A comparative analysis of Parkinson’s disease classification using artificial intelligent techniques for mobile application

• 主要作者: Lee, Teng Hong
• 格式: Thesis
• 語言: 英语; 英语
• 出版: 2025
• 主題: Q Science; QA Mathematics
• 在線閱讀: http://eprints.utem.edu.my/id/eprint/29431/

Parkinson's disease (PD) is a progressive neurodegenerative disorder that severely affects motor functions, particularly gait and balance. The global burden of Parkinson's disease is increasing. At the same time, clinical assessments require specialized settings and are not always accessible. For gait analysis, Detrended Fluctuation Analysis (DFA) has a problem of being too sensitive to noise while the use of Short-Time Fourier Transform (STFT) and Continuous Wavelet Transform (CWT) are not compared extensively. This thesis presents a comprehensive study on classifying Parkinson's disease and its severity using both traditional biosignals such as stride interval and vertical ground reaction force (vGRF), and computer vision-based techniques derived from video analysis. DFA, STFT and CWT were applied to extract meaningful features from biosignals. These features were evaluated using machine learning classifiers, including Support Vector Machine (SVM), k-Nearest Neighbors (KNN), and Random Forest, along with deep learning models such as 1D-Convolutional Neural Networks (1D-CNN), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU). Results showed that SVM, KNN and random forest performed well in classifying both PD and healthy individuals when paired with combination of STFT, CWT and DFA with 100% in precision, recall and F1-score. 1D-CNN demonstrated strong robustness in handling noisy stride interval data, while LSTM and GRU excelled in vGRF classification due to their ability to capture temporal dependencies. Additionally, walking videos were analyzed using the MediaPipe Pose framework, where 33 keypoints were extracted and features computed using the Time Series Feature Extraction Library (TSFEL) library. These features were used to classify PD severity using both machine learning and deep learning models. GRU-LSTM and Random Forest achieved classification precision, recall and F1-score above 80%. A mobile application was developed using the Flutter framework, integrating MediaPipe Pose and Google ML Kit to enable real-time gait analysis from smartphone video. The system classified subjects into healthy, mild PD, or advanced PD categories with high reliability. This work highlights the complementary strengths of signal-based and vision-based approaches for PD assessment and presents a viable framework for remote, non-invasive, and real-time monitoring of gait disorders using mobile technologies.