of infertility: research advances and clinical challenges,  Nat Med ,  2008, vol.  14 (pg.  1197- 1213) Google Scholar CrossRef Search ADS PubMed  Nelson SM,  Yates RW,  Fleming R.  Serum anti-Müllerian hormone and FSH: prediction of live birth and extremes of response in stimulated cycles - implications for individualization of therapy,  Hum Reprod ,  2007, vol.  22 (pg.  2414- 2421) Google Scholar CrossRef Search ADS PubMed  Nelson SM,  Messow MC,  Wallace AM,  Fleming R,  McConnachie A.  Nomogram for the decline in serum antimüllerian hormone: a population study of 9,601 infertility patients,  Fertil Steril ,  2011, vol.  95 (pg.  736- 741) Google Scholar CrossRef Search ADS PubMed  Nelson SM,  Messow MC,  McConnachie A,  Wallace WHB,  Kelsey TW,  Fleming R,  Anderson RA,  Leader B.  External validation of nomogram for the decline in serum anti-Müllerian hormone in women: a population study of 15,834 infertility patients,  Reprod Biomed Online ,  2011, vol.  23 (pg.  204- 206) Google Scholar CrossRef Search ADS PubMed  Ong KK,  Elks CE,  Li S,  Zhao JH,  Luan J,  Andersen LB,  Bingham SA,  Brage S,  Smith GD,  Ekelund U, et al.  Genetic variation in LIN28B is associated with the timing of puberty,  Nat Genet ,  2009, vol.  41 (pg.  729- 733) Google Scholar CrossRef Search ADS PubMed  Richardson SJ,  Senikas V,  Nelson JF.  Follicular depletion during the menopausal transition: evidence for accelerated loss and ultimate exhaustion,  J Clin Endocrinol Metab ,  1987, vol.  65 (pg.  1231- 1237) Google Scholar CrossRef Search ADS PubMed  Tingen CM,  Bristol-Gould SK,  Kiesewetter SE,  Wellington JT,  Shea L,  Woodruff TK.  Prepubertal primordial follicle loss in mice is not due to classical apoptotic pathways,  Biol Reprod ,  2009, vol.  81 (pg.  16- 25) Google Scholar CrossRef Search ADS PubMed  Treloar AE,  Boynton RE,  Behn BG,  Brown BW.  Variation of the human menstrual cycle through reproductive life,  Int J Fertil ,  1967, vol.  12  1 Pt 2(pg.  77- 126) Google Scholar PubMed  van Disseldorp J,  Faddy MJ,  Themmen AP,  de Jong FH,  Peeters PH,  van der Schouw YT,  Broekmans FJ.  Relationship of serum antimüllerian hormone concentration to age at menopause,  J Clin Endocrinol Metab ,  2008, vol.  93 (pg.  2129- 2134) Google Scholar CrossRef Search ADS PubMed  Wallace WH,  Kelsey TW.  Human ovarian reserve from conception to the menopause,  PloS ONE ,  2010, vol.  5 pg.  e8772  Google Scholar CrossRef Search ADS PubMed  Weenen C,  Laven JS,  Von Bergh AR,  Cranfield M,  Groome NP,  Visser JA,  Kramer P,  Fauser BC,  Themmen AP.  Anti-Müllerian hormone expression pattern in the human ovary: potential implications for initial and cyclic follicle recruitment,  Molecular Hum Reprod ,  2004, vol.  10 (pg.  77- 83) Google Scholar CrossRef Search ADS   © The Author 2011. Published by Oxford University Press on behalf of the European Society of Human Reproduction and Embryology. All rights reserved. For Permissions, please email: journals.permissions@oup.com http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Molecular Human Reproduction Oxford University Press

Data-driven assessment of the human ovarian reserve

Loading next page...
 
/lp/oxford-university-press/data-driven-assessment-of-the-human-ovarian-reserve-a6zvFQ4ggs

References (44)

Publisher
Oxford University Press
Copyright
© The Author 2011. Published by Oxford University Press on behalf of the European Society of Human Reproduction and Embryology. All rights reserved. For Permissions, please email: journals.permissions@oup.com
Subject
New Research Horizon Reviews
ISSN
1360-9947
eISSN
1460-2407
DOI
10.1093/molehr/gar059
pmid
21933846
Publisher site
See Article on Publisher Site

Abstract

Abstract Human ovarian physiology is still poorly understood, with the factors and mechanisms that control initiation of follicular recruitment and loss remaining particularly unclear. Conventional hypothesis-led studies provide new data, results and insights, but datasets from individual studies are often small, allowing only limited interpretation. Great power is afforded by the aggregation of data from multiple studies into single datasets. In this paper, we describe how modern computational analysis of these datasets provides important new insights into ovarian function and has generated hypotheses that are testable in the laboratory. Specifically, we can hypothesize that age is the most important factor for variations in individual ovarian non-growing follicle (NGF) populations, that anti-Müllerian hormone (AMH) levels generally rise and fall in childhood years before peaking in the mid-twenties, and that there are strong correlations between AMH levels and both NGF populations and rates of recruitment towards maturation, for age ranges before and after peak AMH levels. ovary, fertility, AMH, oocyte, computational analysis Introduction The ovarian reserve may be defined as the remaining pool of non-growing follicles (NGFs) in the ovary at a given age. NGFs are formed in large numbers in the fetal ovary in humans with peak population occurring at 20–22 weeks gestation (Baker, 1963; Wallace and Kelsey, 2010; Mamsen et al., 2011). The population decreases from this peak, largely due to atresia following follicular recruitment towards maturation. The post-menopausal ovary contains fewer than 1000 NGFs, an insufficient number to support recruitment to ovulation. The study of ovarian function throughout life is important for biological, medical and sociological reasons. Human ovarian physiology is still poorly understood, with the factors and mechanisms that initiate and control follicular recruitment and loss through apoptosis remaining unclear. In medical terms, a more complete understanding of ovarian biology is of fundamental importance in assisted conception, and in the assessment of those young woman with cancer and other conditions who may benefit from fertility preservation. In modern western societies, there is an increasing trend towards women starting a family later in life when ovarian reserve is reduced and the menopause is approaching. A better and more complete understanding of how to