Preoperative prediction of pathological pelvic lymph node metastasis in prostate cancer using an elastic net model integrating clinicopathologic and 68Ga-PSMA PET/CT features: development and temporal validation Authors: Yu Chen, Long Gao, Zhongyi Zheng, Pei Feng, Chenqi Zhang, Guofeng Shi, Shaonan Li, Jingxi Hu, Liqiang Wang, Bo Wu, Xiaoming Cao Published: 2026 ABSTRACT Preoperative decision-making regarding extended pelvic lymph node dissection (ePLND) in prostate cancer remains challenging. Existing nodal-risk prediction tools are mainly based on clinicopathologic variables and may show suboptimal performance across different clinical settings. This study aimed to develop and temporally validate a preoperative prediction model for pathological pelvic lymph node metastasis (pN1) by integrating clinicopathologic and 68Ga-PSMA PET/CT-derived features in a single-center Chinese surgical cohort. In this single-center retrospective study, consecutive patients with prostate cancer who underwent robot-assisted radical prostatectomy with ePLND and preoperative 68Ga-PSMA PET/CT were screened. Of 617 patients assessed for eligibility, 315 met the final analytic criteria. Patients treated from 2020 to 2024 formed the development cohort (n = 210), and patients treated in 2025 served as an independent temporal validation cohort (n = 105). Two elastic net regularized logistic regression models were developed using preoperative variables: a clinicopathologic model and an integrated model combining clinicopathologic and PSMA PET/CT-derived features. Internal robustness was assessed using 50 repeated stratified random splits of the development cohort. Final locked models were then applied, without re-tuning, to the temporal validation cohort. Model performance was evaluated using area under the receiver operating characteristic curve (AUC), calibration, Brier score, and decision curve analysis. The Memorial Sloan Kettering Cancer Center (MSKCC) model was evaluated as a benchmark. In the development cohort, 57/210 (27.1%) patients had pathologically confirmed pN1 disease, compared with 27/105 (25.7%) in the temporal validation cohort. Across 50 repeated stratified splits in the development cohort, the clinicopathologic model achieved a mean test AUC of 0.910 (SD 0.029), whereas the integrated model achieved a mean test AUC of 0.962 (SD 0.022). In the temporal validation cohort, the clinicopathologic model achieved an AUC of 0.877 (95% CI 0.799–0.955), whereas the integrated model achieved an AUC of 0.943 (95% CI 0.898–0.989). The integrated model showed sensitivity of 0.852, specificity of 0.897, negative predictive value of 0.946, accuracy of 0.886, and a Brier score of 0.080. Calibration was acceptable, with more favorable calibration for the integrated model than for the clinicopathologic model. Decision curve analysis showed greater net benefit for the integrated model than for both the clinicopathologic model and the MSKCC benchmark across most clinically relevant threshold probabilities. In this single-center Chinese surgical cohort, an elastic net model integrating clinicopathologic and 68Ga-PSMA PET/CT-derived features showed favorable performance for preoperative pN1 prediction, including in temporal validation. Compared with a clinicopathologic model alone and a locally implemented MSKCC benchmark, the integrated model showed more favorable discrimination and potential clinical utility. Multicenter external validation is required before broader clinical application.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-55589-4. Introduction Prostate cancer (PCa) is one of the most common malignancies affecting men worldwide, and its global incidence and mortality continue to show substantial geographic variation and changing temporal trends1. For patients with localized or locally advanced disease, radical prostatectomy remains a major curative-intent treatment option. In men with an elevated risk of lymph node metastasis (LNM), extended pelvic lymph node dissection (ePLND) is often performed at the time of surgery to provide accurate pathological staging and to guide postoperative management2,3. In China, the epidemiology and clinical context of PCa have also evolved considerably in recent years, with increasing disease burden, distinctive presentation patterns, and growing interest in population-specific management strategies4. However, whether ePLND should be performed, and to what extent, remains a critical issue in surgical decision-making for PCa5,6. Unnecessary or overly extensive lymph node dissection may prolong operative time and increase the risks of bleeding, lymphocele formation, and lower-extremity lymphedema, thereby adversely affecting postoperative recovery and quality of life7,8. Conversely, failure to identify nodal metastasis may result in understaging and may influence decisions regarding adjuvant or salvage therapy9,10. Therefore, accurate preoperative estimation of nodal risk is essential for individualized surgical planning. Traditionally, the risk of LNM has been estimated using clinicopathologic variables such as prostate-specific antigen (PSA), clinical stage, and biopsy Gleason score, often with the aid of nomograms such as the Briganti and MSKCC (Memorial Sloan Kettering Cancer Center) models11,12. However, these tools were primarily developed in European and North American populations, and their performance may not fully translate across institutions and practice settings. In addition, their discriminative ability remains imperfect, particularly when used for individualized decision-making in contemporary clinical workflows. In recent years, prostate-specific membrane antigen positron emission tomography/computed tomography (PSMA PET/CT) has substantially improved the accuracy of preoperative staging in PCa, and broader advances in diagnosis and treatment have further reshaped the management landscape of this disease13. Incorporating PSMA PET/CT-derived imaging information into preoperative prediction models may therefore improve risk stratification beyond conventional clinicopathologic tools alone14,15. At the same time, penalized regression approaches such as elastic net are well suited to clinical prediction modeling because they can accommodate correlated predictors, reduce overfitting, and improve model stability16–18. In this study, we developed and temporally validated a preoperative elastic net-based prediction model for pathological pelvic lymph node metastasis (pN1) in patients with PCa undergoing radical prostatectomy with ePLND. The model