SAT-225 Machine Leaning Models to Predict Personalized Response to Ovarian Stimulation during In Vitro Fertilization (IVF)

Abstract

In Vitro fertilization (IVF) is a widely accepted method for attaining pregnancy in sub-fertile couples. A primary step in IVF is controlled ovarian hyperstimulation (COHS) by reproductive hormones, including Follicle Stimulating Hormone (FSH), to stimulate the production of eggs that can be harvested for IVF. However, there is substantial variation in women’s response to the COHS protocol: almost 40% of patients either respond poorly (<5 eggs) or hyper-stimulate (>15 eggs). Suboptimal response can reduce the chances of pregnancy while ovarian hyperstimulation can be associated with serious health consequences for the patient. Thus, there is an unmet need for algorithms that can help predict response to controlled ovarian hyperstimulation for IVF and help guide treatment decisions. We hypothesized that ovarian response is related to and modulated by several host-level variables, such as age, body mass index (BMI), luteinizing hormone (LH), follicle stimulating hormone (FSH), antral follicular count, anti-Mullerian Hormone (AMH) and estradiol concentrations, that have been shown in previous studies to be associated with the number of oocytes produced (outcome) during COHS. Therefore, we posited that entrained neural networking algorithms that simultaneously take into account these host level variables can be used to predict ovarian response in individual patients. Accordingly, we used a neural network training platform to predict ovarian response using baseline characteristics of IVF patients including age, BMI, smoking status and initial estrogen levels. Linear and multilayer neural network regressors were executed with 80% of data used for training and 20% for prediction testing. From the 441 observations, optimized artificial neural network accurately predicted oocyte number for 71% of women within ±1 eggs SD. The performance of the ANN will be presented in comparison to alternative models. Conclusion: Artificial Neural Networking Algorithms can be entrained and used to predict ovarian responses in subfertile patients undergoing IVF. Additional research is needed to optimize input parameters and predictive value of such algorithms in guiding personalized protocols for controlled ovarian hyperstimulation during IVF.