Integrating digital gait data with metabolomics and clinical data to predict outcomes in Parkinson's disease

Abstract

Parkinson's disease (PD) presents diverse symptoms and comorbidities, complicating its diagnosis and management. The primary objective of this cross-sectional, monocentric study was to assess digital gait sensor data's utility for monitoring and diagnosis of motor and gait impairment in PD. As a secondary objective, for the more challenging tasks of detecting comorbidities, non-motor outcomes, and disease progression subgroups, we evaluated for the first time the integration of digital markers with metabolomics and clinical data. Using shoe-attached digital sensors, we collected gait measurements from 162 patients and 129 controls in a single visit. Machine learning models showed significant diagnostic power, with AUC scores of 83-92% for PD vs. control and up to 75% for motor severity classification. Integrating gait data with metabolomics and clinical data improved predictions for challenging-to-detect comorbidities such as hallucinations. Overall, this approach using digital biomarkers and multimodal data integration can assist in objective disease monitoring, diagnosis, and comorbidity detection.

Fig. 1. Schematic overview of the study workflow.

The input data types used for the cross-validated machine learning analyses are highlighted on the left and the prediction tasks on the right. The prediction tasks focus on the estimation of four groups of outcomes (roughly sorted by increasing complexity): (1) disease diagnosis, (2) motor score severity, (3) gait and mobility impairments, and (4) comorbidities, non-motor outcomes, and progression rate (measured by the average annual change in the MDS-UPDRS III motor score over four years of follow-up and categorized as slow or fast, depending on whether the change falls in the lower or upper quartile, respectively). For the first three data modalities, unimodal machine learning models were built using gait data only, as this was sufficient to achieve satisfactory cross-validation performance, whereas for the more challenging fourth group of tasks, aimed at detecting comorbidities, non-motor outcomes and progression rate subgroups, multimodal models combining gait, omics and clinical data were built in addition to comparing the individual data modalities.

Fig. 2. SHAP value plot of the top-ranked features for predicting low vs. high UPDRS 3 motor score outcomes.

The plot shows the gait-specific digital biomarker features with the highest SHAP values for predicting low vs. high UPDRS 3 motor score outcomes using extreme gradient boosting for machine learning. The color coding from purple to yellow represents the feature value range from low to high. The labels on the left correspond to the individual gait features that were most predictive in terms of the absolute SHAP value, sorted from top to bottom (corresponding absolute SHAP values are shown in bold on the left side of the plot; for a description of the individual features see Supplementary Table 1).

Fig. 3. Bar plot visualization comparing the predictive performance for different PD outcomes and models using different input data.

The plot compares the predictive performance in terms of median cross-validated AUC values with median absolute deviations (MAD) for the PD outcomes considered in the integrative analyses and models using different input data (Clinical Data, Gait Data, Metabolomics Data, and Integrative Model). Each outcome is represented on the horizontal axis, ordered by decreasing average AUC values across all models.

Fig. 4. Data flow for the Luxembourg Parkinson’s Study.

This figure illustrates the distribution of samples and the overlapping subsets of data types in the Luxembourg Parkinson’s Study. The study included 736 Parkinson’s disease (PD) patients and 855 controls, all of whom provided clinical data. Out of these, 162 PD patients and 129 controls had digital gait data collected. In addition, metabolomics data were available for 549 PD patients and 590 controls. For 151 PD patients, all three data types available (gait, metabolomics, and clinical data) were available and used for integrative analyses. Unimodal analyses were performed for all subjects with available gait data. The color-coded boxes indicate the specific datasets used for unimodal and multimodal analyses.