Prostate cancer (PCa) can range from slow-growing tumors that may never cause harm to aggressive cancers requiring immediate treatment. Correctly categorizing a patient into low-risk, intermediate-risk, or high-risk groups is critical because treatments differ substantially: low-risk patients may only need active surveillance, intermediate-risk patients may undergo surgery or radiation, and high-risk patients typically need systemic therapy.
Current standard tools have real limitations. PSA (prostate-specific antigen) blood tests have a high false-positive rate, leading to unnecessary biopsies. Multiparametric MRI is improving but varies significantly between readers and institutions. Prostate biopsy -- still the definitive diagnostic step -- is invasive, uncomfortable, and can miss tumors due to sampling limitations.
There is growing evidence that peripheral blood lymphocytes -- the immune cells circulating in blood -- reflect the immune response to cancer and can carry diagnostic and prognostic information. Lymphocyte subsets including T cells, B cells, and NK cells have been shown to distinguish prostate cancer from benign disease and can predict clinical risk.
This study asks whether a machine learning model built from detailed immune cell profiling of blood samples can accurately classify patients into all three risk categories -- a more clinically useful and nuanced task than the simpler binary classifications attempted in prior studies.
The study screened 2,039 patients with prostate cancer or benign prostate disease from Wuhan Tongji Hospital between 2020 and 2022. After quality control filtering for complete data and adequate imaging, 197 patients were enrolled, including 56 with benign prostate disease. These were classified by European guidelines into 59 low-risk, 48 intermediate-risk, and 90 high-risk cancer patients.
Each patient had 42 clinical characteristics measured, with a particular focus on functional subsets of peripheral lymphocytes analyzed by flow cytometry -- a laboratory technique that identifies and counts specific immune cell types based on protein markers on their surfaces. The measured cell populations included T cells, B cells, natural killer (NK) cells, helper T cells (Th), cytotoxic T cells (Ts), regulatory T cells, and various activated and memory subtypes.
In addition to immune cells, inflammatory proteins called interleukins (IL-1beta, IL-2R, IL-6, IL-8) and tumor necrosis factor-alpha were measured from blood. These cytokines are signaling molecules released during immune activation and can reflect the intensity of the inflammatory response surrounding a tumor.
Key differences between risk groups emerged in 12 features: higher-risk patients tended to have markedly higher PSA levels (averaging 200 ng/mL in high-risk versus 7.5 ng/mL in low-risk patients), elevated alkaline phosphatase and lactate dehydrogenase (markers of bone or cellular stress), higher interleukins, and alterations in the Th/Ts ratio (the balance between helper and cytotoxic T cells).
To identify the most predictive variables and avoid overfitting, LASSO regression was applied to the 41 usable clinical features. This mathematical method penalizes complexity, forcing the model to retain only the features with genuine predictive value. The result was a reduced set of nine key features: Age, Alkaline phosphatase, B cells, Interleukin-1beta, Interleukin-2R, Lactate dehydrogenase, Neutrophil percentage, PSA, and the Th/Ts ratio.
Five different machine learning algorithms were then tested on this nine-feature dataset using 10-fold cross-validation (a technique that trains the model 10 times on different subsets of the data to assess stability): Support Vector Machine (SVM), Decision Tree (DT), Random Forest (RF), XGBoost, and AdaBoost. Each algorithm approaches the classification problem differently.
XGBoost emerged as the best-performing algorithm, achieving an AUC of 0.989 on the training set and 0.842 on the test set. XGBoost is a gradient boosting method that builds many decision trees sequentially, each correcting the errors of the previous one, which makes it particularly effective on structured tabular data.
Despite strong AUC performance, the pure XGBoost model had limited clinical interpretability -- its internal workings are not easily translatable into a clinical workflow. This motivated the development of a nomogram approach that preserves machine learning power while adding transparency.
The researchers developed a novel approach to combine machine learning's predictive power with the practical accessibility of a nomogram -- a visual chart where physicians can input patient values and read off a risk score. Traditional nomograms are built from logistic regression and are easy to use, but less powerful than modern machine learning methods.
The innovation was a Feature Mapping Algorithm (FMA). First, a standard nomogram was built from the ML model's predictions (called an ML nomogram). The FMA then translated the ML model's internal variable weights back into the original clinical variables -- creating a clinic-ML nomogram where the axes are familiar lab measurements rather than abstract machine learning scores.
