We recruited 35 participants with mild-to-moderate PD (Hoehn and Yahr scale score ≤3; mean age 68.3 years, SD 8.0 years; males, n=23; and females, n=12) and 60 healthy participants (mean age 44.1 years, SD 10.7 years; males, n=27; and females, n=33). Participants with PD took regular dopaminergic medication (levodopa-equivalent daily dose, mean 165.5 mg, SD 81.3 mg). Participants with PD were recruited and tested at Tufts Medical Center, Boston, Massachusetts. All procedures were approved by The Tufts Health Sciences Institutional Review Board (#12371). The protocol for the healthy cohort was approved by the Schulman Institutional Review Board (#201500837) and conducted at Pfizer, Andover, Massachusetts.

A participant with PD who self-reported as “On with dyskinesia” was excluded from the analysis since dyskinesia might interfere with gait feature measurements. Additionally, 1 healthy volunteer was removed from the analysis due to technical errors with data capture. The clinical and demographic characteristics of the participants whose data were available for analysis are listed in Table 1.

As illustrated in Multimedia Appendix 1A, participants were instrumented with 6 wearable devices (Opal, APDM Inc) located bilaterally on the wrist and foot, and at the lumbar (approximately at the L5 vertebra) and sternum locations. Each device recorded raw data from 9-axis inertial sensors (triaxial accelerometer, triaxial gyroscope, and triaxial magnetometer) at a sampling rate of 128 Hz.

Participants performed a battery of physical activities and cognitive tasks over the course of 2 visits. Both visits were identical for healthy participants but were randomized for participants with PD so that they were in the On state (about 1 hour after medication intake, confirmed with the patient self-report and clinician report) during 1 visit and in the Off state (about 3 hours after last medication intake, confirmed with the patient self-report and clinician report) during the other visit. Physical activities during each visit included scripted activities of daily living (eg, tying a shoe, opening and closing a door) and motor assessments from the MDS-UPDRS part III (eg, 2-minute gait task, finger tapping). In this paper, we present the analysis based on the data collected during the 2-minute gait task. This is to ensure uniform testing conditions for determining the agreement of postexperiment sensor data processing. During this gait task, participants were instructed to walk back and forth along a straight 10-meter track at a comfortable pace for a period of 2 minutes. Participants with PD were assigned an MDS-UPDRS gait score on an ordinal scale of 0 to 4 by a neurologist to assess the degree of gait impairment. Sample sizes (n) for MDS-UPDRS gait scores of 0, 1, 2, and 3 across both visits were 17, 27, 18, and 6, respectively.

APDM Mobility Lab is a commercially available system widely used for objective assessment of gait and leverages data from 3 to 6 wireless, body-worn Opal inertial devices [13,26]. We used APDM Mobility Lab to derive a set of lower limbs, lumbar, and trunk range of motion and upper limb gait features from 6 wearable devices placed on the lower back, sternum, and bilaterally on the feet and wrists (Multimedia Appendix 2). Using 3 sensors located on the lower back and both feet, APDM Mobility Lab can only derive features related to the lower limb and lumbar range of motion. Therefore, we used only features related to lower limb and lumbar range of motion as the 3-sensor feature set (Multimedia Appendix 2). To derive gait features from a single lumbar-mounted triaxial accelerometer, we developed and implemented a previously published wavelet-based method [14] in a Python v3.6 package called GaitPy (Multimedia Appendix 1) [25]. A complete list of gait features derived from a single lumbar-mounted device and those requiring additional devices can be seen in Multimedia Appendix 2.

Statistical analysis was performed in R, version 3.4.1 (The R Project for Statistical Computing) [27], using the following packages: “psych” for intraclass correlation coefficient (ICC), “BlandAltmanLeh” for Bland-Altman plots, “nlme” for linear mixed-effects model, “car” for type 3 analysis of variance, and “MASS” for stepwise model selection.

