Geautomatiseerde detectie en beoordeling met draagbare sensoren en machinelearning van ‘freezing’ van het stappen bij de ziekte van Parkinson

Parkinson's disease (PD) is a progressive neurodegenerative disorder that affects more than six million people worldwide, with the number projected to double by 2030. One of its most disabling motor symptoms is freezing of gait (FOG), defined as a 'brief, episodic absence or marked reduction of forward progression of the feet despite the intention to walk.' FOG is currently classified into three manifestations based on leg movement: (1) trembling, characterized by tremulous leg oscillations at 3-8 Hz; (2) shuffling, involving short steps with poor foot clearance; and (3) complete akinesia, where there is no visible leg movement. FOG is associated with a significantly increased risk of falls, reduced quality of life, and elevated anxiety. Although treatments such as dopaminergic medication, physiotherapy, and sensory cueing are available, their effectiveness varies across individuals and FOG manifestations. To support the development and personalization of these interventions, it is essential to quantify FOG severity accurately and without subjective bias. Traditional FOG assessment methods rely primarily on self-reported questionnaires, which are limited by recall bias and high test-retest variability. The current gold standard involves expert video annotation of patients performing FOG-provoking tasks to compute objective metrics such as the percentage of time frozen (%TF) and the number of FOG episodes (#FOG). However, this method is time-consuming, labor-intensive, and depends on trained professionals. In response, automated FOG detection methods leveraging machine learning (ML) and deep learning (DL) models applied to wearable sensor data have been developed to provide a more objective, scalable, and efficient alternative to traditional assessment approaches. Despite many studies proposing ML/DL approaches for FOG detection using various data modalities, including motor-related signals such as 3D optical motion capture (MoCap) and wearable inertial measurement units (IMUs), as well as physiological signals such as galvanic skin response (GSR) and electrocardiography (ECG), several limitations persist. This thesis addresses these limitations through six studies. In the first study, we examined how variations in FOG-provoking task type, medication state, and the inclusion/exclusion of volitional stopping periods influence the performance of FOG detection models. Twelve participants with PD performed both Timed Up and Go (TUG) and 360° turning-in-place (360Turn) tasks under ON and OFF medication states, with and without volitional stopping events. A multi-stage temporal convolutional network (MS-TCN) was trained using IMU data to detect FOG episodes. Our results showed that task-specific models failed to generalize to unseen tasks, while models trained on a specific medication state generalized more effectively across medication states. We also found that including stopping events during training was crucial to optimizing detection performance in trials involving volitional stopping. These findings emphasize the importance of incorporating diverse FOG-provoking tasks and stopping events into training datasets to improve model robustness in free-living environments, where human movement patterns are more varied and complex. The second study focused on the fine-grained classification of FOG manifestations, specifically trembling and akinetic FOG, and their differentiation from volitional movement cessations such as sitting and stopping. Using IMU data from eighteen participants with PD performing TUG and 360Turn tasks, we evaluated whether an MS-TCN model could be trained to detect distinct FOG manifestations. Results showed that the model achieved strong agreement with expert annotations for trembling FOG and moderately strong agreement for akinetic FOG, demonstrating the potential of automated detection of FOG manifestations. The model was also able to successfully distinguish FOG episodes from sitting. However, volitional stopping events were reliably detected only in TUG trials without concurrent FOG, and akinetic episodes were frequently misclassified during TUG tasks. These findings highlight the challenge of using IMU data alone to resolve such ambiguities. The third study investigated the added value of physiological data (i.e., GSR and ECG) for improving the differentiation between FOG and volitional stopping during TUG tasks. We proposed and evaluated two multimodal approaches: a two-step pipeline combining a Transformer-based model with an XGBoost classifier, and an end-to-end MS-TCN model. Both approaches integrated IMU, GSR, and ECG data as input features. However, neither multimodal approach significantly outperformed IMU-only models in distinguishing FOG from stopping. This indicates that, although physiological signals may capture autonomic