Explainable Deep Tabular Learning with SHAP-Guided Feature Optimization and Squeeze-Excitation Enhanced TabNet for Polycystic Ovary Syndrome Classification

Résumé fourni par la source

Polycystic Ovary Syndrome (PCOS) is one of the most prevalent reproductive endocrinopathies in women of reproductive age. This requires diagnostic systems that reliable and interpretable. This paper presents a deep tabular framework that like to maintain efficiency in the classification of Polycystic Ovarian Syndrome (PCOS) through effective data pre-processing, synthetic balancing of data, and optimizing for significant features. Initially, the clinical dataset is preprocessed by iterative imputation, outlier handling, feature normalization and synthetic balancing for statistical consistency and robust model learning. The study uses Conditional Tabular GAN (CTGAN) to augment the sample diversity and address class imbalance by generating additional minority samples. Moreover, a Transformer-based feature extractor captures associations between features to encode better reference feature embeddings. Then, performing Shapley Additive Explanations (SHAP)-guided feature selection using an XGBoost explainer to determine what predictors are most important in the model performance, hence improving both interpretability and computational efficiency. The TabNet-SE ahead of running the baseline model, i.e.Random Forest (RF) and XGBoost, in addition both conducting of these steps with the complete feature set and also SHAP-selected features. The experimental results show that TabNet-SE consistently outperforms all baseline models in terms of accuracy, precision, recall and F1-score when trained on the set of features optimized with SHAP. Model performance was further corroborated with confusion matrix, Receiver Operating Characteristic (ROC) curve, and other evaluation metrics. This framework not only facilitates accurate diagnosis of PCOS but also enhances interpretability, thus bridging the gap between clinical requirements and model transparency.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.