Machine learning-based prognostic model integrating preoperative HALP score and lactate dehydrogenase for predicting postoperative recurrence of prostate cancer Authors: Hao Wang, Xuemeng Qiu, Zhen Li, Jiyue Wu, Lijian Gan, Dawei Xie, Yirui Wei, Jianwen Wang, Wei Wang Published: 2026 ABSTRACT AbstractObjectivePostoperative biochemical recurrence (BCR) of prostate cancer (PCa) remains a major clinical challenge, and traditional risk assessment systems show suboptimal predictive performance for PCa recurrence. This study aimed to develop and validate interpretable machine learning (ML) models for predicting PCa postoperative recurrence by integrating multi-dimensional clinical features, and to construct a simplified and practical prognostic model for individualized risk stratification.MethodsA total of 320 PCa patients (125 recurrences vs. 195 non-recurrences) who underwent laparoscopic radical prostatectomy (LRP) at the primary center were retrospectively enrolled as the internal cohort, and 144 patients (50 recurrences vs. 94 non-recurrences) from another campus were included as the external validation cohort. Ten ML algorithms were used to construct prediction models with clinicopathological, preoperative hematological and nutrition-inflammation features. Stratified sampling and ten-fold cross-validation were used for model training and validation, and SHAP analysis was adopted for feature importance evaluation and model interpretability. Recursive feature inclusion was performed to optimize the model, and clinical cutoffs of key indicators were determined.ResultsThe gradient boosting machine (GBM) model achieved the best predictive performance in the internal cohort with an AUC of 0.891, which was significantly superior to the UCSF-CAPRA score (AUC = 0.703) and the D’Amico classification (AUC = 0.610). A simplified 5-feature GBM model [positive surgical margin, preoperative hemoglobin-albumin-lymphocyte-platelet (HALP) score, postoperative Gleason score, preoperative maximum prostate specific antigen, preoperative lactate dehydrogenase (LDH)] achieved an AUC of 0.912 in the internal cohort and 0.895 in the external cohort, with excellent calibration and higher net clinical benefit. The optimal cutoffs were 41.31 for preoperative HALP score and 182.61 U/L for preoperative LDH. Low HALP was associated with shorter recurrence-free survival (HR = 0.30, P < 0.0001), and high LDH indicated increased recurrence risk (HR = 1.37, P = 0.083). A three-tier risk stratification system was established based on the cutoff values to predict postoperative recurrence risk.ConclusionThe ML model integrating preoperative HALP score, LDH and core clinicopathological features has high accuracy and good clinical applicability for predicting PCa postoperative recurrence. Validated successfully in both internal and external cohorts, the 5-feature simplified model can serve as a practical tool for individualized recurrence risk assessment, facilitating optimized clinical management of PCa patients.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12957-026-04423-2.Keywords: Prostate cancer, Biochemical recurrence, Machine learning, HALP score, Lactate dehydrogenase, Prognostic prediction Introduction Prostate cancer (PCa) is one of the most common cancers among middle-aged and elderly men worldwide, and represents a major threat to physical health and quality of life [1]. Radical prostatectomy (RP) remains the gold-standard treatment for localized PCa, which can prolong life expectancy and improve prognosis [2]. Despite continuous advances in surgical techniques, postoperative biochemical recurrence (BCR) continues to be a common and clinically important problem, closely linked to worse long-term survival and poorer quality of life [3]. Traditional risk assessment systems, such as tumor grade, pathological stage, and the D’Amico risk classification, only use limited clinical and pathological variables. Their predictive performance for PCa recurrence remains suboptimal, and is insufficient to support personalized prognostic evaluation and precision clinical decision-making [4]. Artificial intelligence (AI) and big data are changing urological oncology by making diagnosis and outcome assessment more accurate, while also facilitating personalized treatment for prostate and other urological cancers [5]. In recent years, machine learning (ML) has become a common approach for developing prognostic models in oncology [6, 7]. By integrating different types of clinical information, ML methods can detect complex non-linear relationships between variables and improve predictive performance. Unlike conventional approaches, ML algorithms have advantages in handling large, heterogeneous clinical data and identifying latent prognostic patterns. This approach is being commonly used in genitourinary oncology to create more accurate