incorporated preoperatively available clinicopathologic variables together with 68Ga-PSMA PET/CT-derived features. We hypothesized that integrating molecular imaging information with conventional preoperative factors would improve pN1 risk stratification compared with a clinicopathologic model alone and would provide clinically relevant decision-support value in a contemporary single-center Chinese surgical cohort. Materials and methods Study design and population This was a single-center retrospective cohort study designed to develop and temporally validate a preoperative prediction model for pathological pelvic lymph node metastasis (pN1) in patients with prostate cancer undergoing radical prostatectomy (RP) with extended pelvic lymph node dissection (ePLND). Because the model was intended to support real-world preoperative surgical decision-making, only variables available before surgery were considered as candidate predictors. Consecutive patients treated at the First Hospital of Shanxi Medical University between January 2020 and December 2025 were screened. A total of 617 patients were initially assessed for eligibility. Of these, 302 were excluded, including 189 patients with missingness in ≥ 30% of candidate predictors during preprocessing and 113 patients due to failure to meet inclusion criteria, incomplete pathology, missing preoperative PSMA PET/CT, or other predefined exclusion reasons. The final analysis included 315 patients, comprising 210 patients in the development cohort (2020–2024) and 105 patients in the temporal validation cohort (2025). The inclusion criteria were: (1) histologically confirmed prostate cancer; (2) treatment with RP and ePLND; (3) completion of preoperative 68Ga-PSMA PET/CT before surgery; and (4) availability of postoperative pathological lymph node status (pN0 or pN1) as the reference standard. The exclusion criteria were: (1) any antitumor treatment before surgery, including androgen deprivation therapy, radiotherapy, or chemotherapy; (2) an interval of more than 90 days between preoperative evaluation and surgery; and (3) missing key outcome data. Patients with excessive missingness in candidate predictors were further excluded during preprocessing, as described below. Surgical procedure and pathological assessment During the study period, the indication for ePLND was determined according to guideline-based and institutional risk-adapted clinical practice, based on preoperative clinical risk assessment including PSA level, biopsy findings, clinical stage, and imaging findings, at the discretion of the treating surgeon and multidisciplinary team. Patients with suspicious pelvic lymph nodes on preoperative imaging were not routinely excluded from surgery; RP with ePLND was still performed when clinically indicated. Radical prostatectomy was performed using a robot-assisted approach with the da Vinci Surgical System. All included patients underwent ePLND according to institutional surgical practice and guideline-based templates, generally including lymphatic tissue from the obturator, external iliac, internal iliac, and common iliac regions. The total number of resected lymph nodes was recorded from the final pathological report. PSMA PET/CT feature extraction All patients underwent preoperative 68Ga-PSMA PET/CT before surgery. Primary-lesion PSMA SUVmax was measured by experienced nuclear medicine physicians using a dedicated workstation. In patients with multifocal disease, the index lesion or the lesion with the highest uptake was used for analysis. Pelvic lymph node SUVmax was defined as the maximum standardized uptake value among pelvic lymph nodes identified on preoperative imaging. LN SUVmax was selected as the quantitative PET-derived nodal variable, whereas binary PSMA-positive lymph node status on imaging was used descriptively and was not entered simultaneously into the model to avoid redundancy. Preoperative biopsy and MRI (magnetic resonance imaging)-based assessment Preoperative prostate biopsy was performed via a transperineal approach, generally consisting of systematic biopsy with additional targeted cores according to institutional routine practice when clinically indicated. Detailed MRI/PI-RADS (Prostate Imaging Reporting and Data System) -specific variables were not uniformly available as separate structured variables; however, preoperative imaging-based assessments of clinical T stage, suspected extraprostatic extension, and suspected seminal vesicle invasion were available in the medical record and were incorporated as candidate preoperative predictors. Outcome definition and candidate predictors The primary outcome was postoperative pathological lymph node status, dichotomized as pN1 versus pN0. Only variables available before surgery were considered as candidate predictors so that the model would reflect real-world preoperative decision-making. Candidate predictors included demographic, biopsy-related, clinical, and PSMA PET/CT-derived variables. Specifically, the following preoperative variables were collected from the electronic medical record system: age, preoperative prostate-specific antigen (PSA), biopsy International Society of Urological Pathology (ISUP) grade group, biopsy Gleason score, number of positive biopsy cores, total number of biopsy cores, preoperative clinical T stage based on MRI assessment, primary-lesion PSMA SUVmax, pelvic lymph node SUVmax, prostate volume, maximum tumor diameter, suspected extraprostatic extension on preoperative imaging, and suspected seminal vesicle invasion on preoperative imaging. Two derived variables were additionally calculated: positive core ratio (positive cores/total cores) and PSA density (PSA/prostate volume). Detailed MRI/PI-RADS-related variables were not uniformly available across the study period and were therefore not incorporated as separate candidate predictors. However, a preoperative MRI-based clinical T stage variable was included as part of the clinicopathologic feature set. Model development Model development was performed using the development cohort only. Before model fitting, patients with missingness in 30% or more of candidate predictors were excluded. The 30% threshold was chosen pragmatically to avoid excessive information loss and unstable preprocessing in a relatively modest retrospective cohort. In the final analytic dataset, no missing values were present among the retained candidate predictors in either the development or temporal validation cohort. Accordingly, although the preprocessing pipeline allowed median imputation using