estimate ovarian reserve for the individual woman would help inform her choices about the timing of having a family. Delayed child bearing in women who have a reduced ovarian reserve is now a major burden on reproductive services and may result in significant emotional and psychological distress if treatment is unsuccessful. There is a good reason for the paucity of knowledge of human ovarian reserve throughout life: direct longitudinal assessment is currently impossible, and is likely to remain impossible for the foreseeable future. No in vivo technique for counting NGFs exists. All studies involving the estimation of NGF populations for ovaries at various chronological ages have analysed tissue post-mortem or post-oophorectomy (Block, 1952, 1953; Baker, 1963; Richardson et al., 1987; Bendsen et al., 2006; Forabosco and Sforza, 2007; Hansen et al., 2008). These estimates are therefore cross-sectional snapshots of populations at known ages, providing no opportunity to assess rates of NGF loss in the subjects, and hence making highly tentative any attempt to describe population trajectories at different ages and stages of reproductive life. A partial solution to this fundamental problem is to consider indirect indicators of ovarian reserve. Endocrine factors, such as anti-Müllerian hormone (AMH) and FSH, are measurable by available serum assays. Physical factors, such as ovarian volume (OV) and antral follicle counts, are measurable by transvaginal sonography. These measurements are inexpensive, non-destructive and repeatable, allowing the possibility of large-scale longitudinal studies. But the key research question remains: how are these measurable indirect indicators related to NGF populations for individuals of known age? Subsidiary questions include the relationships (if any) between indirect factors and NGF recruitment towards maturation, and which combinations of indirect factors (if any) best predict ovarian reserve for which age ranges. The classical approach to this type of research question is to (i) generate a hypothesis, (ii) design an experiment, (iii) collect data and (iv) analyse and discuss. Examples include (Hehenkamp et al., 2006) and (Brett et al., 2009) which considered variations in AMH across menstrual cycles and precision of OV measurement by 2D and 3D ultrasound, respectively. Another common approach is to (i) recruit a cohort of subjects, (ii) collect data, (iii) report and discuss reference ranges, correlations and centiles. High-quality exemplars include (Hansen et al., 2008) and (Hagen et al., 2010), in which NGF populations and serum AMH levels were obtained post-oophorectomy and from juvenile and adolescent volunteers, respectively. Both types of study are crucial for an incremental increase in our understanding of ovarian reserve. They are the current gold standard for scientific endeavour in this field. Another approach is to use the new genetics including micro-array analysis. This methodology has not addressed NGF numbers in individual women thus far, but has provided both detailed analysis of the involvement of key genes in determining premature ovarian insufficiency (Matzuk and Lamb, 2008), and, more recently, large genome-wide association studies addressing the genetic determinants of age of puberty and menopause (Ong et al., 2009; He et al., 2010). Individual studies will typically have the advantages of standardized data acquisition and cohort selection, but the disadvantages of a restricted cohort age range and non-standard data analysis techniques. In this paper, we report initial results for, and discuss the potential of, an alternative scientific approach: the systematic aggregation and analysis of data from several existing studies. We make no claim of originality for this approach: researchers have successfully aggregated data from distinct NGF population studies to produce the first plausible models of NGF decline from birth (Faddy et al., 1992; Faddy and Gosden, 1996). Our aim is to extend this technique, making full use of recent advances in data identification, data retrieval and data analysis. New developments In recent years the biomedical research environment has changed markedly. There has been an exponential increase in the amount of digitized data archived in searchable repositories. Once extracted, the data can be re-analysed using more modern techniques (often using more computationally intensive methods than those available at time of original publication), and datasets from multiple studies can be combined for analysis as a single dataset. The increase in data availability has been matched by enhanced computational capability. Processing power has roughly doubled every 2 years since 1958, and this trend is expected to continue at least until 2015. Many of the algorithms used to model biomedical data are more than 20 years old, but improved processing power means that not only is it possible to analyse larger and more complex datasets, but also that the traditional method of fitting to a theoretically derived model can be replaced by fitting several hundred models and then ranking these by goodness-of-fit. Moreover, increases in computational power allow us to run sophisticated model validation algorithms in those cases where validation by an external dataset is not practical. These algorithms typically involve the repeated generation of models using some of the data, followed by measurement of model error for data that was held back. This has the advantages that small datasets need not be split into test and training sets (thereby compromising predictive power), and we can guard against overfitting by analysis of the trade-off between goodness-of-fit and generalization error. Taken together, these developments motivate and enable systematic and controlled data aggregation and univariate analysis. Notably this approach has clear similarities with that of individualized patient data meta-analyses—the major strength of which is the ability to adjust for confounders in multivariate analyses, however, the lack of provision of the original raw data can bias the included studies and conclusions. Consequently, the