The final clinic-ML nomogram integrates predictions from four of the five ML algorithms (AdaBoost, Decision Tree, Random Forest, and XGBoost -- all four were identified as independent predictors by logistic regression). A physician uses the nomogram by locating each patient's values for the nine variables on the chart and summing the points to get a total score, which maps to a low-, intermediate-, or high-risk prediction.
Cutoff values of 2.24 and 6.00 on the nomogram score were identified to separate the three risk groups. This creates a practical, step-by-step tool that any clinician can apply using standard blood test results without needing specialized software.
The clinic-ML nomogram outperformed both the standalone clinical nomogram and the best individual ML model (XGBoost) on all key performance metrics. On the test set, it achieved an AUC of 0.864 (versus 0.842 for XGBoost and 0.837 for the clinical nomogram alone), with sensitivity of 0.713 and specificity of 0.869.
The improvement was especially pronounced in the training set (AUC 0.998 vs. 0.897 for the clinical nomogram), demonstrating that the ML models captured patterns the logistic regression approach alone missed. The DeLong statistical test confirmed the differences between the clinic-ML nomogram and other models were statistically significant (p less than 0.05).
Calibration curves showed the clinic-ML nomogram's predicted probabilities closely matched actual observed outcomes, indicating the model is well-calibrated and not just discriminating but also providing accurate probability estimates. This matters clinically: a model that says a patient has a 70% chance of high-risk disease should be correct in roughly 70% of such cases.
Decision Curve Analysis confirmed the clinic-ML nomogram provided greater net clinical benefit than either comparison model across a wide range of decision thresholds, meaning it would be most useful in practice when applied to guide treatment decisions.
The features that best predicted prostate cancer risk reflect both tumor burden and the immune response to it. PSA remains the strongest single predictor, as its levels scale dramatically with disease stage. Elevated alkaline phosphatase and lactate dehydrogenase indicate bone involvement and cellular stress in aggressive disease.
The Th/Ts ratio -- the balance between helper T cells that activate immune responses and cytotoxic T cells that kill targets -- emerged as a significant independent predictor. An elevated Th/Ts ratio in high-risk patients may reflect a shift toward immune regulation and away from effective tumor killing, a pattern consistent with tumor immune escape.
B cell counts were negatively associated with risk (lower B cells correlated with higher-risk disease), and inflammatory markers like Interleukin-2R and IL-6 were elevated in higher-risk groups. IL-6 is a well-established promoter of tumor progression and is known to activate signaling pathways that drive prostate cancer resistance to treatment.
Together, these findings support the concept that prostate cancer of varying risk levels creates distinct, measurable signatures in circulating blood immune cells and inflammatory proteins -- making peripheral blood profiling a biologically meaningful, not merely statistical, predictor.
The most compelling clinical application is as a preoperative screening tool that could help guide biopsy decisions and treatment planning without requiring imaging or tissue sampling. A simple blood draw analyzed by flow cytometry and standard lab tests provides all the inputs needed for the nomogram.
For patients identified as likely low-risk by the nomogram, there may be grounds to defer biopsy or pursue active surveillance rather than immediate treatment. For patients with nomogram scores suggesting high-risk disease, earlier or more aggressive intervention may be warranted, and such patients might be fast-tracked for mpMRI and targeted biopsy.
Notably, the nomogram performs tripartite classification -- low versus intermediate versus high risk -- which is more clinically useful than binary models, since treatment differs between all three groups. Previous blood-based approaches generally only distinguished high-risk from others.
The authors envision this as a complement to, not replacement for, existing diagnostic tools. Integrating immune profiling with MRI imaging data in future iterations could further improve performance, addressing a key limitation of the current blood-only approach.
This study successfully developed and validated a clinic-ML nomogram for three-way risk stratification of prostate cancer using immune and inflammatory blood markers. It represents the largest sample size of its kind for this type of analysis and is the first to combine ML predictions with a feature-mapped nomogram for tripartite risk classification.
The key innovation is the Feature Mapping Algorithm, which bridges the gap between machine learning's predictive power and the clinical usability of a nomogram -- delivering both performance and interpretability rather than sacrificing one for the other.
The main limitation is that all patients came from a single institution in one city, leaving external generalizability unproven. The authors plan a multicenter study to test performance across different hospitals, patient populations, and laboratory protocols.
Future work should also integrate imaging data and explore automation of the flow cytometry analysis to reduce the technical requirements for implementation. If validated externally, this tool could meaningfully improve the accuracy and reduce the invasiveness of prostate cancer risk assessment in clinical practice.