training-derived values, no actual post-inclusion imputation was required for the final retained candidate predictor set. To reduce redundancy and improve model stability, near-zero-variance predictors were removed, and highly correlated predictors were filtered using the training-set correlation matrix with a cutoff of 0.90. Continuous predictors were then centered and scaled using parameters estimated from the training data, and the same preprocessing parameters were applied to the corresponding test or validation data. Two elastic net regularized logistic regression models were developed using the development cohort only: a clinicopathologic model and an integrated model combining clinicopathologic and PSMA PET/CT-derived predictors. Hyperparameters were optimized using repeated 5-fold cross-validation with AUC as the tuning metric. After internal robustness analysis, the final locked models were refit using the full development cohort. For the clinicopathologic model, the final tuned hyperparameters were alpha = 0.6 and lambda = 0.000672, with a locked classification threshold of 0.121. For the integrated model, the final tuned hyperparameters were alpha = 0.2 and lambda = 0.006210, with a locked classification threshold of 0.298. Final model coefficients and detailed model specifications are provided in the Supplementary Material. The locked classification thresholds were optimized for model discrimination within the development pipeline and should not be interpreted as direct substitutes for clinical nomogram thresholds used to recommend ePLND. Internal robustness analysis To assess model robustness within the development cohort, repeated stratified random splitting was performed 50 times. In each repetition, 70% of patients were assigned to the training subset and 30% to the test subset. For every split, the full modeling workflow was repeated, including training-set-based preprocessing, hyperparameter tuning by repeated cross-validation, and evaluation in the held-out test subset. The distributions of test-set AUC, sensitivity, specificity, accuracy, Brier score, and overfitting gap were summarized across the 50 repetitions. The overfitting gap was defined as the difference between the cross-validated AUC in the training subset and the AUC in the corresponding held-out test subset. Final model locking and temporal validation After internal robustness analysis, the final models were refit using the entire development cohort. All preprocessing procedures, retained predictor sets, fitted coefficients, tuning parameters, and classification thresholds were locked using the development cohort only. Final model coefficients are provided in Supplementary Table S3, and detailed model specifications are provided in Supplementary Table S4. The locked models were then applied, without any re-tuning or variable re-selection, to the independent temporal validation cohort consisting of consecutive patients treated in 2025. Model performance in the temporal validation cohort was evaluated in terms of discrimination, calibration, and clinical utility. Discrimination was assessed using AUC with 95% confidence intervals. Additional performance measures included sensitivity, specificity, positive predictive value, negative predictive value, accuracy, and Brier score. Calibration was assessed using calibration intercept, calibration slope, and calibration plots. Clinical utility was evaluated by decision curve analysis across a range of threshold probabilities. Threshold selection and risk stratification The final classification threshold for each study model was determined in the development cohort using out-of-fold predictions and the Youden index, and was then applied unchanged to the temporal validation cohort. These thresholds were used for classification-based performance assessment and model comparison, rather than as direct clinical recommendations for performing or omitting ePLND. For clinical interpretability, an exploratory risk-stratification scheme based on the integrated model was additionally provided. Patients were categorized as low risk (< 10%), intermediate risk (10%–30%), or high risk (≥ 30%) according to the predicted probability of pN1. This stratification was intended to facilitate interpretation of model-predicted risk and was not used for model training, tuning, or validation. The final decision to perform ePLND should remain individualized and should incorporate predicted risk, imaging findings, overall clinicopathologic risk, surgical feasibility, patient preference, and multidisciplinary team judgment. Benchmark comparison with the MSKCC model As a benchmark model, the MSKCC lymph node invasion model was additionally evaluated in the study cohorts. The benchmark was implemented according to the official preoperative MSKCC nomogram using locally available variables. Specifically, biopsy grade information was mapped using the highest Gleason-based grade information available in the dataset, the number of negative biopsy cores was calculated as total biopsy cores minus positive cores, and clinical T stage was approximated using the institution’s preoperative MRI-based clinical staging. Because local variable definitions did not fully replicate those used in the original derivation of the MSKCC tool, this comparison should be interpreted as a pragmatic local benchmark rather than as a strict head-to-head reconstruction of the original model. Statistical analysis Continuous variables are presented as median and interquartile range (IQR), and categorical variables as number and percentage. Baseline characteristics were compared between the development and temporal validation cohorts. Continuous variables were compared using the Wilcoxon rank-sum test, and categorical variables using the chi-square test or Fisher’s exact test, as appropriate. All statistical analyses and model development were performed in R software (version 4.5.1). Model development, validation, and reporting were performed in accordance with TRIPOD principles for multivariable prediction model studies17. The primary focus of model evaluation was temporal validation performance and overall clinical performance, including discrimination, calibration, Brier score, and decision curve analysis, rather than formal pairwise hypothesis testing between models. A two-sided P value < 0.05 was considered statistically significant. Ethical approval This study was approved by the Ethics Committee of the First Hospital of Shanxi Medical University (KYLL-2024-265). Given the retrospective design and use of