structure of a data-led research project might now be (i) perform a structured search for all publications relating to an indicator of ovarian reserve, (ii) import the publications and perform an iterated search based on citation information, (iii) extract raw data from selected publications and combine into a single dataset, (iv) apply knowledge-discovery techniques on the data to produce internally validated models and correlations, (v) report and discuss. Application to the study of the NGF population In an example of the use of this methodology, eight studies were identified that estimated NGF populations for limited age ranges (Table I). Each study used a variation on the standard method for NGF estimation devised by Block in the 1950s: manual stereological counts of NGFs appearing in a small subset of ovarian tissue are integrated to obtain an estimate of the population for the entire ovary. Since the data were collected in a similar manner, they were collated into a single dataset (n= 325) with age range −0.6 through 51.0 years (median 31.0 years). Imaginary but biologically accurate datapoints were added that forced the population at conception to be zero, and non-linear peak models were fitted to the data. After a validation process involving fitting the same models to randomly selected subsets of the data, the best model in terms of goodness-of-fit was reported (Fig. 1; Wallace and Kelsey, 2010). This model is the first that attempts to describe in quantitative terms the dynamics of ovarian reserve for all ages up to menopause. Confidence intervals (CIs) for the model were obtained, together with prediction limits that define normal NGF population ranges. Table I The eight quantitative histological studies forming the combined dataset for the NGF model. Study  Statistics   Number  First author  Year  No. ovaries  Min. age  Max. age  Median age  1  Bendsen  2006  11  −0.6  −0.6  −0.6  2  Baker  1963  11  −0.6  7.0  −0.2  3  Forabasco  2007  15  −0.5  0.5  −0.3  4  Block  1953  19  −0.2  0.0  0.0  5  Hansen  2008  122  0.1  51.0  38.0  6  Block  1951  86  6.0  44.0  28.0  7  Gougeon  1987  52  25.0  46.0  39.5  8  Richardson  1987  9  45.0  51.0  46.0  Overall      325  −0.6  51.0  32.0  Study  Statistics   Number  First author  Year  No. ovaries  Min. age  Max. age  Median age  1  Bendsen  2006  11  −0.6  −0.6  −0.6  2  Baker  1963  11  −0.6  7.0  −0.2  3  Forabasco  2007  15  −0.5  0.5  −0.3  4  Block  1953  19  −0.2  0.0  0.0  5  Hansen  2008  122  0.1  51.0  38.0  6  Block  1951  86  6.0  44.0  28.0  7  Gougeon  1987  52  25.0  46.0  39.5  8  Richardson  1987  9  45.0  51.0  46.0  Overall      325  −0.6  51.0  32.0  Reproduced from Wallace and Kelsey (2010). All ages are in years; the year column refers to the year of publication. Data taken from doi:10.1371/journal.pone.0008772.t001. View Large Figure 1 View largeDownload slide The Wallace–Kelsey model of NGF populations from conception to menopause. The best model for the establishment of the NGF population after conception, and the subsequent decline until age at menopause is described by an Asymmetric Double-Gaussian Cumulative (ADC) model with parameters amplitude = 5.56 (95% CI 5.38–5.74), centre = 25.6 (95% CI 24.9–26.4), width = 52.7 (95% CI 51.1–54.2), growth shape = 0.074 (95% CI 0.062–0.085) and decline shape = 24.5 (95% CI 20.4–28.6). Our model has correlation coefficient r2 = 0.81, fit standard error = 0.46 and F-value = 364. The figure shows the dataset (n= 325), the model, the 95% prediction limits of the model and the 95% CI for the model. The horizontal axis denotes age in months up to birth at age zero, and age in years from birth to 51 years. Reproduced from Wallace and Kelsey (2010); doi:10.1371/journal.pone.0008772. Figure 1 View largeDownload slide The Wallace–Kelsey model of NGF populations from conception to menopause. The best model for the establishment of the NGF population after conception, and the subsequent decline until age at menopause is described by an Asymmetric Double-Gaussian Cumulative (ADC) model with parameters amplitude = 5.56 (95% CI 5.38–5.74), centre = 25.6 (95% CI 24.9–26.4), width = 52.7 (95% CI 51.1–54.2), growth shape = 0.074 (95% CI 0.062–0.085) and decline shape = 24.5 (95% CI 20.4–28.6). Our model has correlation coefficient r2 = 0.81, fit standard error = 0.46 and F-value = 364. The figure shows the dataset (n= 325), the model, the 95% prediction limits of the model and the 95% CI for the model. The horizontal axis denotes age in months up to birth at age zero, and age in years from birth to 51 years. Reproduced from Wallace and Kelsey (2010); doi:10.1371/journal.pone.0008772. This model can then be used to assess NGF recruitment, i.e. the rate of loss of NGFs from the pool—bearing in mind that this does not distinguish between NGF activation and loss for other reasons, the degree and mechanisms of which are debated (Gougeon, 1996; Tingen et al., 2009). Analysis of the model showed that NGF recruitment peaks at 14.2 years in the average case. If it is assumed that NGF population dynamics are similar for the general population (i.e. that the shape of the curve is similar for individuals having different peak values and ages at menopause) then NGF recruitment peaks at an average age 13–15 years in normal women (Fig. 2; Wallace and Kelsey, 2010). This is a research hypothesis derived from the detailed examination of aggregated data, and is potentially testable by reproductive biologists in the laboratory. Figure 2 View largeDownload slide NGF model from conception to 25 years. The best model for the establishment of the NGF population after conception, and the subsequent decline until 25 years of age is described by an ADC model with parameters amplitude = 5.79 (95% CI 5.03–6.55), centre = 28.0 (95% CI 15.8–40.2), width = 57.4 (95% CI 33.1–81.8), growth shape = 0.074 (95% CI 0.067–0.081) and decline shape = 34.3 (95% CI −4.2 to 72.8). This model has correlation coefficient = 0.95, fit standard error = 0.29 and F-value = 585. This figure shows the dataset (n= 126), the model, the 95% prediction limits of the model and the 95% CI