and clinically relevant models [8, 9]. Among novel prognostic indicators, preoperative nutrition-inflammation biomarkers have been confirmed to reflect systemic physiological status, which plays a critical role in tumor growth, invasion, and recurrence. The hemoglobin-albumin-lymphocyte-platelet (HALP) score is a novel composite index that simultaneously reflects systemic nutrition, immune, and inflammatory status. A low HALP score indicates poor nutritional status, weakened immune function, and elevated systemic inflammation, all factors that may promote tumor progression and lead to poor prognosis. Previous studies have demonstrated that the HALP score is closely associated with survival and recurrence across a variety of solid tumors [10, 11]. As a key enzyme in anaerobic glycolysis, lactate dehydrogenase (LDH) has also been recognized as a useful prognostic marker in oncology. Elevated LDH levels are closely associated with increased tumor burden, more aggressive biological behavior, and higher risk of distant metastasis, since tumor cells highly rely on anaerobic metabolism to support rapid proliferation [12, 13]. However, few studies have used ML to combine HALP score, LDH, classic pathological indicators, and serological markers for predicting recurrence in PCa. Moreover, the prognostic value of these multi-dimensional indicators has not been fully clarified. This limits the development of comprehensive predictive models that can simultaneously reflect tumor biological behavior and the overall systemic status of patients. This study aimed to develop and evaluate multiple ML models for postoperative recurrence prediction in PCa by integrating clinicopathological and preoperative nutritional‑inflammatory features, and to identify the optimal predictive model. Our findings could provide a simple and clinically applicable tool for individualized recurrence risk stratification, and improving the optimized clinical management of PCa patients. Methods Study design and study population The internal cohort included PCa patients who underwent laparoscopic radical prostatectomy (LRP) at the Changying Campus of Beijing Chaoyang Hospital, Capital Medical University, from January 2017 to December 2023. Follow-up ended on December 31st, 2024. The external validation cohort included patients from the Shijingshan Campus who had LRP from June 2022 to August 2025, with follow-up ending on February 28th, 2026. Demographic characteristics and laboratory indicators were collected using Jiahe electronic medical record system (Beijing Jiahe Meikang Information Technology Co., Ltd., China). Inclusion criteria were: (1) diagnosis of PCa confirmed by preoperative prostate biopsy and postoperative pathological examination; (2) no radiotherapy and/or endocrine therapy; (3) complete medical records, including routine blood and biochemical test results within one week before surgery and full follow-up data; (4) no blood or immune disorders and no active infections. The principal exclusion criteria included unresectable tumor, other oncological diseases, and contraindications to surgery. The study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of Beijing Chaoyang Hospital (2020-science-299-1). Study variables, outcomes and definitions Variables related to general condition, inflammatory status, and tumor-related characteristics were collected. Overall, 19 features, including age, body mass index (BMI), preoperative highest prostate specific antigen (PSA, ng/mL), prostate-specific antigen density (PSAD) (ng/mL2), prostate volume, free/total PSA (f/t PSA), preoperative and postoperative Gleason score, lymphovascular invasion, positive surgical margin (PSM), perineural invasion (PNI), T stage, biopsy-positive cores (BPC), and D’Amico risk classification were included. Routine blood biochemical indicators included hemoglobin (g/L), albumin (g/L), lymphocyte counts (109/L), platelet counts (109/L), and LDH (U/L). The HALP index was calculated as: Following LRP, patients were categorized into two cohorts: non-recurrence and recurrence. The recurrence group included patients with BCR, bone or distant metastasis, or death. BCR was defined as two consecutive PSA values over 0.2 ng/mL without evidence of metastatic lesions on imaging. The non-recurrence group included patients with no BCR or signs of radiographic progression by the last follow-up. Recurrence-free survival (RFS) was the time from surgery to the first event or the last follow-up date. Data extraction and processing We used the “tidyverse” package in R to collect and clean the data. We only included cases with full documentation for all variables and outcomes in the final analysis. Model development and validation To build a robust model for predicting recurrence, we used a machine learning pipeline with the “caret” and “fastshap” packages in R. The dataset was