Table 3.
Reclassification of individual risk predictions with machine learning and statistical models
| Model | Reclassification in overall testing cohort: No (%)* | |
|---|---|---|
| Reclassified* | Not reclassified | |
| Overall cohort | ||
| QRISK3 10 year risk prediction (reference model): | ||
| ≤7.5% threshold | 94 186 (13.6) | 597 478 (86.4) |
| >7.5% threshold | 129 348 (57.8) | 94 467 (42.2) |
| Logistic model (Caret) 10 year risk prediction (reference model): | ||
| ≤7.5% threshold | 209 221 (25.9) | 597 478 (74.1) |
| >7.5% threshold | 14 313 (13.2) | 94 467 (86.8) |
| Cohort without censoring | ||
| QRISK3 10 year risk prediction (reference model): | ||
| ≤7.5% threshold | 34 607 (54.6) | 28 779 (45.4) |
| >7.5% threshold | 1248 (2.6) | 47 234 (97.4) |
| Logistic model (Caret) 10 year risk prediction (reference model): | ||
| ≤7.5% threshold | 6004 (17.3) | 28 779 (82.7) |
| >7.5% threshold | 29 851 (38.7) | 47 234 (61.3) |
*
Patients were reclassified if they had a risk prediction in any model that crossed the threshold compared with the prediction of the reference model.