for the model. The horizontal axis denotes age in months up to birth at age zero, and age in years from birth to 25 years. Reproduced from Wallace and Kelsey (2010); doi:10.1371/journal.pone.0008772. Figure 2 View largeDownload slide NGF model from conception to 25 years. The best model for the establishment of the NGF population after conception, and the subsequent decline until 25 years of age is described by an ADC model with parameters amplitude = 5.79 (95% CI 5.03–6.55), centre = 28.0 (95% CI 15.8–40.2), width = 57.4 (95% CI 33.1–81.8), growth shape = 0.074 (95% CI 0.067–0.081) and decline shape = 34.3 (95% CI −4.2 to 72.8). This model has correlation coefficient = 0.95, fit standard error = 0.29 and F-value = 585. This figure shows the dataset (n= 126), the model, the 95% prediction limits of the model and the 95% CI for the model. The horizontal axis denotes age in months up to birth at age zero, and age in years from birth to 25 years. Reproduced from Wallace and Kelsey (2010); doi:10.1371/journal.pone.0008772. A second finding based on selective analysis of the aggregated data is that age is the dominant factor for differences in individual NGF populations for ages up to 25 (Fig. 3; Wallace and Kelsey, 2010). The coefficient of determination for this subset of the data is 0.95, indicating that 95% of individual variation is due to age alone. For ages from conception to menopause, this falls to 81%, despite the dataset being larger and thus expected to be more robust, indicating that (i) lifestyle and genetic factors are more important later in life, and (ii) age remains the dominant factor that explains population variations. Figure 3 View largeDownload slide Rates of NGF recruitment towards maturation. Each sub-figure describes the absolute number of NGFs recruited per month, for ages from birth to 55 years, based on population decline predicted by the ADC model. The red curve denotes recruitment for individuals whose decline is in line with the average age at menopause; maximum recruitment of 880 follicles per month occurs at 14 years 2 months. The green curve denotes recruitment for individuals whose decline is in line with early age at menopause (the lower 95% prediction limit of the model); maximum recruitment of 104 follicles per month occurs at 14 years 2 months. The yellow curve denotes recruitment in line with late age at menopause (the upper 95% prediction limit of the model); maximum recruitment of 7520 follicles per month occurs at 14 years 2 months. Reproduced from Wallace and Kelsey (2010); doi:10.1371/journal.pone.0008772. Figure 3 View largeDownload slide Rates of NGF recruitment towards maturation. Each sub-figure describes the absolute number of NGFs recruited per month, for ages from birth to 55 years, based on population decline predicted by the ADC model. The red curve denotes recruitment for individuals whose decline is in line with the average age at menopause; maximum recruitment of 880 follicles per month occurs at 14 years 2 months. The green curve denotes recruitment for individuals whose decline is in line with early age at menopause (the lower 95% prediction limit of the model); maximum recruitment of 104 follicles per month occurs at 14 years 2 months. The yellow curve denotes recruitment in line with late age at menopause (the upper 95% prediction limit of the model); maximum recruitment of 7520 follicles per month occurs at 14 years 2 months. Reproduced from Wallace and Kelsey (2010); doi:10.1371/journal.pone.0008772. Application to the study of serum AMH in healthy females In a recent aggregation study we found 25 studies reporting normative levels of serum AMH for cohorts with limited age ranges. We aggregated the data extracted from these studies with our own data, giving a total dataset of 12 241 AMH measurements for ages from −0.3 to 68 years. Many data points were associated with women attending infertility clinics (for whom normal ovarian reserve indicators could not be assumed) or with individuals with chronic disease, such as lupus or cancer. After careful exclusion of these data, a single dataset of healthy women (n= 3260, −0.3 through 54.3 years, median 28.3 years) was obtained from 20 separate sources (Table II; Kelsey et al., 2011). After fixing AMH levels to zero at conception, a range of non-linear models were fitted to this data, with a cross-validation stage to guard against reporting a model that provides a good fit to the data but which is unlikely to generalize well to unseen data. The best model, in terms of validated goodness-of-fit, is the first quantitative model for AMH levels for the normal population from conception to post-menopausal ages (Fig. 4; Kelsey et al., 2011). As for the NGF population model, the AMH model both supports existing hypotheses and identifies new hypotheses. Consistent with previous reports (Hagen et al., 2010) our model shows that serum AMH falls shortly after birth in girls, with concentrations only increasing again after about 2 years of age. This feature is in line with evidence of a mini-puberty seen in neonatal girls (Chellakooty et al., 2003; Lee, 2003). Secondly, our model demonstrated a non-linear decline from age 24.5, with 99% correlation to the decline reported for large infertility cohorts (n= 25 435; Nelson et al., 2011a, b). Table II Serum AMH data summary. Ref.  First author  Data  Assay  n  Average age  Age range  Detection limits  Intra CV  Inter CV  [35]  Soto  Graph  IBC  58  30.3  ±8.7  0.10  5.3  8.7  [38]  Guibourdenche  Graph  IBC  192  NS  −0.3 to 1.0  0.30  5.3  8.7  [39]  Hudecova  Graph  IBC  64  46.3  ±6.4  0.70  12.3  12.3  [40]  Mulders  Graph  IBC  82  29.9  20–36  NS  5.0  8.0  [41]  Pastor  Graph  IBC  42  NS  18–50  0.10  5.3  7.8  [42]  Piltonen  Graph  IBC  44  31.6  21–44  NS  5.1  6.6  [20]  van Rooilj  Graph  IBC  162  NS  25–46  0.05  5.0  8.0  [43]  Laven  Graph  IBC  41  NS  20–36  0.05  5.0  8.0  [19]  de Vet  Graph  IBC  82  29.0  ±4.0  0.05  5.0  8.0  [44]  Knauf  Graph  IBC  83  34.2  ±3.4  0.03  11.0  11.0  [45]  Lee  Graph  IBC  225  NS  0–51  0.5  9.0  15.0  [36]  La Marca  Graph  IBC  24  44.0  ±2.8  0.24  5.0  8.0  [29]  Hagen  Graph  IBC  891  NS  0–68  0.03  7.8  11.6  [46]  van Beek  Graph  DSL  82  29.0  20–35  NS  5.0  15.0  [47]  Sanders  Graph  DSL  43  24.1  0–51  0.01  NS  11.4  [34]  van Disseldorp  Graph  DSL  144  37.9  25–46  0.03  11.0  11.0  [48]  Tehrani  Graph  DSL  267  27.1  16–44  0.01  5.2  9.1  [49]  Dorgan  Graph  DSL  204  44.7  33–55  0.06  8.0  8.0  [30]  Ahmed  Raw  DSL  128  8.5  1–17  0.50  8.0  8.0  [25]  Nelson  Raw  DSL  441  36.1  22–48  0.03  3.4  8.6    Total IBC      1990    −0.3 to 68          Total DSL      1309    0–55          Total n      3299    −0.3 to 68          Censored total n      3260    −0.3 to 54.3        Ref.  