A novel prognostic model that predicts the risk of prostate cancer–specific mortality—rather than merely cancer diagnosis—following prostate-specific antigen screening may have outperformed an established biopsy risk calculator in both internal and external validation cohorts.
The model, developed from the screening arm of a large cancer trial and validated in a Veterans Affairs population totaling more than 200,000 patients, achieved an area under the receiver operating characteristic curve (AUC) of 78% for 20-year mortality prediction in external validation vs 75% for the comparator model.
In the study, researchers addressed a gap in clinical decision-making tools for the estimated 10 million prostate-specific antigen (PSA) tests performed annually. Existing risk calculators are capable of predicting the risk for prostate cancer on biopsy but don't incorporate time-to-event endpoints or account for competing risk for other-cause mortality—a limitation given prostate cancer’s typically prolonged natural history. As the researchers noted, no existing prediction model calculates risk for prostate cancer–specific mortality from the point of prostate cancer screening while adjusting for the competing risk for other-cause mortality and allowing the end user to specify a time point of relevance.
Study Design and Population
The researchers developed their prognostic model using the screening group of the PLCO Cancer Screening Trial, which enrolled patients aged 55 to 74 years from 1993 to 2001. Following the exclusion of 5,001 patients with missing covariate values, the development cohort comprised 33,339 patients. External validation was performed in 174,787 patients receiving care in the Veterans Affairs (VA) Healthcare System who underwent PSA testing from 2002 to 2006. Survival follow-up was updated through 2022 in both cohorts, yielding median follow-up of 25 years in the development cohort and 18 years in the validation cohort. A total of 501 patients in the PLCO cohort and 1,933 in the VA cohort died of prostate cancer, noted lead study author Patrick Lewicki, MD, MS, of the Department of Urology at the University of Michigan, and colleagues.
Model Architecture
The researchers fit two Weibull models: one for prostate cancer–specific mortality and one for other-cause mortality. Predictors for the prostate cancer–specific mortality component included PSA level (modeled as a cubic B-spline with one internal knot at the median), age, race, and family history of prostate cancer in a first-degree relative. The PSA value was truncated at 100 ng/mL. The other-cause mortality component incorporated age; body mass index; smoking status (never, former, or current); and presence or absence of stroke, diabetes mellitus, or hypertension diagnoses, along with interaction terms between age and each comorbidity. Prostate cancer–specific mortality was modeled by treating other-cause mortality as a competing risk, with the probability of a patient experiencing prostate cancer–specific mortality prior to a specified time calculated using cause-specific hazard functions.
Notably, the researchers omitted digital rectal examination from the model despite its availability in the PLCO data and inclusion in contemporary prediction models, citing “evidence of its limited efficacy in the screening setting and its omission from the U.S. Preventive Services Task Force guideline on [prostate cancer] screening.” PSA velocity was similarly excluded given “its uncertain role in a screening or diagnostic setting.”
Discrimination and Calibration
In the development cohort, time-dependent AUC at 29.5 years from screening was 67% for the PLCO model compared with 64% for the Prostate Biopsy Collaborative Group (PBCG) model. The PLCO model demonstrated progressively higher discrimination at earlier time points, with AUC values of 91% at 5 years, 80% at 10 years, and 77% at 15 years. In the external validation cohort, the AUC was 78% at 20 years for the PLCO model vs 75% for the PBCG model. Among patients in the highest decile of PBCG risk, the PLCO model outperformed the PBCG model in discrimination for 29.5-year prostate cancer–specific mortality in the development cohort (64% vs 58%) and 20-year prostate cancer–specific mortality in external validation (72% vs 70%).
The researchers also evaluated the model’s ability to predict prostate cancer–specific mortality prior to age 85 years—chosen to represent “an optimistic life expectancy for healthy screening-aged male patients.” Discrimination for this endpoint was better in the PLCO model compared with in the PBCG model (71% vs 64%). This analysis addressed the clinical reality that traditional time-to-event predictions aren't equally meaningful for all patients. As the researchers explained, “a 74-year-old man may not be bothered by being at high risk of prostate cancer–specific mortality at age 94 years, whereas a 55-year-old man is likely to care about dying of [prostate cancer] before age 75 years.”