The median value of each gait feature extracted from the data collected during the 2-minute walking task was calculated for each visit separately. Test-retest reliability of gait features was assessed by calculating the ICC on data collected from healthy volunteers during visit 1 and visit 2. ICC was also used in addition to Bland-Altman plots and 95% limits of agreement to evaluate the agreement between gait features derived using GaitPy and APDM Mobility Lab. The results are presented in Figure 1, where values are ICC_2,1 coefficient (2-way random effects, absolute agreement) with lower and upper confidence bounds, reported as ICC coefficient (lower, upper). Test-retest reliability and agreement between features were assessed according to the following benchmarks. ICC ≤ 0.4 indicates “poor,” 0.4 to 0.59 “moderate,” 0.6 to 0.74 “good,” and 0.75 to 1 “excellent” reliability [28]. Variation of gait features with the MDS-UPDRS gait score in patients with PD was assessed using the Kruskal-Wallis test. Posthoc Conover-Iman tests were used for pairwise comparisons, and multiplicity was adjusted using false discovery rate correction.

Gait features derived using a single, 3, and 6 devices were separately used to fit 3 longitudinal mixed-effects regression models to predict the clinician’s MDS-UPDRS gait score (using the lme function in “nlme” R package). Prior to model fitting, pairwise correlation between sensor features was computed, and highly correlated features were removed. Gait features and covariates including age, gender, visit number, BMI, and years since first symptoms were modeled as fixed effects and participant as a random effect. An unstructured correlation matrix was used. Numerical features were standardized to have 0 mean and unit variance. Stepwise model selection was performed using Akaike Information Criterion as a cost function to achieve the optimal model fit (using the stepAIC function in “MASS” R package). Analysis of variance findings were reported as chi-square values and corresponding P values using Type 3 sum of squares (statistics were derived using the “car” R package). Final models were used to predict the clinician’s score using leave-1-subject-out cross-validation. We report the root mean square error (RMSE) and marginal R² representing the variance explained by the model of fixed effects. Paired Wilcoxon signed rank tests were used to compare predicted gait scores between On and Off states.

The study of participants with PD was approved by The Tufts Health Sciences Institutional Review Board (#12371) and conducted at the Tufts Medical Center. The study of healthy participants was approved by the Schulman Institutional Review Board (#201500837), and conducted at Pfizer in Andover, Massachusetts. Written informed consent was obtained from all participants prior to testing. We confirm that we have read the journal’s position on issues involved in ethical publication and affirm that this work is consistent with those guidelines.

Gait features derived during the 2-minute walking task using GaitPy (single lumbar-mounted device) were compared with the same features derived using the APDM Mobility Lab (6 devices) for healthy participants and participants with PD with data from their 2 visits. Gait features derived using the APDM Mobility Lab were used as the reference since the APDM device has been validated against data from an instrumented treadmill and has been extensively used in both healthy populations and populations with PD [26,29]. Excellent agreement was observed between the 2 methods for stride time and step time in both healthy participants and participants with PD (ICC≥0.86), as shown in Figure 1. Furthermore, excellent agreement was observed for stance time, stride length, and gait speed in participants with PD (ICC=0.86, 0.88, and 0.89, respectively), and agreement was good in healthy participants (ICC=0.68, 0.60, and 0.70, respectively) (Figure 1). Bland-Altman analysis showed that mean difference between GaitPy and APDM Mobility Lab was smaller for longer stance times (Multimedia Appendix 3). Good agreement was also observed for swing time in both healthy participants and participants with PD (0.64≤ICC≤0.73) (Figure 1). In contrast, double support showed poor agreement in healthy participants (ICC=0.20) and moderate agreement in participants with PD (ICC=0.46) (Figure 1). Asymmetry and variability features also showed poor agreement (ICC≤0.31) between the 2 methods for both healthy participants and participants with PD (Figure 1).

Test-retest reliability of gait features derived using GaitPy was assessed using the data collected from healthy participants. Excellent test-retest reliability (ICC≥0.85) (Figure 1) was observed for all spatial and temporal gait features. Asymmetry and variability features showed poor-to-excellent test-retest reliability (0.14≤ICC≤0.77) (Figure 1).

We assessed the ability of gait features derived by using methods relying on different device setups (single device, 3 devices, and 6 devices) to distinguish between MDS-UPDRS gait scores. Based on the APDM Mobility Lab documentation [30], we could determine which gait features are available for analysis with a 3-device setup and a 6-device setup. Therefore, for the purpose of this comparison, we limited our analysis to the 1-device setup (using GaitPy) and 3- and 6-device setups (using APDM Mobility Lab).