first partitioned into a training set (70%) and a test set (30%) using stratified sampling to maintain the distribution of the outcome. We then tested ten algorithms: logistic regression (LR), K-nearest neighbors (KNN), support vector machines (SVM), decision tree (DT), random forest (RF), Adaboost, gradient boosting machine (GBM), XGboost, artificial neural networks (ANN) and LightGBM. We used ten-fold cross validation to avoid overfitting. Final model performance was assessed using the test set and external validation cohort. The primary metric was the area under the receiver operating characteristic curve (AUC). Model interpretability and feature optimization To address the “black box” nature of machine learning, we used SHAP values to measure feature importance with the “fastshap” and “shapviz” packages. Feature importance was visualized by SHAP beeswarm plots. We also added a recursive feature inclusion process. Features were added sequentially based on their SHAP importance rank to determine the minimum number of variables required to achieve optimal performance. The final model was compared against the all-feature model and standard tools like the D’Amico Risk Stratification and the UCSF-CAPRA score. Discriminative power was evaluated via the AUC. Classification metrics including sensitivity, specificity, and F1 score were calculated at the optimal threshold determined by the Youden Index. Finally, we used calibration curves and decision curve analysis (DCA) to assess if the model would be useful in clinical practice. Identification of clinical cutoffs and survival outcomes For continuous variables in the final model, we performed a threshold analysis using SHAP dependence plots. We used Locally Weighted Scatterplot Smoothing (LOWESS) method to show the relationship between clinical features and their SHAP values. The optimal clinical cutoff for each variable was identified using linear interpolation at the intersection where the SHAP value crossed the zero baseline (SHAP = 0), representing the transition from a protective effect to an increased risk of recurrence. Since HALP and LDH are promising indicators of predicting prostate cancer recurrence, we further validated the clinical predictive value of HALP and LDH by stratifying patients into high-risk and low-risk groups based on the above cutoff. Kaplan-Meier (K-M) curve was visualized. Differences between groups were assessed using the log-rank test. Hazard Ratios (HR) and 95% confidence intervals (CI) were calculated using Cox proportional hazards regression to quantify the relative risk of recurrence. Statistical analysis All statistical analyses were performed using R software (version 4.2.1). The DeLong test was used to assess the statistical significance of differences in the AUC values between the models. Variables following a normal distribution were reported as mean ± standard deviation (SD) and compared via independent samples t-test. Variables with a non-normal distribution were expressed as median [interquartile range (IQR)] and compared using the Mann-Whitney U test. Categorical variables were analyzed using Chi-square test or Fisher’s exact test. A two-sided P value < 0.05 was considered statistically significant. Results Patients characteristics The internal group included 320 patients: 195 patients in the non-recurrence group and 125 patients in the recurrence group. The median follow-up duration was 25 months (6 months for the recurrence group and 37 months for the non-recurrence group). The external validation cohort consisted of 144 patients (50 recurrences and 94 non-recurrences), with a median follow-up of 12.5 months. Detailed baseline clinicopathological data for the internal cohort are summarized in Table 1, and those for the external validation cohort are presented in Supplementary Table 1. Table 1Comparison of clinicopathological characteristics between non-recurrence and recurrence group in the internal cohortCharacteristicsNon-recurrence group (n = 195)Recurrencegroup (n = 125)t/Z/χ2P valueAge68.71 ± 6.7169.17 ± 6.640.6020.548BMI [M(Q1, Q3)], kg/m225.16(23.29, 27.00)25.51(23.81, 27.70)1.1650.244Preoperative highest PSA [M(Q1, Q3)], ng/mL11.96(7.83, 17.67)17.20(10.07, 32.44)4.436< 0.001*f/t PSA [M(Q1, Q3)]11.78(9.13, 16.48)11.65(8.56, 16.87)0.5640.573PV [M(Q1, Q3)], cm334.32(24.76, 46.80)35.04(24.96, 46.80)0.2280.820PSAD [M(Q1, Q3)], ng/mL20.34(0.23, 0.55)0.56(0.28, 1.02)4.291< 0.001*BPC ratio [M(Q1, Q3)]41.67(21.43, 61.54)58.33(33.33, 83.33)3.881< 0.001*D’Amico risk classification (n, %)5.577< 0.001* Low risk40(20.51)9(7.20) Intermediate risk80(41.03)28(22.40) High risk75(38.46)88(70.40)pre-Gleason score (n, %)5.119< 0.001* ≤ 656(28.72)13(10.40) 776(38.97)38(30.40) ≥ 863(32.31)74(59.20)post-Gleason score (n, %)6.765< 0.001* ≤ 632(16.41)2(1.60) 7119(61.03)52(41.60) ≥ 844(22.56)71(56.80)Surgical margin (n, %)66.280< 