Prostate-specific antigen (PSA) is a protein produced by the prostate gland and measured by a simple blood test. Elevated PSA levels can signal prostate cancer, and PSA testing is the most widely used screening tool for the disease. However, PSA is far from a perfect marker -- it can be elevated in men with benign prostatic hyperplasia (BPH), prostatitis, or after physical activity, none of which are cancer.
This lack of specificity means many men with elevated PSA undergo prostate biopsy -- an invasive procedure with its own risks -- only to find no cancer. The false alarm rate has contributed to significant overdiagnosis of slow-growing, clinically insignificant tumors that may never cause harm. A better pre-biopsy risk assessment tool could reduce unnecessary procedures while ensuring that men with true cancer are identified.
Improvements such as measuring the free-to-total PSA ratio (f/tPSA), PSA density relative to prostate volume, and multiparametric MRI have improved diagnostic accuracy, but these approaches remain imperfect, are not always available, and add cost. The search for better, widely accessible markers continues -- particularly for settings without advanced imaging technology.
This study from Wenzhou Medical University in China developed a machine learning model to distinguish prostate cancer from benign prostatic hyperplasia using only routine pre-biopsy blood test results. The approach deliberately avoids expensive or specialized tests -- no MRI, no novel biomarkers, no physical examination parameters -- relying entirely on measurements that are collected as standard clinical practice at virtually any hospital worldwide.
The algorithm used is XGBoost (eXtreme Gradient Boosting), a powerful and widely used machine learning method that builds predictions by combining hundreds of simple decision trees. XGBoost is particularly well-suited for tabular clinical data because it handles missing values, avoids overfitting through regularization, and consistently outperforms simpler models in medical prediction tasks.
A key feature of the study is its use of the SHAP (SHapley Additive exPlanations) framework to make the model interpretable. SHAP assigns each input variable a numerical score indicating how much it increased or decreased the predicted cancer risk for each individual patient, transforming the model from an opaque black box into a transparent, explainable tool that clinicians can reason about.
Data came from the Chinese National Clinical Medical Science Data Center, drawing specifically from the Prostate Tumor Warning Dataset of the Chinese PLA General Hospital (Beijing). After excluding patients with missing values, 1,915 male patients who had undergone prostate biopsy were included: 823 with confirmed BPH (43%) and 1,092 with confirmed prostate cancer (57%).
Each patient had 23 predictor variables recorded: age, body mass index, and 21 blood chemistry values including PSA parameters (total PSA, free PSA, and the free-to-total ratio), electrolytes (sodium, potassium, calcium, chloride, inorganic phosphorus), kidney function markers (creatinine, uric acid), liver and muscle enzymes (alkaline phosphatase, creatine kinase MB isoenzyme, creatine kinase, lactate dehydrogenase), lipid measurements (triglycerides, HDL cholesterol, LDL cholesterol), and proteins (albumin, apolipoprotein A1, apolipoprotein B).
Data were split 70/30 into training (1,340 patients) and test sets (575 patients). Model performance was measured using accuracy, sensitivity, specificity, and area under the ROC curve (AUC). Decision Curve Analysis (DCA) assessed net clinical benefit -- how much a model would help clinical decision-making across different risk thresholds.
The XGBoost model achieved an AUC of 0.82, significantly outperforming all single-variable PSA-based comparators: the free-to-total PSA ratio alone scored 0.75, total PSA scored 0.68, and free PSA scored only 0.61 (all comparisons P less than 0.001). The overall accuracy was 74%, with 80% sensitivity and 71% specificity.
Decision Curve Analysis showed that the XGBoost model provided meaningfully higher net clinical benefit than any PSA-based test across threshold probabilities above 10%. This means that using the machine learning model to guide biopsy decisions would lead to better outcomes -- fewer unnecessary biopsies in men without cancer, more biopsies in men who truly have cancer -- than relying on any individual PSA measurement alone.
The XGBoost result also compares favorably to published results for newer specialized biomarkers: the prostate health index (PHI) achieves an AUC of approximately 0.70 and prostate cancer antigen 3 (PCA3) achieves around 0.73 in published studies. The XGBoost model at 0.82 matches or exceeds these tests, without requiring the extra cost and specialized laboratory equipment they demand.
SHAP analysis confirmed that the free-to-total PSA ratio (f/tPSA) was by far the most influential variable -- a lower f/tPSA ratio indicates that a higher proportion of PSA is bound to proteins in the blood, which is more typical of cancer than of benign enlargement. The identified risk threshold was 0.13: patients with f/tPSA below this level face significantly elevated cancer risk.
Beyond PSA metrics, five biochemical parameters emerged as important contributors: inorganic phosphorus (P) above 1.29 mmol/L was associated with increased cancer risk, consistent with prior research linking dietary phosphorus to prostate cancer risk. Potassium (K) below 4.29 mmol/L also predicted higher risk -- an unexpected finding with no prior published evidence, suggesting a new research direction.
Creatine kinase MB isoenzyme (CKMB) above 11.6 U/L and LDL cholesterol above 3.05 mmol/L were both associated with increased cancer risk, consistent with growing evidence linking metabolic and lipid abnormalities to prostate cancer development. Creatinine showed a nonlinear relationship, with a risk threshold interval of 74.5 to 99.1 micromol/L -- a novel finding not previously reported.
A recurring criticism of machine learning in medicine is the black box problem: the model gives a prediction but cannot explain why. Clinicians are understandably reluctant to act on predictions they cannot interrogate, especially for decisions as consequential as whether to perform a biopsy. SHAP addresses this by decomposing each prediction into the individual contributions of each input variable, making the reasoning transparent.
For this prostate cancer model, SHAP does more than just rank feature importance -- it reveals the direction and threshold of each variable's effect. Clinicians can see not just that inorganic phosphorus matters, but that risk rises above 1.29 mmol/L, and not just that creatinine is relevant, but that its relationship with cancer risk is nonlinear and specific to a window of values. This kind of actionable threshold information translates AI insights into clinical decision rules.
The authors are careful to note that SHAP values describe patterns within the model and dataset -- they do not prove causation. The finding that low potassium is associated with prostate cancer risk, for example, is a correlation that requires biological validation before potassium can be considered a true cancer biomarker. SHAP identifies candidates for further investigation rather than established facts.
A central argument of this paper is that prostate cancer diagnostic tools based on MRI or novel biomarkers, while effective, are impractical for widespread use in resource-limited settings. In rural China and other developing regions, advanced MRI scanners are scarce, expensive, and require trained specialist readers. Standard blood chemistry analyzers, by contrast, are available at virtually every hospital and clinic.
The XGBoost model described here could be implemented as a simple software tool that takes standard laboratory values and outputs a cancer probability score. Unlike PSA alone, which requires physician judgment and additional tests to interpret, the model provides a single integrated risk estimate that directly informs the biopsy decision. This could meaningfully improve cancer detection rates in settings where sophisticated imaging is unavailable.
The model's deliberate exclusion of digital rectal examination (DRE) results and ultrasound findings -- subjective parameters with high inter-observer variability -- is an additional strength for broad applicability. By relying only on objective, automated laboratory measurements, the model produces reproducible results regardless of operator skill or experience.