O-246 A cutting-edge artificial intelligence system tracking embryo development and pinpointing the optimal day for blastocyst utilization

Abstract

Abstract Study question Can an AI model predict at which specific day (D) of development a blastocyst will reach expansion at a usable stage for biopsy/cryopreservation/transfer? Summary answer Our model achieved a high accuracy tracking embryos’ developmental times while forecasting the day when they would reach a usable blastocyst stage. What is known already Forecasting the optimal day for blastocyst utilization is crucial. It can facilitate informed decisions on transfer/biopsy/cryopreservation, while streamlining the laboratory workflow. Previous prediction models on this task relied solely on embryo morphological and morphokinetic data. However, emerging approaches, incorporating computer image analysis, offer an opportunity to enhance their accuracy and objectivity. To our knowledge, none has specifically predicted the precise day of embryo development when the blastocyst has expanded enough to become usable. This information is invaluable to plan workload distribution on D5/6/7 of development. Moreover, with the rise of PGT-A, optimizing the time allocated for biopsy becomes essential. Study design, size, duration Retrospective study conducted at a tertiary IVF centre, utilizing data from June 2015 to January 2023. A total of 14,704 time-lapse videos (TLV) from embryos of 1,805 patients, cultured across 2,475 cycles, were analyzed split into train/validation/test by patient (to avoid data leaking) with 60/20/20 proportion. Outcome parameters encompassed the times for all embryo developmental stages or morphokinetic parameters, blastocyst formation and utilization, as well as the specific day of blastocyst utilization. Participants/materials, setting, methods Embryos from patients with primary/secondary infertility were included. Laboratory culture conditions followed normal clinical routine. All embryos were monitored using embryoscope time-lapse systems and annotated by experienced embryologists. Blastocysts graded ≥BL3CC (Gardner) were biopsied/cryopreserved/transferred on D5/6/7. Only videos depicting embryos with normal fertilization were considered in the model. Predictions for the utilization day were made 48-hours and 24-hours ahead of the blastocyst usable stage. Main results and the role of chance We developed a customized variant of the Temporal Fusion Transformer model, modified for analyzing time-sequence-images from TLV. This model architecture allows the analysis of the sequence and clinical variables, identifying its most influential aspects from the model’s perspective. In the first two tasks of blastocyst prediction and forecasting blastocyst utilization day, data was split into training-set: 8,802 TLV (1,083 patients); validation-set: 2,925 TLV (361 patients); and test-set: 2,977 TLV (361 patients). Our trained model achieved a strong performance on ‘early warning’ 48-hour ahead forecasting of embryo reaching a usable blastocyst stage. Utilizing only unlabelled time-lapse image observations of the embryo, it reached 90 AUROC score on the test-set. For ‘late warning’, 24-hour ahead, we reached 97 AUROC score on the test-set, almost ideal prediction. Finally, 91 AUROC score was achieved on the binary blastocyst stage classification task, just considering the blastocyst formation at any expansion. In the embryo morphokinetic stage-labelling task, data was split into training-set: 8,420 TLV (1,078 patients); validation-set: 2,777 TLV (360 patients); and test-set: 2,780 TLV (360 patients). We achieved competitive results with 95.5 macro-averaged AUROC score for multi-class (17 classes) prediction for each time-lapse-image. Each individual parameter from tPB2 to tHB reached a AUROC higher than 90. Limitations, reasons for caution The model needs to undergo an extensive cross-validation, to evaluate its performance, followed by a prospective validation. Wider implications of the findings This model belongs to the realm of IVF lab automation, offering substantial assistance in daily operations. It aids embryologists not only in annotating developmental stages with high accuracy, but also predicts the blastocyst utilization day, thereby providing crucial support for efficient workload planning and streamlined operations. Trial registration number NA