Leveraging Parkinson’s Disease patient’s lower back accelerations to classify Freezing of Gait events

11/26/2024

This thesis investigates the feasibility of classifying FOG events in PD patients using accelerometry data collected from a lower back-mounted sensor. FOG, characterized by sudden inability to initiate or continue walking, significantly impacts patients' quality of life and mobility. The study analyzes data from 833 recording sessions across 62 patients, aiming to identify and classify three distinct types of freezing episodes: Start Hesitation, Turn, and Walking. The methodology encompasses comprehensive signal processing, feature engineering, and the development of a machine learning model combining convolutional and recurrent neural networks. Novel approaches include the optimization of the freeze index for lower back accelerometry and the implementation of an unsupervised gait cycle detection algorithm. The research addresses three major challenges: time-series classification, integration of time-variant and invariant features, and class imbalance in the dataset. Results indicate that accelerometer data from the lower back alone contains insufficient information for reliable FOG type classification, achieving a mean Average Precision of 0.354 on the test set. This finding aligns with previous research suggesting that lower extremity sensor placement provides more accurate gait event detection. The study concludes that while lower back accelerometry can detect FOG episodes, accurate classification of the FOG types requires additional sensor inputs or alternative sensor placement strategies, contributing valuable insights to the development of monitoring systems.

Authors in the community:

• 

  H

  Humberto Francisco Antunes dos Santos

  ist193265

Supervisors of this institution:

• 

  S

  Susana Margarida da Silva Vieira

  ist45346

mechanical-engineering - Mechanical engineering

eng - English

Instituto Superior Técnico