A machine learning system with reinforcement capacity for predicting the fate of an ART embryo

The aim of this work was to construct a score issued from a machine learning system with self-improvement capacity able to predict the fate of an ART embryo incubated in a time lapse monitoring (TLM) system. A retrospective study was performed. For the training data group, 110 couples were included and, 891 embryos were cultured. For the global setting data group, 201 couples were included, and 1186 embryos were cultured. No image analysis was used; morphokinetic parameters from the first three days of embryo culture were used to perform a logistic regression between the cell number and time. A score named DynScore was constructed, the prediction power of the DynScore on blastocyst formation and the baby delivery were tested via the area under the curve (AUC) obtained from the receiver operating characteristic (ROC). In the training data group, the DynScore allowed the blastocyst formation prediction (AUC = 0.634, p < 0.001), this approach was the higher among the set of the tested scores. Similar results were found with the global setting data group (AUC = 0.638, p < 0.001) and it was possible to increase the AUC of the DynScore with a regular update of the prediction system by reinforcement, with an AUC able to reach a value above 0.9. As only the best blastocysts were transferred, none of the tested scores was able to predict delivery. In conclusion, the DynScore seems to be able to predict the fate of an embryo. The reinforcement of the prediction system allows maintaining the predictive capacity of DynScore irrespective of the various events that may occur during the ART process. The DynScore could be implemented in any TLM system and adapted by itself to the data of any ART center.

Automated summary

The aim of this study was to construct a machine learning system-generated score to predict the fate of ART embryos incubated in a time lapse monitoring system. The DynScore showed promising predictive power for blastocyst formation in both training and global setting data groups, with the potential for further improvement through system reinforcement. The system could be adaptable to any ART center's data, maintaining its predictive capacity despite potential events during the ART process.