Surgical prediction models using NSQIP data have significantly lower discrimination ability compared to those using non-NSQIP data, particularly for infection outcomes.

• Mean area under the receiver-operating characteristic curve (AUROC) was 0.71 for NSQIP models versus 0.80 for non-NSQIP models (p<0.001).
• The discrepancy in infection prediction was striking, with a median AUROC of 0.68 for NSQIP versus 0.98 for non-NSQIP (p<0.001).

Surgeons should consider the source of prediction models when counseling patients on infection risks, as reliance on NSQIP data may underestimate risk for certain outcomes.