de-identified data, the requirement for informed consent was waived. The study was conducted in accordance with the Declaration of Helsinki. Results Patient cohorts and baseline characteristics A total of 315 patients were included in the final analysis, comprising 210 patients in the development cohort and 105 patients in the temporal validation cohort. The prevalence of pathological lymph node metastasis (pN1) was similar between cohorts, with 57/210 (27.1%) in the development cohort and 27/105 (25.7%) in the temporal validation cohort. The median lymph node yield was also comparable between the two cohorts, at 17.5 [14.0, 21.0] in the development cohort and 17.0 [14.0, 20.0] in the temporal validation cohort. Baseline characteristics of the development and temporal validation cohorts are summarized in Table 1, and the patient selection process is summarized in Figure S3. Table 1Baseline characteristics of the development and temporal validation cohorts.VariableDevelopment (N = 210)Temporal validation (N = 105)P valueNodal status pN1 rate57 (27.1)27 (25.7)0.8925 LNs resected, median [IQR]17.50 [14.00, 21.00]17.00 [14.00, 20.00]0.3932Demographics and clinical variables Age, years68.00 [64.00, 74.00]68.00 [65.00, 74.00]0.9157 Preoperative PSA, ng/mL17.94 [11.00, 36.81]15.00 [8.00, 41.23]0.3004 PSA density0.58 [0.30, 1.19]0.55 [0.22, 1.21]0.279 Prostate volume, mL30.92 [21.68, 45.97]33.77 [21.99, 48.69]0.3687Biopsy-related variables ISUP grade group3.00 [3.00, 4.00]3.00 [3.00, 4.00]0.6552 Biopsy Gleason score7.00 [7.00, 8.00]7.00 [7.00, 8.00]0.7141 Positive biopsy cores7.00 [4.00, 9.00]8.00 [5.00, 10.00]0.2403 Total biopsy cores12.00 [12.00, 14.00]12.00 [12.00, 13.00]0.3578 Positive core ratio0.56 [0.31, 0.79]0.62 [0.38, 0.83]0.1653Imaging variables MRI-based clinical T stage2.00 [2.00, 3.00]2.00 [2.00, 3.00]0.5985 Prostate PSMA SUVmax23.37 [16.17, 37.35]23.31 [17.42, 36.54]0.894 LN PSMA SUVmax17.74 [11.64, 21.98]15.35 [9.90, 22.59]0.4584 Tumor Dmax, mm18.00 [14.00, 25.00]17.00 [14.00, 25.00]0.2945 Suspected extraprostatic extension on preoperative imaging47 (22.4)23 (21.9)1 Suspected seminal vesicle invasion on preoperative imaging75 (35.7)41 (39.0)0.6496 PSMA-positive lymph nodes on imaging66 (31.4)31 (29.5)0.8292Values are presented as median (interquartile range [IQR]) for continuous variables and number (percentage) for categorical variables. The development cohort included patients treated between 2020 and 2024, and the temporal validation cohort included patients treated in 2025. Comparisons between cohorts were performed using the Wilcoxon rank-sum test for continuous variables and the chi-square test or Fisher’s exact test for categorical variables, as appropriate. Clinical T stage was based on preoperative MRI assessment. Suspected extraprostatic extension and suspected seminal vesicle invasion refer to preoperative imaging-based assessments rather than postoperative pathological findings. PSMA-positive lymph node status refers to pelvic lymph node positivity on preoperative 68Ga-PSMA PET/CT. pN1 indicates pathological pelvic lymph node metastasis. Abbreviations: PSA, prostate-specific antigen; ISUP, International Society of Urological Pathology; PSMA, prostate-specific membrane antigen; SUVmax, maximum standardized uptake value; LN, lymph node; pN, pathological nodal stage; IQR, interquartile range. Table 1Baseline characteristics of the development and temporal validation cohorts.VariableDevelopment (N = 210)Temporal validation (N = 105)P valueNodal status pN1 rate57 (27.1)27 (25.7)0.8925 LNs resected, median [IQR]17.50 [14.00, 21.00]17.00 [14.00, 20.00]0.3932Demographics and clinical variables Age, years68.00 [64.00, 74.00]68.00 [65.00, 74.00]0.9157 Preoperative PSA, ng/mL17.94 [11.00, 36.81]15.00 [8.00, 41.23]0.3004 PSA density0.58 [0.30, 1.19]0.55 [0.22, 1.21]0.279 Prostate volume, mL30.92 [21.68, 45.97]33.77 [21.99, 48.69]0.3687Biopsy-related variables ISUP grade group3.00 [3.00, 4.00]3.00 [3.00, 4.00]0.6552 Biopsy Gleason score7.00 [7.00, 8.00]7.00 [7.00, 8.00]0.7141 Positive biopsy cores7.00 [4.00, 9.00]8.00 [5.00, 10.00]0.2403 Total biopsy cores12.00 [12.00, 14.00]12.00 [12.00, 13.00]0.3578 Positive core ratio0.56 [0.31, 0.79]0.62 [0.38, 0.83]0.1653Imaging variables MRI-based clinical T stage2.00 [2.00, 3.00]2.00 [2.00, 3.00]0.5985 Prostate PSMA SUVmax23.37 [16.17, 37.35]23.31 [17.42, 36.54]0.894 LN PSMA SUVmax17.74 [11.64, 21.98]15.35 [9.90, 22.59]0.4584 Tumor Dmax, mm18.00 [14.00, 25.00]17.00 [14.00, 25.00]0.2945 Suspected extraprostatic extension on preoperative imaging47 (22.4)23 (21.9)1 Suspected seminal vesicle invasion on preoperative imaging75 (35.7)41 (39.0)0.6496 PSMA-positive lymph nodes on imaging66 (31.4)31 (29.5)0.8292Values are presented as median (interquartile range [IQR]) for continuous variables and number (percentage) for categorical variables. The development cohort included patients treated between 2020 and 2024, and the temporal validation cohort included patients treated in 2025. Comparisons between cohorts were performed using the Wilcoxon rank-sum test for continuous variables and the chi-square test or Fisher’s exact test for categorical variables, as appropriate. Clinical T stage was based on preoperative MRI assessment. Suspected extraprostatic extension and suspected seminal vesicle invasion refer to preoperative imaging-based assessments rather than postoperative pathological findings. PSMA-positive lymph node status refers to pelvic lymph node positivity on preoperative 68Ga-PSMA PET/CT. pN1 indicates pathological pelvic lymph node metastasis. Abbreviations: PSA, prostate-specific antigen; ISUP, International Society of Urological Pathology; PSMA, prostate-specific membrane antigen; SUVmax, maximum standardized uptake value; LN, lymph node; pN, pathological nodal stage; IQR, interquartile range. Baseline characteristics of the development and temporal validation cohorts. Values are presented as median (interquartile range [IQR]) for continuous variables and number (percentage) for categorical variables. The development cohort included patients treated between 2020 and 2024, and the temporal validation cohort included patients treated in 2025. Comparisons between cohorts were performed using the Wilcoxon rank-sum test for continuous variables and the chi-square test or Fisher’s exact test for categorical variables, as appropriate. Clinical T stage was based on preoperative MRI assessment. Suspected extraprostatic extension and suspected seminal vesicle invasion refer to preoperative imaging-based assessments rather than