AtlasSession 4L12: Introduction to measurement errors and diagnostic test accuracy studiesp.31–32
Thresholds and ROC curves
- 2 earlier topics
- Accuracy vs precision
- Random vs systematic error
- Assessing measurement error
- Diagnostic test accuracy (2×2)
- Thresholds and ROC curves
Fig. 1
Gold-nanoparticle fluorescence (a.u.) in 30 carcinoma and 30 non-tumour samples, histopathology as reference (PBL 4; intensities illustrative)
Fig. 1. Glow readings for 30 carcinomas and 30 non-tumour samples, one dot each, with a glow line at 105.0 a.u.: a dot at or above the line counts as a yes. Caught 25 of 30 carcinomas (5 missed); 12 of 30 non-tumour samples are false alarms (18 cleared). Lowering the line can only catch more, and can only add false alarms.
- Glow line
- 105.0 a.u.
- Caught (carcinomas)
- 25 of 30, about 83%
- Missed
- 5 of 30
- False alarms (non-tumour samples)
- 12 of 30, about 40%
- Cleared
- 18 of 30
Threshold & ROC · First look
Level 1 · First look · Lecture 12 · p.31–32
Explain how a cut-off line turns a glow reading into yes/no and why catching more cancers costs more false alarms.
Dichotomise a quantitative test at a threshold (at or above = positive), see the sensitivity–specificity trade-off, and read an ROC curve of Sn against 1 − Sp.
Detector dial (the step you are on)
An airport scanner set too touchy beeps at belt buckles; too lax, it misses things.
Like an airport scanner’s dial: set it too touchy and it beeps at belt buckles, too lax and it misses things. Drag the glow line: carcinomas caught 25 of 30, about 83%; false alarms among non-tumour samples 12 of 30, about 40%.