First author  Data  Assay  n  Average age  Age range  Detection limits  Intra CV  Inter CV  [35]  Soto  Graph  IBC  58  30.3  ±8.7  0.10  5.3  8.7  [38]  Guibourdenche  Graph  IBC  192  NS  −0.3 to 1.0  0.30  5.3  8.7  [39]  Hudecova  Graph  IBC  64  46.3  ±6.4  0.70  12.3  12.3  [40]  Mulders  Graph  IBC  82  29.9  20–36  NS  5.0  8.0  [41]  Pastor  Graph  IBC  42  NS  18–50  0.10  5.3  7.8  [42]  Piltonen  Graph  IBC  44  31.6  21–44  NS  5.1  6.6  [20]  van Rooilj  Graph  IBC  162  NS  25–46  0.05  5.0  8.0  [43]  Laven  Graph  IBC  41  NS  20–36  0.05  5.0  8.0  [19]  de Vet  Graph  IBC  82  29.0  ±4.0  0.05  5.0  8.0  [44]  Knauf  Graph  IBC  83  34.2  ±3.4  0.03  11.0  11.0  [45]  Lee  Graph  IBC  225  NS  0–51  0.5  9.0  15.0  [36]  La Marca  Graph  IBC  24  44.0  ±2.8  0.24  5.0  8.0  [29]  Hagen  Graph  IBC  891  NS  0–68  0.03  7.8  11.6  [46]  van Beek  Graph  DSL  82  29.0  20–35  NS  5.0  15.0  [47]  Sanders  Graph  DSL  43  24.1  0–51  0.01  NS  11.4  [34]  van Disseldorp  Graph  DSL  144  37.9  25–46  0.03  11.0  11.0  [48]  Tehrani  Graph  DSL  267  27.1  16–44  0.01  5.2  9.1  [49]  Dorgan  Graph  DSL  204  44.7  33–55  0.06  8.0  8.0  [30]  Ahmed  Raw  DSL  128  8.5  1–17  0.50  8.0  8.0  [25]  Nelson  Raw  DSL  441  36.1  22–48  0.03  3.4  8.6    Total IBC      1990    −0.3 to 68          Total DSL      1309    0–55          Total n      3299    −0.3 to 68          Censored total n      3260    −0.3 to 54.3        Reproduced from Kelsey et al. (2011). The references relate to the bibliography section of the original paper. Age information is given as median and range, or as mean ± SD depending on which form was reported in the referenced study. Detection limits are given in nanogram/millilitre. Intra- and inter-assay coefficients of variation (CV) are percentages. NS denotes not stated. For longitudinal studies we report the average age of participants at first measurement. The censored total excludes any values >54.3 years (i.e. one standard deviation above the average age at menopause). Data taken from doi:10.1371/journal.pone.0022024.t001. View Large Figure 4 View largeDownload slide A validated model of serum AMH from conception to menopause. The red line is the model that best fits the 3260 datapoints shown as triangles. The coefficient of determination, r2, is 0.34, indicating that 34% of variation in serum AMH concentrations is due to age alone. Peak serum AMH is at 24.5 years. Reproduced from Kelsey et al. (2011); doi:10.1371/journal.pone.0022024. Figure 4 View largeDownload slide A validated model of serum AMH from conception to menopause. The red line is the model that best fits the 3260 datapoints shown as triangles. The coefficient of determination, r2, is 0.34, indicating that 34% of variation in serum AMH concentrations is due to age alone. Peak serum AMH is at 24.5 years. Reproduced from Kelsey et al. (2011); doi:10.1371/journal.pone.0022024. We have performed two analyses to compare NGF and AMH models, one from birth to the average age of peak serum AMH concentration (24.5 years), and a second from this age until average age at menopause (51 years). We found that from birth to peak AMH, NGF population is strongly negatively correlated with AMH (r = −0.93), since NGF populations are falling while AMH concentrations are rising in general (Fig. 5). Over the same age range AMH is positively but less closely correlated with rate of NGF recruitment (r = 0.52). Close relationships were also found between AMH concentrations and NGF populations from peak AMH until the menopause. From 24.5 to 51 years both NGF population and NGF recruitment correlate well and positively with AMH (r = 0.83 and r = 0.88, respectively). Figure 5 View largeDownload slide Comparison of serum AMH concentrations with NGF population and with NGF recruitment. The red line is the log-unadjusted validated AMH model (Kelsey et al., 2011), peaking at 24.5 years. The blue line denotes the decline in NGF population (Wallace and Kelsey, 2010), with peak population at 18–22 weeks gestation. The green line denotes the numbers of NGFs recruited towards maturation population (Wallace and Kelsey, 2010), with peak numbers lost at age 14.2 years on average. Each quantity has been normalized so that the peak occurs at 100%. Correlation coefficients (r) are given for AMH concentrations against the other two curves for birth to 24.5 years and for 24.5–51 years. Figure 5 View largeDownload slide Comparison of serum AMH concentrations with NGF population and with NGF recruitment. The red line is the log-unadjusted validated AMH model (Kelsey et al., 2011), peaking at 24.5 years. The blue line denotes the decline in NGF population (Wallace and Kelsey, 2010), with peak population at 18–22 weeks gestation. The green line denotes the numbers of NGFs recruited towards maturation population (Wallace and Kelsey, 2010), with peak numbers lost at age 14.2 years on average. Each quantity has been normalized so that the peak occurs at 100%. Correlation coefficients (r) are given for AMH concentrations against the other two curves for birth to 24.5 years and for 24.5–51 years. Our validated models have thus enabled us to examine the relationship of AMH with the dynamics of the primordial follicle pool. We can demonstrate that circulating AMH concentrations in adult women fall in line with the rate of loss of NGFs. AMH concentrations also broadly rise in line with this rate in children and young adults, but with a difference of