The spatial features of gait (ie, gait speed, stride length, and step length) varied most significantly with MDS-UDPRS gait scores in participants with PD (Multimedia Appendix 4). Using leave-1-subject-out cross-validation, the longitudinal mixed-effects regression model based on gait features derived using a single lumbar-mounted device predicted the clinician’s gait score with an RMSE=0.64 and an R²=0.53. The predicted score significantly distinguished between scores of 1 and 2 (P<.001), and marginally distinguished between scores of 0 and 1 (P=.07), and 2 and 3 (P=.18) (Figure 2A). Stance time (χ²₁=12.8; P<.001), step length (χ²₁=49.2; P<.001), and step length asymmetry (χ²₁=6.7; P=.01) had significant effects on describing the MDS-UPDRS gait score.

Comparable performance was observed for a model based on gait features derived using data from 3 devices (RMSE=0.64; R²=0.54). The R² value for the 3-device model was only slightly higher than the single-device model. The predicted gait score could significantly distinguish between MDS-UPDRS gait scores of 0 and 1 (P=.02), 1 and 2 (P<.001), and 2 and 3 (P=.03) (Figure 2B). Pitch at initial contact (χ²₁=7.3; P=.007), maximum pitch (χ²₁=10.5; P=.001), cadence (χ²₁=14.9; P<.001), initial mid-swing duration (χ²₁=4.5; P=.03), and pitch at toe off variability (χ²₁=6.4; P=.011) had a significant effect on describing the MDS-UPDRS gait score.

The model based on gait features derived using data from 6 devices achieved better performance than the single and 3-device model at predicting clinician gait score (RMSE=0.54; R²=0.65). The predicted gait score significantly distinguished between MDS-UPDRS gait scores of 0 and 1 (P=.001), 1 and 2 (P<.001), and 2 and 3 (P=.02) (Figure 2C). Pitch at initial contact (χ²₁=8.9; P=.003), maximum pitch (χ²₁=5.4; P=.02), cadence (χ²₁=19.8; P<.001), initial mid-swing duration asymmetry (χ²₁=5.7; P=.02), trunk sagittal average angle (χ²₁=18.9; P<.001), upper limb foot phase difference (χ²₁=5.0; P=.03), maximum pitch variability (χ²₁=8.6; P=.003), trunk sagittal average angle variability (χ²₁=5.0; P=.025), BMI (χ²₁=10.0; P=.002), and years since first symptom (χ²₁=7.0; P=.008) had significant effects on describing the MDS-UPDRS gait score.

We assessed the ability of predicted gait scores derived using methods relying on different device setups (single device, 3 devices, and 6 devices) to discriminate between On and Off motor states. As shown in Figure 3A, the clinician-rated MDS-UPDRS gait score was significantly different (P=.001) between the patient-reported On and Off state. Similarly, predicted gait scores, estimated from 1-, 3-, and 6-device models (Figure 3B-D), all significantly differentiated between On and Off states (P=.004, P=.004, and P=.045 respectively).

In this exploratory, noninterventional study involving healthy participants and participants with PD, we derived gait features from participants during 2 clinic visits using wearable devices. We found that gait features derived from a single lumbar-mounted accelerometer could predict the clinician-rated gait impairment score to a similar degree as gait features derived from 3 or 6 sensors. Additionally, analogous to clinician-rated scores, predicted gait scores using gait features derived from either 1, 3, or 6 devices all significantly distinguished between the On and Off medication states. Our results suggest that a subset of gait features, derivable using a single lumbar-mounted accelerometer, may be sufficient to measure the degree of gait impairment and the effects of treatment in patients with PD.

Agreement of GaitPy with the reference system (APDM Mobility Lab) was assessed in both healthy participants and participants with PD. While we observed moderate-to-excellent agreement between the most temporal and spatial features of gait derived using GaitPy from a single device and gait features provided by APDM Mobility Lab using 6 devices, the agreement was poor for asymmetry and variability features. Notably, agreement was better in participants with PD for 4 out of 7 temporal and spatial features. The differences between ICC values for healthy participants and participants with PD were significant for stance time (0.86 vs 0.68), stride length (0.88 vs 0.60), and gait speed (0.89 vs 0.70). This result contrasts a prior study [14] where a good agreement with the reference system (gait mat) for both participants with PD and age-matched healthy controls was observed. However, unlike the prior study [14], the 2 groups in our study were not age-matched. To this end, patients with PD showed a wider range of values for gait speed and stride length compared to healthy participants, especially in the lower range, which may have contributed to a better agreement (Multimedia Appendix 5B and C).

