Skip to main content

Table 6 Computation time and model performance for select feature sets

From: Role of data measurement characteristics in the accurate detection of Parkinson’s disease symptoms using wearable sensors

TREMOR (Binary)

BRADYKINESIA (Binary)

Features

Computation Time (ms)

AUROC

p

Features

Computation Time (ms)

AUROC

p

C

4.41

0.68

< 0.001*

C

4.41

0.61

0.001*

D

6.47

0.73

< 0.001*

D

6.47

0.64

0.006*

T

8.15

0.73

0.001*

CD

7.55

0.65

0.005*

DT

11.29

0.74

0.018†

T

8.15

0.67

0.099

DF

18.42

0.75

0.022†

DT

11.29

0.68

0.174

FT

20.09

0.75

0.004*

ET

63.02

0.68

0.222

CDFT

24.31

0.76

0.040†

DET

66.17

0.68

–

CE

59.29

0.77

0.126

    

EF

70.15

0.77

0.155

    

EFT

74.98

0.77

0.071

    

CEFT

76.05

0.77

–

    

TREMOR (Multiclass)

BRADYKINESIA (Multiclass)

Features

Computation Time (ms)

AUROC

p

Features

Computation Time (ms)

AUROC

p

C

4.41

0.67

0.001*

C

4.41

0.59

< 0.001*

D

6.47

0.71

< 0.001*

D

6.47

0.62

0.010*

DT

11.29

0.72

0.001*

CD

7.55

0.62

0.001*

F

15.28

0.73

0.014†

T

8.15

0.64

0.011*

CF

16.35

0.73

0.015†

DT

11.29

0.65

0.152

DF

18.42

0.74

0.037†

FTE

74.98

0.65

0.189

EF

70.16

0.75

0.234

DEFT

78.11

0.66

–

DEF

73.30

0.75

0.226

    

CEFT

76.05

0.75

–

    
  1. Total computation time and average AUROC for each combination of feature categories (includes tri-axial and magnitude features). Only features combinations showing improved performance with increasing computation time were included. T = Time, F = Frequency, C = Correlation, D = Derivative, E = Entropy. Asterisk (*) indicates significant difference after Holm-Bonferroni correction (α = 0.05) from the best performing feature set, marked in italic. Dagger (†) indicates additional significant differences when not controlling the family-wise error rate
Back to article page

Journal of NeuroEngineering and Rehabilitation

ISSN: 1743-0003

Contact us