Development and Internal Validation of a Prediction Model For Biochemical Recurrence Following Radical Prostatectomy

Summary

This retrospective study developed and internally validated a predictive model for biochemical recurrence after radical prostatectomy in 240 patients. Pathological stage, lymph node metastasis, positron emission tomography–positive lesions, and SUV_max were independent predictors. The model demonstrated strong discrimination, calibration, and clinical utility for postoperative risk stratification.

Abstract

This study aimed to develop and validate a predictive model for biochemical recurrence (BCR) after radical prostatectomy in patients with prostate cancer, incorporating clinical, pathological, inflammatory, and 18F-PSMA-1007 positron emission tomography/computed tomography (PET/CT) imaging parameters. This retrospective study included 240 patients with histopathologically confirmed prostate adenocarcinoma who underwent radical prostatectomy between June 2022 and July 2025. BCR was defined as a postoperative serum prostate-specific antigen (≥0.2 ng/mL). Preoperative clinical variables, systemic immune-inflammation index (SII), PET/CT lesion status, and maximum standardized uptake value (SUV_max) were collected along with postoperative pathological staging. Patients were divided into BCR (n = 64) and non-BCR (n = 176) groups. Univariate and multivariate logistic regression analyses identified independent predictors of BCR. A multivariable predictive model was developed and internally validated using a 70/30 training–validation split. Model performance was evaluated using receiver operating characteristic curves, area under the curve (AUC), calibration, and decision curve analysis. Patients with BCR showed more advanced pathological stage (pT3–4: 96.9% vs. 52.3%, p < 0.001), higher lymph node metastasis rates (59.4% vs. 29.5%, p < 0.001), higher SII (682.45 ± 118.23 vs. 637.08 ± 92.40, p = 0.002), and higher SUV_max values (6.59 ± 1.34 vs. 4.92 ± 1.49, p < 0.001). Multivariate analysis identified pathological stage (OR = 36.814, p < 0.001), lymph node status (OR = 7.286, p < 0.001), SUV_max (OR = 2.732, p < 0.001), and PET-positive lesions (OR = 27.929, p < 0.001) as independent predictors of BCR. The combined predictive model achieved excellent discrimination in training (AUC = 0.952) and validation cohorts (AUC = 0.927). Calibration curves showed agreement between predicted and observed outcomes, and decision curve analysis demonstrated superior net clinical benefit compared with treat-all and treat-none strategies. The integrated model demonstrates excellent discrimination and clinical utility, supporting individualized postoperative risk stratification.

Introduction

Prostate cancer remains one of the most common malignancies affecting men worldwide, and radical prostatectomy represents a key treatment option for patients with localized disease^1,2. Despite improvements in surgical techniques and patient selection, biochemical recurrence (BCR) occurs in approximately 20%–40% of cases, indicating persistent or recurrent disease and increased risks of metastasis and mortality^2,3. Accurate prediction of BCR after prostatectomy remains challenging, necessitating more effective prognostic tools that combine clinical, immunological, and advanced imaging parameters to guide postoperative surveillance and adjuvant treatments.

Inflammation and immune responses play essential roles in tumor progression and recurrence. The systemic immune-inflammation index (SII), a novel inflammatory biomarker calculated from peripheral neutrophil, lymphocyte, and platelet counts, has demonstrated prognostic significance in various cancers^4,5,6. Additionally, 18F-PSMA-1007 positron emission tomography/computed tomography (PET/CT), a prostate-specific membrane antigen-based imaging modality, has shown promise in detecting recurrent disease earlier than conventional imaging methods^7,8, thereby potentially enhancing postoperative prognostication. This integrative approach extends beyond conventional clinicopathological models by incorporating both systemic inflammatory biomarkers and advanced molecular imaging parameters. Unlike established risk stratification tools such as the CAPRA-S score or D’Amico classification, which rely primarily on baseline clinicopathological parameters, the proposed approach integrates dynamic systemic inflammatory indices with functional molecular imaging findings.

