IN-VITRO FERTILIZATION (IVF) OUTCOMES USING EMBRYO MORPHOLOGY AND PATIENT FERTILITY DATA FROM INDIAN COHORTS: A FEASIBILITY STUDY AND PROPOSED MULTIMODAL FRAMEWORK

Abstract

In-vitro fertilization (IVF) outcome assessment depends heavily on the manual morphological grading of embryos, a process that is subjective, prone to inter-observer variability, and difficult to reproduce across clinics. Artificial intelligence (AI) can support more objective and consistent embryo assessment, yet many models remain image-only, opaque, and validated on populations that under-represent Indian cohorts. This paper is organised in two clearly separated parts. The implemented part is a feasibility study: a transfer-learning baseline based on a ResNet34 convolutional neural network pretrained on ImageNet was trained to classify embryo morphology images as good or not-good using the publicly available Hung Vuong Hospital embryo dataset (840 labelled images, a single stratified 80/20 split). The baseline reached approximately 98% training accuracy and 86% validation accuracy (95% confidence interval 80–90%), with an area under the receiver operating characteristic curve of 0.77. Because the validation set is markedly imbalanced (143 not-good versus 25 good), we report metrics that remain informative under imbalance: the majority not-good class is classified well (F1-score 0.92) while the minority good class is weak (F1-score 0.45), giving a balanced accuracy of 0.67, a macro-averaged F1-score of 0.69, a Matthews correlation coefficient of 0.38, and a Cohen’s κ of 0.37; the diagnostic summaries derived from the same confusion matrix (specificity 0.94, Youden’s J 0.34, geometric mean 0.61, positive and negative likelihood ratios 6.36 and 0.64) are reported so that the baseline is characterised beyond headline accuracy. The trained model was deployed in an interactive web interface that returns the predicted class, a confidence score, and the class-probability distribution. The proposed part, which is not yet implemented, outlines a multimodal and explainable framework that would fuse embryo images with patient fertility data from Indian cohorts, add gradient- and Shapley-based explanations, and deliver predictions through a clinician facing decision-support interface; Grad-CAM and Shapley attribution are described only as components of that proposed methodology, and the trained explanations, additional architectures, and cross-validated multimodal results all belong to the same stage. By keeping the implemented feasibility baseline distinct from the proposed framework, the study confirms the practical feasibility of AI-assisted embryo assessment and sets out a concrete, transparent path toward clinically meaningful IVF outcome prediction.