Accelerometry-Based Digital Gait Characteristics for Classification of Parkinson's Disease: What Counts?

Objective: Gait may be a useful biomarker that can be objectively measured with wearable technology to classify Parkinson's disease (PD). This study aims to: (i) comprehensively quantify a battery of commonly utilized gait digital characteristics (spatiotemporal and signal-based), and (ii) identify the best discriminative characteristics for the optimal classification of PD. Methods: Six partial least square discriminant analysis (PLS-DA) models were trained on subsets of 210 characteristics measured in 142 subjects (81 people with PD, 61 controls (CL)). Results: Models accuracy ranged between 70.42-88.73% (AUC: 78.4-94.5%) with a sensitivity of 72.84-90.12% and a specificity of 60.3-86.89%. Signal-based digital gait characteristics independently gave 87.32% accuracy. The most influential characteristics in the classification models were related to root mean square values, power spectral density, step velocity and length, gait regularity and age. Conclusions: This study highlights the importance of signal-based gait characteristics in the development of tools to help classify PD in the early stages of the disease.

PLS-DA is used for the data reduction and to classify the subjects with Parkinson's disease from controls based on the latent variables (capturing high variance) used as training data in discriminant analysis for classification. This method can provide the contribution of the independent variables (gait characteristics) in the classification modeling based on the variable importance in the projection (VIP) [4], which is calculated with the following equation 1.
is a total number of gait characteristics in the model. N is the number of latent variables in the PLS-DA model.
is the sum of variance of the latent variable, quantify the contribution of gait characteristics j according to the kth latent variable. is the contribution of the kth latent variable. If the VIP is above 1 then the particular gait characteristic is highly influential in the model and higher VIP value indicates the higher contribution of the gait characteristic. The model quality is based on the standard goodness of fit indices such as Q2, R2X and R2Y ( Figure S1). Number of components (latent variables) in the PLS-DA are based on these indices. Q2 explains the predictive capability of the model, and with its value the suitable number of components in the PLS-DA can be identified. If the value of Q2 is higher than 0 then model prediction capability is better, if its value is below 0 then model performance will be poor. Therefore, based on its value we selected the number of components in the PLS-DA. Other indices, such as R2X and R2Y explain the captured variance by independent and dependent variables respectively.

Domain
Feature Definition Spatialtemporal [5] Step Length Distance between two heel strikes Step Velocity Step length/step time Step Estimated for each direction as described by Menz et al., 2003 [10, 11]. Harmonic ratios of acceleration signals in VT and AP directions were calculated as the sum of even harmonics divided by the sum of odd harmonics, since these signals have two phases per stride. Harmonic ratios from ML acceleration were calculated as the sum of odd harmonics divided by the sum of even harmonics, since acceleration signals in mediolateral (ML) direction are monophasic per stride. This measurement reflects the rhythmicity of periodic patterns and relates to gait symmetry [12]. Thus, higher values of this feature are related to more rhythmic, paced and symmetric gait patterns [13].
Harmonic Ratio -Stride Harmonic Ratio (V, ML, & AP) The step-to-step symmetry within a stride from calculating a ratio of the odd and even harmonics of a signal following fast Fourier transformation [10]. Estimated for each direction as the PSD of the fundamental frequency (first harmonic) divided by the cumulative sum of the power spectral density of the first six harmonics [13,14]. This measure, proposed by Lamoth et al., 2002 reflects gait smoothness [13]. Thus, values approaching the maximum value of 1.0 indicate a smoother gait pattern, which may reflect a less vigorous/more cautious movement pattern, whereas smaller values might indicate movements that are more erratic. The calculation of the Root mean square of the acceleration signal [15].

Jerk RMS:
The calculation of the root mean square of the first time derivative of the acceleration signal (jerk) [16].

Jerk ratio:
A logarithmic ratio of either the AP or ML acceleration RMS over the V acceleration RMS [16,17] Signal Regularity Step Regularity (V, AP, ML & R) Estimated as the normalized unbiased auto-covariance for a lag of one step time [17]. This feature thus reflects the similarity between subsequent steps of the acceleration pattern over a step. Values of this feature close to 1.0 (maximum possible value) reflect repeatable patterns between subsequent steps.
Stride Regularity (V, AP, ML, & R) Estimated as the normalized unbiased auto-covariance for a lag of one stride time [17]. This feature thus reflects the similarity between subsequent strides of the acceleration pattern over a stride cycle.

Symmetry Autocorrelation Ratio (V, AP, ML, & R)
A ratio between step and stride regularity designed to quantify the level of symmetry between them and indicative of symmetry during a straight walk [17].

Symmetry Autocorrelation Difference (V, AP, ML, & R)
Difference between step and stride regularity designed to quantify the level of symmetry between them and indicative of symmetry during a straight walk [17].

Gait Symmetry Index
Calculated based upon the concept of the summation of the biased autocorrelation from all three components of movement and a subsequent calculation of step and stride timing asymmetry [18].

Signal Complexity
Lyapunov exponent Estimated as the exponential rate of divergence or convergence after a small disturbance of nearby orbits in state space. Since nearby orbits correspond to nearly identical states, a positive value indicates that systems with initial differences will soon behave quite differently, and stability is low [19]. The Lyapunov exponent was calculated using the Rostein and Wolf method [20] for each detrended acceleration signal (VT, ML & AP). We used an embedding dimension of 5 and a delay of 12 samples [21]. As local dynamic stability estimates based on a 6D and 12D state space correlated highly with those of the employed 9D state space in a previous study [22] (respectively r≥0.94 and r≥0.81), the number of embedding dimensions has been considered to have minor effects [22]. Thus, to reduce computational cost, we explored the use of a 5D state reconstruction (with two less dimensions than used in the reference) [23].
Phase plots features [24]: Features from ellipses fitted to full cycles/orbits of phase plots: Full orbit eccentricity

Relative orbit inclination
Full orbit area Full orbit minor axis SD Full orbit major axis SD Features from ellipses fitted to half-cycles/orbits following ellipse segmentation along the minor (short) axis: Short half orbit eccentricity asymmetry Short half orbit segment angle Short half orbit area asymmetry Features from ellipses fitted to half-cycles/orbits following ellipse segmentation along the major (long) axis: Long half orbit eccentricity asymmetry Long half orbit segment angle Long half orbit area asymmetry

Intra-step correlation
Average eccentricity of all fully fitted ellipses.
Average angle subtended by alternating fitted ellipses within a bout of gait.