Advancing Freezing of Gait Heterogeneity Modeling through Subtype-aware Detection, Generative Augmentation, and Adaptive Prediction

Abstract

Freezing of Gait (FOG) is a disabling symptom of Parkinson’s Disease (PD) that varies in manifestations and motion contexts. Its heterogeneity motivates subtype categorization such as manifestation-specific subtypes (akinesia, trembling, or shuffling) and motion-specific subtypes (gait-initiation, walking, or turning), with occurrence and frequency of subtype varying across patients. FOG detection and prediction have attracted significant research interest for their applications in daily monitoring, automated FOG dataset labeling, and on-demand activation of intervention devices.

With respect to FOG detection, despite numerous promising Deep Learning (DL) FOG detection studies, few consider FOG heterogeneity. It remains unclear whether different subtypes require distinct detection strategies, and whether tailoring subtype-specific models could enhance detection generalizability across subtypes. Additionally, training a DL detection model with robustness and generalization across subtypes is limited by data scarcity and imbalances between FOG/non-FOG classes and among subtypes, while FOG generative augmentation is considered a promising solution. However, subtype-conditioned FOG generative augmentation has not been developed, and its effectiveness and advantages compared to simpler, cheaper classical augmentation methods on detection model performance remain unknown. Regarding FOG prediction, one gap lies in the limited adaptability and complexity of available labeling approaches for pre-FOG (transition state leading to FOG), which exhibits heterogeneity across subjects and FOG episodes. Analyzing pre-FOG heterogeneity with respect to FOG subtypes may help better interpret it, but is currently underexplored. Another gap with respect to existing prediction model design is the lack of a multi-horizon prediction function, which could specify FOG onset while simultaneously enabling both short- and long-term alarms.

These gaps are addressed in this thesis through three projects, each detailed in a methodology chapter, focusing on subtype-aware FOG detection, subtype-conditioned FOG generative augmentation, and multi-horizon FOG prediction incorporating a soft, data-driven, adaptive pre-FOG labeling.

The FOG detection chapter first categorizes FOG data into manifestation- or motion-specific subtypes via classifier or clustering methods and then derives their corresponding detection strategies as interpretable feature masks. This chapter then proposes a feature-mask-based Convolutional Neural Network (CNN) that explicitly embeds the identified strategies. Using waist-mounted 3D accelerometer data, a general CNN and subtype-specific CNNs are trained. The results show that according to feature-mask analysis, motion-specific subtypes share a common detection strategy, whereas manifestation-specific subtypes require distinct strategies. Manifestation models exhibit enhanced generalizability across subtypes compared to the general model, boosting the overall average FOG detection sensitivity by 10.95% ± 9.24% and specificity by 32.08% ± 9.01%. Conversely, motion models reduce the overall FOG sensitivity by 1.89% ± 8.74% and specificity by 5.17% ± 10.76%. Consequently, the detection strategy is mainly driven by manifestation composition of the data. The general model favors the dominant manifestation-specific subtype group(s), a bias corrected by tailored manifestation-specific strategies. No comparable benefit arises from motion models due to their similar manifestation compositions. This chapter reveals the detection strategies required by different FOG subtypes and demonstrates the effectiveness of subtype-specific tailoring in improving FOG detection generalizability.

The FOG augmentation chapter proposes a subtype-aware FOG augmentation technique enabling training of DL models to perform consistently across subtypes. Specifically, it introduces Hi-CF cGAN, a two-stage model that generates subtype-conditioned FOG-like ankle accelerations that are realistic and diverse, as verified through visualization, UAMPs, and MMD comparison against real signals. This chapter evaluates Hi-CF cGAN’s effectiveness by training CNNs for FOG detection with both general (subtype-stratified) and personalized (subtype-variant, based on patient-specific subtype composition) augmentation via Hi-CF cGAN, benchmarking against classical augmentations and baseline (no augmentation). Compared to baseline, general augmentation with Hi-CF cGAN effectively improves average detection rates of FOG, trembling FOG, and especially the previously overlooked minor subtypes, shuffling FOG (from 66.8% to 81.6%) and akinesia FOG (from 58.7% to 77.9%). These improvements exceed those of classical augmentations, demonstrating superior realism, richness, and adaptability of Hi-CF cGAN -generated data in addressing FOG/non-FOG and subtype imbalances. Personalized augmentation further enhances accuracy on targeted subtype(s) compared to general augmentation, highlighting its potential for tailored model optimization.

The FOG prediction chapter first proposes a soft, data-driven, and adaptive pre-FOG labeling approach that identifies potential pre-FOG windows using statistical signal properties, including Shannon entropy and auto mutual information, and data-driven features via a CNN-predicted FOG probability. This adaptive labeling effectively captures intensifying pre-FOG characteristics while approaching a FOG episode and generalizes effectively across subjects. The labeling results reveal that for motion-specific subtypes, turning shows the strongest and most statistically reliable pre-FOG trends, while gait-initiation lacks a clear pre-FOG pattern. For manifestation-specific subtypes, trembling exhibits the most statistically consistent pre-FOG trend, while shuffling has the weakest trend. Some subjects display strong general pre-FOG trends, while others only show strong pre-FOG trend with specific subtype(s), highlighting the value of subtype-specific pre-FOG labeling and the interpretability of pre-FOG heterogeneity via subtypes. Additionally, this chapter also proposes a sequence-to-sequence, multi-horizon CNN-transformer that predicts the FOG state for each of the next six seconds. Combined with the proposed adaptive labeling, the model predicts both a discrete FOG state and a soft FOG Score representing FOG probability. It achieves a low mean error of 11.4% ± 4.1% and above-benchmark Prediction horizons of 3.19 ± 0.34 s. Comparisons across labeling methods show that the adaptive labeling improves both window- and sequence-wise prediction accuracy and stability relative to fixed labeling, confirming its higher clarity and flexibility in pre-FOG identification. However, compared to no-pre-FOG labeling, the adaptive labeling demonstrates improved Prediction horizons and prediction success rate on transition sequence but reduced accuracy on non-transition sequence due to increased false alarms, which is a trade-off to consider in practical application.

Collaboratively, these three chapters demonstrate the necessity and benefits of tailoring with respect to manifestation-specific subtypes for cross-subtype detection generalization, manifestation-conditioned FOG augmentation for data imbalance correction, and episode-adaptive pre-FOG labeling for reliable prediction, while also proposing innovative deep learning solutions for each specific FOG modeling problem.