TABLE 3.
Model performance metrics for predicting death from melanoma within 7 years
| Model | AUC | Proportion of sample with low risk of death, No. (%) | No. in subset that died | Risk of melanoma death in subset (95% CI) |
|---|---|---|---|---|
| A. Model performance and AUC for identifying patients at low risk of death from melanoma within 7 yearsa | ||||
| Training data using 2010–2011 cohort (N = 7652)b | ||||
| Model 1A: CART model with 3 leaves | 0.73 | 2707 (35) | 12 | 0.44% (0.25%, 0.77%) |
| Model 1B: CART model 5 leaves | 0.74 | 1950 (25) | 3 | 0.15% (0.05%, 0.45%) |
| Model 2: logistic model, risk of death <0.5% | 0.80 | 1896 (25) | 9 | 0.47% (0.25%, 0.90%) |
| Testing data using 2010–2011 cohort (N = 3942)c | ||||
| Model 1A: CART model with 3 leaves | 0.64 | 1381 (35) | 8 | 0.58% (0.29%, 1.14%) |
| Model 1B: CART model with 5 leaves | 0.61 | 993 (25) | 4 | 0.40% (0.16%, 1.03%) |
| Model 2: logistic model, risk of death < 0.5% | 0.78 | 969 (25) | 5 | 0.52% (0.22%, 1.20%) |
| 7-year follow-up using 2004–2009 cohort (N = 47,171)d | ||||
| Model 1A: CART model with 3 leaves | 0.67 | 19,834 (42) | 126 | 0.64% (0.53%, 0.76%) |
| Model 1B: CART model with 5 leaves | 0.63 | 14,428 (31) | 77 | 0.53% (0.43%, 0.67%) |
| 10-year follow-up using 2004–2008 cohort (N = 35,526)e | ||||
| Model 1A: CART model with 3 leaves | 0.67 | 15,706 (44) | 165 | 1.05% (0.90%, 1.22%) |
| Model 1B: CART model with 5 leaves | 0.63 | 11,498 (32) | 97 | 0.84% (0.69%, 1.03%) |
| Proportion of sample with high risk of death, No. (%) | No. in subset that died | Risk of melanoma death in subset (95% CI) | ||
| B. Model performance for identifying patients at high risk of death from melanoma within 7 yearsf | ||||
| Training data using 2010–2011 cohort (N = 7652) | ||||
| Model 2: logistic model, risk of death ≥20% | 65 (0.8) | 18 | 27.7% (18.3%, 39.6%) | |
| Testing data using 2010–2011 cohort (N = 3942) | ||||
| Model 2: logistic model, risk of death ≥20% | 28 (0.7) | 9 | 32.1% (17.9%, 50.7%) | |
Note: Logistic regression model results are based on recalibrated predicted risks. Training and testing data for the primary analysis include patients diagnosed in 2010–2011.
Abbreviations: AUC, area under the receiver-operating characteristic curve; CAET, classification and regression tree; CI, confidence interval.
Training and testing data for the primary analysis include patients diagnosed in 2010–2011. As a secondary analysis, we applied Models 1A and 1B to the cohort of patients diagnosed from 2004 to 2009. Model 2 is not evaluated in the 2004–2009 cohorts because data on mitotic rate are not available for those cohorts.
For reference, recall that 2.3% of patients in the training data overall died.
For reference, recall that 2.9% of patients in the testing data overall died.
For reference, at 7 years, 2.8% of the 2004–2009 cohort had died.
For reference, at 10 years, 3.9% of the 2004–2008 cohort had died.
Results are shown only for the logistic model (Model 2) because only this model gives a continuous risk estimate for every patient.