Deep Inception‑ResNet: A Novel Approach for Personalized Prediction of Cumulative Pregnancy Outcomes in Vitro Fertilization Treatment (IVF)

• 
• 
• 
• 

Infertility can be described as the inability of a couple to achieve pregnancy, even after trying to conceive for over 12 months. Numbers suggest that approximately 48 million couples are currently facing fertility-related problems globally [1]. To overcome infertility issues, advanced assisted reproductive technology (ART) procedures are being used extensively. Additionally, the willingness to achieve pregnancy later in life by preserving eggs or embryos has also contributed to the growth of the adoption of ART. One such most effective and common type of ART is in vitro fertilization (IVF), which involves the fertilization of retrieved eggs and viable sperm in a laboratory. As per an article, around 2–2.5, lakh IVF cycles are being performed in India every year [2]. With the objective of forecasting IVF outcomes using ML and DL algorithms, many papers have been published in the past. One of the recent live birth prediction models was given by Goyal et al. [3]. An open-access dataset which was provided by Human Fertilization and Embryology Authority (the year 2010–2016) was analyzed. Brás de Guimarães et al. [4] constructed an artificial neural network (ANN) to estimate the likelihood of live birth. In another study, Raef et al. [5] developed a model where β-hCG (Human chorionic gonadotropin hormone) was the target variable in clinical data (April 2016 to February 2018) to filter out insignificant attributes. Hassan et al. [6] proposed a hill climbing-based feature selection approach. Qiu et al. [7] developed a pre-treatment parameter- based live birth rate predictor. However, consideration of treatment parameters limits the use of the first four above-mentioned models, making them unsuitable when only pre-treatment parameters are known. Additionally, for tabular datasets, the applied deep Inception-Residual Network-based approach [8, 9] has remained unexplored in these research studies. Also, the major limiting aspect of all these models is that they cannot be used for Indian sub-population due to socio-cultural discrepancies and differences in the data distribution when compared to countries in the West.

Our study presents a patient-centric artificially intelligent system, that predicts IVF success rate by understanding the complex relationship between multiple significant pre-treatment parameters specific to the Indian context. In this study, the historic data of 2268 Indian patients with 79 features, who opted for IVF-ICSI (intracytoplasmic sperm injection) procedure from January 2018 to December 2020, were acquired from a well-known medical center. Pre-treatment parameters such as maternal age (patient age during the IVF procedure), number of IVF cycle (course of IVF treatment, that starts from the first day of the menstrual period till β-hCG test), type of infertility (primary or secondary infertility), duration of infertility (time span over which the efforts of a couple trying to conceive were futile), AMH (ng/ml), indication for IVF (inducing factor behind infertility), sperm type (on the basis of mortality and sperm characteristics it was divided into Normal or Male Factor), BMI, embryo transfer (Only Fresh cycle, Fresh cycle + Subsequent Frozen Embryo Transfer (FET) and Freeze All Cycles) were considered as the most significant features to predict the positive β-hCG rate.

• 
• 
• 
•