Mathematical Modelling and Explainable Multimodal Deep Learning for ICSI Outcome Prediction in Assisted Reproductive Technology

Abstract

The intracytoplasmic sperm injection (ICSI) is a well-known assisted reproductive technology for treating infertile couples. Predicting ICSI outcomes can be difficult due to several factors affecting the treatment success rate. Existing prediction frameworks usually consider one or a few independent clinical variables or utilize unimodal machine learning algorithms. Thus, their effectiveness is limited by the inability to detect the interconnections between heterogeneous reproductive data. In this paper, we present a novel approach based on mathematical modeling and explainable multimodal deep learning for the prediction of ICSI outcome in assisted reproductive technology. The proposed method combines clinical information, sperm characteristics, oocyte characteristics, embryological information imaging representation into an integrated multimodal architecture. We introduce a mathematical fusion algorithm which learns the importance of each modality a probabilistic prediction layer, which predicts the probability of achieving ICSI success. We use explainable artificial intelligence approaches to define the significant features that influence the probability of success and help to interpret the model better. The proposed approach allows to structure the process of integrating heterogeneous data while preserving mathematical transparency and interpretability of predictions.