postoperative pathological findings. PSMA-positive lymph node status refers to pelvic lymph node positivity on preoperative 68Ga-PSMA PET/CT. pN1 indicates pathological pelvic lymph node metastasis. Abbreviations: PSA, prostate-specific antigen; ISUP, International Society of Urological Pathology; PSMA, prostate-specific membrane antigen; SUVmax, maximum standardized uptake value; LN, lymph node; pN, pathological nodal stage; IQR, interquartile range. Postoperative pathological characteristics, including pathological T stage, final pathological ISUP grade group, and positive surgical margin status, are summarized in Supplementary Table S5. These postoperative variables were reported descriptively and were not used as candidate predictors because the model was intended for preoperative decision-making. Overall, baseline clinicopathologic and imaging variables were well balanced between the two cohorts, and no statistically significant between-cohort differences were observed for age, preoperative PSA, ISUP grade, biopsy Gleason score, number of positive cores, total cores, positive core ratio, PSA density, MRI-based clinical stage, prostate PSMA SUVmax, lymph node PSMA SUVmax, prostate volume, tumor maximum diameter, extraprostatic extension, seminal vesicle invasion, PSMA-positive lymph nodes on imaging, total lymph nodes resected, or pN1 rate (all P > 0.05). The proportion of patients with PSMA-positive lymph nodes on imaging was similar between the two cohorts, at 31.4% in the development cohort and 29.5% in the temporal validation cohort. Lymph node yield also did not differ significantly between the development and temporal validation cohorts (P = 0.393), nor between pN0 and pN1 patients within either cohort (Supplementary Table S1). Internal robustness in the development cohort Internal repeated-split validation results in the development cohort are summarized in Table 2. Across 50 stratified random splits, the integrated model achieved a higher mean test AUC than the clinicopathologic model (0.962 vs. 0.910), along with higher mean sensitivity (0.892 vs. 0.844) and specificity (0.867 vs. 0.776), as well as a lower mean Brier score (0.070 vs. 0.112). Both models showed minimal overfitting gap (0.001 for the clinicopathologic model and − 0.001 for the integrated model), suggesting limited optimism and good internal stability. Table 2Internal repeated-split validation performance of the clinicopathologic and integrated models in the development cohort.MetricClinicopathologic modelIntegrated modelSuccessful splits/requested splits50/5050/50Mean test AUC0.9100.962SD of test AUC0.0290.022Mean sensitivity0.8440.892Mean specificity0.7760.867Mean Brier score0.1120.070Mean overfitting gap0.001−0.001Internal robustness was assessed by 50 repeated stratified random splits of the development cohort. In each repetition, 70% of patients were assigned to the training subset and 30% to the test subset. The overfitting gap was defined as the difference between the cross-validated AUC in the training subset and the AUC in the corresponding held-out test subset. AUC, area under the receiver operating characteristic curve; SD, standard deviation. Table 2Internal repeated-split validation performance of the clinicopathologic and integrated models in the development cohort.MetricClinicopathologic modelIntegrated modelSuccessful splits/requested splits50/5050/50Mean test AUC0.9100.962SD of test AUC0.0290.022Mean sensitivity0.8440.892Mean specificity0.7760.867Mean Brier score0.1120.070Mean overfitting gap0.001−0.001Internal robustness was assessed by 50 repeated stratified random splits of the development cohort. In each repetition, 70% of patients were assigned to the training subset and 30% to the test subset. The overfitting gap was defined as the difference between the cross-validated AUC in the training subset and the AUC in the corresponding held-out test subset. AUC, area under the receiver operating characteristic curve; SD, standard deviation. Internal repeated-split validation performance of the clinicopathologic and integrated models in the development cohort. Internal robustness was assessed by 50 repeated stratified random splits of the development cohort. In each repetition, 70% of patients were assigned to the training subset and 30% to the test subset. The overfitting gap was defined as the difference between the cross-validated AUC in the training subset and the AUC in the corresponding held-out test subset. AUC, area under the receiver operating characteristic curve; SD, standard deviation. As shown in Figure S1, the AUC distribution of the integrated model was shifted to the right compared with that of the clinicopathologic model, indicating consistently stronger discrimination across repeated resampling. Final locked model specification In the final locked clinicopathologic model, all 10 prespecified clinicopathologic predictors were retained with non-zero coefficients, including age, preoperative PSA, ISUP grade group, number of positive biopsy cores, total biopsy cores, MRI-based clinical T stage, prostate volume, maximum tumor diameter, extraprostatic extension, and seminal vesicle invasion. In the final locked integrated model, all 12 prespecified predictors were retained, with the addition of primary-lesion PSMA SUVmax and pelvic lymph node SUVmax. Final model coefficients, intercepts, tuning parameters, retained predictors, and locked thresholds are provided in Supplementary Tables S3–S4. Temporal validation performance In the temporal validation cohort, comparative model performance is summarized in Table 3. The integrated model showed the best overall performance, with an AUC of 0.943 (95% CI, 0.898–0.989), sensitivity of 0.852, specificity of 0.897, positive predictive value of 0.742, negative predictive value of 0.946, accuracy of 0.886, and a Brier score of 0.080. The clinicopathologic model achieved an AUC of 0.877 (95% CI, 0.799–0.955), whereas the MSKCC benchmark showed weaker discrimination, with an AUC of 0.651 (95% CI, 0.504–0.798). Overall, the integrated model showed the most favorable combination of discrimination, classification performance, and overall prediction error in the temporal validation cohort. Table 3Comparative performance of the clinicopathologic, integrated, and MSKCC models in the temporal validation cohort.MetricClinicopathologic modelIntegrated modelMSKCC modelN105105105pN1, n (%)27 (25.7)27 (25.7)27 (25.7)AUC (95% CI)0.877 (0.799–0.955)0.943 (0.898–0.989)0.651 (0.504–0.798)Sensitivity0.8150.8520.481Specificity0.7310.8970.846PPV0.5120.7420.520NPV0.9190.9460.825Accuracy0.7520.8860.752Brier