Constructing the crystal ball: how to get reliable prognostic information for the management of subfertile couples

Abstract

Couples in whom the results of an initial fertility workup fail to identify the presence of any obvious barriers to conception are diagnosed with unexplained subfertility. Couples who have tried to conceive for a relatively short time have a good chance of natural conception and thus may not benefit from immediate access to ART. As fertility decreases over time, the main dilemma that clinicians and couples face is when to abandon an expectant approach in favour of active treatment. Several prognostic or predictive models have been used to try to discriminate between couples with high and low chances of conception but have been unable to compare individualized chances of conception associated with ART relative to chances of natural conception at various time points. These models are also unable to recalculate the chances of pregnancy at subsequent time points in those who return after a period of unsuccessful expectant management. In this paper, we discuss currently available models. We conclude that in order to provide accurate, individualized and dynamic fertility prognoses associated with and without treatment at different points in time, we need to develop, validate and update clinical prediction models which are fit for purpose. We suggest several steps to move the field forwards.

unexplained subfertility, prognosis, natural conception, ART, prediction models, dynamic prediction

Introduction

Subfertility is defined as the inability to conceive over a period of 12 months despite regular, unprotected intercourse (Zegers-Hochschild et al., 2009). This definition implies that those who have not conceived within that period belong to a relatively poor prognosis group, but it is worth noting that poor prognosis in this context does not equate to sterility. Couples with unexplained subfertility, who have no obvious barriers to conception, have chances of natural conception ranging from approximately 0–63% over the course of 1 year after the fertility workup (te Velde et al., 2000; Hunault et al., 2004; van der Steeg et al., 2007; van Eekelen et al., 2017). Subfertility is thus a probabilistic concept (te Velde et al., 2000; McLernon et al., 2014).

In clinical practice, a fertility workup is initiated in all couples to establish a diagnosis, as a prerequisite for targeted treatment, such as anovulation, bilateral tubal disease and azo- or severe oligospermia. However, in a large proportion of couples subfertility is not attributable to a specific aetiology and these couples are put into less distinct and often overlapping diagnostic categories such as mild endometriosis, cervical factor and mild/moderately impaired semen quality (Evers, 2002). Within these categories, it is unclear which pathophysiologic mechanisms underlie the inability to conceive and, consequently, which treatment would be appropriate. Unexplained subfertility is the ultimate non-diagnosis in which the fertility workup has not identified any abnormalities. No disease is present according to the current medical paradigm and counselling couples ‘diagnosed’ as such about various treatment options is intrinsically a non-sequitur (Tjon-Kon-Fat et al., 2016). Nevertheless, even in the absence of proven effectiveness, ART is extensively utilized in these couples leading authors to suggest that we are overusing ART (Kamphuis et al., 2014; Pandian et al., 2015; Kupka et al., 2016; Veltman-Verhulst et al., 2016).

The superiority of ART over an expectant management approach remains unproven in couples with unexplained subfertility of short duration—as does the question whether active treatment becomes more effective at a certain point in time, and as such, when it should be offered (Kamphuis et al., 2014).

It is in these couples that the probabilistic concept of subfertility may provide a solution to this apparent enigma. Many clinicians assess the prognosis of couples intuitively before planning treatment, but this approach is associated with a large variation in clinical practice and an uneven quality of care (Wiegerinck et al., 1999; van der Steeg et al., 2006). The explicit use of an individual prognosis obtained from clinical prediction models aids in a shared decision making process and reduces inconsistency of treatment decisions (van der Steeg et al., 2006; Dancet et al., 2011, 2014). In addition, a recent opinion article states the inability to distinguish between unexplained subfertility and age-related subfertility, an issue that is also addressed by prediction models by quantifying the effect of age (CAPRI, 2017).

In this paper we aim to summarize the current literature on validated prediction models in fertility, their clinical utility and their limitations. We then describe the type of research required to advance this field and discuss the phases a model must pass through before it can be used in clinical practice. Although commonly used interchangeably, in this paper, we define prognostic models as those used to estimate the chance of a live birth after natural conception, whereas we define predictive models as those which estimate chances of live birth after treatment or the added benefit of treatment over expectant management.

Current validated models and clinical utility

In most countries, couples with unexplained subfertility who have tried to conceive for at least 12–24 months continue with ART based on availability (NVOG, 2010; NICE, 2013; Kupka et al., 2016). Several prognostic models to aid in deciding whom to treat have been published, but most have found little popularity in routine clinical practice (Leushuis et al., 2009; Kleinrouweler et al., 2016).

