9.2.1 Time-Lapse vs. Morphological Selection

As there is a well-documented correlation between embryo morphology at given time points and developmental competence, the quality and viability of preimplantation embryos are traditionally evaluated by microscopic inspection at a few, well-defined discrete time points. Morphological grading is simple, relatively easy to learn, and cost-effective. The method, however, has several limitations. As the collected information has been limited to a few time points, it gives an incomplete picture of the dynamic embryo development. Several transient aspects of potential importance, such as number of pronuclei, are easily missed. Furthermore, morphological scoring is relatively subjective, as demonstrated by large inter- as well as intra-observer variability. This variability has implications for the decision to transfer, cryopreserve, or discard the embryos, with a rather low reproducibility. Finally, the standard morphological evaluation requires that embryos are removed from the incubator for assessment, thereby exposing the embryos to changes in temperature, humidity, and gas composition that may be potentially harmful.

There are two obvious benefits to TLT that may be translated into improved clinical outcome: Improved culture conditions and improved embryo selection. Firstly, images for embryo assessment are obtained without removing embryos from the incubator. Although TLT necessitates periodical light exposure, use of moving devices, and magnetic fields, which may all be of potential harm to the embryos, continuous culture made possible by TLT eliminates the exposure of embryos to the stress of mechanical moving and changes in temperature, humidity, and gas composition and is therefore most likely to improve culture conditions.

Secondly, TLT captures the dynamics of embryo development more extensively. In addition to a more precise registration of events, such as timing of cellular divisions until compaction, timing of the morula stage, formation of the blastocoel, blastocyst expansion, and – in some cases – collapse and reformation and eventually hatching, temporary events missed by static evaluation are easily recorded by TLT. Examples that have all been evaluated in several publications are time for formation of the male pronucleus, extrusion of the second polar body followed by formation of the male and female pronuclei, and syngamy/abuttal and disappearance of the pronuclei. From the time points recorded, duration of cellular stages can be calculated and deviations from normal development can be observed, e.g., extremely short duration of the first division cycle (the 2-cell stage), referred to as a direct cleavage from one to three cells. This information may be used to improve embryo selection. Numerous observational studies have been published demonstrating higher implantation rates using a diversity of time-lapse markers and selection models. There are, however, very few studies that have validated their findings in independent populations, and even fewer randomized trials. Given the scope of the present chapter, only prospective studies will be discussed.

In principle, prospective studies can be designed to evaluate the effect of uninterrupted culture, improved embryo selection, or both. Only very few studies have evaluated the isolated effect. Most studies have evaluated the combined effect of uninterrupted culture and time-lapse selection. A recent Cochrane meta-analysis has summarized the evidence from published RCTs (Armstrong et al., 2019).

Six randomized trials have compared culture in a standard incubator to a time-lapse incubator. All six trials were conducted using an integrated device. Two studies were performed randomizing oocytes rather than patients, and evaluated embryo development on day 3 as primary endpoint, where no difference was found. The evidence from these two trials is low quality, and they were excluded from a recent Cochrane analysis. The remaining four studies were, however, included in the Cochrane analysis (Armstrong et al., 2019). The designs in these four studies were heterogenous, in particular day of embryo transfer was different in all four studies. Three studies provided data on ongoing pregnancy or live births. Based on a meta-analysis of these three studies (3 RCTs, N = 826), there is no evidence for any difference between uninterrupted culture and standard incubation (OR 0.91, 95% CI 0.67–1.23).

Only two randomized studies have evaluated the isolated effect of time-lapse selection vs. standard morphological selection under identical culture conditions. One study evaluated ongoing pregnancy following day 3 single embryo transfer with TLT vs. day 5 single embryo transfer with or without adjunctive TLT (Kaser et al., 2017). The study found a nonsignificant higher clinical pregnancy rates in the D5 arm without TLT (OR 0.76 (0.34–1.66), 0.76 (0.34–1.66), and 1.00 (referent) for D3 + TLT, D5 + TLT, and D5 standard evaluation alone, respectively). The authors concluded that although the study was not powered (due to premature termination by the sponsor, Eeva) the higher clinical pregnancy rate in the D5 standard evaluation alone arm, suggested that use of early time-lapse markers may not improve embryo selection. The second randomized study reported clinical pregnancy rates and found no difference between TLT and standard evaluation (Goodman et al., 2016).

Three studies were included in the meta-analysis on clinical pregnancy and live birth evaluating the combination of continued culture and selection using TLT. Based on a meta-analysis of these studies (3 RCTs, N = 826), no difference between TLT and standard evaluation was found between interventions in rates of live birth (OR 1.12, 95% CI 0.92–1.36, 3 RCTs, N = 1617). The authors stress that there was a high risk of bias in the included studies and the high level of heterogeneity in study design. Of particular importance is that randomization was broken in the largest of the studies, as some of the patients received the intervention based on request rather than randomization.

Overall, the Cochrane review concludes that there is insufficient good‐quality evidence of differences in live birth or ongoing pregnancy or clinical pregnancy to choose between TLT, with or without embryo selection software, and conventional incubation. In contrast, another meta-analysis including five RCTS concludes that TLT with the prospective use of a morphokinetic algorithm for selection of embryos improves overall clinical outcome (Pribenszky et al., 2017). Concerns were, however, raised based on inconsistency of inclusion and exclusion of studies, data analysis, classification of bias and heterogeneity, and conflict of interest, as all authors were employees of a company supplying time-lapse incubators.

