Skip to main content

Table 2 The diagnostic performance for detecting pathological HSIL+ at different hypothetical biopsy thresholds

From: Development and validation of an artificial intelligence system for grading colposcopic impressions and guiding biopsies

 

Accuracy, % (95% CI)

Sensitivity, % (95% CI)

Specificity, % (95% CI)

Positive predictive, value % (95% CI)

Negative predictive, value % (95% CI)

Analysis of images alone by CAIADS

 Normal/benign versus low-grade or worse

63.6 (62.1–65.1)

87.3 (85.5–88.9)

48.9 (46.8–50.9)

51.6 (49.6–53.6)

86.0 (84.1–87.8)

 Less severe impressionsa versus high-grade or worse

80.7 (79.4–81.9)

65.8 (63.3–68.2)

90.0 (88.7–91.2)

80.4 (78.0–82.6)

80.8 (79.3–82.3)

Analysis of both images and non-image informationb by CAIADS

 Normal/benign versus low-grade or worse

66.7 (65.2–68.2)

90.5 (88.9–91.4)

51.8 (49.8–53.8)

54.0 (52.0–55.9)

89.7 (88.0–91.3)

 Less severe impressionsa versus high-grade or worse

85.5 (84.3–86.6)

71.9 (69.5–74.2)

93.9 (92.9–94.9)

88.1 (86.2–89.9)

84.3 (82.8–85.6)

Analysis of both images and non-image informationb by colposcopists

 Normal/benign versus low-grade or worse

64.1 (62.6–65.6)

83.5 (81.5–85.3)

52.0 (50.0–54.1)

52.1 (50.0–54.1)

83.5 (81.5–85.3)

 Less severe impressionsa versus high-grade or worse

81.6 (80.4–82.8)

60.4 (57.9–62.9)

94.9 (93.9–95.7)

88.0 (85.9–89.9)

79.3 (77.8–80.8)

  1. Abbreviations: CAIADS Colposcopic Artificial Intelligence Auxiliary Diagnostic System
  2. aLess severe impressions included normal/benign and low-grade
  3. bNon-image information included patient’s primary screening findings (cytology, HPV status)