Skip to main content

Table 4 Best twenty models using feature selection and best ten all variable (AV) models using a single 75:25 train:test stratified holdout. Feature subset numbers are defined in Table 3. For AV, feature set indicates the sensor and number of variables (in parentheses) in the subset

From: Feature selection for elderly faller classification based on wearable sensors

Method

Feature Set

Modela

Accuracyb (%)

Sensitivity (%)

Specificity (%)

PPV (%)

NPV (%)

F1

MCC

SR

Relief-F

1

SVM-7

96.0 [80.4: 99.3]

100.0

94.7

85.7

100.0

0.923

0.901

33

Relief-F

1

SVM-6

92.0 [75.0: 97.8]

83.3

94.7

83.3

94.7

0.833

0.781

43

Relief-F

2

NN-15

88.0 [70.0: 95.8]

50.0

100.0

100.0

86.4

0.667

0.657

44

Relief-F

3

NN-21

88.0 [70.0: 95.8]

50.0

100.0

100.0

86.4

0.667

0.657

44

Relief-F

3

NN-23

88.0 [70.0: 95.8]

50.0

100.0

100.0

86.4

0.667

0.657

44

Relief-F

3

NN-25

88.0 [70.0: 95.8]

50.0

100.0

100.0

86.4

0.667

0.657

44

Relief-F

1

NN-21

88.0 [70.0: 95.8]

50.0

100.0

100.0

86.4

0.667

0.657

44

Relief-F

4

NN-9

88.0 [70.0: 95.8]

50.0

100.0

100.0

86.4

0.667

0.657

44

Relief-F

4

NN-21

88.0 [70.0: 95.8]

50.0

100.0

100.0

86.4

0.667

0.657

44

Relief-F

1

SVM-5

92.0 [75.0: 97.8]

100.0

89.5

75.0

100.0

0.857

0.819

52

Relief-F

5

SVM-4

88.0 [70.0: 95.8]

66.7

94.7

80.0

90.0

0.727

0.656

65

Relief-F

6

SVM-4

88.0 [70.0: 95.8]

66.7

94.7

80.0

90.0

0.727

0.656

65

Relief-F

7

NN-21

88.0 [70.0: 95.8]

66.7

94.7

80.0

90.0

0.727

0.656

65

Relief-F

3

SVM-3

88.0 [70.0: 95.8]

83.3

89.5

71.4

94.4

0.769

0.693

68

Relief-F

8

NB-Q

84.0 [65.3: 93.6]

83.3

84.2

62.5

94.1

0.714

0.618

102

AV

H(29)

SVM-4

84.0 [65.3: 93.6]

33.3

100.0

100.0

82.6

0.500

0.525

104

AV

I(30),H(29)

SVM-4

84.0 [65.3: 93.6]

33.3

100.0

100.0

82.6

0.500

0.525

104

AV

I(30),P(29), LS(29)

SVM-2

84.0 [65.3: 93.6]

33.3

100.0

100.0

82.6

0.500

0.525

104

AV

H(29),P(29), LS(29),RS(29)

NN-5

84.0 [65.3: 93.6]

33.3

100.0

100.0

82.6

0.500

0.525

104

CFS/FCBF

9

NN-8

84.0 [65.3: 93.6]

33.3

100.0

100.0

82.6

0.500

0.525

104

CFS/FCBF

9

NN-10

84.0 [65.3: 93.6]

33.3

100.0

100.0

82.6

0.500

0.525

104

AV

H(29)

SVM-2

84.0 [65.3: 93.6]

66.7

89.5

66.7

89.5

0.667

0.561

107

AV

I(30),P(29), LS(29),RS(29)

NB-Q

80.0 [60.9: 91.1]

83.3

78.9

55.6

93.8

0.667

0.554

120

AV

I(30),P(29)

SVM-2

84.0 [65.3: 93.6]

50.0

94.7

75.0

85.7

0.600

0.521

121

AV

I(30),H(29), P(29)

SVM-3

84.0 [65.3: 93.6]

50.0

94.7

75.0

85.7

0.600

0.521

121

AV

I(30),P(29)

NN-9

84.0 [65.3: 93.6]

50.0

94.7

75.0

85.7

0.600

0.521

121

AV

I(30),H(29), P(29),LS(29)

NN-20

84.0 [65.3: 93.6]

50.0

94.7

75.0

85.7

0.600

0.521

121

CFS/FCBF

9

NB-Q

76.0 [56.6: 88.5]

66.7

78.9

50.0

88.2

0.571

0.418

157

CFS/FCBF

9

SVM-2

80.0 [60.9: 91.1]

33.3

94.7

66.7

81.8

0.444

0.369

176

CFS/FCBF

9

SVM-3

80.0 [60.9: 91.1]

33.3

94.7

66.7

81.8

0.444

0.369

176

  1. AV all variables, I pressure-sensing insole measures, H head accelerometer measures, P pelvis accelerometer measures, LS left shank accelerometer measures, RS right shank accelerometer measures, NN neural network, NB naïve Bayesian model, SVM support vector machine, SR summed rank
  2. aNN-a, where a is the number of nodes in the hidden layer; SVM-b, where b is the polynomial degree; NB-Q is quadratic naïve Bayesian
  3. bAccuracy [95% Confidence Interval]