Motor assessment of X-linked dystonia parkinsonism via machine-learning-based analysis of wearable sensor data

Abstract

X-linked dystonia parkinsonism (XDP) is a neurogenetic combined movement disorder involving both parkinsonism and dystonia. Complex, overlapping phenotypes result in difficulties in clinical rating scale assessment. We performed wearable sensor-based analyses in XDP participants to quantitatively characterize disease phenomenology as a potential clinical trial endpoint. Wearable sensor data was collected from 10 symptomatic XDP patients and 3 healthy controls during a standardized examination. Disease severity was assessed with the Unified Parkinson's Disease Rating Scale Part 3 (MDS-UPDRS) and Burke-Fahn-Marsden dystonia scale (BFM). We collected sensor data during the performance of specific MDS-UPDRS/BFM upper- and lower-limb motor tasks, and derived data features suitable to estimate clinical scores using machine learning (ML). XDP patients were at varying stages of disease and clinical severity. ML-based algorithms estimated MDS-UPDRS scores (parkinsonism) and dystonia-specific data features with a high degree of accuracy. Gait spatio-temporal parameters had high discriminatory power in differentiating XDP patients with different MDS-UPDRS scores from controls, XDP freezing of gait, and dystonic/non-dystonic gait. These analyses suggest the feasibility of using wearable sensor data for deriving reliable clinical score estimates associated with both parkinsonian and dystonic features in a complex, combined movement disorder and the utility of motion sensors in quantifying clinical examination.

Keywords: Digital health; Dystonia; Dystonia parkinsonism; Machine Learning; Parkinsonism; Wearable sensors.

Figure 1

Sensor-based data feature projections color-coded by MDS-UPDRS clinical scores. Three-dimensional data feature projections for the (A) finger-to-nose (item 3.16), (B) hand pronation/supination (item 3.6), (C) leg agility (item 3.8), (D) toe-tapping (item 3.7), and (E) gait (item 3.10) motor tasks. The points in each plot correspond to the representation in the reduced dimensionality space of the data features derived from the sensor signals and are color-coded according to the clinical labels (control and MDS-UPDRS scores). For sample numbers for each task, please see Supplementary Table S3.

Figure 2

Sensor-based data feature projections color-coded by presence/absence of dystonia. Three-dimensional data feature projections for the (A) finger-to-nose (item 3.16), (B) hand pronation/supination (item 3.6), (C) leg agility (item 3.8), (D) toe-tapping (item 3.7), and (H) gait (item 3.10) motor tasks, as well as for the non-MDS-UPDRS dystonia provoking maneuvers (E) heel-toe alternate movement, (F) heel walking, and (G) toe walking. The points in each plot correspond to the representation in the reduced dimensionality space of the data features derived from the sensor signals and are color-coded according to the clinical labels (control and XDP with/without dystonia). For sample numbers for each task, please see Supplementary Table S4.

Figure 3

Projections of the gait spatio-temporal parameters color-coded by clinical characteristics. Three-dimensional data feature projections of gait spatio-temporal parameters color-coded by (A) presence/absence of freezing of gait (FoG) on examination, (B) MDS-UPDRS FoG score (Item 3.11), (C) MDS-UPDRS gait score (Item 3.10), and (D) presence/absence of dystonia. The points in each plot correspond to the representation in the reduced dimensionality space of the aggregated statistics (mean, standard deviation, coefficient of variation, and right/left ratio) of the gait parameters extracted from each trial. For sample numbers for each task, please see Supplementary Table S5.

Figure 4

Cadence, stride length, and stride velocity boxplots for different clinical characteristics. Boxplots of cadence, stride length, and stride velocity for presence/absence of freezing of gait (FoG) on clinical examination, MDS-UPDRS FoG scores (Item 3.11), MDS-UPDRS gait scores (Item 3.10), and presence/absence of dystonia on clinical gait examination. The boxplots visually summarize the distribution of data. Each boxplot displays the median (central line), interquartile range (box edges), and overall range excluding outliers (whiskers). Outliers are marked with individual points. Pairwise significant differences were assessed with a mixed regression model and are indicated by a horizontal red line. *** indicates p-value < 0.001, while * indicates p-value < 0.01.

Figure 5

Experimental set-up and data processing pipelines. (A) Two wearable motion sensors were placed on the wrists during the performance of upper-limb tasks and repositioned on the ankles when performing lower-limb tasks. An enlarged view of a motion sensor (Shimmer3 by Shimmer Research Ltd, Dublin, Ireland) and its reference system are also shown. (B) Sensor-based data feature analysis pipeline. The diagram illustrates the processing steps used to derive data feature projections from the raw motion sensor data when participants performed standardized motor tasks. (C) Gait spatio-temporal parameter analysis pipeline. The diagram depicts the steps of the algorithm used to extract gait spatio-temporal parameters from the raw motion sensor signals recorded during gait and visualize them via projections.