about 10 years between the two peaks. The relationship of AMH to the total number of primordial follicles remaining in the pool for future recruitment therefore has two distinct phases, with the direction of association changing in early adulthood. In adult women serum AMH concentrations correlate with the size of the stereologically determined primordial follicle pool (Hansen et al., 2011), despite AMH production being restricted to later stages of follicular development. Similar data have been obtained in rodents (Kevenaar et al., 2006). The fate of the majority of NGFs (i.e. primordial and transitional stages) in the human is unclear. However, there is evidence that the greatest contribution to this is activation of growth, following which almost all follicles undergo atresia with only very few resulting in ovulation, rather than primordial follicles undergoing atresia directly (Gougeon, 1996). Thus the rate of loss of follicles from the non-growing population largely reflects the rate of initiation of follicle growth. It has recently been shown that initial follicular recruitment rises during childhood and adolescence (Wallace and Kelsey, 2010), despite an ongoing decline in the primordial follicle pool (Faddy et al., 1992; Faddy and Gosden, 1996; Hansen et al., 2008; Wallace and Kelsey, 2010). In this study, we have demonstrated that this increasing recruitment during childhood and adolescence is accompanied by a general increase in circulating AMH. In contrast, initial follicular recruitment rates decline in adult life (Wallace and Kelsey, 2010), as do circulating AMH concentrations (Kelsey et al., 2011) in accordance with previous cross-sectional and longitudinal analyses (de Vet et al., 2002; van Disseldorp et al., 2008; La Marca et al., 2005; Nelson et al., 2011a, b). These relationships between AMH and the age dependent changes in rate of follicular recruitment are entirely consistent with AMH being produced from the earliest stages of follicular development (Weenen et al., 2004). Increasing age and lower AMH both independently predict reduced oocyte (i.e. growing follicle) yield following ovulation induction for assisted conception (Nelson et al., 2007; La Marca et al., 2010), which is consistent both with this study and with the mathematical models suggesting declining follicular recruitment with increasing age (Hansen et al., 2008; Wallace and Kelsey, 2010). For these reasons, we speculate that the strong correlations between AMH and NGF dynamics do reflect a causal link, and hence that AMH can be used as an indirect approximation of age-related NGF populations and rates of recruitment. Our data intriguingly indicate a substantial distinction between peak follicle recruitment at puberty and peak AMH 10 years later. It is possible that this is an artefact reflecting the limitations of the data, as both sets are derived from cross-sectional sampling with limited numbers of young adults. However, if confirmed by more robustly designed studies, this may reflect that the organization of the adult follicle hierarchy within the ovary continues to mature after menarche as ovulatory cycles are initiated and become more prevalent over several years (Treloar et al., 1967; Flug et al., 1984). The finding that during childhood AMH tends to rise with time/age while NGF number falls does not mean that individuals with highest AMH will have lowest NGF number, although you might expect them to have lowest NGF recruitment based on the known biology of AMH (Durlinger et al., 2002). Speculatively, this might be associated with a later menopause (Broer et al., 2011) but our study finds the opposite, that there is a positive relationship between AMH and NGF recruitment during childhood. The relationship between AMH and NGF population seems to change in early adulthood. This may reflect the absence of larger AMH-producing follicles in childhood, which will change post-puberty thus causing a rise in AMH independent of NGF number. However, the relationship between AMH and NGF recruitment is positive both before and after the peak. Implications A model obtained by analysis of aggregated data serves as a series of reference points, which generate testable hypotheses. For our NGF study, the hypotheses are (i) that age is the dominating factor for variations in NGF populations [being responsible for about 80% of the variation for all ages from conception to menopause (Fig. 1), and for about 95% for ages up to 25 years (Fig. 2)], and (ii) that, in general, the recruitment of NGF towards maturation peaks at 13–15 years of age (Fig. 3). The hypotheses that arise from our AMH study are (i) that, in general, AMH levels in females fall shortly after birth, rise sharply during childhood before levelling out or even declining in pubertal years, thereafter rising to a peak level in the mid-twenties, and (ii) that about one-third of the variation in AMH levels for the healthy population is due to age alone (Fig. 4). By considering both models together, we can speculate that (i) NGF populations are strongly and negatively correlated with AMH from birth to peak AMH, (ii) for the same age range, NGF recruitment is positively correlated with AMH but much less strongly, and (iii) for ages from peak AMH to menopause both NGF population and NGF recruitment are strongly and positively correlated with AMH, with NGF recruitment having the higher rate of correlation (Fig. 5). Our NGF and AMH models are testable, either directly through histological analysis or indirectly through markers of ovarian reserve. Future in vitro estimates of NGF populations can be matched against the model, either confirming accuracy or identifying systematic bias. Future measurements of serum AMH can also be matched against the model, giving an indication of how well the model predicts unseen data. The results of these tests increase our understanding of the dynamics of human ovarian biology in either case. Evidence for inaccuracies in the models will lead to improved models; supporting evidence will lead to increased confidence in the models, and, in all likelihood, another set of testable hypotheses. In