Background: Obstructive sleep apnea (OSA) is a common but often overlooked comorbidity in patients with Parkinson’s disease (PD), impacting both motor and non-motor symptoms and potentially affecting gait control. Early identification of PD with comorbid OSA (PD-OSA) is crucial for improving patient outcomes.
Methods: This study employed a wearable inertial sensor system to systematically collect spatiotemporal gait parameters and dynamic stability indicators during a standardized Timed Up and Go (TUG) test in 61 PD patients. Patients were categorized into PD-OSA and PD-NOSA groups based on apnea-hypopnea index (AHI). Linear mixed-effects models analyzed gait differences, and an Extreme Gradient Boosting (XGBoost) algorithm was used to develop a screening model integrating gait features and non-motor symptom scores.
Results: The PD-OSA group showed significant reductions in key gait parameters including step length, stride velocity, arm swing range, and Trunk Max Sagittal Angular Velocity compared to the PD-NOSA group. The XGBoost model demonstrated excellent performance in an independent test set with an AUC of 0.944, sensitivity of 83.3%, and specificity of 83.3%. The most predictive features aligned closely with underlying pathophysiological mechanisms.
Conclusion: A machine learning model based on wearable-derived gait parameters and clinical scores can effectively identify PD patients with comorbid OSA noninvasively at an early stage, offering potential clinical utility to enhance comprehensive management and prognosis of PD.
Supplementary Information: The online version contains supplementary material available at 10.1186/s12883-026-04700-3.
Keywords: Gait; Machine learning; Obstructive sleep apnea (OSA); Parkinson's disease (PD); Wearable inertial sensor.