Bi-cephalic self-attended model to classify Parkinson's disease patients with freezing of gait

Shomoita Jahid Mitin, Rodrigue Rizk, Maximilian Scherer, Thomas Koeglsperger, Daniel Lench, Kc Santosh, Arun Singh

Abstract

Parkinson’s Disease (PD) is frequently characterized by motor and cognitive impairments, with Freezing of Gait (FOG) being one of the most debilitating symptoms. Traditional FOG detection relies on subjective clinical assessments or specialized gait analysis that requires physical movement. This study presents a novel, data-driven classification model—the Bi-cephalic Self-Attention Model (BiSAM)—designed to identify FOG using resting-state EEG signals and clinical variables. Utilizing a dataset of 124 participants (42 PDFOG+, 41 PDFOG-, and 41 healthy controls), we evaluated signal-only, descriptive-only, and multi-modal configurations. While unimodal models showed limited efficacy, the multi-modal BiSAM achieved a peak classification accuracy of 88% using as few as four EEG channels. These findings suggest that integrating neural signatures with objective clinical data provides a robust, scalable alternative for the early diagnosis and monitoring of gait dysfunction in PD.

Introduction

Freezing of Gait (FOG) is a paroxysmal phenomenon where Parkinson’s patients experience a sudden, brief inability to start or continue walking, often described as their feet being "glued to the floor." This symptom is a primary contributor to falls, loss of independence, and reduced quality of life.

Currently, the gold standard for FOG assessment involves clinical questionnaires or gait lab observations. However, these methods are often subjective or impractical for patients with advanced motor impairment. There is a critical clinical need for objective, resting-state biomarkers that can identify "freezers" without requiring active movement. This study explores whether the neural "fingerprint" of FOG is detectable in resting-state brain activity when combined with demographic and clinical variables.

Methodology

The BiSAM Architecture

The researchers developed the Bi-cephalic Self-Attention Model (BiSAM), a deep learning architecture featuring two distinct branches designed to process heterogeneous data types:

  1. The Signal Branch: This branch extracts spectral-temporal embeddings from resting-state EEG signals. It specifically analyzes power spectral density (PSD) across the theta, alpha, beta, and gamma frequency bands.
  2. The Descriptive Branch: This branch processes objective clinical and demographic variables, including age, years of education, and disease duration.

At the core of BiSAM is a Self-Attention mechanism, similar to those used in Transformer architectures. This allows the model to capture complex, long-range dependencies within the brainwave data that traditional linear models might overlook.

Channel Optimization

To enhance clinical utility, the study sought to minimize the number of EEG sensors required. Using a "grouped permutation-importance" method, the team ranked the 63 EEG channels by their contribution to model accuracy. This allowed for the testing of reduced configurations: BiSAM-16, BiSAM-8, and BiSAM-4.

Experimental Design

The study utilized a multi-modal dataset of 124 participants categorized into three groups: Parkinson’s patients with FOG (PDFOG+), Parkinson’s patients without FOG (PDFOG-), and age-matched healthy controls (HC). EEG data was recorded for 120–180 seconds while participants remained seated with their eyes open. The model's performance was benchmarked across three modalities: signal-only, descriptive-only, and multi-modal fusion.

Results

Classification Performance

The multi-modal approach significantly outperformed unimodal models ($p < 0.001$). While EEG-only models reached a maximum accuracy of 76% (using all 63 channels) and clinical data alone achieved only 68%, the fused BiSAM models reached 88% accuracy.

Efficiency and Reliability

Notably, the model’s performance did not degrade with channel reduction; in fact, BiSAM-8 and BiSAM-4 both maintained the peak 88% accuracy. This suggests that a full EEG cap introduces redundant noise, and the most relevant signals for FOG are concentrated in specific regions, primarily the temporoparietal and occipital areas. The models also achieved a Cohen’s Kappa score of 0.76, indicating substantial agreement and high reliability for clinical classification.

Discussion

The Power of Multi-modal Fusion

The primary insight of this research is the synergy between neural and clinical data. EEG signals provide a snapshot of current brain state, while variables like disease duration provide the necessary context to interpret those signals. This fusion allows the model to overcome the instabilities inherent in low-channel EEG recordings.

Resting-State Validity

The success of BiSAM demonstrates that FOG-related pathology is encoded in the brain's resting state. This challenges the traditional assumption that FOG must be physically provoked to be detected. By identifying these signatures while a patient is stationary, clinicians can potentially screen for FOG risk earlier and more comfortably.

Conclusion

The Bi-cephalic Self-Attention Model offers a highly accurate, objective, and efficient method for classifying FOG in Parkinson’s patients. By achieving high performance with only four EEG channels, this approach paves the way for portable, low-cost diagnostic tools that can be integrated into routine clinical workflows. Future research will focus on validating the model across different medication states ("OFF" vs. "ON") and testing its predictive power in larger, longitudinal cohorts.

Summary updated

Abstract

Parkinson Disease (PD) often results in motor and cognitive impairments, including gait dysfunction, particularly in patients with freezing of gait (FOG). Current detection methods are either subjective or reliant on specialized gait analysis tools. This study aims to develop an objective, data-driven, and multi-modal classification model to detect gait dysfunction in PD patients using resting-state EEG signals combined with demographic and clinical variables. We utilized a dataset of 124 participants: 42 PD patients with FOG (PDFOG+), 41 without FOG (PDFOG-), and 41 age-matched healthy controls. Features extracted from resting-state EEG and descriptive variables (age, education, disease duration) were used to train a novel Bi-cephalic Self-Attention Model (BiSAM). We tested three modalities: signal-only, descriptive-only, and multi-modal, across different EEG channel subsets (BiSAM-63, -16, -8, and -4). Signal-only and descriptive-only models showed limited performance, achieving a maximum accuracy of 55% and 68%, respectively. In contrast, the multi-modal models significantly outperformed both, with BiSAM-8 and BiSAM-4 achieving the highest classification accuracy of 88%. These results demonstrate the value of integrating EEG with objective descriptive features for robust PDFOG+ detection. This study introduces a multi-modal, attention-based architecture that objectively classifies PDFOG+ using minimal EEG channels and descriptive variables. This approach offers a scalable and efficient alternative to traditional assessments, with potential applications in routine clinical monitoring and early diagnosis of PD-related gait dysfunction.
BibTeX
@misc{mitin2025bi,
  author = {Shomoita Jahid Mitin and Rodrigue Rizk and Maximilian Scherer and Thomas Koeglsperger and Daniel Lench and KC Santosh and Arun Singh},
  title = {{Bi-cephalic self-attended model to classify Parkinson's disease patients with freezing of gait}},
  year = {2025},
  eprint = {2507.20862},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.20862},
}
Figure