responses associated with FOG, their predictive contribution was limited under the tested conditions. We further applied explainable AI techniques to examine the relative importance of each input modality. The analysis revealed that IMU features contributed more prominently to model predictions than physiological features. These findings highlight the need for future research to explore alternative physiological modalities or more advanced multimodal integration strategies to distinguish FOG from stopping. In the fourth study, we developed a simulation pipeline to generate synthetic IMU data from optical MoCap recordings for FOG detection. This approach enables the simulation of IMU signals at various body locations, helping to overcome the lack of standardized sensor configurations, which currently limits fair comparisons and reusability of FOG detection datasets involving MoCap. We proposed an MS-TCN model and used it to train on synthetic IMU data generated from MoCap recordings of ten individuals with PD and FOG. This setup allowed for objective comparisons against two state-of-the-art FOG detection models across five simulated IMU configurations, ranging from one to seven sensors. The proposed model consistently outperformed both baselines in most configurations, with the highest performance achieved using a seven-sensor lower-body setup. Notably, models trained on synthetic IMU data outperformed those trained solely on real IMU recordings. These findings demonstrate that existing MoCap datasets can be effectively reused to develop IMU-based FOG detection models, potentially reducing the need for extensive data collection in future research. In the fifth study, we investigated whether marker trajectories derived from Skinned Multi-Person Linear (SMPL) models reconstructed from video could reliably support FOG detection, serving as a first step toward validating SMPL as a unified, modality-agnostic motion representation. We used the ViTPose model to extract 2D keypoints from video frames and lifted them to 3D SMPL meshes using the WHAM model. Virtual marker trajectories were then generated by placing markers on the SMPL surface. These trajectories were used to train a multi-stage spatial-temporal graph convolutional neural network (MS-GCN) for FOG detection. The model was evaluated on two datasets: one lab-based dataset with 38 participants (12 of whom were also recorded at home) performing TUG and 360Turn tasks, and a second dataset with 12 participants performing walking tasks involving 180°/360° turns. Results showed that the model trained on SMPL-derived markers from video achieved comparable performance to models trained on raw IMU and optical MoCap data, supporting its reliability. Cross-dataset and cross-environment evaluations demonstrated that the model achieved moderately strong to strong agreement with expert annotations for %TF and #FOG, confirming its generalizability. These results represent an important first step toward establishing SMPL-derived features as a unified motion representation for FOG detection. The sixth study addressed inter-center generalization by developing a DL model trained on KU Leuven datasets and validating it on five external cohorts, which varied in FOG annotation criteria, task protocols, IMU configurations, and patient characteristics. While the model showed strong agreement with expert annotations internally, its performance declined on external datasets, underscoring the challenges posed by cross-center variability. Fine-tuning with a small portion of external data significantly improved performance, matching models trained from scratch with more data and demonstrating a data-efficient strategy for enhancing generalizability. To support clinical adoption, this study introduced AID-FOG, a web-based platform that displays automated FOG detection results and allows users to interactively review and correct model annotations. Based on the input of three clinical raters at KU Leuven, AID-FOG outperformed the commonly used ELAN tool in system usefulness, information quality, and interface design. While further refinement is warranted, these initial results highlight AID-FOG's potential for integration into clinical workflows. Finally, to support long-term impact and maximize the valorization potential of this research, an exploitation plan has been developed outlining a roadmap for the short-term deployment and long-term commercialization of the AID-FOG platform. Supported by KU Leuven's technology transfer office and funding instruments such as IOF Type 1a and the VLAIO Innovation Mandate, the plan identifies early-stage challenges related to regulation, product development, and market access. Through a phased implementation strategy, a revenue model, and use-case design, AID-FOG aims to deliver a clinically valuable and scalable solution for researchers and clinicians, bridging the gap between academic innovation and real-world clinical adoption.