Developing AI-Driven Personalized Early Detection and Clinical Decision Support for Prostate Cancer
death. Doctors use a blood test called PSA (prostate-specific antigen) to help decide who needs
more testing or treatment, but current rules for interpreting PSA are not perfect. Many guidelines
rely on a single PSA value above a cutoff to decide if a man should have a biopsy. Unfortunately,
this can lead to unnecessary procedures for men with harmless conditions and can also miss
aggressive cancers that need early treatment.
Our project aims to make PSA testing smarter and more personal. Instead of looking at just one
PSA result, we will use artificial intelligence (AI) to study how a man’s PSA changes over time. By
combining these PSA patterns with personal health information like age, family history, and past
treatments, we hope to build a tool that helps doctors make better choices about who needs
further testing and how to monitor men who have already been treated.
This research focuses on two important groups: (1) men who have not yet been diagnosed with
prostate cancer, helping decide whether they should get a biopsy sooner, and (2) men who have
already been treated for prostate cancer, to catch any signs of cancer returning as early as
possible.
Our approach uses health information that is already collected in clinics, so it will be cost-effective
and widely available. By applying advanced computer techniques to PSA testing, we believe we
can help doctors find dangerous cancers earlier and avoid unnecessary procedures for men who
do not need them. This project supports the mission of the Prevent Cancer Foundation by making
cancer screening and follow-up more accurate, personal, and effective, improving outcomes and
peace of mind for men at risk for prostate cancer.
Aim 1: Curate and structure a comprehensive longitudinal PSA and clinical metadata resource: We will
extract and harmonize PSA time-series data from electronic health records, link them to clinical decision and
outcomes such as biopsy results, treatment modality, recurrence, and integrate patient metadata including age, race/ethnicity, and family history. Two cohorts will be defined: a) a screening cohort comprising patients without a prior prostate cancer diagnosis, and b) a surveillance cohort of patients monitored following curative-intent treatment. We will quantify PSA velocity, doubling time, and non-linear temporal trends to generate an enriched feature set for model development.
Aim 2: Develop AI models to characterize PSA trajectory patterns associated with early prostate cancer
detection and improve risk stratification to guide biopsy decisions: We will train a probabilistic, patient-
specific risk model that integrates the full PSA time series plus key covariates (age, race/ethnicity, family history, prostate volume, MRI, medication use) to identify abnormal PSA dynamics that precede biopsy-confirmed cancer for the patients in the screening cohort. These models will be benchmarked against conventional threshold and velocity-based rules. Our goal is to demonstrate that modeling individualized PSA kinetics improves detection of clinically significant cancer and reduces unnecessary biopsies. Models will be designed for real-time use and generalizability across diverse populations.
Aim 3: Predict biochemical recurrence during post-treatment surveillance by modeling PSA trends in
the preclinical window: We will apply similar AI models to patients who have undergone surgery or radiation
therapy in our surveillance cohort. By identifying trajectory patterns that precede biochemical recurrence, we aim to define a preclinical window of risk and support earlier intervention. This will directly inform surveillance
strategies and optimize follow-up intervals.
Ken Chen University of Texas MD Anderson Cancer Center
Merve Dede University of Texas MD Anderson Cancer Center
Yukun Tan University of Texas MD Anderson Cancer Center
Vakul Mohanty University of Texas MD Anderson Cancer Center