The Netherlands is the only country where the use of a validated prognostic model is recommended in the national guideline (Hunault et al., 2004; van der Steeg et al., 2007; NVOG, 2010). The factors on which this model for natural conception is based are female age, duration of subfertility, primary subfertility, referral status and sperm motility. The model calculates a prognosis after completing the fertility workup in terms of the estimated chance to conceive naturally over the course of the next 12 months leading to a live birth. The prognosis is then used to counsel couples and to identify subgroups that are expected to benefit from early access to treatment. Couples with a favourable prognosis are advised expectant management while those with an unfavourable prognosis are advised ART. In case of an intermediate prognosis, management will usually depend on the personal preferences of couples and their clinicians.

The only validated model for estimating success after IUI is not mentioned in any clinical guideline (Steures et al., 2004; Custers et al., 2007; NVOG, 2010). Four models estimating the chances of live birth after IVF have been developed using data from the Human Fertilisation and Embryology Authority (HFEA), of which three were validated (Templeton et al., 1996; Nelson and Lawlor, 2011; Smith et al., 2015; McLernon et al., 2016). The model by Nelson and Lawlor (‘IVFPredict’) has been used to predict success after IVF as part of a cost-effectiveness analysis in the generation of the National Institute for Health and Care Excellence (NICE) guideline in 2013 (NICE, 2013). This led to the recommendation for IVF in unexplained subfertile couples after at least 2 years of trying to conceive, but IVFPredict or other IVF models are not recommended so far to use for counselling couples in daily clinical practice.

Limitations of current models

Models which predict the chances of live birth associated with IUI or IVF are, on their own, unhelpful for clinical decision making unless the prognosis associated with expectant management is also considered. Likewise, using the ‘Hunault’ model for natural conception to identify unexplained subfertile couples who are expected to benefit from treatment merely represents an indirect approach that is equally unhelpful as the added value of treatment is not taken into account in the model (Hunault et al., 2004).

Another approach involves using separate models for natural conception, IUI and IVF side by side to determine the benefit of ART in a particular couple. This approach was taken by NICE when they used the Hunault model alongside the IVFPredict model by Nelson and Lawlor to conduct a cost-effectiveness analysis comparing strategies for access to IVF over a woman's reproductive life (Hunault et al., 2004; Nelson and Lawlor, 2011; NICE, 2013). However, as acknowledged in those guidelines, this approach has limitations. It is based on low-grade evidence because the chances of having a naturally conceived or an ART-conceived live birth were evaluated in separate populations that are not comparable in terms of female age, duration of subfertility, severity of infertility, etc. (Leushuis et al., 2009; McLernon et al., 2014). So, there are currently no predictive models that allow comparison of chances of both natural conception and chances after ART.

A second limitation of current models is that they cannot be reapplied for couples who return to the clinic after a period of unsuccessful expectant management. Time is the key factor for prognosis because successful conception is inherently based on probabilities in every consecutive ovulatory cycle. As time progresses, the probability of natural conception in the population of couples who have not yet conceived, which is initially heterogeneous and varies from 0 to ~60% per cycle, gradually becomes more homogeneous with a lower average chance of conception (te Velde et al., 2000). This is because couples with good prognoses are more likely to conceive first, leaving behind a cohort of couples whose prognosis becomes increasingly poorer over time (Bongaarts, 1975; te Velde et al., 2000; van Eekelen et al., 2017). This selection process must be taken into account when we wish to update predictions as time progresses in order to make informed decisions at multiple points in time during the management of unexplained subfertile couples.

Why can a prediction model not be reused at a later time point?

This selection process that occurs over time, and which is critical in the management of subfertile couples, cannot accurately be captured by commonly used statistical techniques for time-to-pregnancy data, most notably Cox proportional hazards models that underlie current models. Unfortunately, this also holds true for models that incorporate the duration of subfertility as a covariate. This information cannot fully explain the selection that occurs over time. The issue with such models is depicted in Fig. 1, in which the analysis of a large prospective cohort comprising 5184 couples with unexplained subfertility from 38 hospitals in the Netherlands is summarized (van der Steeg et al., 2007; Bensdorp et al., 2017). We compared predicted chances from a conventional Cox model developed on this cohort, with female age (separate effect below and above 31 years) and duration of subfertility as covariates, to the observed cumulative pregnancy rates after natural conception. To recreate the situation of a couple returning to the clinic and making an updated prediction using the same model, we reapplied the Cox model at 6-month intervals of expectant management after increasing the female age and duration of subfertility by the elapsed period. These predictions were compared with Kaplan–Meier estimates for observed pregnancy rates after natural conception using only the patients still in follow-up at those time points. As shown in Fig. 1, reapplying the Cox model overestimates the probability of conceiving when used for repeated predictions over time. Since the model was developed and evaluated in the same dataset, such a structural overestimation cannot be assigned to random variance or limited generalizability but instead indicates model misfit.

