Table 2.
Sensitivity, specificity, positive predictive and negative values for machine-learning models for detecting familial hypercholesterolaemia in the validation cohort (n = 1,006,943).
| Machine-learning models | % High probability (>1/250) (%) | Sensitivity | Specificity | Positive predictive value (PPV) | Negative predictive value (NPV) |
|---|---|---|---|---|---|
| Logistic regression | 3.38 | 37.6% (35.5–39.8) | 96.7% (96.6–96.7) | 4.4% (4.1–4.6) | 99.7% (99.7–99.8) |
| Random forest | 8.09 | 69.1% (67.0–71.2) | 92.0% (92.0–92.1) | 3.4% (3.3–3.5) | 99.9% (99.9–99.9) |
| Gradient boosting | 4.27 | 58.3% (56.1–60.5) | 95.8% (95.8–95.9) | 5.3% (5.1–5.5) | 99.8% (99.8–99.8) |
| Deep learning | 10.16 | 72.6% (70.6–74.6) | 90.0% (89.9–90.0) | 2.8% (2.8–2.9) | 99.9% (99.9–99.9) |
| Ensemble learning | 0.73 | 30.5% (28.4–32.6) | 99.3% (99.3–99.3) | 15.5% (14.5–16.4) | 99.7% (99.7–99.7) |
aAssumes population frequency of familial hypercholesterolaemia of 1 in 2504.