Prostate biopsy remains the definitive test for confirming prostate cancer, but it carries real risks: bleeding, infection, pain, urinary retention, and psychological distress. A critical challenge is that the standard screening tool -- the PSA (prostate-specific antigen) blood test -- has poor specificity. PSA can be elevated due to benign enlargement, inflammation, or other non-cancerous conditions, leading to biopsies in men who do not have cancer.
Multiparametric MRI (mpMRI) with PI-RADS scoring has significantly improved risk stratification, and studies show that about 30% of low-risk mpMRI cases can safely avoid biopsy. However, mpMRI still misses approximately 15% of clinically significant cancers, and interpretation varies between radiologists, limiting its reliability as a standalone decision tool.
Novel molecular tests -- including the Prostate Health Index, 4K score, PCA3, and SelectMDx -- have been developed to improve biopsy decision-making, but their high cost and suboptimal accuracy have limited widespread adoption. Simpler, more affordable tools using routinely collected data are needed.
An underexplored resource is the routine blood count, which yields inflammatory markers like the neutrophil-to-lymphocyte ratio (NLR), lymphocyte-to-monocyte ratio (LMR), and systemic immune-inflammation index (SII). These indices reflect systemic immune responses increasingly associated with cancer development and progression, and they are available at essentially no additional cost since blood counts are already standard pre-biopsy testing.
This retrospective study analyzed records from 244 male patients who underwent prostate biopsy at Eskisehir City Hospital in Turkey between 2019 and 2023. All patients had undergone mpMRI with PI-RADS scoring and had PSA and complete blood count measurements within one month of biopsy.
Strict exclusion criteria removed patients with prior biopsies, endoscopic procedures, 5-alpha-reductase inhibitor use (which artificially lowers PSA), pre-existing malignancies, autoimmune diseases, or significant comorbidities like diabetes, cardiovascular disease, or obesity -- all of which could confound inflammatory markers. This resulted in a clean cohort of 244 patients.
Biopsy approach varied by PI-RADS score: patients with PI-RADS 1-2 received a standard 12-core transrectal systematic biopsy, while those with PI-RADS 3-5 received an additional cognitive fusion biopsy targeting the MRI-identified suspicious area. Pathological results classified patients as either Group 1 (benign, n=171, 70.1%) or Group 2 (prostate cancer, n=73, 29.9%).
Variables collected included age, prostate volume, total and free PSA, the free/total PSA ratio (lower ratios indicate higher cancer risk), PSA density, PI-RADS score, and blood count parameters (neutrophils, lymphocytes, monocytes, platelets). Calculated inflammatory indices included NLR, LMR, platelet-to-lymphocyte ratio (PLR), and SII (neutrophils x platelets / lymphocytes).
Five machine learning algorithms were compared: ExtraTrees classifier, Light Gradient-Boosting Machine (LGBM), eXtreme Gradient Boosting (XGB), Logistic Regression, and Random Forest. This diverse selection spans classical statistical models, ensemble tree methods, and gradient boosting approaches, enabling a comprehensive comparison.
The dataset was split 80/20 for training and testing, and cross-validation was applied to assess generalizability and mitigate overfitting. Model performance was evaluated on accuracy, sensitivity, specificity, and AUC-ROC (area under the receiver operating characteristic curve).
Feature importance was assessed using the permutation feature importance method, which measures how much a model's accuracy drops when the values of each input variable are randomly shuffled. This isolates each feature's individual contribution to predictive performance without assuming linear relationships.
The LGBM classifier achieved the best performance with an accuracy of 81.6%, AUC-ROC of 78.4%, sensitivity of 66.7%, and specificity of 88.2%. The high specificity is particularly meaningful clinically: it means 88.2% of men without prostate cancer were correctly identified as not needing biopsy.
Feature importance analysis ranked the free/total PSA ratio as the most influential predictor, followed by age and platelet count. PSA density and total PSA ranked fourth and fifth. The PI-RADS score ranked sixth -- noteworthy because it is the current clinical standard but ranked lower than PSA-derived variables and platelet count in this model.
Men with prostate cancer (Group 2) were significantly older, had smaller prostate volumes, higher PI-RADS scores, higher total PSA, lower free/total PSA ratios, higher PSA density, and higher monocyte counts than men with benign findings (Group 1). Most individual inflammatory markers (NLR, PLR, SII) did not differ significantly between groups on univariate analysis, but contributed to the ML model through complex interactions.
Among the five algorithms, logistic regression achieved competitive accuracy (79.6%) but a lower AUC (74.1%), while tree ensemble methods (ExtraTrees, Random Forest, XGB) performed in a middle range. The LGBM's advantage likely reflects its efficiency in capturing non-linear interactions among the mixed input features.
The biological rationale for including inflammatory markers lies in the well-established role of the immune system in cancer development. Neutrophils and lymphocytes are both participants in tumor immunity: elevated NLR, which reflects a shift toward pro-tumor inflammation and reduced anti-tumor immune surveillance, has been associated with poor prognosis in multiple cancers including prostate cancer.
The systemic immune-inflammation index (SII), calculated as neutrophils x platelets divided by lymphocytes, captures a broader picture of the inflammatory-immune balance. Platelets can promote tumor growth and metastasis by releasing growth factors and shielding circulating cancer cells from immune detection, making platelet count a potentially informative oncological signal.
Interestingly, platelet count ranked among the top three most important features in this model despite not showing a statistically significant difference between groups on standard univariate analysis -- highlighting how machine learning can detect complex interaction effects that traditional statistical comparisons miss.
This study represents one of the few published investigations combining hematologic inflammatory parameters with PSA and MRI findings in an ML biopsy prediction model, building on earlier work by Chen et al. and the more complex 36-feature PCAIDS system. The use of only routine, universally available clinical variables makes the model potentially easier to implement than systems requiring specialty biomarkers.
All variables used in this model -- PSA, prostate volume, PI-RADS score, and complete blood count -- are already collected in routine urological practice before a biopsy decision. The model does not require any additional testing, making it immediately practical for real-world clinical deployment as a decision-support tool.
A model with 88.2% specificity means that for every 100 men without prostate cancer who undergo evaluation, about 88 could be correctly identified and spared from biopsy. Given that roughly 70% of biopsied patients in this cohort had benign pathology, the potential reduction in unnecessary procedures is substantial.
The sensitivity of 66.7% -- meaning the model correctly detected about two-thirds of true cancer cases -- represents a current limitation: a clinically deployable tool would ideally need higher sensitivity to avoid missing significant cancers. Future multi-feature models incorporating additional biomarkers or higher-resolution imaging may address this gap.
This study demonstrates that a machine learning model combining PSA metrics, mpMRI findings, and routine blood count inflammatory parameters can predict prostate biopsy outcomes with meaningful accuracy, achieving an AUC of 78.4% using a simple, readily available feature set.
Key limitations include the single-center retrospective design with only Turkish patients, the lack of external validation, and the use of systematic biopsy for low PI-RADS cases (which may miss some cancers compared to MRI-targeted biopsy). These factors limit the generalizability of current findings.
Future research should pursue prospective multicenter validation, standardized MRI protocols, and broader demographic inclusion. Incorporating additional biomarkers, genetic data, or more detailed imaging features in future model iterations could further improve both sensitivity and specificity, moving toward clinically deployable tools that reduce the burden of unnecessary prostate biopsies globally.