Table 2.
Comparison of the performance of the three regression models
|
No. of flagged patients
|
Model
|
|||
|---|---|---|---|---|
| Measure | A | B | C | |
| 250 000 | Sensitivity | 27.4% | 85.7% | 27.3% |
| Specificity | 92.9% | 45.1% | 92.4% | |
| PPV | 29.6% | 19.9% | 23.4% | |
| 150 000 | Sensitivity | 19.5% | 65.9% | 19.7% |
| Specificity | 96.1% | 69.5% | 95.7% | |
| PPV | 35.1% | 25.6% | 28.2% | |
| 50 000 | Sensitivity | 9.1% | 32.7% | 9.4% |
| Specificity | 99.0% | 91.7% | 98.8% | |
| PPV | 48.8% | 38.4% | 40.6% | |
| ROC `c' statistic | 0.72 | 0.75 | 0.70 | |
PPV, positive predictive value; model A, all index spells, adjusting for conditions most amenable to case management, predicting 2+ further spells in next 365 days; model B, index spells for only conditions most amenable to case management, predicting 2+ further spells in next 365 days; model C, all index spells, predicting 2+ further spells and survival in next 365 days; ROC, receiver operating characteristic (see methods section)