most studies involving the measurement of serum AMH, detection limits and inter- and intra-assay variations are reported as a matter of course, and are broadly similar across studies although the assay methodology also needs to be noted as there remains no recognized international standard. This supports the assumption that an aggregated dataset can be analysed as a single entity. However, the exclusion of data so that the final dataset approximates the general healthy population is not a simple process. Each paper has to be studied in detail to identify biases for inclusion or exclusion of subsets of the data. This process cannot be automated at any level: there is a requirement for intellectual input from experts in the field. This research methodology is limited to data that can be found and retrieved. Articles published in languages other than English, or published in older, more obscure, journals that have not yet been digitized will be unobtainable and/or uninterpretable. Naming conventions can also cause problems: AMH is referred to as MIS (Müllerian inhibiting substance) in many publications. Often data are summarized by descriptive statistics, as opposed to being reported in tabular or chart form. There are well-known techniques for the re-creation of datasets from summary descriptions, but another validation step is needed before such data can be used as a reasonable approximation to the data from the original study. The overall implication is that studies based on aggregated data add to our collective ability to accurately model ovarian reserve using both direct and indirect factors. The models and correlations obtained from these studies provide hypotheses that can be tested by traditional studies. Moreover, the results from data aggregation studies provide the estimates of event sizes needed to determine how many subjects are needed to test the hypotheses to a reasonable level of statistical significance. The resulting scientific process then becomes iterative: data from hypothesis-led studies are combined and analysed, the results lead to further hypothesis-led studies, the data from which can be aggregated anew. Authors' roles All authors contributed extensively to the work presented in this paper. Each author contributed to the data aggregation phase, the assessment of data quality, the modelling and the writing of the manuscript. All authors discussed the results and implications and commented on the manuscript at all stages. Funding This work was supported by UK Engineering & Physical Sciences Research Council grant EP/H004092/1. References Baker TG.  A quantitative and cytological study of germ cells in human ovaries,  Proc R Soc Lond B Biol Sci ,  1963, vol.  158 (pg.  417- 433)  Containing papers of a Biological character. Royal Society (Great Britain) Google Scholar CrossRef Search ADS PubMed  Bendsen E,  Byskov AG,  Andersen CY,  Westergaard LG.  Number of germ cells and somatic cells in human fetal ovaries during the first weeks after sex differentiation,  Hum Reprod ,  2006, vol.  21 (pg.  30- 35) Google Scholar CrossRef Search ADS PubMed  Block E.  Quantitative morphological investigations of the follicular system in women; variations at different ages,  Acta Anat ,  1952, vol.  14 (pg.  108- 123) Google Scholar CrossRef Search ADS PubMed  Block E.  A quantitative morphological investigation of the follicular system in newborn female infants,  Acta Anat ,  1953, vol.  17 (pg.  201- 206) Google Scholar CrossRef Search ADS PubMed  Brett S,  Bee N,  Wallace WHB,  Rajkhowa M,  Kelsey TW.  Individual ovarian volumes obtained from 2-dimensional and 3-dimensional ultrasound lack precision,  Reprod Biomed Online ,  2009, vol.  18 (pg.  348- 351) Google Scholar CrossRef Search ADS PubMed  Broer SL,  Eijkemans MJ,  Scheffer GJ,  van Rooij IA,  de Vet A,  Themmen AP,  Laven JS,  de Jong FH,  Te Velde ER,  Fauser BC, et al.  Anti-müllerian hormone predicts menopause: a long-term follow-up study in normoovulatory women,  J Clin Endocrinol Metab ,  2011, vol.  96 (pg.  2532- 2539) Google Scholar CrossRef Search ADS PubMed  Chellakooty M,  Schmidt IM,  Haavisto AM,  Boisen KA,  Damgaard IN,  Mau C,  Petersen JH,  Juul A,  Skakkebaek NE,  Main KM.  Inhibin A, inhibin B, follicle-stimulating hormone, luteinizing hormone, estradiol, and sex hormone-binding globulin levels in 473 healthy infant girls,  J Clin Endocrinol Metab ,  2003, vol.  88 (pg.  3515- 3520) Google Scholar CrossRef Search ADS PubMed  de Vet A,  Laven JS,  de Jong FH,  Themmen AP,  Fauser BC.  Antimüllerian hormone serum levels: a putative marker for ovarian aging,  Fertil Steril ,  2002, vol.  77 (pg.  357- 362) Google Scholar CrossRef Search ADS PubMed  Durlinger AL,  Gruijters MJ,  Kramer P,  Karels B,  Ingraham HA,  Nachtigal MW,  Uilenbroek JT,  Grootegoed JA,  Themmen AP.  Anti-Müllerian hormone inhibits initiation of primordial follicle growth in the mouse ovary,  Endocrinology ,  2002, vol.  143 (pg.  1076- 1084) Google Scholar PubMed  Faddy MJ,  Gosden RG.  A model conforming the decline in follicle numbers to the age of menopause in women,  Hum Reprod ,  1996, vol.  11 (pg.  1484- 1486) Google Scholar CrossRef Search ADS PubMed  Faddy MJ,  Gosden RG,  Gougeon A,  Richardson SJ,  Nelson JF.  Accelerated disappearance of ovarian follicles in mid-life: implications for forecasting menopause,  Hum Reprod ,  1992, vol.  7 (pg.  1342- 1346) Google Scholar PubMed  Flug D,  Largo RH,  Prader A.  Menstrual patterns in adolescent Swiss girls: a longitudinal study,  Ann Hum Biol ,  1984, vol.  11 (pg.  495- 508) Google Scholar CrossRef Search ADS PubMed  Forabosco A,  Sforza C.  Establishment of ovarian reserve: a quantitative morphometric study of the developing human ovary,  Fertil Steril ,  2007, vol.  88 (pg.  675- 683) Google Scholar CrossRef Search ADS PubMed  Gougeon A.  Regulation of ovarian follicular development in primates: facts and hypotheses,  Endocr Rev ,  1996, vol.  17 (pg.  121- 155) Google Scholar CrossRef Search ADS PubMed  Hagen CP,  Aksglaede L,  Sørensen K,  Main KM,  Boas M,  Cleemann L,  Holm K,  Gravholt CH,  Andersson AM,  Pedersen AT, et al.  