KEY POINTS
- PREDICT is a machine-learning risk model developed using 73,203 prostate cancer patients diagnosed in the SEER database between 2018 and 2022. It incorporates age, PSA, primary and secondary Gleason patterns, positive and examined biopsy cores, and clinical T stage to estimate pretreatment nodal and distant dissemination risk. Nodal predictions were standardized to a hypothetical 15-node dissection. External validation used four independent surgical cohorts comprising 1,923 patients, including 106 with pathologically positive lymph nodes.
- PREDICT outperformed the Roach formula in the SEER development population. Among 17,164 patients undergoing surgery and lymph node dissection, including 1,595 node-positive patients, the cross-validated AUC for nodal involvement was 0.793 versus 0.749 with Roach, a paired difference of +0.043 (95% CI +0.035 to +0.052). Performance was comparable to the MSKCC/Kattan nomogram, with an AUC difference of +0.001 (95% CI −0.003 to +0.006) in the relevant subset.
- The discrimination advantage did not persist in external validation. PREDICT achieved AUCs of 0.724, 0.799, 0.892, and 0.765 in the Rochester, Seoul, MSKCC, and San Raffaele cohorts, respectively. The pooled paired AUC difference versus Roach was −0.006 (95% CI −0.046 to +0.035), indicating no demonstrated external improvement. Calibration also varied substantially, with observed-to-expected nodal event ratios ranging from 0.44 to 1.72, limiting transportability of absolute risk estimates.
- A lower PREDICT threshold selected fewer patients for hypothetical pelvic irradiation in SEER. A cutoff of 15.3% (95% CI 14.4–16.2%) matched the 71.2% sensitivity of the POP-RT eligibility criterion of Roach risk ≥20%, while selecting 7.0 fewer pelvic fields per 100 patients (95% CI 5.0–8.5). However, this reduction was not reproduced externally. At a fixed 15% PREDICT threshold, 11.9% of external-cohort patients would receive pelvic RT versus 11.1% using the POP-RT criterion.
- Observed nodal involvement increased consistently across predicted risk categories. In SEER, pathological node positivity rose from 1.6% at PREDICT risk below 5% to 25.1% at risk ≥20%; corresponding external-cohort rates were 1.8% and 27.9%. In a secondary analysis of 8,106 SEER patients approximating NRG/RTOG 0924 eligibility criteria, 56.3% had a PREDICT risk below 15%, with node positivity of 5.5% versus 18.7% among those above the threshold.
- PREDICT may complement PSMA PET, but evidence for combined selection remains exploratory. In the San Raffaele cohort, 7/41 patients (17.1%) with negative PSMA PET nevertheless had positive nodes at dissection. A Bayesian update combining modeled nodal risk with assumed PET performance achieved AUC 0.868, compared with approximately 0.758 for the nodal model alone and 0.766 for PET alone. This analysis included only 62 patients and assumed fixed PET sensitivity and specificity.
- The model predicts pathological dissemination, not benefit from elective pelvic irradiation. Development and validation relied predominantly on surgical populations, and the external cohorts tested different predictor-specific model versions. Neither the proposed 15% threshold nor the apparent reduction in pelvic treatment fields was prospectively validated, and no clinical outcomes were available to determine whether PREDICT-guided irradiation improves disease control or reduces toxicity.
CLINICAL TAKEAWAY
PREDICT provides more individualized pretreatment risk stratification than the historical Roach formula and demonstrated better nodal discrimination in a large SEER cohort. However, external validation did not confirm superiority, and the proposed 15% threshold cannot currently be used as a universal criterion for elective pelvic irradiation. Prospective evaluation after modern staging, particularly PSMA PET, is needed before the model can guide treatment-field selection.