Background: Prostate-specific antigen (PSA) kinetics, the change in PSA levels over time, may help predict prostate cancer. Traditionally, PSA velocity (PSAV) is used, with values above 0.35 or 0.75 ng/ml/yr considered abnormal. Machine learning models might improve risk assessment using PSA kinetics.
Objective: The study aimed to enhance the utility of PSA kinetics by developing a generalizable machine learning model.
Design, Setting, and Participants: Data were collected from the PLCO and PCPT trials and a contemporary Australian cohort. PSA data were interpolated using a modified Gaussian process, and a machine learning model based on a two-headed approach with multivariable input was designed using a one-dimensional ResNet18 model.
Outcome Measures and Statistical Analysis: Model performance was compared to PSA levels and PSA velocity using the area under the receiver operator characteristic curve (AUC).
Results and Limitations: The analysis included 10,719 patients. The machine learning model achieved an AUC of 0.886 for diagnosing grade group ≥2, compared to 0.807 for PSA levels and 0.627 for PSA velocity.
Conclusions: Machine learning models significantly improve the diagnostic utility of PSA kinetics for prostate cancer over traditional methods like PSA velocity and thresholds.
Patient Summary: Machine learning algorithms can enhance the accuracy of prostate cancer diagnosis using PSA blood tests, offering better screening without additional costs or tests.
