Predicting Human Embryonic Development Potential Based On Artificial Intelligence (AI)

In human assisted reproductive technology(ART),accurate evaluation and screening of the single embryo with the most developmental potential for transfer to patients are the key to a good clinical outcome.Currently,the most widely used method of embryo screening worldwide is still the traditional morphological assessment,which is subjective and limited in assessing the consistency of embryo development.In recent years,using artificial intelligence(AI)technology to assist embryo selection has become a research hotspot.In this study,we collected developmental videos of embryos cultured in a time-delayed incubator at Chengdu Xinian Women’s Hospital between 1 January 2019 and 31 December2021,and conducted the development and validation of an AI embryonic developmental potential prediction model,with a view to compensating for the shortcomings of embryo assessment by embryo laboratory operators at the current stage and reducing the influence of subjective factors,and improving the accuracy of the assessment of embryonic developmental potential.The main research contents and results are as follows:(1)A total of 6334 embryonic development videos labelled as usable and non-usable blastocysts were included as data for model development and validation,of which 2800 were used for model construction(experimental group)and 6334 for final model performance evaluation(validation group).Deep convolutional neural networks,long and short-term memory networks and attentional mechanisms were selected to construct the AI embryo developmental potential prediction model,and the accuracy,area under the curve of the subject’s working characteristic curve,F1 score,accuracy,sensitivity and specificity calculated by confusion matrix were used to measure the performance of the model.(2)The 5-fold cross validation conducted on the experimental dataset showed that the average accuracy,AUC,F1 score,accuracy,sensitivity,and specificity of the AI embryo development potential prediction model in predicting available blastocyst formation were0.7171,0.7749,0.7356,0.6905,0.7872,and 0.6471,respectively.(3)By comparing with six other machine learning models,the AI embryonic developmental potential prediction model proposed in this study showed highe accuracy in predicting usable blastocyst formation(0.7129 vs 0.5311/ 0.5761/ 0.6629/ 0.5100/ 0.5579/0.6750).(4)By exploring the effect of image characteristics on model performance showed that the model performed best in predicting available blastocyst formation when using an image size of 150 × 150 pixels as input,but with very little difference from when using 100 × 100 pixels and 200 × 200 pixels sizes.In addition,the model showed the best performance when using 32 frames of images as input,but the difference in performance with using 16 and 64 frames of image frames was very small.(5)Experimental results from validation of the AI embryonic developmental potential prediction model on the validation set dataset showed that the average accuracy,AUC,F1 score,precision,sensitivity and specificity of the model in predicting usable blastocyst formation were 0.7169,0.7860,0.6924,0.6575,0.7375 and 0.6354 respectively,which differ ed very little from those on the experimental set dataset.In summary,this study for the first time combines attention mechanism with artificial intelligence deep learning to establish an AI embryo development potential prediction model based on convolutional neural network + short-term memory network + attention mechanism(CNN-LSTM Attention).It can predict the formation of usable blastocysts by analyzing videos from the first three days of embryo development and has relatively good performance.With further analysis of clinical data in subsequent work,this technology is expected to be applied in clinical practice to assist embryologists in selecting quality embryos for transfer,thereby improving the success rate and efficiency of ART.