Forecasting Outcomes of Assisted Reproductive Treatments Using Artificial Networks (FORTUNE) classification system: a...

Can a prediction model classify IVF patients into distinct prognostic groups based on their expected yield of euploid blastocysts?

Five distinct prognostic groups were identified, with chance of obtaining at least one euploid blastocyst ranging from <1% to 2% in very poor to ∼95% in very good prognosis groups.

Euploid blastocyst yield is a critical determinant of IVF success. While female age strongly influences embryo euploidy, other factors like ovarian reserve markers, partner age, and BMI may also contribute. Current approaches rely on basic and somewhat arbitrary classification of ovarian reserve markers and patient age, which are unable to represent the granular multidimensional relationship between them. A systematic approach to classify patients based on their expected euploid yield would enable better treatment planning and patient counseling.

A retrospective analysis of 10 774 IVF cycles from 8256 couples undergoing pre-implantation genetic testing for aneuploidy (PGT-A) of all available blastocysts from years 2020 to 2023 and a temporal validation cohort of 2089 cycles of 2089 patients from year 2024. The prediction model was developed using generalized additive models for location, scale, and shape using a development cohort from 2020 to 2023, and cross-validated through exhaustive 5-fold cross-validation. Temporal validation was performed using an entirely separate cohort from 2024. Model performance was assessed through calibration plots and discrimination metrics.

Couples undergoing IVF with PGT-A of all their available blastocysts were included. Model included female age, partner age, anti-Müllerian hormone (AMH), antral follicle count, and BMI as predictors. Non-linear associations were captured using neural networks and restricted cubic splines. Missing data were handled using multivariate imputation by chained equations.

The median female age was 36.3 years (IQR: 33.3-39.5) and AMH was 2.0 ng/ml (IQR: 1.0-3.8). Models for predicting ≥1, ≥2, and ≥3 euploid blastocysts yield achieved very good discrimination performance in 5-fold cross-validation samples with mean AUCs of 0.834, 0.849, and 0.861, respectively. Models showed negligible shrinkage (<1%) between training and cross-validation sets with near-perfect calibration slopes (mean: 1.00, IQR: 0.99-1.01) and intercepts (mean: 0.015, IQR: 0.00-0.03). Using predicted absolute counts of euploid blastocysts, five distinct prognostic groups were created based on predicted euploid blastocyst yield. Patients in the very poor prognosis group had 98.3% probability of obtaining no euploid blastocysts after stimulation while the probabilities were 80.2%, 47.5%, 15.8%, and 4.7% in poor, borderline, good, and very good prognosis groups. The chances of obtaining ≥3 euploid blastocysts were 79.8%, 43.7%, 8.9%, 0.2%, and 0% in very good, good, borderline, poor, and very poor prognosis groups. In the temporal validation set (n = 2089), which constituted the first cycles of patients that were treated, the rates of no euploid blastocysts obtained at the end of stimulation were 100.0%, 82.6%, 47.4%, 13.1%, and 4.4% in the very poor, poor, borderline, good, and very good prognosis groups. The rates of three or more euploid blastocysts obtained at the end of stimulation in the temporal validation set were 83.5%, 51.0%, 10.7%, 0.6%, and 0.0% for very good, good, borderline, poor, and very poor prognosis groups. The FORTUNE (Forecasting Outcomes of Reproductive Treatments Using artificial Networks) model is available for use and further validation at https://epsilonkappa-analytics.shinyapps.io/FORTUNE and App Store for iOS mobile devices https://apps.apple.com/en/app/fortune-ivf/id6747190429.

Single-center study design may limit generalizability. The model does not account for laboratory-specific factors or stimulation protocols.

This novel classification system provides objective, personalized counseling for IVF patients regarding expected euploid yield, enabling better-informed decision-making about treatment options and number of planned stimulation cycles.

There was no funding needed for this study. The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

N/A.