Skip to main content

Table 6 External Validation Results on Radiology Reports from MIMIC-III and NHS Tayside

From: Ontology-driven and weakly supervised rare disease identification from clinical notes

 

MIMIC-III Radiology (n=53+/198)

Tayside Brain Imaging (n=79+/283)

Text to UMLS

P

R

\(F_{1}\)

P

R

\(F_1\)

SemEHR [15]

26.7

100.0

42.2

26.9

94.9

41.9

+ WS (transfer)

54.4

92.5

68.5

56.3

91.1

69.6

+ SS (transfer)

89.4

79.2

84.0

69.0

62.0

65.3

+ rules (tuned)

87.5

92.5

89.9

56.8

94.9

71.1

+ WS (in-domain)

72.9

96.2

82.9

48.0

92.4

63.2

+ WS (+ tuning R)

81.5

100.0

89.8

58.1

94.9

72.1

+ WS (+ tuning \(F_1\))

89.1

92.5

90.7

75.3

88.6

81.4

  1. The column statistics (n=\(N_+\)+/N) show number of positive data \(N_+\) and all samples N in the dataset. WS, weak supervision; SS, strong supervision. The original parameters for WS were \(p=0.005\) and \(l=3\). The new parameters for best recall (R) were \(p=0.01\) and \(l=4\) and for best \(F_1\) were \(p=0.0005\) and \(l=4\), for both datasets. For SemEHR+rules, we present the results of rules, where \(p=0.0005\) and \(l=4\), with an OR operation. The best scores for the metrics are bolded