Figure 1

Summary analyses of natural conception leading to ongoing pregnancy in a large prospective cohort comprising 5184 couples with unexplained subfertility from 38 hospitals. The black lines represent the observed fraction of ongoing pregnancy and the corresponding confidence limits, the green lines represent estimated probabilities from a statistical (Cox) model that contains information on female age and duration of subfertility. The leftmost black and green lines follow all couples who completed their fertility workup, the lines second from left only those who did not conceive in the first 6 months of follow-up, etc.

To understand why this overestimation takes place, we have to look at the methodology in more detail. A Cox proportional hazards model applies the following formula:

$$S\left(t,x\right)={S_0(t)}^{\exp (\mathit{PI})},$$

where S(t, x) is the chance of not conceiving by time t for a couple with specific characteristics x, S₀(t) is commonly called the baseline survival, i.e. the chance of not conceiving until time t for all patient characteristics at their reference level, which can be seen as the ‘common’ chance pattern of conceiving over time and PI is the Prognostic Index, i.e. multiplying patient characteristics with their respective effects and taking the sum. This formula thus combines two aspects: first, individual patient characteristics and second, the baseline chance or underlying prevalence of conception in the cohort.

Reapplying this formula for couples who return after a given period of expectant management is based upon three implicit assumptions. First, any within-couple change in predicted chances over time is due to increased female age and duration of subfertility, which are estimated based on the differences between couples; second, the baseline chance pattern at this new ‘starting’ point in time should be the same as before; and third, the effects of these characteristics remain constant over time. The latter is known as the proportional hazards assumption.

The first assumption states that reapplying the Cox model assumes that ‘between’-couple age and duration differences at time zero translate directly to ‘within’-couple changes in predictions when we update predictions over time and this principle does not hold, as can be seen by the distance between the observed and predicted lines in Fig. 1. These time-related characteristics do not fully capture the latent heterogeneity in underlying chances of conception that results in the selection process that is occurring over time. That this selection occurs, which is a violation of the first assumption regarding baseline chance, can be seen by the decrease over time in observed fractions of ongoing pregnancy according to Kaplan–Meier estimates. We have no evidence that suggests the third assumption regarding proportional hazards does not hold.

The inability of female age and duration of subfertility to account for decreasing chances as time progresses for a couple highlights the necessity of a new class of models, called dynamic prediction models (van Houwelingen and Putter, 2012). These models should enable us to predict the chances of pregnancy at various time points, for instance when a couple returns after 6 or 12 months of unsuccessful expectant management. As a first step along this route, we recently developed a dynamic model capable of giving repeated predictions of natural conception but this model has not yet been validated and does not allow direct comparison with chances for alternative scenarios involving ART (van Eekelen et al., 2017).

How to move forward?

To be clinically useful, we require novel models that can predict at different points in time and not just at diagnosis. We must be able to compare the chances of live birth following natural conception with the chances following ART at multiple points in time. Only this approach will allow clinicians and couples to take an individualized, well-informed decision at various time points after the fertility workup.

To develop prediction models that are clinically more useful, we invite stakeholders to continue in the following two steps.

Step 1: estimation of individual treatment effects

The first step is to identify datasets of comparable groups. Ideally, these datasets would be from large RCTs in which couples were randomized to expectant management or to a treatment modality, guaranteeing that the couples in both arms are really comparable. The second best alternative is to use large observational datasets of couples on expectant management and on treatment, statistically adjusting for the non-random allocation. To enhance comparability, observationally followed couples would ideally be within the same cohort, but in absence of such a sufficiently large cohort, separate cohorts can also be compared, yielding higher risks of bias.

These RCTs and/or observational datasets need not only contain the straightforward, easily available characteristics of the couple such as female age and duration of subfertility, but may also incorporate biomarkers, for example for ovarian reserve, or other measurements that may be related to prognosis and/or moderate the effect of treatment.

Using these approaches, a main or overall effect for treatment can be estimated. Estimating patient-specific treatment effects, by including interactions between patient characteristics and treatment, would be the ideal additional step that increases the clinical utility of these models but which would require a larger sample size (Hingorani et al., 2013; McLernon et al., 2014).