However, the clinical value of integrating inflammatory markers such as SII, pathological tumor staging, and advanced imaging modalities, including 18F-PSMA-1007 PET/CT, in predicting prostate cancer recurrence remains unclear. Therefore, this study aimed to construct and internally validate a comprehensive predictive model for BCR following radical prostatectomy by combining these factors to improve risk stratification, optimize postoperative patient management, and facilitate early intervention strategies. Clinically, this model is intended to assist clinicians in tailoring personalized surveillance protocols and identifying high-risk candidates who may benefit from early adjuvant interventions or intensified follow-up.

Protocol

This study protocol was reviewed and approved by the Ethics Committee of First Affiliated Hospital of Xinjiang Medical University in accordance with the Declaration of Helsinki (Approval No.: 230608-7). Informed consent was waived for this retrospective study due to the exclusive use of de-identified patient data, which posed no potential harm or impact on patient care.

Study Design and Patient Selection
This retrospective study included 240 patients with histopathologically confirmed prostate adenocarcinoma who underwent radical prostatectomy between June 2022 and July 2025. Patients were identified through a systematic search of the electronic medical record system at the First Affiliated Hospital of Xinjiang Medical University. A consecutive sampling method was employed to minimize selection bias. Patients were categorized into a BCR group (n = 64) and a non-BCR group (n = 176) based on postoperative serum prostate-specific antigen (PSA) levels, with BCR defined as a serum PSA level of ≥0.2 ng/mL confirmed by two consecutive measurements. The inclusion criteria were as follows: (1) histopathologically confirmed prostate adenocarcinoma; (2) radical prostatectomy performed; and (3) availability of complete clinical, pathological, and imaging data. These criteria were independently applied by two experienced researchers (Yubin Li and Qizhou Zhang). Any disagreements regarding patient eligibility were resolved through discussion or consultation with a third senior investigator. The exclusion criteria were: (1) Gleason score of <7; (2) receipt of preoperative hormonal therapy, chemotherapy, or radiotherapy, or incomplete medical records; and (3) receipt of neoadjuvant therapy, including androgen deprivation therapy, chemotherapy, or radiotherapy, prior to PET/CT or surgery. Exclusion criteria were applied sequentially. Patients who received neoadjuvant or preoperative therapies were excluded first, followed by patients with Gleason scores of <7 and finally those with missing key data points. For patients meeting multiple exclusion criteria, the primary reason for exclusion was documented. Patients with missing or incomplete clinical, pathological, or imaging data were excluded from the final analysis to ensure the robustness of the predictive model. No data imputation methods were applied owing to the retrospective study design.

Clinical and Pathological Data Collection
Baseline clinical data, including patient age, Gleason score, preoperative serum PSA levels, pathological tumor stage, lymph node involvement, and postoperative follow-up details, were retrospectively extracted from electronic medical records. Data extraction was performed through manual review of the electronic medical records by two independent investigators using a standardized data collection form to ensure consistency. Any discrepancies between the two investigators were resolved through consensus or consultation with a third senior investigator who reviewed the original medical records. Pathological tumor staging was classified according to the 8th edition of the American Joint Committee on Cancer TNM staging system. Staging data were obtained directly from the original postoperative pathology reports issued by the hospital pathology department. To ensure accuracy, all pathological reports and representative slides were secondarily reviewed by an experienced uropathologist who was blinded to the imaging findings and clinical outcomes.

Systemic Immune-Inflammation Index (SII)
Preoperative peripheral blood samples were obtained from all patients. Blood samples were collected via venipuncture within 7 days prior to surgery. All patients were required to fast for at least 8 h before sample collection to ensure baseline stability. Laboratory parameters, including neutrophil, lymphocyte, and platelet counts, were processed and analyzed in a single centralized clinical laboratory at our institution using standardized automated hematology analyzers. The SII was calculated using the following equation:

The optimal cutoff value for SII was determined using r  eceiver operating characteristic (ROC) analysis and Youden’s index. The cutoff value was data-driven and derived from the entire dataset using the “pROC” package in R statistical analysis software (version 4.2.2). ROC analysis was performed to maximize the combined sensitivity and specificity, and the resulting threshold was used to categorize patients into high-SII and low-SII groups.

