Freezing of gait detection in Parkinsons Disease using relative power from wearable sensors: a Random Forest approach
Freezing of Gait (FoG) is one of the most debilitating motor symptoms of Parkinsons Disease (PD), characterized by the sudden and temporary inability to start or continue walking. Detecting and monitoring these episodes is fundamental to improving diagnoses and intervention strategies. This paper pr...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Estado: | Versión publicada |
| Fecha de publicación: | 2025 |
| País: | Brasil |
| Institución: | Universidade de São Paulo (USP) |
| Repositorio: | Biblioteca Digital de Teses e Dissertações da USP |
| Idioma: | inglés |
| OAI Identifier: | oai:teses.usp.br:tde-14052025-142659 |
| Acceso en línea: | https://www.teses.usp.br/teses/disponiveis/18/18153/tde-14052025-142659/ |
| Access Level: | acceso abierto |
| Palabra clave: | análise de marcha Doença de Parkinson freezing of gait gait analysis machine learning parkinson's disease processamento de sinais signal processing |
| Sumario: | Freezing of Gait (FoG) is one of the most debilitating motor symptoms of Parkinsons Disease (PD), characterized by the sudden and temporary inability to start or continue walking. Detecting and monitoring these episodes is fundamental to improving diagnoses and intervention strategies. This paper proposes an approach based on spectral analysis of acceleration signals to identify patterns associated with FoG. The methodology employs the extraction of relative power characteristics using Welchs method, as well as the definition of a metric called Frequency Spectral Bands Ratio (FSBR). The data analyzed came from the Daphnet Freezing of Gait Dataset, which contains records from inertial sensors positioned on the ankle, thigh and trunk of patients with PD. The Random Forest algorithm was used to classify the events, evaluating different sensor positions and time window lengths (2s, 3s and 4s). The results indicate that longer windows improve FoG detection, with the trunk sensor showing the highest recall rate (0.918) for a 4-second window, making it the ideal configuration for minimizing false negatives. Confusion matrix analysis shows that the proposed approach captures critical motor transitions with high precision, making it a promising alternative for applications in continuous monitoring and real-time interventions. Additionally, the investigation of the most relevant spectral bands revealed that low-frequency oscillations in the Z-axis (1.5-2.0 Hz) and high-frequency components in the X-axis (20.0-30.0 Hz) play a key role in distinguishing between FoG episodes and normal gait. These findings reinforce the potential of spectral analysis in characterizing gait dynamics in PD patients, contributing to the development of more accurate and individualized detection systems. |
|---|