score0.1140.0800.182Performance was evaluated in the independent temporal validation cohort. Discrimination was assessed using AUC with 95% CIs. Additional metrics included sensitivity, specificity, PPV, NPV, accuracy, and Brier score. The study-model thresholds were locked using the development cohort and applied unchanged to the temporal validation cohort. AUC, area under the receiver operating characteristic curve; CI, confidence interval; PPV, positive predictive value; NPV, negative predictive value; MSKCC, Memorial Sloan Kettering Cancer Center. Table 3Comparative performance of the clinicopathologic, integrated, and MSKCC models in the temporal validation cohort.MetricClinicopathologic modelIntegrated modelMSKCC modelN105105105pN1, n (%)27 (25.7)27 (25.7)27 (25.7)AUC (95% CI)0.877 (0.799–0.955)0.943 (0.898–0.989)0.651 (0.504–0.798)Sensitivity0.8150.8520.481Specificity0.7310.8970.846PPV0.5120.7420.520NPV0.9190.9460.825Accuracy0.7520.8860.752Brier score0.1140.0800.182Performance was evaluated in the independent temporal validation cohort. Discrimination was assessed using AUC with 95% CIs. Additional metrics included sensitivity, specificity, PPV, NPV, accuracy, and Brier score. The study-model thresholds were locked using the development cohort and applied unchanged to the temporal validation cohort. AUC, area under the receiver operating characteristic curve; CI, confidence interval; PPV, positive predictive value; NPV, negative predictive value; MSKCC, Memorial Sloan Kettering Cancer Center. Comparative performance of the clinicopathologic, integrated, and MSKCC models in the temporal validation cohort. Performance was evaluated in the independent temporal validation cohort. Discrimination was assessed using AUC with 95% CIs. Additional metrics included sensitivity, specificity, PPV, NPV, accuracy, and Brier score. The study-model thresholds were locked using the development cohort and applied unchanged to the temporal validation cohort. AUC, area under the receiver operating characteristic curve; CI, confidence interval; PPV, positive predictive value; NPV, negative predictive value; MSKCC, Memorial Sloan Kettering Cancer Center. The MSKCC benchmark showed substantially weaker discrimination than the two study models in the temporal validation cohort; however, because the benchmark was implemented using locally mapped variables rather than a fully matched reconstruction of the original derivation setting, this comparison should be interpreted as a pragmatic local benchmark rather than as a strict external head-to-head validation. For clinical interpretability, patients in the temporal validation cohort were further stratified into low-, intermediate-, and high-risk groups according to the predicted pN1 probability from the integrated model (Supplementary Table S6). The observed pN1 rate increased from 4.6% in the low-risk group to 11.1% in the intermediate-risk group and 74.2% in the high-risk group. This exploratory stratification was intended to facilitate interpretation of model-predicted risk rather than to define a stand-alone rule for performing or omitting ePLND. ROC analysis Receiver operating characteristic analysis in the temporal validation cohort confirmed superior discrimination of the integrated model. As shown in Fig. 1, the integrated model achieved the highest AUC (0.943), followed by the clinicopathologic model (0.877), whereas the MSKCC model showed substantially lower discriminatory ability (0.651). The ROC curves demonstrated a clear separation between the integrated model and the other two models across most of the operating range. Fig. 1Receiver operating characteristic curves of the clinicopathologic, integrated, and MSKCC models in the temporal validation cohort. AUC values with 95% confidence intervals are shown in the legend. AUC, area under the receiver operating characteristic curve; CI, confidence interval; MSKCC, Memorial Sloan Kettering Cancer Center. Receiver operating characteristic curves of the clinicopathologic, integrated, and MSKCC models in the temporal validation cohort. AUC values with 95% confidence intervals are shown in the legend. AUC, area under the receiver operating characteristic curve; CI, confidence interval; MSKCC, Memorial Sloan Kettering Cancer Center. Calibration analysis Calibration analysis in the temporal validation cohort showed acceptable agreement between predicted and observed risks for both study models. The clinicopathologic model yielded a calibration intercept of −0.05 and a calibration slope of 0.58, whereas the integrated model showed a calibration intercept of −0.03 and a calibration slope of 0.80. The integrated model therefore demonstrated calibration closer to the ideal line than the clinicopathologic model. Detailed calibration intercepts and slopes are provided in Supplementary Table S2. As illustrated in Fig. 2, both models showed reasonable calibration overall, while the integrated model displayed better agreement between predicted probabilities and observed event rates, particularly in the intermediate-to-high risk range. When the MSKCC model was additionally plotted, its calibration appeared less consistent than that of the two study models (Figure S2). Fig. 2Calibration plots of the clinicopathologic and integrated models in the temporal validation cohort. The x-axis shows the predicted pN1 risk, and the y-axis shows the observed pN1 rate. The dotted 45-degree line represents ideal calibration. Calibration plots of the clinicopathologic and integrated models in the temporal validation cohort. The x-axis shows the predicted pN1 risk, and the y-axis shows the observed pN1 rate. The dotted 45-degree line represents ideal calibration. Decision curve analysis Decision curve analysis demonstrated that the integrated model provided the greatest clinical net benefit across a broad range of threshold probabilities in the temporal validation cohort (Fig. 3). The integrated model consistently yielded higher net benefit than the clinicopathologic model, the MSKCC model, and the treat-all or treat-none strategies over most clinically relevant thresholds. The clinicopathologic model also outperformed the MSKCC benchmark but remained inferior to the integrated model throughout the majority of the evaluated range. Fig. 3Decision curve analysis of the clinicopathologic, integrated, and MSKCC models in the temporal validation cohort. Treat-all and treat-none strategies are shown as reference strategies. The shaded area indicates the clinically relevant