A Survey on Diagnosing Parkinson's Disease Severity in Patients

Abstract

Parkinson’s Disease (PD) is a progressive neurodegenerative disorder with diverse symptom trajectories, complicating the accurate assessment of disease severity. Precise evaluation is critical for treatmentplanning and personalized care. This survey presents a comprehensive review of computational approaches developed between 2021 and 2025 for PD severity diagnosis, focusing on machine learning and deep learning techniques across multiple data modalities, including gait, voice, neuroimaging, and multimodal sensors. Methods range from traditional classifiers like SVM and Random Forests to advanced models such as CNNs, BiLSTMs, and Transformers. Gait-based models using wearable data have achieved over 97% accuracy, while voice-based systems using LightGBM and XGBoost report up to 96% accuracy. Imaging models leveraging CNN variants on MRI and EEG data exceed 98% performance, particularly when combined with demographic normalization. Multimodal frameworks integrating clinical, demographic, and physiological data show strong diagnostic capability, with AUC values reaching 1.0 in severity classification. Despite promising results, challenges remain in data diversity, sample size limitations, model interpretability, and clinical scalability. This paper identifies current trends such as mobile-basedmonitoring, longitudinal modeling, and explainable AI, offering insights for the development of robust and generalizable PD severity assessment tools.