Search for a command to run...
Abstract Study question Can a machine learning (ML) model predict embryo aneuploidy and live births using morphokinetic meta-variables and clinical data? Summary answer The novel ML model LIFE Predict demonstrated robust predictive performance (AUC = 0.824), discerning between embryos leading to live birth (LB) or carrying aneuploidies. What is known already Time-lapse imaging and artificial intelligence have shown promise for improving embryo selection, yet their predictive accuracy remains limited, with AUC values not surpassing 0.75. Deep learning models tend to offer lower transparency and explainability than morphokinetic-based machine learning approaches could potentially provide. However, the use of interpretable variables that aggregate morphokinetic features in predictive models for ploidy and live birth outcomes remains underexplored. Study design, size, duration This is a multicentre, retrospective case-control study that analysed data from 882 blastocysts obtained from nine different fertility clinics. Embryos were cultured under hypoxic conditions, using different media (G-TL, Vitrolife, Göteborg, Sweden; SAGE 1-Step, Cooper Surgical, Trumbull, USA; CSCM-C, Fujifilm Irvine Scientific, Santa Ana, USA) and time-lapse incubators (Embryoscope, Vitrolife, Göteborg, Sweden; Geri, Genea Biomedx, Sydney, Australia). Data collection included ICSI cycles from 2017 to 2024. Participants/materials, setting, methods A total of 882 embryos (487LB, 395 diagnosed as aneuploid by PGT-A trophectoderm biopsies) were used for model training and testing (V-fold and leave-one-out cross-validation). The predictive features consisted of clinical data, morphokinetic variables, and two meta-variables derived from morphokinetics. Model performance included AUC, accuracy, sensitivity, and specificity. Odds ratios (OR) and 95% confidence intervals (CI, p < 0.05) were used to assess the association between aneuploidy outcomes and both meta-variables and model output (LIFE Predict Score). Main results and the role of chance Significant differences in morphokinetic features were observed between embryos that resulted in live births and those reported as aneuploids. Two novel meta-variables were strongly associated with the embryo outcomes. The first meta-variable, representing the sum of morphokinetic features with values outside the reference range (using LB as the benchmark), showed a notable association with embryo outcomes (OR = 4.36, 95% CI:2.30–6.25). The second meta-variable, denoting the sum of morphokinetic features with Mean Absolute Error (MAE) values above a specific cut-off (based on the MAE of LB) between the morphokinetic values recorded by embryologists and those calculated by a regression machine learning model designed for each event, also obtained significant results (OR = 2.43, 95% CI:1.61–3.72). The predictive model achieved the following metrics: AUC=0.824, accuracy=0.750, sensitivity=0.814, specificity=0.662. The numerical output of the predictive model was converted into the LIFE Predict Score by scaling it from 0 to 10. Higher accuracy was found with the score values at the lower and higher ends. Specifically, scores below 3 contained 84% of aneuploid embryos (OR = 13.3, 95% CI:3.01-19.8). In contrast, this percentage decreased to 22.8% for ratings between 7 and 9 (OR = 0.315, 95% IC:0.197-0.458) and further dropped to 9.7% for ratings exceeding 9 (OR = 0.105, 95% CI: 0.006-0.212). Limitations, reasons for caution As a retrospective study, the findings require prospective validation (in process) across broader populations and clinical settings to confirm generalizability. Wider implications of the findings This ML model offers an interpretable, data-driven approach to enhance embryo selection and optimize IVF outcomes, highlighting the importance of morphokinetic annotations and clinical data integration. The use of meta-variables allows for a broader analysis, identifying variables that fall outside the reference range and enhancing explainability. Trial registration number No