For decades, prostate-specific antigen (PSA) testing has been one of the most common—and controversial—cancer screening tools in medicine. Nearly 10 million men undergo PSA screening each year in the U.S., yet clinicians and patients are often left interpreting results using blunt thresholds that fail to reflect individual risk, life expectancy, or the likelihood that prostate cancer will ever become life-limiting.
A new prediction model published in Annals of Internal Medicine seeks to change that calculation. Developed and validated with long-term data from over 200,000 men, the tool estimates an individual’s risk of dying from prostate cancer within a specific time frame while considering competing causes of death—a crucial aspect that most existing PSA-based risk calculators tend to overlook.
“This tool uses all of the information that is already available,” said first author Patrick Lewicki, MD, of the department of urology at the University of Michigan. “There are no novel biomarkers, no diagnostic imaging. This is entirely a prediction model.”
The work draws on longitudinal data from the Prostate, Lung, Colorectal, and Ovarian (PLCO) Cancer Screening Trial, which tracked over 33,000 men aged 55 to 74 who had PSA screening, along with an external validation group of nearly 175,000 Veterans Affairs patients in the same age range. By combining PSA levels with factors like age, race, family history, and comorbidities, the researchers created a model that predicts prostate cancer–specific mortality (PCSM) rather than just cancer detection on biopsy.
That distinction is central to the study’s motivation.
“Most of the models in this space are related to prostate cancer detection,” Lewicki explained. “That’s not bad, but as most men get older, they are likely to get prostate cancer. The real question is whether that cancer is going to impact their life expectancy.”
Moving beyond diagnosis as a surrogate endpoint
Current PSA interpretation tools generally estimate the likelihood of finding prostate cancer—or so-called “clinically significant” disease—on biopsy. But biopsy results themselves depend heavily on imaging strategies, sampling methods, and evolving pathology definitions. More importantly, diagnosis is only a surrogate for the outcomes that matter most to patients: metastasis and death.
“Diagnosis as an endpoint has a lot of problems,” Lewicki said. “It doesn’t always mean the same thing depending on how the diagnosis is arrived at—whether by MRI, what the biopsy strategy was, even what the biopsy result was.”
The new model instead predicts the probability of dying from prostate cancer while explicitly adjusting for other-cause mortality. That feature reflects a reality clinicians intuitively recognize but often fail to quantify.
“There’s this phenomenon where patients who are not very healthy—even if they have slightly elevated prostate cancer risk—aren’t dying of prostate cancer because they die of other causes,” Lewicki noted. “We understand that intuition as clinicians, but we probably only act on it at the extremes.”
At the other end of the spectrum are younger, healthier men with long life expectancy whose PSA levels may fall within conventional “normal” ranges but are elevated relative to population baselines.
“The model is meant to provide a convenient individual risk estimate for clinicians and patients alike, to interpret risk in the context of longevity,” Lewicki said.
Time-to-event risk, chosen by the patient
One of the model’s defining features is that it allows users to specify the time horizon of interest—an approach Lewicki describes as essential for shared decision-making.
“If you’re a 70-year-old guy and you get a risk prediction that says you’re going to die of prostate cancer in 25 years, that may not be that bothersome,” he said. “But if you’re 50, that’s definitely meaningful.”
Lewicki emphasizes that user specifies the time point of interest. “If you’re 75, that might be 15 years. If you’re 50, it might be 35 years.”
The researchers envision future iterations that align risk estimates directly with life expectancy—calculating the probability of dying from prostate cancer before a patient would otherwise be expected to die.
“That’s implicitly what patients want,” Lewicki said. “They’re asking about their risk in the context of how long they otherwise expect to live.”
Implications for overtreatment—and screening hesitancy
Overdiagnosis and overtreatment remain persistent problems in prostate cancer, particularly among older men. According to Lewicki, the harms are twofold.
“One is the direct harm to the patient who is overtreated,” he said. “The other is when reports of overdiagnosis are used to weaken the evidentiary basis for prostate cancer screening.”
That dynamic, he noted, can make primary care physicians hesitant to offer PSA testing even to patients who may benefit most.
The authors argue that a mortality-focused risk model could help resolve this tension by aligning screening intensity and follow-up decisions with meaningful outcomes rather than detection alone. In their analysis, the new model outperformed existing PSA interpretation strategies—including fixed cutoffs such as labeling PSA values above 4 ng/mL as “abnormal”—in predicting long-term prostate cancer death.
“We’re providing one estimate,” Lewicki said. “Right now, clinicians are doing a multivariate calculus in their heads—thinking about prostate cancer risk, all-cause mortality, and life expectancy, and trying to make the math work. This tool puts that information together in one place.”
From research tool to clinical workflow
The researchers are actively working to integrate the model into electronic medical records, with the goal of embedding risk estimates directly alongside PSA results.
“Right now, a patient gets a PSA value and a reference range,” Lewicki said. “But the reference ranges are somewhat arbitrary. There is no inherent normal or abnormal level of PSA.”
Ultimately, he envisions a future in which PSA results are accompanied by personalized mortality risk estimates displayed directly in patient portals and clinician inboxes—supporting informed conversations at the point of care.
“Like any other risk estimate, you use it to inform your next steps,” Lewicki said.
If adopted broadly, the approach could represent a shift away from one-size-fits-all screening thresholds toward precision risk assessment—grounded not in diagnosis alone, but in outcomes that matter most to patients.