18F-PSMA-1007 PET/CT Imaging Protocol and Analysis
All patients underwent preoperative imaging with 18F-PSMA-1007 PET/CT within four weeks prior to surgery. The timing of imaging relative to surgery was standardized, and patients with an imaging-to-surgery interval exceeding 30 days were excluded to ensure consistency between imaging findings and pathological status. Imaging was performed following intravenous administration of 4.0 MBq/kg of 18F-PSMA-1007. The radiotracer was synthesized onsite using a cyclotron and an automated synthesis module. Quality control procedures were performed before each injection to ensure a radiochemical purity >95%. Radiotracer dosing was standardized according to body weight, and no major dosing deviations were recorded. PET/CT scans were acquired from the skull base to the mid-thigh region approximately 60 min post-injection using a dedicated PET/CT system. Scans were performed using a Discovery VCT PET/CT scanner. CT acquisition parameters included 120 kV, automatic tube current modulation (150–200 mA), and a slice thickness of 3.0 mm. PET images were reconstructed using an ordered subset expectation maximization algorithm with a 192 × 192 or 256 × 256 matrix size, incorporating time-of-flight and point spread function corrections.

PET/CT images were independently analyzed by two experienced nuclear medicine physicians who were blinded to the clinical and pathological outcomes. Interobserver agreement was assessed using Cohen’s kappa coefficient. Any discrepancies in lesion localization or categorization were resolved through consensus review or consultation with a third senior nuclear medicine physician. Visual analysis categorized lesions as positive (local recurrence, pelvic lymph nodes, or distant metastases) or negative. Lesions were classified as positive according to the Prostate Cancer Molecular Imaging Standardized Evaluation criteria. Specifically, focal tracer uptake exceeding the surrounding background activity and not attributable to physiological distribution (e.g., ureter or bladder activity) was considered indicative of recurrence or metastasis. Additionally, the maximum standardized uptake value (SUV_max) of suspected lesions was recorded. Regions of interest were manually defined using a three-dimensional spherical volume encompassing the focal uptake on PET images. In patients with multiple lesions, the lesion with the highest uptake was selected for SUV_max measurement. Quantitative image analyses were performed using Advantage Workstation software (version 4.7).

Follow-up and Biochemical Recurrence Definition
Follow-up included regular clinical visits every three to six months postoperatively, with serum PSA measurements. The total follow-up duration for the entire cohort ranged from 8 to 19 months. Follow-up intervals were generally consistent across all patients. For patients who missed scheduled visits, follow-up information was updated through telephone interviews or review of the most recent outpatient records available at our institution. Patients who were lost to follow-up before the minimum 6-month threshold or before documentation of a BCR event were excluded from the final analysis to maintain data integrity. BCR was defined as a postoperative serum PSA level of ≥0.2 ng/mL confirmed by two consecutive measurements. The confirmatory PSA measurements were typically obtained 4–8 weeks apart. This definition of BCR was applied consistently throughout the manuscript and served as the primary endpoint of the predictive model.

Subgroup Analysis
To assess the robustness of the predictive model and reduce potential bias associated with short follow-up durations, a subgroup analysis was performed in patients with follow-up durations >12 months. A total of 133 patients met this criterion, whereas 107 patients with follow-up durations of ≤12 months were excluded from the subgroup analysis. Multivariate logistic regression analysis was subsequently performed in this subgroup using the same statistical methodology applied in the primary analysis.

