A machine learning model using prostate-specific antigen (PSA) values predicted the risk of clinically meaningful recurrence in patients with prostate cancer who had undergone definitive radiotherapy, according to findings presented at the American Society for Radiation Oncology (ASTRO) 2026 Annual Meeting (Abstract 287).  

The model outperformed standard approaches to monitoring for disease recurrence following radiotherapy.  

Background

After patients with prostate cancer undergo treatment, options for monitoring disease during follow-up are limited, primarily consisting of PSA testing and the application of the Phoenix criteria. PSA is a dynamic biomarker, and several factors, including those unrelated to prostate cancer, may contribute to rising PSA levels.  

The Phoenix criteria are typically used to assess biochemical recurrence following radiotherapy. They use a standardized threshold for recurrence—a PSA level of 2.0 ng/mL above the nadir—for all patients, making them useful as a clinical endpoint. However, their utility in clinical decision-making is limited, particularly because they do not capture heterogeneous patterns of disease progression, explained lead author Mariana Borras-Osorio, MD, a research fellow in the Department of Radiation Oncology at Mayo Clinic in Rochester, Minnesota.

Dr. Borras-Osorio also noted that recent evidence suggests the Phoenix criteria may delay the detection of clinically meaningful recurrence in the context of newer imaging and treatment modalities for prostate cancer. 

The research team sought to use AI to personalize prostate cancer care following radiotherapy. 

Study and Model Methods

The researchers conducted a retrospective study at a single academic medical center with multiple practice sites. Eligible patients had biopsy-proven prostate cancer and had received radiotherapy with curative intent. Patients were excluded if they had fewer than three post-radiotherapy PSA measurements, less than 3 years of follow-up, or incomplete data.  

The primary endpoint was clinically meaningful recurrence within a 5-year window, defined as biochemical recurrence or clinical progression.  

To train the model, the researchers used post-radiotherapy PSA levels, PSA kinetics, and clinical and treatment data to predict the occurrence of an event of interest within 5 years using a time-anchored binary classification. The model was also evaluated on a per-patient basis, with patients classified as positive if any prediction exceeded the prespecified threshold.  

Data were censored after an event of interest and at the visit level when follow-up was shorter than 5 years and no event had occurred during that period following a given PSA measurement.  

The Extreme Gradient Boosting (XGBoost) model was trained using an 80/20 split of the cohort, fivefold cross-validation, and hyperparameter tuning involving 30 random combinations and prespecified parameters.  

From an initial cohort of 10,480 patients with prostate cancer and 63,258 PSA measurements, the study population was narrowed to 1,742 patients with 13,377 PSA measurements following chart review. Of these, 1,384 patients were included in the training set and 358 in the test set, with demographics well-balanced between the groups. Most patients had T1 or T2 disease and received androgen-deprivation therapy in addition to radiotherapy.  

Study Findings

To assess the model’s performance, Dr. Borras-Osorio explained that the researchers used a PSA control tower with a prespecified threshold of 0.84, selected to target 90% specificity in the cross-validation folds based on published estimates of the specificity of the Phoenix criteria.  

At this threshold, the model achieved a patient-level area under the curve of 0.965 (95% confidence interval [CI] = 0.950–0.980), with a sensitivity of 91% (95% CI = 86%–95%) and a specificity of 90%. In comparison, the Phoenix criteria had a sensitivity of 81% and a specificity of 95%.  

At the visit level, the model demonstrated a calibration slope of 0.998, an intercept of –0.106, and a Brier score of 0.139. “The prediction of the model closely matched the observed events,” Dr. Borras-Osorio noted. In the visit-level decision curve analysis, the model demonstrated a net benefit of 0.133 (95% CI = 0.083–0.179).  

Among the 154 patients with true-positive results, the model achieved a median lead time of 9.7 months before recurrence, compared with 3.3 months using the Phoenix criteria. The PSA level at the first warning was 1.93 ng/mL with the XGBoost model vs 3.26 ng/mL with the Phoenix criteria.  

The researchers also conducted a SHapley Additive exPlanations (SHAP) analysis to explore the model’s interpretability. They found that the main drivers of the predictions were PSA velocity, the most recent PSA value, primary Gleason score, patient risk category, and time since completion of radiotherapy. “This is very important because this points out that the model is not looking at PSA value in isolation, but it is actually considering how fast the PSA is increasing and what is the baseline risk for the patient,” Dr. Borras-Osorio commented.  

The model is currently undergoing external validation and has thus far achieved an area under the curve of 0.858. The research team also hopes to conduct prospective validation and has developed a dashboard that can be integrated into electronic health records.  

DISCLOSURES: The study authors reported no conflicts of interest. 

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.