28 March 2025 at 9:00 AM
Jakob Rostovski, "Development of Methods for Real-time In-Step Anomaly Detection in Gait Analysis"
Supervisor: Professor Muhammad Mahtab Alam, Thomas Johann Seebeck Department of Electronics, School of Information Technologies, Tallinn University of Technology
Tallinn, Estonia
Co-supervisor: Alar Kuusik, PhD, Thomas Johann Seebeck Department of Electronics, School of Information Technologies, Tallinn University of Technology, Tallinn, Estonia
Opponents:
- Professor Maurizio Magarini, Politecnico di Milano, Dipartimento di Elettronica e Informazione, Milan, Italy
- Professor Matti Hämäläinen, University of Oulu, Faculty of Information Technology and Electrical Engineering, Oulu, Finland
Neurological diseases affect millions, often impairing mobility. Gait disturbances reduce independence, making gait analysis and rehabilitation crucial. Assistive devices like exoskeletons and Functional Electrical Stimulation (FES) help correct gait abnormalities, but they rely on precise gait phase detection.
Exoskeletons are often bulky and impractical for daily use. FES devices offer a lighter alternative by stimulating muscles to restore movement. However, current FES systems lack personalized gait anomaly detection, often stimulating every step, leading to fatigue and skin irritation. Ideally, stimulation should only occur when needed, requiring real-time gait deviation detection.
Detecting gait anomalies within a single step remains a challenge due to limited research. This thesis addresses this gap by evaluating popular gait detection algorithms and introducing a novel dataset and framework for fair comparison.
A key contribution is a new gait dataset collected under a clinical trial protocol approved by the Estonian National Institute for Health Development. It includes 155 recordings from 22 individuals, uniquely combining normal and abnormal steps in a single session and covering eight common gait deviations.
The thesis also adapts and evaluates real-time gait analysis algorithms. Traditional machine learning models, like support vector machines (SVMs), showed promising results for classification, but had limited accuracy in real-time application. Deep learning methods, such as long short-term memory (LSTM) networks and one-dimensional convolutional neural networks (1D-CNNs), performed better, with 1D-CNN achieving 95% accuracy and an F1 score of 88.2%. However, deep learning models require high computational power, limiting their use in real-time wearable applications.
To overcome this, a heuristic algorithm—Signal Shape Tracking Anomaly Detection (SST-AD)—was developed. SST-AD achieved 91% accuracy with an F1 score of 81%, matching deep learning models but with significantly lower computational demands. It also was among the fastest algorithms, detecting anomalies just 0.4 seconds after the mid-swing phase started.
This thesis lays the groundwork for real-time gait deviation detection in assistive devices. It presents high-performing algorithms and explores practical steps for integration into wearable technology. Future work will optimize hardware, refine stimulation parameters, and assess real-world performance, paving the way for more effective, user-friendly gait assistive devices.
Join the public defence in MS Teams
Meeting ID: 311 161 969 971
Passcode: sn9hM9z3