Spatiotemporal Gait Parameters from Wrist-Worn Accelerometers

Supervised by: Aaron Colombo and Dr. Michael Single
Institute: Gerontechnology and Rehabilitation Group, ARTORG Center for Biomedical Engineering Research, University of
Bern
Workplace: ARTORG Center (SITEM, Freiburgstrasse 3) on the Insel hospital campus.
Start Date: October 2025 or upon agreement

If you are interested in this topic or have further questions, do not hesitate to contact aaron.colombo@unibe.ch

Background

Gait is an important biomarker for neurodegenerative diseases (e.g., Parkinson’s disease,
Alzheimer’s disease). Since most consumer and research wearables contain accelerometers and large datasets
are available, developing robust methods that work across settings (indoor/outdoor) could enable scalable gait
monitoring. Walking indoors and outdoors represents different neurological aspects. Outdoor walking, typically
involving longer and more continuous strides, reflects core gait mechanics, whereas indoor walking often
includes dual- or even triple-task elements, requiring complex neurological networks to work in tandem. Deep
learning (DL) models for predicting spatiotemporal gait parameters are a relatively new development (Brand et
al., 2024, 2025; Yuan et al., 2024). Many questions, such as model performance in different settings (e.g., indoor
vs. outdoor), incorporating a biomechanical model, or including demographic information to improve prediction,
remain unanswered.

Aim

This thesis aims to develop and validate methods for predicting spatiotemporal gait parameters from wrist
accelerometers by combining classical signal processing with DL models, trained with foundation models. The
study will investigate whether these daily-life trained models generalize to indoor and outdoor settings, and
whether performance can be improved by training separate models for each condition. The influence of
demographic variables (sex, height, weight) on prediction accuracy will also be assessed.

Material and Methods

The project will draw on existing large wrist accelerometer datasets, complemented by
a small custom dataset collected indoors and outdoors for validation. After preprocessing and feature extraction,
hybrid models will be trained using both signal-based features and AI-based time-series representations.
Comparisons will be made between general and condition-specific models, with and without demographic
variables, and evaluated against reference gait measures using standard performance metrics.

Nature of the Thesis

Method development (signal processing algorithms, machine learning models): 60%
Data collection and analysis (acquisition, preprocessing, evaluation): 40%

Required Skills

• Strong programming skills in Python
• Experience with Numpy, Pandas/Polars, PyTorch
• Familiarity with wearable sensor data and time-series analysis is advantageous
• Basic knowledge of statistics.