Statistical Analysis
Continuous variables were presented as median and range, whereas categorical variables were expressed as frequencies and percentages. Data distribution was assessed using the Shapiro–Wilk test. Continuous variables with a normal distribution were presented as mean ± standard deviation (SD) and compared using Student’s t-test. Variables with a non-normal distribution were expressed as median and range and compared using the Mann–Whitney U test. Categorical variables were compared using the chi-square test or Fisher’s exact test, as appropriate. Prior to analysis, assumptions for the chi-square test were verified, including confirmation that no more than 20% of cells had expected frequencies <5. When these assumptions were violated, Fisher’s exact test was applied to ensure statistical validity.

A multivariate logistic regression model was developed to predict BCR, incorporating clinically relevant factors such as SII, PET/CT SUV_max, imaging positivity, and pathological tumor stage. Variables with p-values <0.05 in univariate analysis, together with variables considered clinically relevant based on prior literature, were entered into the multivariate model. Multicollinearity among predictors was assessed using the variance inflation factor, with values of <5 considered acceptable.

Internal validation of the predictive model was performed using a 70/30 train–test split method. The dataset was randomly divided into training (70%) and testing (30%) cohorts using stratified sampling to preserve the BCR event distribution across both cohorts. A fixed random seed was applied to ensure reproducibility. Model performance was evaluated using ROC curves and corresponding area under the curve (AUC) values. Calibration was assessed using the Hosmer–Lemeshow goodness-of-fit test. Decision curve analysis was additionally performed to evaluate the clinical net benefit of the integrative model.

Statistical analyses were performed using SPSS software (version 26.0) and R statistical software (version 4.2.2). The statistical analysis software packages “pROC” and “rms” were used for ROC and calibration analyses, respectively. Statistical significance was defined as p < 0.05. As the primary outcome was binary BCR status during follow-up, model performance was evaluated using ROC curves and AUC metrics. Time-dependent ROC analysis was not performed because recurrence time points were not uniformly recorded across the cohort. Prior to analysis, data were screened for outliers, and no extreme values requiring transformation were identified. Continuous variables were analyzed in their original form to preserve clinical interpretability.

Results

Baseline Characteristics
Patients with BCR exhibited significantly more advanced pathological tumor stages (pT3–4: 96.9% vs. 52.3%, p < 0.001), higher rates of lymph node metastasis (59.4% vs. 29.5%, p < 0.001), elevated SII values (682.45 ± 118.23 vs. 637.08 ± 92.40, p = 0.002), and higher maximum standardized uptake values (SUV_max) on PET/CT imaging (6.59 ± 1.34 vs. 4.92 ± 1.49, p < 0.001) compared with patients without BCR. PET-positive lesio...

Discussion

In this study, we successfully developed a comprehensive predictive model for BCR following radical prostatectomy in prostate cancer patients, integrating the SII, pathological tumor stage, lymph node status, and preoperative 18F-PSMA-1007 PET/CT imaging findings. The model demonstrated excellent discriminative performance, with an AUC of 0.95, underscoring the benefit of combining multiple biological and imaging markers for precise risk stratification. Pathological tumor stage emerged as a powerful and independent predi...

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
18F-PSMA-1007	Shaanxi Zhengze Biotechnology Co., Ltd., China	18F-PSMA-1007	Radiotracer used for PET/CT imaging of prostate cancer lesions
Advantage Workstation software	GE Healthcare, Chicago, IL, USA	Version 4.7	Used for PET/CT image processing, region-of-interest definition, and SUVmax quantification
Blood analyzer	Mindray Bio-Medical Electronics Co., Ltd., China	6000PLUS	Used for measurement of neutrophil, lymphocyte, and platelet counts for SII calculation
PET/CT scanner	GE Healthcare, Chicago, IL, USA	Discovery VCT	Used for whole-body PET/CT image acquisition
ROI/image analysis software	GE Healthcare, Chicago, IL, USA	Advantage Workstation software (version 4.7)	Used for image review and quantitative lesion analysis
R software environment	R Foundation for Statistical Computing, Vienna, Austria	Version 4.2.2	Used for statistical analysis, ROC analysis, and model calibration
SPSS statistical software	IBM Corp., Armonk, NY, USA	Version 26.0	Used for statistical analysis