How to Select Balance Measures Sensitive to Parkinson’s Disease from Body-Worn Inertial Sensors—Separating the Trees from the Forest

Abstract

This study aimed to determine the most sensitive objective measures of balance dysfunction that differ between people with Parkinson’s Disease (PD) and healthy controls. One-hundred and forty-four people with PD and 79 age-matched healthy controls wore eight inertial sensors while performing tasks to measure five domains of balance: standing posture (Sway), anticipatory postural adjustments (APAs), automatic postural responses (APRs), dynamic posture (Gait) and limits of stability (LOS). To reduce the initial 93 measures, we selected uncorrelated measures that were most sensitive to PD. After applying a threshold on the Standardized Mean Difference between PD and healthy controls, 44 measures remained; and after reducing highly correlated measures, 24 measures remained. The four most sensitive measures were from APAs and Gait domains. The random forest with 10-fold cross-validation on the remaining measures (n = 24) showed an accuracy to separate PD from healthy controls of 82.4%—identical to result for all measures. Measures from the most sensitive domains, APAs and Gait, were significantly correlated with the severity of disease and with patient-related outcomes. This method greatly reduced the objective measures of balance to the most sensitive for PD, while still capturing four of the five domains of balance.

Turning Detection During Gait: Algorithm Validation and Influence of Sensor Location and Turning Characteristics in the Classification of Parkinson’s Disease

“…The performance was further improved when turning characteristics were combined from other IMU locations; with a combination of the neck, lower back, head and inner ankle resulting in an optimal accuracy of 98% (sensitivity 95%, specificity 100%). Our findings are in line with previous work [ 27 , 28 ]. Using a sensor on the pelvis and ankles, Shah et al [ 27 ] reported the AUC of 0.89 when using turn angle only (29 PD and 27 CL).…”

“…The turning characteristics which were most influential in the classification modeling were related to RMS of angular velocity (radians/second), angular velocity (turn angle/turn duration), turn duration, jerk, angular acceleration, number of steps/transitions per turn, and RMS of acceleration. The results from this study are in line with previous findings [ 28 ] confirming that turn velocity was the most important feature given by the random forest classifier. In another study [ 27 ], the variability of turn duration, jerk and turn angle gave better performance using logistic regression.…”

“…In addition, PLS-DA trained on the gait characteristics gave a classification accuracy of 70.42–88.73% (AUC: 78.4–94.5%) with a sensitivity of 72.84–90.12% and specificity of 60.3–86.89% [ 30 ]. The studies [ 27 , 28 ], which include both turning, gait and other characteristics confirm the value of including turning characteristics with gait characteristics for PD classification.…”

“…Using a sensor on the pelvis and ankles, Shah et al [ 27 ] reported the AUC of 0.89 when using turn angle only (29 PD and 27 CL). Similarly, Hasegawa et al [ 28 ] used eight sensors (attached at sternum, wrists, lumbar region, shank and feet) to quantify different domains of balance and reported an overall accuracy of 82.4% with a random forest classifier (144 PD and 79 CL). Of the gait and turning characteristics included in the random forest classifier, angular velocity was the most important in the classification of PD.…”

“…Previous work has shown that turning characteristics, such as turn velocity, angle and jerk extracted from the lower back, are useful for PD classification [ 27 , 28 ]. People with PD display a lack of dissociation, i.e., a lack of segment coordination, in the rotation of head, pelvis, and feet (turning en bloc) [ 29 ].…”