14August2025
14:00 Master's Defense IC3 Auditorium
Topic on
Machine learning techniques for detecting biometric patterns associated with Parkinson's disease from wearable sensor data
Student
Jesús Francisco Paucar Escalante
Advisor / Teacher
Esther Luna Colombini - Co-supervisor: Aurea Rossy Soriano Vargas
Brief summary
Parkinson’s disease (PD) is a neurodegenerative disorder that severely affects motor and non-motor functions, with resting tremor, bradykinesia, and rigidity as the main symptoms. Early and accurate diagnosis is crucial, but traditional methods, especially subjective clinical scales such as the Unified Parkinson’s Disease Rating Scale (UPDRS), often suffer from variability and inconsistencies. Wearable sensors, such as accelerometers and gyroscopes, have emerged as valuable tools for monitoring PD symptoms in real-world settings. However, challenges remain in applying machine learning (ML) to PD diagnosis, especially with signal corruption due to motion artifacts, environmental noise, and hardware limitations. These issues degrade model performance and hinder accurate classification, highlighting the need for improved preprocessing techniques. This research explores the potential of machine learning for tremor classification in PD by reviewing several models, including classical approaches (e.g., Support Vector Machines) and advanced deep learning methods (e.g., Convolutional Neural Networks, Long Short-Term Memory Networks). The study addresses key challenges of ML for PD, including data partitioning, overfitting, and generalization across diverse patient populations. A key contribution of this work is the introduction of Stable Diffusion Models (SDMs) for denoising inertial sensor signals, improving training stability and control over signal characteristics. By using SDMs to recover clean signals, the aim is to improve the extraction of tremor-related features, focusing on subtle tremor frequencies that are often hidden by noise. The study demonstrates that integrating SDMs with deep learning classification models significantly improves tremor classification performance, with higher accuracy and robustness. This work offers an innovative method for dealing with noisy sensor data and presents a path towards more reliable and clinically applicable assessments of tremor in PD.
Examination Board
Headlines:
| Esther Luna Colombini | IC / UNICAMP |
| Sandra Eliza Fontes de Avila | IC / UNICAMP |
| Agma Juci Machado Traina | ICMC / USP |
Substitutes:
| Fátima de Lourdes dos Santos Nunes Marques | EACH / USP |
| Anderson de Rezende Rocha | IC / UNICAMP |