Since the first commercial time-lapse instruments were introduced 10 years ago, several instruments have become available, and TLT and continuous culture are implemented in large numbers. Clinical RCTs are time-consuming, but without question the gold standard of evidence-based medicine. Based on the available low-quality evidence, it seems reasonable to comply with the recommendation of ESHRE’s recent guideline (2020) that more RCTs with adequate design and sufficient power be conducted, reporting on live births and perinatal outcomes, in order to firmly establish a putative beneficial effect of TLT.

9.2.2 Time-Lapse Prediction Models

A crucial part of implementing time-lapse in the laboratory is to decide on which parameters to annotate and evaluate. Some instruments offer automated annotation, and predefined selection models. Ideally, each clinic should evaluate their own data in order to develop an in-house algorithm or selection strategy designed for their particular patient population and culture conditions. This is, however, very time-consuming and requires a large dataset. Therefore, clinics often refer to the literature for guidance. ESHRE recommendations for practical use of TLT have recently been published (2020).

A large amount of observational studies has suggested different parameters and time-intervals. A non-exhaustive overview is presented in Table 9.1. The most thoroughly tested model is the hierarchical model presented by Meseguer et al. in 2011 (Meseguer et al., 2011), which has been tested in a RCT and subsequently refined using large datasets. The model ranks embryos based on a combination of de-selection (poor morphology, direct cleavage from 1 to 3 cells, uneven blastomere size at 2-cell stage, and multinucleation at 4-cell stage) and selection based on timings (t5 (48.8–56.6 hours), t4–t3 (<0.76 hours), t3–t2/cc2 (<11.9 hours). When tested in an RCT (n = 856), an OR of 1.23 (1.06–1.43) in favor of the study group for the primary outcome of ongoing pregnancy rate along was reported. The embryos in the study group were cultured in a standard incubator, therefore the individual contribution of the effect of TL selection cannot be separated from the effect of different incubators. In addition, the randomization was broken, as some of the patients were assigned to TLT by request after randomization, introducing potential bias. Finally, nearly half of the patients received embryos from oocyte donors, which limits the external validity. Another prediction model tested in several studies is the model suggested by Eeva, where embryos are ranked in high/low categories based on early time intervals (Table 9.1; VerMilyea, Adamson). The model has been tested (with modifications) in one RCTs (Kaser et al., 2017), where no difference was found, however the study was terminated early.

All of the models have proved difficult to transfer from one clinical setting to another. A plausible explanation is that timing of development is influenced by a variety of patient- and treatment-related factors, such as oxygen tension, fertilization method, culture media, smoking, age, baseline androgen levels, and type of gonadotropin used for stimulation (ESHRE 2020). Consequently, a prediction model that is developed from one dataset will most likely perform well only on the dataset it was developed from, or on datasets from similar clinical conditions, in particular if the dataset is small and the clinical characteristics narrow. An important limitation of the majority of studies is, therefore, that they have not been validated on independent datasets, which limits the transferability. Furthermore, several of the observational studies have not taken into account that timing of embryo development not only reflects viability, but is heavily influenced by patient characteristics. Consequently, several observational studies tend to overestimate the predictive power of the identified parameters.

Another important consideration when developing prediction models is to weight sensitivity against specificity. Several of the early models defined very narrow time intervals for optimal division to achieve high specificity at the expense of a low sensitivity. As they were developed on small datasets, this subsequently limited transferability and introduced a high risk of discarding viable embryos if transferred uncritically. On the other hand, using too broad intervals lowers the value of the algorithm, as no real ranking is performed. In general, a model that aims at de-selecting embryos with low viability would be expected to be more transferable to other settings than a model based on narrow intervals for optimal division. Recently, an algorithm based on large datasets of heterogeneous origin has been developed, which theoretically should be applicable to different culture conditions and patient-groups (Petersen et al., 2016). The algorithm has not been tested in randomized studies.

9.2.3 Time-Lapse and Computer Assisted Technology/Machine Learning

TLT generates a large amount of images. Manual evaluation of all the available information is unattainable in daily practice and potentially valuable information is therefore lost; consequently, time-lapse data are currently underutilized in clinical decision making. Furthermore, traditional statistical methods often prove insufficient for the large amount of data generated. The rapidly developing field of artificial intelligence (AI) may prove to be the missing link that will utilize the full potential of the technology in daily clinic. Overall, AI refers to various methods where machines can learn to perform intelligent tasks like humans, allowing for automation of various procedures, or better than humans, resulting in improved outcome. In particular, deep learning that is part of a broader family of machine learning methods based on artificial neural networks has been applied for analysis of imaging in medicine. A deep learning process can directly analyze all the time-lapse images and correlate to outcome, without the need for manual annotation parameters. The obvious benefit lies in the potential for full automation, use of all available information instead of a few time-points, and objectiveness. The technology is still in its infancy, but several projects are being undertaken to establish the appropriate use. In a recent publication, the technology has been used to classify embryos and predict fetal heart beat (Tran et al., 2019). With an impressing area under the curve (AUC) of the method to predict pregnancy on the dataset of 0.93 (95% CI 0.92–0.94), full reproducibility, and transferability, the approach is extremely promising. Apart from the automation and objectivity, it may render the need for developing prediction models based on few parameters redundant.

9.3.1 Types and Characteristics of TLT devices

Stay updated, free articles. Join our Telegram channel

Join