Development of a generally applicable morphokinetic algorithm capable of predicting the implantation potential of embryos transferred on Day 3

Abstract

Study question: Can a generally applicable morphokinetic algorithm suitable for Day 3 transfers of time-lapse monitored embryos originating from different culture conditions and fertilization methods be developed for the purpose of supporting the embryologist's decision on which embryo to transfer back to the patient in assisted reproduction?

Summary answer: The algorithm presented here can be used independently of culture conditions and fertilization method and provides predictive power not surpassed by other published algorithms for ranking embryos according to their blastocyst formation potential.

What is known already: Generally applicable algorithms have so far been developed only for predicting blastocyst formation. A number of clinics have reported validated implantation prediction algorithms, which have been developed based on clinic-specific culture conditions and clinical environment. However, a generally applicable embryo evaluation algorithm based on actual implantation outcome has not yet been reported.

Study design, size, duration: Retrospective evaluation of data extracted from a database of known implantation data (KID) originating from 3275 embryos transferred on Day 3 conducted in 24 clinics between 2009 and 2014. The data represented different culture conditions (reduced and ambient oxygen with various culture medium strategies) and fertilization methods (IVF, ICSI). The capability to predict blastocyst formation was evaluated on an independent set of morphokinetic data from 11 218 embryos which had been cultured to Day 5. PARTICIPANTS/MATERIALS, SETTING,

Methods: The algorithm was developed by applying automated recursive partitioning to a large number of annotation types and derived equations, progressing to a five-fold cross-validation test of the complete data set and a validation test of different incubation conditions and fertilization methods. The results were expressed as receiver operating characteristics curves using the area under the curve (AUC) to establish the predictive strength of the algorithm.

Main results and the role of chance: By applying the here developed algorithm (KIDScore), which was based on six annotations (the number of pronuclei equals 2 at the 1-cell stage, time from insemination to pronuclei fading at the 1-cell stage, time from insemination to the 2-cell stage, time from insemination to the 3-cell stage, time from insemination to the 5-cell stage and time from insemination to the 8-cell stage) and ranking the embryos in five groups, the implantation potential of the embryos was predicted with an AUC of 0.650. On Day 3 the KIDScore algorithm was capable of predicting blastocyst development with an AUC of 0.745 and blastocyst quality with an AUC of 0.679. In a comparison of blastocyst prediction including six other published algorithms and KIDScore, only KIDScore and one more algorithm surpassed an algorithm constructed on conventional Alpha/ESHRE consensus timings in terms of predictive power.

Limitations, reasons for caution: Some morphological assessments were not available and consequently three of the algorithms in the comparison were not used in full and may therefore have been put at a disadvantage. Algorithms based on implantation data from Day 3 embryo transfers require adjustments to be capable of predicting the implantation potential of Day 5 embryo transfers. The current study is restricted by its retrospective nature and absence of live birth information. Prospective Randomized Controlled Trials should be used in future studies to establish the value of time-lapse technology and morphokinetic evaluation.

Wider implications of the findings: Algorithms applicable to different culture conditions can be developed if based on large data sets of heterogeneous origin.

Study funding/competing interests: This study was funded by Vitrolife A/S, Denmark and Vitrolife AB, Sweden. B.M.P.'s company BMP Analytics is performing consultancy for Vitrolife A/S. M.B. is employed at Vitrolife A/S. M.M.'s company ilabcomm GmbH received honorarium for consultancy from Vitrolife AB. D.K.G. received research support from Vitrolife AB.

Keywords: algorithm; blastocyst; decision support tool; embryo selection; implantation; morphokinetic; prediction model; time-lapse.

Figure 1

Classification tree diagram illustrating the KIDScore algorithm, which was based on morphokinetic information from embryo culture up until Day 3. The Scores 1–5 are indicated for each tree split with the corresponding number of embryos and implantation rate (Impl.). tPNf, time (h) from insemination to pronuclei fading at the 1-cell stage. t2, t3, t5, time (h) from insemination to the 2, 3 or 5-cell stage, respectively. cells66h, number of cells 66 h after insemination.

Figure 2

Percentage of embryos that developed into blastocysts (light grey) and top or GQ blastocysts (Gardner: 3AA, 4AA, 5AA, 3/4/5BB, AB or BA; dark grey) shown according to their assigned score groups for all evaluated algorithms. GQ, good quality.

Figure 3

The predictive capability of the evaluated algorithms is illustrated for blastocyst formation (A) and blastocyst quality (B). Predictive capabilities are shown as RPV (Gini Index; [2 × AUC]-1) with a theoretical range from 0 to 1, with 95% confidence intervals illustrated by error bars. The Alpha/ESHRE timings algorithm (open circle) was used as common reference. Meseguer (Meseguer et al., 2011); Basile (Basile et al., 2015); Eeva I, (Conaghan et al., 2013); Eeva II (VerMilyea et al., 2014); Milewski (Milewski et al., 2015); Liu (Liu et al., 2016); Alpha/ESHRE (Alpha Scientists in Reproductive Medicine and ESHRE Special Interest Group of Embryology, 2011). RPV, relative predictive value; AUC, area under the curve.