Utility of PSA density velocity in diagnosing prostate cancer on prostate MR

To assess if PSA density (PSAD) velocity is a predictor of clinically significant prostate cancer (csPCa).

Patients undergoing prostate MR for suspected csPCa between 2017 and 2023 were included. Of these, patients who underwent two or more MR exams within 0.3 to 2 years to each other and pathological confirmation within 1 year after the last MR were identified. PI-RADS score, patient age, no-prior-PCa cohort or active surveillance (AS), PSA, PSAD, PSA-velocity and PSAD-velocity were assessed if they were predictor of csPCa. For univariate analysis, all variables were assessed using student t or chi-square tests. When variables were considered significant, variables were entered into multivariate logistic regression model. To assess whether PSAD-velocity adds diagnostic value beyond established predictors, we compared a base model (age, PI-RADS score, clinical cohort, and current PSAD) with the same model plus PSAD-velocity, using optimism-corrected AUC from 1,000 bootstrap resamples.

23,244 MRs were performed (17,703 for no-prior-PCa, 5,541 for AS). 1,174 patients (mean age: 65.9 years) met the inclusion criteria, and 550 (46.8%) were diagnosed with csPCa. On univariate analysis, PI-RADS score (p<.0001), age (67.0 vs. 64.9 years, p<.0001), AS status (67% vs. 59%, p = .004), PSA (7.5 vs. 6.6 ng/mL, p<.001), PSAD (p<.0001), and PSAD-velocity (0.023 vs. 0.012 ng/mL/mL/year, p = .005) were associated with csPCa, but PSA-velocity (1.4 vs. 1.2 ng/mL/year, p=.38) was not. On multivariable analysis, PI-RADS score (p<.0001), age (p<.0001), AS status (p = .002), PSA (p<.0001), and PSAD (p<.0001) remained significant predictors, but PSAD-velocity was not (p = .13). Adding PSAD-velocity to the base model produced no significant improvement in optimism-corrected AUC (0.7605 vs. 0.7611; ΔAUC = + 0.0006, 95% CI -0.0002 to +0.0019).

Higher PSAD-velocity was associated with presence of csPCa on univariate analysis but not on multivariate analysis, and on fixed-model analysis did not meaningfully improve diagnostic discrimination beyond established predictors.