Integrating digital gait sensor data with metabolomics and clinical data to predict clinically relevant outcomes in Parkinson’s disease#

Authors#

Cyril Brzenczek, Quentin Klopfenstein, Holger Fröhlich, Enrico Glaab

Abstract#

Background: Parkinson’s disease (PD) is characterized by heterogeneous motor and non-motor symptoms and a wide variety of comorbidities that complicate diagnosis and management. Digital biomarkers, such as gait sensor data, along with omics and clinical data, offer potential for early and objective prediction of clinical outcomes in PD. Objective: This study used digital gait sensor data to predict clinically relevant outcomes in PD, including the integration with complementary metabolomics and clinical data to enhance diagnosis, severity assessment, and forecasting of comorbidity occurrences. Methods: We employed the gait sensor technology eGaIT (embedded Gait Analysis using Intelligent Technologies) to collect gait measurements for standardized walking exercises from 291 subjects, including 162 PD patients and 129 controls. A representative selection of different types of machine learning methods, including Deep Boosting, Support Vector Machines and Random Forest, among others, were used to analyze extracted features for predicting PD vs. control status, motor severity (MDS-UPDRS 3 scores), specific gait impairments such as freezing of gait, and comorbidities. Model interpretability was studied through SHAP value analysis. Results: Using generic gait features as predictor, the machine models demonstrated significant predictive power, with area under the curve (AUC) scores ranging from 0.83 to 0.92 for PD vs. control predictions, and up to 0.76 for distinguishing between low or high motor score performance among patients. Significant predictive performances were also achieved for gait- and mobility-specific outcomes, such as freezing of gait, the PDQ39 mobility sub-score, and general prediction of gait disorder occurrence. For the more challenging task of predicting the occurrence of comorbidities in PD, the integration of gait data with omics and clinical variables improved predictions as compared to using only single data modalities, with notable advances in forecasting dopamine dysregulation syndromes, dyskinesias and hallucinations. Conclusions: Digital gait sensor data, and in particular its integration with metabolomics and clinical data, can effectively predict clinically relevant outcomes in PD. The approach can not only assist in objective diagnosis and severity assessment but also help in forecasting the occurrence of comorbidities, where the synergy of different data modalities provides an added value.