Step 2: development of models capable of giving repeated predictions

The decreasing chance of conception over time described above holds for both natural conception and conception after treatment, but possibly not at the same speed. A dynamic prediction model accounts for the decline in pregnancy rates and can re-evaluate treatment decisions based on the current information of the couple at each new time point (van Houwelingen and Putter, 2012; McLernon et al., 2014). In addition to preferable designs as mentioned in Step 1, the development of such a model ideally requires a trial design in which couples are randomly assigned to start treatment at different time points. In the absence of trials with such a design, the best available approach is to use data from a large or population-based cohort of couples who wish to conceive and who are subsequently followed-up after receiving fertility treatment. Preferably, there should be much variation in this cohort in terms of when treatment was initiated and follow-up for a sufficient length of time to enable modelling at different, extended time intervals (McLernon et al., 2014).

Implementing models in clinical practice

Before any prognostic or predictive model can be used in clinical practice, it should go through the appropriate number of phases including model development, followed by internal and external validation (Coppus et al., 2009; Leushuis et al., 2009; Steyerberg et al., 2013). Discrimination and calibration are important aspects of the validation process after developing a model to find out how well it is able to predict pregnancy resulting in live birth (Steyerberg, 2009). Discrimination assesses the ability of a model to distinguish between patients who do and do not conceive. Calibration assesses the level of agreement between the predicted and observed chances of pregnancy. Internal validation refers to investigating whether the model performs correctly in the setting of the development data (Steyerberg, 2009). External validation refers to repeating the steps of internal validation on a separate dataset, preferably in more than one setting including a different location or country. External validation is essential as prognostic or predictive models tend to perform better on their ‘own’ data, where they were developed (Steyerberg, 2009). This step is infrequently performed (Leushuis et al., 2009; Kleinrouweler et al., 2016). Note that in the external validation of a dynamic model, discrimination and calibration can be applied at several points in time to investigate the ability of the model to predict as time progresses. In the third phase, impact analysis, the consequences of using the model in everyday clinical practice, is assessed (Steyerberg et al., 2013). Finally, cost-effectiveness analyses and decision curve analyses can be applied to establish thresholds for predicted probabilities after which expectant management or treatment will be the recommended course of action (McLernon and Bhattacharya, 2016).

The four major areas in model development mentioned (prognostic research, individual benefit of treatment over expectant management, dynamic modelling and implementation) are outlined in Table I.

Table I

Scientific areas of interest for prognostic research in reproductive medicine.

Research area	Characteristic	Activities
Prognostic modelling	Estimate outcome after expectant management	Development and validation of prognostic models in observational data of couples on expectant management
Predictive modelling of individual benefit	Estimate benefit of treatment alternatives given the individual prognosis of expectant management	Perform trials and observational studies comparing expectant management and treatment modalities to estimate effects Integrate results in prediction models enabling to estimate the individual benefit of treatment at time zero Include patient-specific treatment effects
Dynamic modelling	Estimate the effect of progressing time for couples on expectant management on their natural chance of live birth and their chance of conception leading to live birth after treatment	Develop and validate dynamic prediction models, preferably on studies involving long follow-up Apply dynamic models to determine the optimal time to start treatment for an individual couple
Implementation and impact analysis	Translate these models for use in clinical practice	Define easily applicable decision tools based on prognostic, predictive and dynamic models Implement these models into clinical practice Cost-effectiveness analysis Decision curve analysis Evaluate the impact of these models on clinical decision making and outcome

Conclusion

The probabilistic concept of subfertility does not translate well into a pathology driven clinical environment where unexplained subfertile couples are categorized into fertile or infertile. Current prognostic and predictive models in reproductive medicine have a number of limitations which restrict their clinical utility. We invite all stakeholders to join us in the proposed scientific initiative to support clinical decision making based on high quality predictive models which are able to provide individualized chances of pregnancy associated with different treatment strategies at varying points in time.

Acknowledgements

The author group would like to thank Prof. Egbert te Velde for all of his efforts regarding the present manuscript, including his work on the first draft and his insightful remarks during the process.

Authors’ roles

Coming up with the initial concept, drafting of the manuscript and preparation of the final manuscript was done by all authors.

Funding

B.W.M. is supported by a NHMRC Practitioner Fellowship (GNT1082548).

Conflict of interest

B.W.M. reports consultancy for ObsEva, Merck and Guerbet.