Decoding Parkinson’s Disease through Network Metrics of Inertial Sensor Data: A Machine Learning approach
Camilla calomino1, Maria Giovanna Bianco1, Salvatore Mazzeo2, Fabiana Novellino1, maria salsone2, alessia cristofaro1, Marianna Crasà1, Giulia Sgrò2, andrea quattrone1, aldo Quattrone1, Rita Nistico1
1University of Catanzaro, 2Policlinico San Donato S.p.A.
Objective:

  • To develop a framework that integrates statistical and network features from multi-segment IMU recordings to classify Parkinson’s disease (PD) versus healthy controls (HC).

 

Background:

Parkinson’s disease (PD) is a progressive neurodegenerative disorder characterized by both motor and non-motor impairments. Early and objective identification of subtle gait alterations is essential for accurate monitoring and therapeutic management. Conventional clinical scales often fail to capture fine motor abnormalities, whereas wearable inertial measurement units (IMUs) allow for quantitative characterization of movement dynamics during gait. While sensor-derived temporal and spatial features have been extensively used in PD classification, the integration of network-based descriptors derived from inertial data remains largely unexplored.

Design/Methods:
 Participants performed standardized gait tasks wearing sensors on the pelvis, spine, neck, head, limbs, and feet. Preprocessed signals (low-pass Butterworth filter) were analyzed to extract skewness, kurtosis, and interquartile range, along with network metrics derived from Jensen–Shannon distance matrices. Feature selection employed a voting strategy, and classification was performed using nested cross-validation with an XGBoost model. SHAP analysis was used to interpret feature importance.
Results:

The model achieved robust discriminative performance (ROC AUC = 0.88; accuracy = 0.81; sensitivity = 0.79; specificity = 0.84). Features with the highest relevance included both statistical variability indices and network measures such as clustering strength and connectivity density. These network patterns reflected biomechanical and neurophysiological alterations typical of PD, including increased segmental co-activation, reduced independence of trunk-limb coordination, and overall axial rigidity. Such findings suggest that network-based analysis captures higher-order interdependencies among body segments, providing a more integrative representation of motor dysfunction.

Conclusions:

This approach extends beyond traditional gait metrics, offering a biologically interpretable and objective framework for digital biomarker development in Parkinson’s disease. Future work will address validation in larger cohorts and longitudinal tracking of disease progression.

10.1212/WNL.0000000000217163
Disclaimer: Abstracts were not reviewed by Neurology® and do not reflect the views of Neurology® editors or staff.