Detecting freezing of gait in Parkinson’s Disease (PD) with functional near-infrared spectroscopy (fNIRS), surface electromyography, and inertial sensors

Parkinson’s disease (PD) is a progressive neurodegenerative disorder affecting the basal ganglia and resulting in motor symptoms such as bradykinesia, rigidity, and tremor. Among its most disabling manifestations, freezing of gait (FoG) is characterized by brief episodes of gait arrest despite the intention to move, thus significantly increasing fall risk and compromising patients’ autonomy. Its pathophysiology remains unclear, involving disrupted cortical-subcortical communication and impaired motor control. The transient and context-dependent nature of FoG makes objective, multimodal assessments essential to improve its characterization and clinical management. This study aims to investigate FoG in PD using a multimodal design in which fNIRS, electromyography (EMG), and inertial measurement units (IMU) are equally valued and synchronously acquired. From each modality, we extract features capturing cortical activation, muscle activity, and gait kinematics, which are then fused and fed into machine-learning models trained to detect and predict freezing episodes by automatically labeling gait segments as “FoG” or “noFoG”. Twelve patients diagnosed with PD were enrolled in this preliminary study, including four individuals who reported experiencing FoG episodes on daily basis. Participants completed seven motor tasks, two of which with dual tasks (DT), specifically designed to elicit FoG episodes. The present analysis focused on six motor tasks, including turning and walking under different conditions, during which fNIRS, EMG and IMU data were simultaneously acquired. All recorded signals underwent dedicated preprocessing pipelines to ensure data quality and synchronization across modalities: fNIRS data were processed using the NIRStorm toolbox, including artifact and non-physiological components removal, 3D reconstruction of cortical activation maps, and GLM-based analysis; IMU data were processed in python to identify gait events (turn and walk phases) and to detect steps, and in MATLAB for the machine learning (ML) pipeline; EMG data underwent standard EMG processing, with filtering, normalization and envelope extraction, all performed in MATLAB. For the ML analysis, fNIRS, EMG, and IMU data were segmented into fixed-length epochs, from which modality-specific feature vectors were extracted. These features were then concatenated to form multimodal representations and used to train and validate ML models through a leave-one-subject-out cross-validation strategy. The binary classification aimed to distinguish between “FoG” and “noFoG” epochs, while a multi-class configuration, including a “preFoG” class, was implemented to enable early prediction of upcoming freezing events.