Table 1 Performance comparison of our proposed multi-task, multi-stage deep transfer learning (DTL) versus multi-task, single-stage DTL models for the joint prediction of abnormal cognitive, language, and motor outcomes at 2 years corrected age in very preterm infants.

From: A multi-task, multi-stage deep transfer learning model for early prediction of neurodevelopment in very preterm infants

Model

Cognition

BA (%)

Sen (%)

Spe (%)

LR + 

FPR (%)

AUC

Single-stage

74.2 ± 6.1

68.0 ± 6.0

80.3 ± 6.6

3.5 ± 1.6

19.7 ± 6.6

0.77 ± 0.05

Multi-stage

81.5 ± 3.2

74.0 ± 4.9

88.9 ± 3.1

6.6 ± 1.9

11.1 ± 3.1

0.86 ± 0.05

Model

Language

BA (%)

Sen (%)

Spe (%)

LR + 

FPR (%)

AUC

Single-stage

58.0 ± 4.6

56.0 ± 6.0

60.0 ± 5.6

1.4 ± 0.2

40.0 ± 5.6

0.63 ± 0.05

Multi-stage

68.9 ± 2.3

60.0 ± 4.0

77.8 ± 3.2

2.6 ± 0.3

22.2 ± 3.2

0.66 ± 0.03

Model

Motor

BA (%)

Sen (%)

Spe (%)

LR + 

FPR (%)

AUC

Single-stage

66.4 ± 4.7

62.0 ± 4.2

70.7 ± 6.6

2.0 ± 0.4

29.3 ± 6.6

0.74 ± 0.04

Multi-stage

73.9 ± 2.4

76.0 ± 4.9

71.7 ± 3.4

2.6 ± 0.2

28.3 ± 3.4

0.84 ± 0.02

  1. BA, balanced accuracy; Sen, sensitivity; Spe, specificity; LR+, likelihood ratio positive; FPR, false positive rate; AUC, area under the receiver operating characteristic curve.
Back to article page