Table 3 Model performance metrics across EEG models.

From: Predicting treatment response using EEG in major depressive disorder: A machine-learning meta-analysis

(a)

Authors

Sensitivity

2.5%

97.5%

Specificity

2.5%

97.5%

Bailey [26]

0.731

0.460

0.896

0.946

0.798

0.988

Bailey [26]

0.700

0.448

0.870

0.914

0.758

0.973

Corlier [28]

0.607

0.494

0.709

0.643

0.477

0.780

Erguzel [29]

0.919

0.772

0.975

0.827

0.643

0.927

Erguzel [31]

0.841

0.665

0.945

0.938

0.769

0.985

Hasanzadeh [33]

0.854

0.665

0.945

0.938

0.769

0.985

Cao [34]

0.794

0.558

0.922

0.886

0.694

0.964

Cook [35]

0.731

0.576

0.845

0.542

0.383

0.692

Salle [36]

0.696

0.511

0.834

0.929

0.741

0.983

Jaworska [37]

0.768

0.585

0.886

0.980

0.834

0.998

Mumtaz [38]

0.921

0.719

0.982

0.763

0.539

0.899

Zhdanov [41]

0.791

0.666

0.878

0.846

0.742

0.913

Average

0.776

0.600

0.892

0.846

0.678

0.923

Test for equality of sensitivities: X2 = 23.09, p-value = 0.017

Test for equality of specificities: X2 = 46.23, p-value = 0.00000294

Correlation of sensitivities and false-positive rates: Rho = −0.203 (−0696 to 0.420)

Total DOR: 23.49 (95% CI: 10.40–52.02), τ2 = 1.395 (95% CI: 0.00–2.13)

posLR: 5.232 (95% CI: 3.15–8.67), τ2 = 0.502 (0.00–1.24)

negLR: 0.271 (95% CI: 0.195–0.376), τ2 = 0.190 (0.00–0.495)

AUC: 0.850 (95% CI: 0.747–0.890); pAUC: 0.777

(b)

Authors

Mean accuracy

95% CI

%W (random)

Bailey [26]

91.0

81.34–100

8.3

Bailey [26]

86.6

82.23–91.16

12.5

Corlier [28]

68.5

59.96–78.24

7.1

Erguzel [29]

89.0

80.49–98.59

9.0

Erguzel [31]

86.4

80.86–92.31

11.5

Hasanzadeh [33]

91.3

82.57–100

9.1

Cao [34]

81.3

70.04–94.36

6.3

Cook [35]

64.4

53.89–76.94

5.0

Salle [36]

80.9

69.79–93.91

6.3

Jaworska [37]

88.2

79.05–98.49

8.4

Mumtaz [38]

87.50

75.77–100

6.5

Zhdanov [41]

82.4

75.58–89.83

10.0

Random effects model

Mean = 83.93% (95% CI: 78.90–89.29)

  1. A summary of performance metrics across all predictive models of treatment response using EEG.
  2. (a) The madad function in the “mada” package was used to calculate the sensitivity, specificity, and partial Area-Under-The-Curve (AUC) across studies, while the maduani function was used to calculate the Diagnostic Odds Ratio (DOR), positive likelihood ratio (posLR), and negative likelihood ratio (negLR). AUC was calculated using the AUC_boot function in dmetatools, with an alpha of 0.95 and 2000 bootstrap iterations. Overall, the balanced accuracy (sensitivity + specificity/2) was 81.1%.
  3. (b) The metamean function in the “meta” package was used to pool accuracy across studies in a random effects model using an inverse variance method with Knapp–Hartung adjustments to calculate the confidence interval around the pooled effect. Across models, overall model accuracy was 83.93% (95% CI: 78.90–89.29).
Back to article page