Table 3.
Performance of the four predictive models (validation cohort, 90% sensitivity)
| AUC (95% CI) | P | Sen (%)a | Spe (%) | NPV (%) | PPV (%) | Accuracy (%) | Bx avoided (%)b | |
|---|---|---|---|---|---|---|---|---|
| Model I: Benign vs Cancer (Marker number in model = 26) | ||||||||
| 5 clinical factorsc | 0.75 (0.70–0.80) | < 10–4 | 90 | 30 | 75 | 58 | 61 | 16 |
| Marker panel | 0.87 (0.83–0.91) | < 10–4 | 90 | 44 | 81 | 63 | 68 | 23 |
| Marker panel + PSA | 0.88 (0.84–0.92) | < 10–4 | 90 | 44 | 81 | 63 | 68 | 23 |
| Combinedd | 0.89 (0.86–0.93) | < 10–4 | 90 | 57 | 85 | 69 | 74 | 30 |
| Model II: (Benign + VLR/LR) vs (FIR + UIR + HR/VHR + mPC) (Marker number in model = 24) | ||||||||
| 5 clinical factorsc | 0.81 (0.76–0.86) | < 10–4 | 90 | 42 | 85 | 55 | 63 | 26 |
| Marker panel | 0.93 (0.90–0.95) | < 10–4 | 90 | 77 | 91 | 75 | 83 | 49 |
| Marker Panel + PSA | 0.94 (0.92–0.97) | < 10–4 | 90 | 81 | 91 | 78 | 85 | 51 |
| Combinedd | 0.95 (0.93–0.97) | < 10–4 | 90 | 87 | 92 | 84 | 88 | 55 |
| Model III: (Benign + VLR/LR + FIR) vs (UIR + HR/VHR + mPC) (Marker number in model = 26) | ||||||||
| 5 clinical factorsc | 0.84 (0.80–0.89) | < 10–4 | 90 | 51 | 91 | 50 | 65 | 36 |
| Marker panel | 0.88 (0.84–0.92) | < 10–4 | 90 | 70 | 93 | 63 | 78 | 49 |
| Marker Panel + PSA | 0.92 (0.92–0.96) | < 10–4 | 90 | 82 | 94 | 74 | 85 | 57 |
| Combinedd | 0.93 (0.90–0.96) | < 10–4 | 90 | 84 | 94 | 76 | 86 | 59 |
| Model GS: (Benign + GS < 7) vs (GS ≥ 7) (Marker number in model = 22) | ||||||||
| 5 clinical factorsc | 0.78 (0.72–0.83) | < 10–4 | 90 | 25 | 82 | 41 | 49 | 16 |
| Marker panel | 0.89 (0.86–0.93) | < 10–4 | 90 | 76 | 93 | 69 | 81 | 48 |
| Marker Panel + PSA | 0.91 (0.88–0.94) | < 10–4 | 90 | 79 | 94 | 71 | 83 | 50 |
| Combinedd | 0.91 (0.88–0.94) | < 10–4 | 90 | 80 | 94 | 72 | 84 | 51 |
P: p value for AUC (null hypothesis: AUC = 0.5)
Sen sensitivity, Spe specificity, NPV negative predictive value, PPV positive predictive value, Bx biopsy, CI confidence interval, GS Gleason score
aSensitivity set at 90% for clinical relevance
bThe percent biopsy avoided was calculated after the cohort was normalized to the original risk group composition of the entire cohort enrolled during the study period
cFive clinical risk factors, including age, PSA value, family history of PC, previous negative biopsy for PC, abnormal DRE
dMetabolite marker panel plus 5 clinical factors. Table S3 (Additional file 1) shows similar statistics at 95% sensitivity