Exploring the feature prioritization and data sampling of PCOS diagnosis via densely connected attention based squeeze deep learning detection model.

Accurate and timely diagnosis remains a challenge due to the complexity of Polycystic Ovary Syndrome (PCOS) symptoms and data imbalance issues in the existing datasets. This research aims to develop a robust PCOS detection model that addresses these challenges by introducing a novel hybrid methodology with effective feature prioritization while handling data balancing issues. The research involves three major phases: pre-processing, feature selection, and PCOS detection. In the pre-processing stage, dataset balancing is emphasized by the combination of Synthetic Minority Oversampling Techniques (SMOTE) and Edited Nearest Neighbor (ENN). Under this stage, replacing null values, balancing the dataset, and dropping unnecessary columns are accomplished to increase PCOS detection accuracy. The second stage is feature selection, where a distinct hybrid bionic strategy named the Gorilla Salp Swarm Troop Model (GS2TM) is proposed to pick the optimal set of dominant features. The GS2TM algorithm reduces the feature set by 51.1 %, retaining only 23 features while achieving a state-of-the-art accuracy of 98.7 %. In addition, the Densely Connected Attention-Based Squeeze Convolutional Detection Model (DASCD) is proposed for the prediction of PCOS, in which multiple layers are adjusted in a feed-forward manner. The novelty of this work lies in the unified pipeline that simultaneously addresses three major challenges in PCOS detection, such as dataset imbalance (SMOTE-ENN), feature redundancy (GS2TM), and overfitting (DASCD with attention), providing both high accuracy and enhanced interpretability. As a result, the proposed detection model greatly improves accuracy compared to other existing ML-based strategies. Specifically, by utilizing 23 characteristics with GS2TM, the proposed model outperforms with an accuracy of 98.7 % in categorizing PCOS and non-PCOS.