threshold probability range of 5%–30%. Decision curve analysis of the clinicopathologic, integrated, and MSKCC models in the temporal validation cohort. Treat-all and treat-none strategies are shown as reference strategies. The shaded area indicates the clinically relevant threshold probability range of 5%–30%. Discussion In this single-center retrospective study, we developed and temporally validated a preoperative elastic net-based model for predicting pathological pelvic lymph node metastasis in prostate cancer using clinicopathologic and 68Ga-PSMA PET/CT-derived variables. Several findings merit emphasis. First, both study models showed good internal robustness in the development cohort, whereas the integrated model consistently outperformed the clinicopathologic model across repeated split-based validation. Second, after model locking in the development cohort, the integrated model maintained favorable performance in the independent temporal validation cohort, with an AUC of 0.943, a Brier score of 0.080, and a high negative predictive value. Third, compared with the clinicopathologic model alone and the locally implemented MSKCC benchmark, the integrated model showed more favorable discrimination, calibration, and decision-curve performance in this cohort. Taken together, these findings suggest that incorporating PSMA PET/CT-derived information may improve preoperative nodal-risk stratification in a contemporary surgical setting. This finding is also consistent with recent evidence supporting the diagnostic and staging value of PSMA PET/CT in prostate cancer. In particular, PSMA PET/CT has shown high accuracy in nodal staging, although its performance may vary across tumor grade groups and clinical subtypes. These data support the rationale for integrating PSMA PET/CT-derived features with conventional clinicopathologic variables when estimating the risk of pathological nodal involvement19,20. From a clinical perspective, accurate preoperative estimation of nodal risk remains important because the decision to perform ePLND must balance staging benefit against operative time, lymphocele formation, vascular or nerve injury, and other procedure-related morbidity. Widely used tools such as the Briganti nomograms and the MSKCC model estimate nodal risk primarily on the basis of clinicopathologic variables11,21. The Briganti 2019 nomogram represented an important step forward by incorporating multiparametric MRI and MRI-targeted biopsy information22. However, these tools were mainly derived in Western cohorts, and their performance may be influenced by differences in patient selection, imaging pathways, biopsy strategies, and surgical practice across institutions and populations23–27. In the present cohort, the findings suggest that a model developed within the local diagnostic and surgical context, while incorporating PSMA PET/CT-derived information, may better capture preoperative predictors of pN1 in this setting than conventional clinicopathologic assessment alone. We also considered comparison with the Briganti nomogram, given its widespread use in identifying candidates for ePLND11,21. However, formal calculation of the Briganti 2019 score was not performed because several required inputs, particularly detailed MRI-targeted/systematic biopsy information and standardized MRI-derived variables, were not uniformly available as structured fields in this retrospective database22. Because partial or approximate reconstruction of the Briganti score could introduce additional measurement error, we did not report a “Briganti-like” estimate. Therefore, the MSKCC model was retained as a pragmatic local benchmark, while we acknowledge that established nomograms may vary in performance across populations, imaging pathways, and clinical settings21,23. Clinically, the integrated model may be most useful as a decision-support tool to refine preoperative selection for ePLND, particularly in patients with equivocal or intermediate baseline risk, rather than as a replacement for multidisciplinary clinical judgment. Its high negative predictive value in the temporal validation cohort suggests potential utility for identifying patients at relatively low risk of pathological nodal involvement and thereby supporting more selective use of ePLND in appropriate settings. The model should therefore be viewed as an adjunct to multidisciplinary assessment rather than as a stand-alone determinant of whether ePLND should be performed. From a practical perspective, PSMA PET/CT may also be attractive because it can provide diagnostic and staging information in a single preoperative procedure. Recent evidence has suggested favorable cost-effectiveness of PSMA PET/CT in the diagnostic and staging pathway of high-grade prostate cancer. Although formal health-economic evaluation was beyond the scope of the present study, our findings support the efficient secondary use of PSMA PET/CT information already obtained during routine preoperative staging28. More recently, data-driven prediction approaches, including machine-learning methods, have also been explored for nodal-risk prediction in prostate cancer29,30. Among these methods, elastic net regularization is attractive for clinical prediction because it can accommodate correlated predictors, shrink unstable coefficients, and reduce overfitting while retaining a comparatively interpretable modeling framework31. In the present dataset, the primary value of elastic net was not aggressive variable elimination, but coefficient shrinkage and improved stability in the presence of correlated preoperative predictors. This was particularly relevant in the present study, where preoperative variables related to biopsy tumor burden, local tumor extent, lesion size, and PSMA PET/CT uptake may contain overlapping yet complementary information. The superior performance of the integrated model is biologically and clinically plausible, as pelvic nodal involvement is unlikely to depend on a single domain alone. Rather, nodal spread likely reflects a combination of overall tumor burden, local aggressiveness, and tumor biology captured by molecular imaging. Our findings are therefore broadly consistent with previous studies highlighting the prognostic importance of local stage, biopsy burden, and PSMA PET/CT-derived features in nodal-risk assessment32–34. An important strength of the present study is that model evaluation did not rely solely on a single random split or internal cross-validation estimate. In addition to repeated split-based internal robustness analysis, we