Clinical Utility
Decision curve analysis for the risk of prostate cancer–specific mortality prior to age 85 years demonstrated a net benefit across a range of clinically relevant thresholds compared with strategies of discontinuing screening in all patients, continuing in all patients, and continuing based on individual or age-based PSA cutoffs. The researchers proposed a threshold of 0.5% risk for prostate cancer–specific mortality prior to age 85 years as a clinically relevant decision point for screening discontinuation. “This threshold identifies the 17% of patients experiencing 4% of the observed prostate cancer–specific mortality before age 85 years,” stated the study authors.
The median predicted risk of prostate cancer–specific mortality at 30 years from screening in the development cohort was 2%, while the median predicted risk at 20 years in the external validation cohort was 1%.
Life Expectancy and Cancer Risk Interaction
Stratification by decile of prostate cancer risk (derived from PBCG estimates) and quartile of life expectancy (derived from the other-cause mortality component of the PLCO model) revealed that the association between life expectancy and prostate cancer–specific mortality was strongest among patients with high prostate cancer risk. “The lowest life expectancy quartile identifies a group of patients at ‘high risk’ for [prostate cancer] diagnosis whose observed prostate cancer–specific mortality rate is only modestly elevated compared with the general population,” the study authors noted. Among patients in the highest PBCG risk decile and shortest life expectancy quartile, observed 30-year prostate cancer–specific mortality was 9%, whereas those in the highest risk decile with the longest life expectancy experienced 4% prostate cancer–specific mortality over the same period.
The comparable performance of the PLCO and PBCG models in the development cohort was attributed to the overall low comorbidity burden expected in a clinical trial population, while the modest outperformance in the external validation cohort may reflect that population’s higher comorbidity burden. Additionally, allowing the end user to specify a time-to-event of interest permits prediction individualized to a relevant time frame. “Most contemporary models, including the PBCG model, handle age as a positive term; increasing age leads to increasing risk. This is counter to the intuition that a 55-year-old man with a PSA level of 4.0 ng/mL is at risk for a greater number of life-years lost to [prostate cancer] than a 75-year-old man with the same PSA level.”
Clinical Applications
The researchers envisioned several applications for the risk calculator. Physicians ordering PSA tests can use calculated risk estimates to calibrate their clinical judgment, noting that “a PSA value of 3.6 ng/dL may at first suggest that the patient is at high risk for harboring [prostate cancer] (more so if accompanied by a ‘warning sign’ or an ‘abnormal/elevated flag’), but how does it compare with a positive colorectal or lung cancer screening result?” The model also enables comparison of current patients’ risk with that of index patients for whom screening decisions are clear, and institutional thresholds could be amended to incorporate risk rather than PSA level alone.
Limitations
The researchers acknowledged several limitations. PLCO trial patients were treated under a more aggressive management paradigm, with more definitive treatment for low-risk patients, “which may hamper the model’s contemporary calibration (risk may be higher than calculated) but also offers advantages for model discrimination.” Just 4% of the development cohort identified as non-Hispanic Black patients compared with 16% of those in the external validation cohort. Family history data were unavailable in the VA validation population; missing smoking status data were imputed to “nonsmoker,” which “would bias against our model’s performance by making the other-cause mortality half of our equation less informative.” The model may not be generalizable to more contemporary PSA screening practices given the periods studied.
Baseline Characteristics
In the development cohort, the median age at first PSA test was 62 years, 4% of the patients self-reported non-Hispanic Black race, 8% had a family history of prostate cancer, and the median PSA level was 1.13 ng/dL. Diabetes was present at screening in 9% of the patients, hypertension in 33%, and stroke history in 3%. In the external validation cohort, the median age was 62 years, 16% self-reported non-Hispanic Black race, and the median PSA level was 0.97 ng/dL. Current smoking status was more prevalent (37% vs 11%), as were diabetes (19% vs 9%) and hypertension (41% vs 33%).
“PSA screening for [prostate cancer] will reach its greatest efficacy when conducted in a risk-adjusted manner,” The study authors concluded. “To date, few tools exist to implement risk-adjusted screening. We designed and externally validated a prediction model for risk stratification of patients at the point of PSA screening. Future work will determine the clinical impact of integrating this tool within primary care workflows.”
The study received no external funding. Dr. Lewicki is supported by the National Cancer Institute. Disclosure forms are available with the article online.
Source: Annals of Internal Medicine
