How Well Do AI Risk Estimates Match Real-World Prostate Cancer Outcomes?

Daniel Spratt, MD, describes the core findings of the benchmarking analysis.¹ For 10-year distant metastasis, he says, the MMAI-predicted risk estimates from a real-world commercial cohort were strikingly similar to the outcomes actually observed in STAR-CAP across National Comprehensive Cancer Network (NCCN) low-, intermediate-, and high-risk disease, differing on the order of approximately 1%, with confidence intervals around both estimates. In the data presented, MMAI predicted 10-year distant metastasis rates of 2.4%, 3.4%, and 12.9% for NCCN low-, intermediate-, and high-risk disease, compared with observed rates of 1.0%, 4.0%, and 13.5% in STAR-CAP. Spratt notes the agreement is notable given that some STAR-CAP patients were treated in a more historical era.

He then turns to prostate cancer–specific mortality (PCSM), an outcome he says patients want to understand, particularly with more aggressive disease, because a reduction in metastasis that does not change survival is an important distinction to discuss. Here, too, Spratt reports strong concordance between predicted and observed outcomes. MMAI predicted 10-year PCSM of 1.1%, 1.6%, and 6.7% for NCCN low-, intermediate-, and high-risk disease, vs observed rates of 0.5%, 2.0%, and 8.2%; the analysis also compared PCSM across STAR-CAP stage groups.

Spratt also addresses a common concern. MMAI was initially validated in randomized trials of patients treated with radiotherapy with or without hormone therapy, raising questions about patients who choose surgery. The majority of STAR-CAP patients underwent surgery, he explains, and calibration remained accurate in this mixed surgery and radiation cohort, consistent with the principle that a prognostic tool should perform regardless of the treatment a patient receives.

In the next segment, Spratt walks through how he weighs MMAI results alongside NCCN risk groups, including when the two point in opposite directions.