Serum Levels of Anti-Müllerian Hormone as a Marker of Ovarian Function in 926 Healthy Females from Birth to Adulthood and in 172 Turner Syndrome Patients,  J Clin Endocrinol Metab ,  2010, vol.  95 (pg.  1- 8) Google Scholar CrossRef Search ADS PubMed  Hansen KR,  Knowlton NS,  Thyer AC,  Charleston JS,  Soules MR,  Klein NA.  A new model of reproductive aging: the decline in ovarian non-growing follicle number from birth to menopause,  Hum Reprod ,  2008, vol.  23 (pg.  699- 708) Google Scholar CrossRef Search ADS PubMed  Hansen KR,  Hodnett GM,  Knowlton N,  Craig LB.  Correlation of ovarian reserve tests with histologically determined primordial follicle number,  Fertil Steril ,  2011, vol.  95 (pg.  170- 175) Google Scholar CrossRef Search ADS PubMed  He C,  Kraft P,  Chasman DI,  Buring JE,  Chen C,  Hankinson SE,  Paré G,  Chanock S,  Ridker PM,  Hunter DJ.  A large-scale candidate gene association study of age at menarche and age at natural menopause,  Hum Genet ,  2010, vol.  128 (pg.  515- 527) Google Scholar CrossRef Search ADS PubMed  Hehenkamp WJ,  Looman CW,  Themmen AP,  de Jong FH,  Te Velde ER,  Broekmans FJ.  Anti-Müllerian hormone levels in the spontaneous menstrual cycle do not show substantial fluctuation,  J Clin Endocrinol Metab ,  2006, vol.  91 (pg.  4057- 4063) Google Scholar CrossRef Search ADS PubMed  Kelsey TW,  Wright P,  Nelson SM,  Anderson RA,  Wallace WH.  A validated model of serum anti-Müllerian hormone from conception to menopause,  PLoS ONE ,  2011, vol.  6 pg.  e22024  Google Scholar CrossRef Search ADS PubMed  Kevenaar ME,  Meerasahib MF,  Kramer P,  van de Lang-Born BM,  de Jong FH,  Groome NP,  Themmen AP,  Visser JA.  Serum anti-müllerian hormone levels reflect the size of the primordial follicle pool in mice,  Endocrinology ,  2006, vol.  147 (pg.  3228- 3234) Google Scholar CrossRef Search ADS PubMed  La Marca A,  De Leo V,  Giulini S,  Orvieto R,  Malmusi S,  Giannella L,  Volpe A.  Anti-Müllerian hormone in premenopausal women and after spontaneous or surgically induced menopause,  J Soc Gynecol Investig ,  2005, vol.  12 (pg.  545- 548) Google Scholar CrossRef Search ADS PubMed  La Marca A,  Sighinolfi G,  Radi D,  Argento C,  Baraldi E,  Artenisio AC,  Stabile G,  Volpe A.  Anti-Müllerian hormone (AMH) as a predictive marker in assisted reproductive technology (ART),  Hum Reprod Update ,  2010, vol.  16 (pg.  113- 130) Google Scholar CrossRef Search ADS PubMed  Lee MM.  Reproductive hormones in infant girls—a harbinger of adult reproductive function?,  J Clin Endocrinol Metab ,  2003, vol.  88 (pg.  3513- 3514) Google Scholar CrossRef Search ADS PubMed  Mamsen LS,  Lutterodt MC,  Andersen EW,  Byskov AG,  Andersen CY.  Germ cell numbers in human embryonic and fetal gonads during the first two trimesters of pregnancy: analysis of six published studies,  Hum Reprod ,  2011, vol.  26 (pg.  2140- 2145) Google Scholar CrossRef Search ADS PubMed  Matzuk MM,  Lamb DJ.  The biology of infertility: research advances and clinical challenges,  Nat Med ,  2008, vol.  14 (pg.  1197- 1213) Google Scholar CrossRef Search ADS PubMed  Nelson SM,  Yates RW,  Fleming R.  Serum anti-Müllerian hormone and FSH: prediction of live birth and extremes of response in stimulated cycles - implications for individualization of therapy,  Hum Reprod ,  2007, vol.  22 (pg.  2414- 2421) Google Scholar CrossRef Search ADS PubMed  Nelson SM,  Messow MC,  Wallace AM,  Fleming R,  McConnachie A.  Nomogram for the decline in serum antimüllerian hormone: a population study of 9,601 infertility patients,  Fertil Steril ,  2011, vol.  95 (pg.  736- 741) Google Scholar CrossRef Search ADS PubMed  Nelson SM,  Messow MC,  McConnachie A,  Wallace WHB,  Kelsey TW,  Fleming R,  Anderson RA,  Leader B.  External validation of nomogram for the decline in serum anti-Müllerian hormone in women: a population study of 15,834 infertility patients,  Reprod Biomed Online ,  2011, vol.  23 (pg.  204- 206) Google Scholar CrossRef Search ADS PubMed  Ong KK,  Elks CE,  Li S,  Zhao JH,  Luan J,  Andersen LB,  Bingham SA,  Brage S,  Smith GD,  Ekelund U, et al.  Genetic variation in LIN28B is associated with the timing of puberty,  Nat Genet ,  2009, vol.  41 (pg.  729- 733) Google Scholar CrossRef Search ADS PubMed  Richardson SJ,  Senikas V,  Nelson JF.  Follicular depletion during the menopausal transition: evidence for accelerated loss and ultimate exhaustion,  J Clin Endocrinol Metab ,  1987, vol.  65 (pg.  1231- 1237) Google Scholar CrossRef Search ADS PubMed  Tingen CM,  Bristol-Gould SK,  Kiesewetter SE,  Wellington JT,  Shea L,  Woodruff TK.  Prepubertal primordial follicle loss in mice is not due to classical apoptotic pathways,  Biol Reprod ,  2009, vol.  81 (pg.  16- 25) Google Scholar CrossRef Search ADS PubMed  Treloar AE,  Boynton RE,  Behn BG,  Brown BW.  Variation of the human menstrual cycle through reproductive life,  Int J Fertil ,  1967, vol.  12  1 Pt 2(pg.  77- 126) Google Scholar PubMed  van Disseldorp J,  Faddy MJ,  Themmen AP,  de Jong FH,  Peeters PH,  van der Schouw YT,  Broekmans FJ.  Relationship of serum antimüllerian hormone concentration to age at menopause,  J Clin Endocrinol Metab ,  2008, vol.  93 (pg.  2129- 2134) Google Scholar CrossRef Search ADS PubMed  Wallace WH,  Kelsey TW.  Human ovarian reserve from conception to the menopause,  PloS ONE ,  2010, vol.  5 pg.  e8772  Google Scholar CrossRef Search ADS PubMed  Weenen C,  Laven JS,  Von Bergh AR,  Cranfield M,  Groome NP,  Visser JA,  Kramer P,  Fauser BC,  Themmen AP.  Anti-Müllerian hormone expression pattern in the human ovary: potential implications for initial and cyclic follicle recruitment,  Molecular Hum Reprod ,  2004, vol.  10 (pg.  77- 83) Google Scholar CrossRef Search ADS   © The Author 2011. Published by Oxford University Press on behalf of the European Society of Human Reproduction and Embryology. All rights reserved. For Permissions, please email: journals.permissions@oup.com

Journal

Molecular Human ReproductionOxford University Press

Published: Sep 20, 2011

Keywords: ovary fertility AMH oocyte computational analysis

There are no references for this article.