Prostate-specific antigen (PSA) is the most widely used blood test for prostate cancer screening, but it has a fundamental limitation: PSA is organ-specific rather than cancer-specific. Levels can be elevated by benign prostate enlargement, prostatitis, or other non-cancerous conditions, leading to a large number of unnecessary biopsies.
At the standard cutoff of 4 ng/mL, PSA sensitivity is only around 50%, meaning that roughly half of all prostate cancers are missed. The low specificity generates even more concern: fewer than 40% of men with elevated PSA will actually have prostate cancer on biopsy, exposing many to an invasive procedure without clinical benefit.
Derived PSA metrics such as PSA density (PSA divided by prostate volume) and PSA velocity (rate of PSA change over time) were developed to improve on raw PSA accuracy, but neither achieves sufficient diagnostic performance on its own. Multiple national guidelines recommend annual PSA testing, resulting in many patients accumulating two or more consecutive PSA measurements over time before any biopsy decision is made.
These serial PSA values represent a richer longitudinal signal than a single measurement, yet conventional approaches use only simple cutoffs rather than exploiting the full temporal pattern of PSA change. Machine learning methods (MLMs) that can learn complex, nonlinear relationships from multiple input variables offer a promising path toward extracting more diagnostic information from routine PSA test series.
Data from 512 patients who underwent prostate biopsy at Hitachi General Hospital were analyzed retrospectively. All patients had two consecutive annual PSA measurements before biopsy, and 304 patients had three annual measurements available. The outcome was prostate cancer diagnosis on first biopsy (positive in 37.7% of patients), with subsequent biopsies also analyzed.
Three supervised machine learning algorithms were evaluated: artificial neural networks (ANNs), support vector machines (SVM), and random forest (RF). Input variables included patient age, PSA-level statistics derived from the serial measurements (maximum, minimum, median, mean, and variance), prostate volume measured by ultrasound, and white blood cell (WBC) count in urinalysis.
WBC count in urinalysis was deliberately included to help the models distinguish true cancer-related PSA elevation from PSA elevation caused by prostatitis or urinary infection -- a systematic attempt to reduce false positives. Digital rectal examination findings were intentionally excluded to avoid subjective, examiner-dependent input.
All three models were trained using 10-fold cross-validation with grid search hyperparameter optimization, ensuring that model performance was evaluated on held-out data and that the best parameter configuration was systematically identified for each algorithm without overfitting.
An artificial neural network (ANN) -- specifically a multilayer perceptron -- passes input data through one or more hidden layers of weighted nonlinear functions. Training iteratively adjusts these weights to minimize prediction error. Key tuning parameters include the number of nodes per hidden layer and the activation function type.
A support vector machine (SVM) finds the decision boundary (hyperplane) that maximally separates cancerous from non-cancerous cases in the feature space. Nonlinear kernels (such as radial basis functions) allow the boundary to take complex curved shapes. The penalty parameter C controls the trade-off between margin width and classification error, while gamma controls the influence radius of each training point.
A random forest (RF) builds an ensemble of many independent decision trees, each trained on a random subset of the data and features. Individual trees tend to overfit, but averaging their predictions cancels out idiosyncratic errors. This ensemble averaging makes RF inherently more robust to overfitting than single-tree methods, an important property when the number of input variables is large relative to the sample size.
Model performance was assessed using four metrics: AUC (area under the ROC curve, measuring overall discriminatory ability), sensitivity (proportion of cancers correctly identified), specificity (proportion of non-cancers correctly excluded), and accuracy (overall proportion of correct predictions).
Using two annual PSA measurements, all three machine learning models substantially outperformed conventional PSA-based parameters. ANNs achieved the highest AUC at 0.69, followed by RF at 0.64 and SVM at 0.63. By comparison, the AUCs for PSA level, PSA density, and PSA velocity were only 0.53, 0.41, and 0.55, respectively -- all close to the threshold of random prediction.
Accuracy was similarly superior: all three MLMs reached 71.6% to 72.1% accuracy, compared to only 39.1% for PSA level, 49.7% for PSA density, and 54.9% for PSA velocity. The MLMs also showed better sensitivity (56.4% to 66.7%) and specificity (56.2% to 76.6%) than conventional PSA-related parameters.
When PSA density and PSA velocity were added as additional input variables to the MLMs, model performance did not further improve. This suggests that the machine learning algorithms had already learned the underlying patterns that make PSA density and velocity informative -- effectively internalizing these derived metrics from the raw PSA, age, and volume inputs already provided.
Using three annual PSA measurements instead of two yielded slightly higher AUCs (0.70 to 0.71), but the improvement was modest and not clearly clinically meaningful, suggesting that PSA data from the past two years captures most of the available predictive signal and additional historical measurements add diminishing value.
The fundamental advantage of machine learning over simple PSA cutoffs is the ability to learn nonlinear, multivariable patterns. A patient's cancer risk is not captured by any single PSA value -- it depends on the pattern of change over time, the size of the prostate, the patient's age, and signals like urinary inflammation. MLMs can weight and combine these factors in ways that simple formulas cannot.
The incorporation of prostate volume (enabling implicit PSA density calculation) and WBC count (flagging inflammation) as inputs allowed the models to account for the two major confounders of PSA specificity. Excluding digital rectal examination results -- which are subjective and examiner-dependent -- made the model more reproducible and suitable for standardized screening applications.
Among the three algorithms, ANN showed the highest AUC in the two-measurement dataset, while SVM edged slightly ahead with three measurements. The authors note that RF may be preferable in practice because its ensemble averaging makes it less prone to overfitting when variable counts are high relative to sample size -- an important consideration for deployment in new patient populations.
A key limitation is the modest overall AUC range of 0.63 to 0.71, which reflects the inherent difficulty of the prediction task: even with machine learning, predicting prostate cancer from PSA-based data alone is challenging. The small sample size (512 patients) and retrospective single-institution design also limit the generalizability of the findings and require external validation before clinical deployment.
The practical value of these models lies in improving the biopsy decision: rather than applying a simple PSA cutoff, a machine learning model could help clinicians decide whether a patient with mildly elevated PSA should proceed directly to biopsy or continue with watchful PSA surveillance, reducing unnecessary procedures while catching more cancers earlier.
National guidelines in Japan (JUA) and internationally (NCCN, European Association of Urology) already recommend serial annual PSA testing for men in intermediate PSA ranges, meaning that the two-year PSA history required by these models is routinely available in clinical practice -- there is no additional data collection burden for implementation.
The finding that three annual PSA values add little over two suggests that current guideline-recommended screening intervals of 1 to 2 years are appropriate, and that extending to longer historical PSA series is unlikely to improve prediction. This also means the models could be deployed even for patients who are relatively new to PSA screening.
Future directions include prospective validation in larger and more diverse populations, incorporation of additional features such as free PSA ratio or genetic risk scores, and development of individualized rescreening interval recommendations that reduce the psychological burden and cost of annual testing for low-risk men.