Data-driven Clinical Decision Support System Using Neural Network Topology Optimization for PCOS Diagnosis

Abstract

Polycystic ovarian syndrome (PCOS) is a prevalent endocrine disorder that affects women of reproductive age worldwide. It is characterized by chronic oligo- or anovulation, clinical hyperandrogenism, and polycystic ovaries. Long-term complications include endometrial cancer, infertility, obesity, and type 2 diabetes mellitus. Early diagnosis and treatment are crucial to reduce the incidence of these complications. Objectives: The aim of this work is to develop a Clinical Decision Support System (CDSS) utilizing Multi-Objective Particle Swarm Optimization (MOPSO) - backpropagation (BP) with artificial neural networks (ANN) to determine the presence of PCOS. Methods: CDSS comprises three subsystems: preprocessing, training, and classification. The preprocessing subsystem manages the missing values and performs correlation-based feature selection with a threshold of 0.25. The training subsystem employs an ANN trained with BP and MOPSO. The local best values are obtained from the BP training seed MOPSO, which has two objective functions: minimizing the mean square error and achieving faster convergence without stagnation. The PCOS dataset from the Kaggle repository is used in the experiments. Results: The developed CDSS achieved an overall accuracy of 92.02%, with 87.27% sensitivity and 94.44% specificity. Conclusion: The CDSS is a valuable second-opinion tool for junior gynaecologists in diagnosing PCOS, offering a robust and accurate method for early detection and intervention.