We evaluated the test-retest reliability of gait parameters derived using GaitPy in a sample of 59 healthy participants. Test-retest reliability for GaitPy was excellent for all spatial and temporal features, whereas it was poor-to-excellent for asymmetry and variability features. These results suggest that the temporal and spatial features of gait can be reliably measured using a single accelerometer mounted on the lower back. However, as has been reported previously [14], the agreement and reliability of variability and asymmetry features might be sensitive to the employed measurement technique (eg, sensing modality or device location). This is partially because asymmetry and variability are small measurements, which are significantly affected by noise or error in the measurement of temporal or spatial features. Potential sources of measurement error for GaitPy include (1) biomechanical approximation of the inverted pendulum model, (2) error in the estimation of vertical displacement from vertical acceleration, and (3) distal location of the sensing device relative to the feet.

We assessed the criterion and discriminative validity of MDS-UPDRS gait scores using linear mixed-effects models based on gait features derived using data from a single device, 3 devices, and 6 devices. Although a single device provides substantially fewer features of gait compared to either 3- or 6-device models, 17 of the 34 features that varied most significantly (P=.004) with MDS-UPDRS gait score can be derived using a lumbar-mounted sensor (Multimedia Appendix 4). This includes many gait features known to be affected in PD, including stance time, gait speed, step-to-step asymmetries, and gait variability [3,4].

While the 6-device model (RMSE=0.54; R²=0.65) performed slightly better at estimating MDS-UPDRS gait score, performance of the 3-device model (RMSE=0.64; R²=0.54) was comparable to the single device model (RMSE=0.64; R²=0.53). However, unlike the 3 and 6 device models, the single device model was unable to significantly distinguish between the adjacent scores such as 0 and 1 or 2 and 3. A potential reason for this could be the small number of observations for class 3 (n=6). Additionally, gait features related to the pitch and mid-swing duration that were significant for both the 3 and 6-device models could not be derived using the single device model. This indicates features derived from the lower extremity (eg, foot) might have a higher predictive power. Indeed, 3 of 10 features in the 6-device model and 3 of 5 features in the 3-device model that were significant were related to the pitch of foot.

When we assessed the ability of gait scores predicted by the linear mixed-effects models to differentiate between On and Off motor states, we found significant differences (P=.045) for gait scores derived using 1-, 3-, and 6-device models. Additionally, the Off to On gait score directionality was largely consistent between those produced by each model and the clinician-rated gait score. In 10 of the 12 subjects, the clinician score and the predicted score differences between On and Off states were in the same direction (Multimedia Appendix 5A). This was comparable with the 3-device model (10/12) (Multimedia Appendix 5B) and 6-device model (10/12) (Multimedia Appendix 5C).

Data analyzed in this study were collected during performance of motor assessments in the laboratory settings and could be affected by the observer effect and heightened awareness of the patient. Gait features derived using wearable devices were not validated against a gold-standard reference (eg, an instrumented walkway or a motion capture system). This limitation of our work is mitigated to some extent by prior work in which the authors evaluated the algorithm implemented in GaitPy against an instrumented walkway [31]. Another limitation of our work is that the healthy participants and participants with PD were not age matched. Therefore, the results for accuracy and reliability in our healthy cohort might be different in healthy older adults. Additional work is required to validate the results presented herein on an independent data set as well as to confirm the ability of GaitPy to accurately assess gait impairment in free-living conditions.

Our results suggest that a single triaxial accelerometer on the lower back may be sufficient to characterize gait impairments in patients with PD. Algorithms that estimate gait features from a lumbar-mounted sensor, such as GaitPy, could provide clinically meaningful measures of changes in the severity of gait impairments and changes in motor state associated with the effects of treatment in patients with PD. The long battery life of an accelerometer-only device and high degree of utility associated with a single device worn on the lower back enable further investigations to assess the validity of this approach for monitoring gait under free-living conditions. Comparing sensor-derived gait features with classical patient-reported motor diary–based approaches in their ability to detect treatment-related effects may provide an insight into the utility of a single lumbar-mounted sensor in free-living environments. Our ongoing efforts are focused on performing a clinical validation in a semisupervised setting as an intermediate step between the clinic and at-home environment.