Gait analysis provides essential insights into an individual’s movement patterns, balance, and
overall health. It plays a crucial role in diagnosing neurological disorders, monitoring disease
progression, and assessing rehabilitation outcomes. Traditionally, gait analysis relies on motion
capture systems (MoCaps) constrained to controlled lab environments. While highly accurate, these
systems are expensive, stationary, and require specialized setups. To overcome these limitations,
inertial measurement units (IMUs) - small wearable sensors that measure acceleration and angular
velocity - enable continuous gait analysis in real-world settings. Their portability and cost-effectiveness
allow for real-time monitoring outside the lab, making them ideal for long-term health tracking.
Building on gait analysis, gait classification focuses on categorizing movement patterns - typically into
healthy vs. pathological groups - based handcrafted features or deep learning models applied directly
to time series data. However, these approaches often fail to capture the intrinsic geometric structure
of the signals.
Topological Data Analysis (TDA) provides a complementary geometry-focused method for analysing
IMU signals by capturing the shape and structure of data using methods like persistent homology.
While traditional pipelines have been proven to be discriminative [1][2], they inherently rely on
handcrafted fixed-length features. Topological Deep Learning (TDL) [3] allows the model to
automatically learn which topological features are most informative for the task, creating an end-toend
framework that potentially improves robustness and generalization. However, the practical
benefits of TDL methods in IMU-based gait classification remain largely unexplored.
In this thesis, the student will explore existing TDL architectures, such as DeepSets [4], PersLay [5] and
Persformer [6] to classify gait patterns recorded with a single IMU. To the best of our knowledge, these
architectures have not yet been specifically investigated for single-IMU gait analysis. Depending on
the findings, the networks will therefore be adapted to the characteristics of gait signals and the
targeted gait-analysis task.

To investigate whether TDL methods can improve the robustness and generalization of
IMU-based gait classification compared to traditional TDA-ML and DL approaches. This includes
comparing TDL pipelines on benchmark IMU dataset(s) and investigating how the different
approaches influence classification performance.

Literature Review (ETOC: 2 months): Familiarize with existing literature on TDA and TDL for
IMU signal analysis in gait classification.
- Dataset Familiarization (ETOC: 1 month) - Understand the structure and characteristics of the
IMU datasets to be used. Perform preprocessing if needed and prepare data for the pipelines.
- Implementation of TDL Pipelines (ETOC: 6 months) - Optimize the different TDL architectures
and compare performance versus traditional baselines (raw IMUs, TDA features).

(Mandatory) qualifications:
• Following a MSc in a field related to one or more of the following: Computer Science,
Biomedical Engineering, Applied Computer Science - Digital Health.
• Strong programming skills (Python).
• Ability to write scientific reports and communicate research results at conferences in English.