Dados do Trabalho

Title

Prospective performance assessment of an artificial intelligence software for embryo selection and clinical pregnancy prediction

Objective

New technologies are emerging to assist embryologists on selecting the best embryo for transfer. Shortening time to pregnancy and preventing extra emotional distress on ART patients is a crucial step in any IVF treatment. The development of artificial intelligence (AI) algorithms based on machine learning methods to increase reproductive medicine outcomes is a new frontier worldwide. The use of time-lapse technology incubators provides new data regarding embryo morphokinetics and morphological details that were unnoticed. The aim of this study is report the performance of an AI software developed in our center for embryo selection and clinical pregnancy prediction.

Methods

Prospective cohort study including patients undergoing a single embryo transfer (sET, blastocyst stage) after an in vitro fertilization (IVF) treatment between June/2022 and June/2023 (n=230 patients). Embryo selection and clinical pregnancy (presence of gestational sac and heartbeat) prediction software was previously built through artificial neural networks technique and genetic algorithms using the input of morphological data of 1.000 blastocyst with known reproductive outcomes. Blastocyst digital image processing were analyzed considering 33 mathematic variables. All embryos were cultured in a time-lapse incubator (Embryoscope Plus, Vitrolife). Input data were randomized for training, validation and test (70, 15 and 15%, respectively). Software performance in the validation test achieved an area under curve of 0,62 for positive clinical pregnancy (CP) and 0,52 for negative CP. Prediction rates were assessed in elective (2 or more embryos available to transfer, n=148 patients) and non-elective (1 available embryo to transfer, n=82 patients) sET, prospectively analyzed in the software, which classified the embryos accordingly to its chance (%) for clinical pregnancy. Embryos were also analyzed by embryologists according to morphological and morphokinetics features.

Results

In 58.1% of the patients from the elective sET group (n=86/148), the top embryo choice for transfer was concordant between software and embryologists. In this group, CP rate was 62.8%, with a software performance for positive CP of 96.3% (the software may predict none of the embryos in a cohort would achieve a CP). Overall, the performance of the software in cohort (positive + negative CP prediction) was 64%. In non-concordant cases for top embryo choice for elective transfers (n=62, 41.9%), the embryo chosen by the embryologist was transferred and CP rate was 64.8%. The software performance for negative CP was 83.9%. In non-elective sET, CP was achieved by 47.6% of the cohort. Software performance for negative CP was 72.1% and overall positive and negative prediction was 62.2%.

Conclusion

Our software was able to predict 64% of CP outcomes for concordant elective sET and 62.2% for non-elective sET, similar to the performance of validation test (62%). Proper validation of new AI tools for embryo selection and clinical pregnancy prediction is crucial for potentially use in the IVF laboratory. Currently, the software may be used as a support tool for embryologist’s decision.

Keywords

Artificial Neural Networks, Genetic Algorithm, Time-lapse system, embryo selection, clinical pregnancy prediction, non-invasive diagnostic.