An XAI-Integrated BiLSTM-GRU-CNN Framework for Early Polycystic Ovary Syndrome Detection

Polycystic Ovary Syndrome (PCOS) is a chronic hormonal disorder affecting women of reproductive age. Its incurable nature and associated health complications empha-size the dire need for early detection. This study presents a systematic approach for early PCOS Detection (PCOSD) employing Machine Learning (ML) and Deep Learning (DL) models on a dataset of 541 patients. Following pre-processing, feature reduction was performed with Mutual Information (MI) emerging as the ideal technique, validated by Analysis of Variance (ANOVA) test, selecting 15 significant features. Synthetic data augmentation was conducted where Adaptive Synthetic Sampling (ADASYN) identified as the most effective technique based on Kolmogorov–Smirnov (KS) test, increasing the dataset size to 1520. Subsequently, the data was split into training, validation, and test sets, allowing the evaluation of diverse ML classifiers and DL architectures, alongside ensemble and hybrid models. After hyper parameter tuning and Cross-Validation (CV), followed by performance evaluation, a hybrid model was proposed. The framework integrated Bidirectional Long Short-Term Memory (BiLSTM), Gated Recurrent Unit (GRU), and Convolutional Neural Network (CNN). It demonstrated superiority over other models by securing 96.71% accuracy with 97% F1 score, indicating its efficacy in PCOSD. Consequently, Explainable Artificial Intelligence (XAI) techniques were availed for gaining deeper insights into model’s predictions.