Table 2.
Risk score performance in C3PO versus Convenience Samples.
| Model | Hazard ratio (per 1-SD increase) | C-index3 (95% CI) | GND | Recalibrated GND | ICI6 (95% CI) | Recalibrated ICI5,6 (95% CI) | Calibration slope7 (95% CI) |
|---|---|---|---|---|---|---|---|
| C3PO | |||||||
| PCE (White women)1 | 2.51 (2.43–2.59) | 0.768 (0.760–0.775) | 487 | 1689 | 0.018 (0.017–0.020) p < 0.01 | 0.034 (0.031–0.037) p < 0.01 | 0.67 (0.65–0.70) p = 0.02 |
| PCE (Black women)1 | 2.39 (2.17–2.64) | 0.724 (0.702–0.746) | 69 | 257 | 0.030 (0.023–0.036) p < 0.01 | 0.057 (0.050–0.064) p < 0.01 | 0.60 (0.53–0.67) p = 0.36 |
| PCE (White men)1 | 2.17 (2.11–2.24) | 0.738 (0.730–0.746) | 361 | 618 | 0.024 (0.022–0.027) p < 0.01 | 0.032 (0.029–0.035) p < 0.01 | 0.70 (0.68–0.73) p = 0.59 |
| PCE (Black men)1 | 2.04 (1.85–2.25) | 0.725 (0.698–0.751) | 21 | 183 | 0.012 (0–0.025) p = 0.06 | 0.010 (0–0.024) p = 0.83 | 0.88 (0.77–1.00) p = 0.84 |
| CHARGE-AF1 | 2.56 (2.50–2.61) | 0.782 (0.777–0.787) | 1856 | 1367 | 0.028 (0.027–0.030) p < 0.01 | 0.019 (0.018–0.021) p < 0.01 | 0.77 (0.75–0.79) p < 0.01 |
| Convenience Samples | |||||||
| PCE (White women)2 | 2.44 (2.39–2.49) | 0.770 (0.764–0.775) | 1797 | 4923 | 0.032 (0.031–0.034) | 0.047 (0.044–0.049) | 0.64 (0.62–0.65) |
| PCE (Black women)2 | 2.29 (2.13–2.46) | 0.732 (0.716–0.748) | 213 | 562 | 0.046 (0.040–0.053) | 0.074 (0.067–0.081) | 0.56 (0.51–0.61) |
| PCE (White men)2 | 2.18 (2.13–2.22) | 0.744 (0.739–0.749) | 1291 | 1493 | 0.041 (0.038–0.043) | 0.039 (0.037–0.042) | 0.70 (0.68–0.71) |
| PCE (Black men)2 | 2.01 (1.88–2.15) | 0.727 (0.705–0.749) | 36 | 13 | 0.028 (0.018–0.037) | 0.012 (0.0026–0.022) | 0.87 (0.79–0.95) |
| CHARGE-AF2 | 2.40 (2.38–2.43) | 0.781 (0.778–0.784) | 7188 | 8322 | 0.036 (0.035–0.036) | 0.028 (0.027–0.029) | 0.69 (0.68–0.70) |
1PCE (White women): 4231, 107,998, 7.1 (2.8, 10); PCE (Black women): 617, 8450, 7.7 (2.9, 10); PCE (White men): 4928, 76,304, 6.2 (2.3, 10); PCE (Black men): 425, 5432, 6.7 (2.5, 10); CHARGE-AF: n events = 7877, N total = 174,644, median follow-up, years (Q1,Q3): 5.0 (2.3,5.0).
2PCE (White women): 10,259, 182,349, 7.5 (3.6, 10); PCE (Black women): 1119, 12,690, 7.2 (3.2, 10); PCE (White men): 12,891, 136,629, 6.2 (2.6, 10); PCE (Black men): 843, 8558, 6.0 (2.6, 10); CHARGE-AF: n events = 26,907, N total = 501,272, median follow-up, years (Q1,Q3): 5.0 (2.0,5.0).
3C-index calculated using the inverse probability of censoring weighting method28.
4Greenwood-Nam-D’Agostino (GND) test, a test of calibration30. Lower chi-squared values suggest better calibration (across equally sized samples). Significant p-values indicate evidence of miscalibration. Corresponding p-values are all p < 0.01 except for C3PO PCE Black men (p = 0.02), C3PO PCE Black men recalibrated (p = 0.03), Convenience Sample PCE Black men recalibrated (p = 0.17).
5Values after recalibration to the baseline hazard of the sample (see text).
6Integrated calibration index, a quantitative measure of the average difference between predicted event risk and observed event incidence, weighted by the empirical distribution of event risk29. Smaller values indicate better calibration. P-values indicated pairwise comparison of ICI with the corresponding Convenience Sample.
7A measure of calibration applicable to models that are calibrated in the large31,44. A calibration slope equal to one is optimally calibrated. P-values indicated pairwise comparison of calibration slope with corresponding Convenience Sample.
SD standard deviation, CI confidence interval.