ID 2452: Neural Network-Based Detection of Gait Battery Tests in Unsupervised Home Recordings

One of the primary methods for assessing gait and mobility in patients with Parkinson’s Disease (PD) is through gait battery tests. These tests are typically conducted in clinical settings under the supervision of experts who interpret the symptoms and assess disease severity. However, these evaluations are brief and occur in controlled conditions, which may not reflect the patient’s overall mobility and daily behavior.

The advent of wearable devices, such as inertial measurement units (IMUs), has enabled long-term monitoring of patients’ mobility in unsupervised environments. To better understand gait behavior and facilitate comparisons with laboratory measurements, patients are now asked to perform the same gait tests at home in an unsupervised setting.

This project aims to detect and label these unsupervised gait tests using a neural network approach for further analysis. The identified tests will be excluded from the patient’s general mobility data to focus on distinct gait patterns. The results will be compared to the baseline method proposed by Ulrich et al..

This work is part of the MobilityApp project.

Dataset

The dataset comprises 3,600 gait tests performed during home monitoring and 600 tests conducted in laboratory settings. The data spans three types of Parkinson’s Disease: Progressive Supranuclear Palsy (PSP), Multiple System Atrophy (MSA), and Idiopathic Parkinson’s Disease (IPD).

Requirements

• Proficiency in Python programming language
  • Proficiency in basic libraries such as Pandas and Numpy.
• Understanding of signal processing. Familiarity with gait analysis is a plus.
• Familiarity with the concepts of Neural Networks, creating and testing them. Experience in implementing a NN is a plus.
• Great scientific writing skills.
• Communication skills (English).

Tasks

• Literature search.
• Data pre-processing and cleaning.
• Manual Label checking. The labeling has been done, but checking the labels might be required.
• Data Augmentation.
• Implementing different NNs using different layers.
• comparison to the baseline method.
• Writing.

Supervisor:

Hamid Moradi, M. Sc.

Researcher & PhD Candidate