A novel machine-learning framework based on early embryo morphokinetics identifies a feature signature associated with blastocyst development

BACKGROUND

Artificial Intelligence entails the application of computer algorithms to the huge and heterogeneous amount of morphodynamic data produced by Time-Lapse Technology. In this context, Machine Learning (ML) methods were developed in order to assist embryologists with automatized and objective predictive models able to standardize human embryo assessment. In this study, we aimed at developing a novel ML-based strategy to identify relevant patterns associated with the prediction of blastocyst development stage on day 5.

METHODS

We retrospectively analysed the morphokinetics of 575 embryos obtained from 80 women who underwent IVF at our Unit. Embryo morphokinetics was registered using the Geri plus® time-lapse system. Overall, 30 clinical, morphological and morphokinetic variables related to women and embryos were recorded and combined. Some embryos reached the expanded blastocyst stage on day 5 (BL Group, n = 210), some others did not (nBL Group, n = 365).

RESULTS

The novel EmbryoMLSelection framework was developed following four-steps: Feature Selection, Rules Extraction, Rules Selection and Rules Evaluation. Six rules composed by a combination of 8 variables were finally selected, and provided a predictive power described by an AUC of 0.84 and an accuracy of 81%.

CONCLUSIONS

We provided herein a new feature-signature able to identify with an high performance embryos with the best developmental competence to reach the expanded blastocyst stage on day 5. Clear and clinically relevant cut-offs were identified for each considered variable, providing an objective tool for early embryo developmental assessment.

P–241 Construction of a Machine Learning algorithm based on early morphokinetics for human blastocyst development prediction: a retrospective analysis of 575 cleavage-stage embryos

Study question

Can morphokinetic features included into Machine Learning (ML) algorithms identify cleavage-stage embryos with the best chance to reach the expanded blastocyst stage on day 5?

Summary answer

A ML algorithm based on early morphokinetic features can identify cleaving embryos that will reach the expanded blastocyst stage on day 5.

What is known already

To date, the conventional morphology assessment of cleaving human embryos has a limited predictive power on further embryo developmental potential. The morphokinetic analysis using Time-Lapse systems (TLS) was introduced in order to provide a new tool to identify dynamic biomarkers of embryo quality. More recently, ML approach has been applied for the analysis of specific embryo-related features, aiming at developing predictive algorithms to assess the embryo development potential.

Study design, size, duration

We retrospectively analysed 575 embryos obtained from 80 women aged 25–42 years, with normal BMI, AFC≥8, day 3 FSH<12 IU/l, AMH>2.5 ng/ml, no diagnosis of polycystic ovary syndrome or endometriosis. These patients underwent IVF at our IVF Unit between March 2018 and March 2020; their embryos were cultured using the Geri plus® TLS and a single blastocyst transfer was performed.

Participants/materials, setting, methods

A total number of 29 morphological and morphokinetic parameters were considered to build six different ML algorithms. The performance to assess which was the best-fitting algorithm was calculated using the ROC curve considering accuracy (% of embryos correctly classified by the algorithm), Cohen-kappa coefficient (measurement of the agreement among features), mean number of TP (embryos correctly classified as undergoing developmental arrest), mean number of TN (embryos uncorrectly classified as undergoing developmental arrest).

Main results and the role of chance

Overall, 210 embryos progressed to the expanded blastocyst stage on day 5 (BL group), whereas 365 displayed developmental delay or arrest at any stage (nBL group). Among the six different algorithms, the best-fitting algorithm was obtained using the Kbest features selection approach combined with a Random Forrest evaluation strategy. This algorithm was based on 7 variables: embryo morphological score on day 2, pronuclear fading time (tPNf), completion time of cleavage to two, four and eight cells (t2, t4, and t8 respectively), time intervals t4-t3 and t8-t4. The algorithm showed an AUC of 0.78, with an accuracy of 0.73, a Cohen-kappa of 0.41, a mean TP number of 302/365 embryos in the nBL group and a mean TN number of 120/210 embryos in the BL group. Mean false positive (FP) and false negative (FN) numbers were of 63 and 90.2, respectively.

Limitations, reasons for caution

The results obtained in this study may not be generalizable to patients with other clinical characteristics, to other time-lapse systems or different laboratory settings. The predictive power of the algorithm should be validated prospectively on a larger number of embryos.

Wider implications of the findings: The current study represents a preliminary analysis for the development of hierarchical predictive models for embryo assessment based on their developmental potential, that embryologists will be able to apply as a support for decision-making.

Trial registration number

Not applicable