applied locked models to an independent temporal validation cohort consisting of patients treated in a later period. This design better reflects real-world deployment than repeated random splitting alone. The integrated model maintained strong discrimination in temporal validation, whereas its calibration was also more favorable than that of the clinicopathologic model, with a calibration slope closer to 1.0 (0.80 vs. 0.58) and a slightly smaller absolute calibration intercept. Moreover, decision curve analysis showed that the integrated model provided greater net benefit than the clinicopathologic model across most clinically relevant threshold probabilities. These findings suggest that adding PSMA PET/CT-derived information may improve not only discrimination but also the potential clinical usefulness of preoperative nodal-risk assessment. Several features of the present study strengthen its clinical relevance. First, the development and temporal validation cohorts were broadly comparable in baseline characteristics, pN1 prevalence, and lymph node yield, which supports the validity of the temporal validation framework. Second, the model was developed exclusively from variables available before surgery, enhancing its applicability to real-world preoperative decision-making. Third, all included patients underwent ePLND according to institutional practice and guideline-based templates, using a robot-assisted approach, and suspicious pelvic lymph nodes on preoperative imaging were not routinely considered an absolute contraindication to surgery. Although institutional practice followed a standard template, some variation in surgical decision-making and extent of dissection may still have occurred in routine clinical care. These features improve the internal clinical consistency of the cohort, although they do not eliminate the risk of selection bias inherent in retrospective surgical series. The present cohort should also be interpreted as a relatively high-risk surgical population, as reflected by the high baseline prevalence of pathological lymph node metastasis (> 25%), the high proportion of adverse pathological ISUP grade groups, and the routine clinical indications for ePLND in this surgical cohort. The present study should also be interpreted in light of several limitations. First, this was a retrospective single-center analysis performed in a Chinese surgical cohort, and the findings should not be assumed to be directly generalizable to broader populations or other clinical settings. Although temporal validation strengthens the study beyond simple resampling, it does not replace true external validation. Second, the overall sample size remained modest, particularly in the temporal validation cohort, which may limit the precision of performance estimates, especially calibration-related measures. Third, because the study included only patients who ultimately underwent RP with ePLND and preoperative PSMA PET/CT, the cohort reflects a surgically selected population rather than an unselected diagnostic population, and this introduces the possibility of selection bias. Fourth, exclusion of patients with substantial missingness may also have introduced selection bias. We did not perform a formal comparison between included patients and those excluded for substantial missingness, which limits assessment of the magnitude of potential selection bias. Fifth, although preoperative MRI-based clinical T stage was incorporated, detailed MRI/PI-RADS-related variables were not uniformly available as standardized candidate predictors across the full study period. Sixth, PSMA PET/CT interpretation is inherently reader-dependent, and variation in image acquisition, reconstruction, and reporting across centers may influence broader applicability. In addition, pelvic lymph node SUVmax may partly reflect overt imaging suspicion of nodal disease rather than latent biological risk alone; therefore, the integrated model should be interpreted as a preoperative decision-support tool incorporating molecular imaging evidence rather than as a purely clinicopathologic risk nomogram. Seventh, uncommon histological variants, particularly ductal adenocarcinoma of the prostate, were not specifically analyzed. Previous evidence suggests that conventional imaging and PSMA PET/CT may be insufficient for clinical staging of ductal adenocarcinoma, and additional imaging information such as mpMRI and/or FDG PET may be required in selected cases. Therefore, the present model should not be directly extrapolated to patients with ductal adenocarcinoma or other uncommon histological variants without further validation35. Finally, as with other regularized multivariable models, individual coefficient directions should be interpreted cautiously, particularly when correlated predictors are entered simultaneously and coefficients are shrunk under penalization. Some individual coefficient directions may appear counterintuitive after penalization because correlated predictors were entered simultaneously and the model was optimized for prediction rather than for etiologic interpretation. Despite these limitations, the present study provides a clinically relevant framework for integrating conventional clinicopathologic variables with PSMA PET/CT-derived features in preoperative nodal-risk prediction. In this single-center Chinese surgical cohort, the integrated model showed favorable temporal validation performance, with strong discrimination, acceptable calibration, and greater net benefit than the clinicopathologic model across most clinically relevant thresholds. Future multicenter external validation studies are warranted to determine whether this modeling strategy can be reliably applied across institutions for preoperative nodal-risk assessment and ePLND decision support. Conclusion In this single-center retrospective surgical cohort, an elastic net model integrating clinicopathologic and 68Ga-PSMA PET/CT-derived features showed favorable performance for preoperative prediction of pathological pelvic lymph node metastasis, including in temporal validation. These findings support the potential value of integrating molecular imaging information into preoperative nodal-risk assessment, pending external multicenter validation. Conference presentation and redundant publication statement This manuscript has not been presented at any meeting. We confirm that this work is original, has not been published elsewhere, and is not under consideration by any other journal. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1