Feature extraction optimization using particle swarm optimization for voice pathology detection with online sequential extreme learning machine

Voice pathology analysis offers a non-invasive approach for early disease detection. In such systems, machine learning plays a crucial role in accurate classification, yet feature redundancy often reduces efficiency and performance. Th is study presents a voice pathology detection system using the Malaysia Voice Pathology Database (MVPD) and the Online Sequential Extreme Learning Machine (OSELM). Voice signals of the sustained vowel /a/ from both healthy and pathological speakers are processed using Mel-Frequency Cepstral Coefficients (MFCC) for feature extraction. To address redundancy, Particle Swarm Optimization (PSO) is applied for feature reduction, ensuring retention of only the most relevant features. The optimized features are then classified using OSELM. Model performance is evaluated based on accuracy, sensitivity, specificity, precision, and F1-score. Results show that PSO reduces 69.23% of features while preserving essential information, enabling the optimized OSELM model to achieve 96.22% accuracy using only 30.77% of the total features. This research is significant as it utilizes a unique Malaysian dataset with consistent duration and language context, contributing to advancements in voice-based pathology detection and supporting early diagnosis and improved healthcare services in Malaysia.