AI-Based Quality of Voice Analysis Models for Clinical Use, Insights of Quality of Models from 19 Parkinson’s Disease Studies (2013-2023)

Voice analysis, powered by Artificial Intelligence (AI) and Machine Learning (ML), has emerged as a valuable tool for detecting and monitoring voice disorders. By identifying vocal biomarkers, AI-driven models can facilitate early diagnosis, track disease progression and support clinical decision-making. This study systematically evaluates the effectiveness and quality of various ML models applied in the 19 studies of AI-related voice analysis in Parkinson’s’ Disease retrieved from The Royal Society of Medicine Library UK, spanning the period from 2013 to 2023. The models assessed include Support Vector Machines (SVM), Convolutional Neural Networks (CNNs), Random Forest (RF) and hybrid CNN-LSTM architectures. Their performance is examined based on accuracy, sensitivity, specificity and error metrics such as Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). Findings indicate that SVM consistently delivers high accuracy (up to 96%) and is particularly effective for small to medium-sized voice-related datasets with pre-engineered datasets. CNNs achieve superior performance (up to 97%) on large, feature-rich datasets; however, their computational demands and limited validation constrain scalability. Random forest models demonstrate robustness in handling imbalanced datasets, while CNN-LSTM hybrids show potential by integrating spatial and temporal feature extraction, though they require further validation. A critical limitation identified in the analyzed studies is the lack of detailed dataset descriptions, diversity and real-world applicability, which restricts comparison with other studies and generalizability. This paper highlights the strengths and limitations of current models for AI-driven voice analysis approaches and emphasizes the need for standardized, diverse datasets and enhanced evaluation metrics to advance AI applications in voice disorder diagnostics and monitoring.