PREDICTING POLYCYSTIC OVARY SYNDROME THROUGH MACHINE LEARNING TECHNIQUE USING PATIENTS’ SYMPTOM DATA AND OVARY ULTRASOUND IMAGES

Abstract

Polycystic ovary syndrome (PCOS) is the most prevalent endocrinological abnormality & one of the primary causes of anovulatory infertility in women globally. The real-world clinical PCOS detection technique is critical since the accuracy of interpretations being substantially dependent on vast numbers of symptoms & physician’s expertise. An artificially intelligent PCOS detection system might be a feasible alternative to the typical diagnostic procedure. Thus, the objectives of this study are: to propose intelligent computer-aided PCOS detection techniques based on patient symptom data & ovary ultrasonography(USG) images; as well as to compare the performances of the proposed techniques with the existing machine learning (ML) based methodologies. To achieve these objectives, firstly, a modified ensemble ML classification technique has been proposed for PCOS detection with patients’ symptom data utilizing state-of-the-art stacking technique with five traditional ML models as base learners & one bagging or boosting ensemble model as meta-learner. At this phase, three distinct types of feature selection methods are applied to explore the minimal & optimal features for PCOS detection. Secondly, for PCOS prediction using ovary USG images, an extended ML classification technique has been proposed, trained & tested over 594 ovary USG images; where the Convolutional Neural Network (CNN) with different stateof-the-art techniques & transfer learning has been employed for feature extraction from the images; then stacking ensemble machine learning technique using conventional models as base learners & bagging or boosting ensemble model as meta-learner have been used on that reduced feature set to classify between PCOS & non-PCOS ovaries. Finally, the comparative analysis has revealed that the proposed techniques for both cases significantly enhances the predictive performances in comparison to the existing ML based techniques. In case of symptom data, the proposed ensemble technique with ‘Gradient Boosting’ classifier as meta learner outperforms others with 95.7% accuracy while using the features selected using PCA method. Using the proposed technique with USG images, the best performing results are obtained by incorporating the ‘VGGNet16’ pre-trained model with CNN architecture as feature extractor and then stacking ensemble model with the meta-learner being ‘XGBoost’ model as image classifier with an accuracy of 99.89% for classification.