0.001* Negative118(60.51)18(14.40) Positive77(39.49)107(85.60)PNI (n, %)5.2960.021* No49(25.13)18(14.40) Yes146(74.87)107(85.60)Pathological lymph node metastasis0.0950.758 No125(64.10)78(62.40) Yes70(35.90)47(37.60)T stage (n, %)16.740< 0.001* T2138(70.77)60(48.00) T3 ~ T457(29.23)65(52.00)LDH [M(Q1, Q3)]166.00(148.50, 189.50)173.0(158.00, 204.00)2.6130.009*HALP [M(Q1, Q3)]50.07(42.74, 63.99)36.79(29.88, 46.82)7.377< 0.001**means the p-value < 0.05 is considered statistically significantBMI Body mass index, PSA Prostate specific antigen, PV Prostate volume, PSAD Prostate specific antigen density, BPC Biopsy positive cores, pre-Gleason score preoperative biopsy Gleason score, post-Gleason score postoperative Gleason score, PNI Perineural invasion Comparison of clinicopathological characteristics between non-recurrence and recurrence group in the internal cohort *means the p-value < 0.05 is considered statistically significant BMI Body mass index, PSA Prostate specific antigen, PV Prostate volume, PSAD Prostate specific antigen density, BPC Biopsy positive cores, pre-Gleason score preoperative biopsy Gleason score, post-Gleason score postoperative Gleason score, PNI Perineural invasion In the internal cohort, there were no significant differences between the two groups regarding age, BMI, prostate volume, or f/t PSA ratio. However, several pathological markers demonstrated discrepancies. Patients in the recurrence group exhibited substantially higher preoperative and postoperative Gleason score compared to the non-recurrence group (P < 0.001). Notably, a significantly higher proportion of patients in the recurrence group had PSM, lymph node invasion, BPC and PNI. Additionally, the D’Amico risk classification revealed that 70.40% of the recurrence group were categorized as high risk, compared to only 38.46% in the non-recurrence cohort (P < 0.001). For hematological and inflammatory markers, patients with recurrence presented with higher preoperative PSA levels, higher LDH levels and lower HALP scores, suggesting nutritional and inflammatory status distinct between the two groups (Table 1). Consistent findings were observed in the external validation cohort, with similar between-group differences detected in clinicopathological and hematological indicators (Supplementary Table 1). Model development and performance comparison Among the ten ML models evaluated, GBM achieved the highest discriminative performance with an AUC of 0.891 (95%CI: 0.855–0.935), followed by AdaBoost (0.890, 95%CI: 0.826–0.955) and XGBoost (0.844, 95%CI: 0.764–0.924) in the test set (Fig. 1A; Table 2). Detailed performance metrics, including accuracy, sensitivity, and specificity for the test sets across all ten models, were summarized in Table 3. The SHAP analysis revealed that PSM, preoperative HALP score, postoperative Gleason score, preoperative highest PSA level, and preoperative LDH level were the top five important features in the GBM model (Fig. 1B). To assess the contribution of each variable to this predictive model, we analyzed the ROC curves for the five most influential ones. While all variables contributed to the prediction, preoperative HALP score emerged as the most potent predictor (AUC = 0.843), followed by PSM (AUC = 0.765). Other markers, including preoperative LDH, postoperative Gleason score, preoperative maximum PSA, showed moderate accuracy (Fig. 1C). The machine learning model demonstrated higher predictive accuracy than traditional tools. The D’Amico risk stratification system yielded a considerably lower AUC of 0.61. The DeLong test showed that the GBM model was significantly better than the D’Amico classification (P = 3.27e-9) (Fig. 1D). Performance evaluation and feature importance of 10 machine learning algorithms. (A) Predictive performance of 10 machine learning models evaluated on the test set. (B) SHAP summary plot illustrating feature importance within the GBM model. (C) Individual AUCs for the top five important variables. (D) Comparison of AUCs between the GBM model and the D’Amico risk classification system. Abbreviations: AUC, area under the curve; LR, logistic regression; RF, random forest; KNN, K-nearest neighbors; SVM, support vector machines; DT, decision tree; GBM, gradient boosting machine; BMI, body mass index; PSA, prostate specific antigen; V_prostate, prostate volume; PSAD, prostate specific antigen density; BPC, biopsy positive cores; pre-gleason, preoperative biopsy Gleason score; post-gleason, postoperative Gleason score; PNI, perineural invasion.; LN status, lymph node status; TPR, true positive rate; FPR, false positive rate; 95% CI, 95% confidence interval Table 2Analysis results of AUC for different models in training and test setsModelTraining_AUC (95%CI)Test_AUC (95%CI)GBM0.895 (0.855–0.935)0.891 (0.829–0.954)AdaBoost0.888 (0.846–0.93)0.890 (0.826–0.955)XGBoost1 (1–1)0.844 (0.764–0.924)LR0.878 (0.833–0.923)0.841 (0.761–0.921)RF1 (1–1)0.829 (0.746–0.913)DT0.783 (0.724–0.843)0.823 (0.741–0.905)LightGBM1 (1–1)0.807 (0.716–0.898)SVM0.903 (0.867–0.94)0.784 (0.693–0.875)ANN0.934 (0.901–0.967)0.768 (0.667–0.87)KNN0.976 (0.961–0.99)0.714 (0.608–0.821) Analysis results of AUC for different models in training and test sets Table 3Comparison results of discrimination ability indicators for 10 modelsModelAUCSensitivitySpecificityF1 ScoreAccuracyGBM0.8910.9730.6720.7830.789AdaBoost0.890.8920.7930.8050.832XGBoost0.8440.8920.6720.7420.758LR0.8410.7570.7760.7180.768RF0.8290.7570.7930.7270.779DT0.8230.7030.8450.7220.789LightGBM0.8070.8110.7760.750.789SVM0.7840.9190.5860.7160.716ANN0.7680.730.7070.6670.716KNN0.7140.4590.8790.5570.716 Comparison results of discrimination ability indicators for 10 models These findings indicate that while individual markers are strong indicators of recurrence, the integration of these features into GBM model is more robust in predictive accuracy. A five-feature subset maintains optimal predictive performance We then evaluated the impact of each feature inclusion based on SHAP importance ranking. The AUC reached a near-optimal plateau of 0.912 after including only the top five features (PSM, HALP, postoperative Gleason score, preoperative maximum PSA, and LDH). This suggests that these five variables provide the majority of the predictive value (Fig. 2A). When compared to the “all-feature” model, the top 5 model performed almost the same in terms of sensitivity, specificity, and F1 score (Fig. 2B). Therefore, the top 5 model was selected as the optimal model. Additionally, the top 5 model showed an AUC of 0.895 in the external cohort, and 0.875 in the internal training set (Fig. 2C). Regarding clinical utility, the calibration curves for the 5-feature GBM model demonstrated high concordance between predicted probabilities and observed recurrence proportions in both the internal test set and the external cohort (Fig. 2D). Furthermore, DCA indicated that the model provided superior net clinical benefit (Fig. 2E). The GBM model also outperformed both the UCSF-CAPRA score and the D’Amico classification (Fig. 2F-G). Finally, a sensitivity analysis restricted to BCR was conducted, confirming satisfactory model performance for predicting BCR alone (Supplementary Fig. 1). Feature optimization. (A) Impact of feature inclusion on model AUC. (B) Comparison of discrimination metrics between the top 5-feature model and the all-feature model. (C) Area under the curve of the training, test and external cohorts. (D) Calibration curve of the top-5 feature model of the training, test, and external cohorts. (E) DCA of the training, test, and external cohorts. (F-G) Area under the curve and DCA comparing the top 5 GBM model with the UCSF-CAPRA score and the D’Amico classification system. Abbreviations: AUC, area under the curve; SHAP, Shapley Additive exPlanations.; TPR, true positive rate; FPR, false positive rate Preoperative HALP and LDH based prognostic stratification For clinical translation, we identified thresholds for key continuous variables using SHAP dependence plots (Fig. 3A). The cutoffs were 182.61 U/L for preoperative LDH, 41.31 for HALP score, and 31.61 ng/mL for preoperative maximum PSA (Fig. 3A). For categorical variables, the dependence plots showed discrete SHAP value distribution across different levels of PSM and postoperative Gleason score (Fig. 3B-C). Clinical stratification and prognostic value of HALP and LDH. (A) Cutoff determination for the three indicators. (B-C) SHAP dependence plots for categorical variables. (D-E) Kaplan-Meier survival curves comparing high- and low-risk groups stratified by HALP and LDH levels. Abbreviations: AUC, area under the curve; SHAP, Shapley Additive exPlanations; post gleason, post-operative gleason score; PSM, positive surgical margin; HR, hazard ratio; pre_max_PSA, pre-operative maximum prostate specific antigen Prognostic stratification based on these cutoffs implicated them as potential predictors for RFS. Patients in the high HALP group (≥ 41.31) exhibited a significantly higher survival probability compared to the low HALP cohort, with a HR of 0.30 (95% CI: 0.21–0.43; P < 0.0001). Conversely, patients with LDH levels ≥ 182.61 U/L demonstrated an increased risk of recurrence with a HR of 1.37 (95% CI: 0.96–1.96; P = 0.083). These findings indicated the clinical utility of the preoperative HALP score and LDH levels. Correlations of preoperative HALP and LDH with clinicopathological features Correlation analysis revealed that patients with preoperative HALP < 41.31 had significantly higher age, preoperative highest PSA and PSAD levels relative to those with HALP ≥ 41.31 (all P < 0.05). Notably, preoperative HALP < 41.31 group presented a remarkably higher proportion of high D’Amico risk, elevated post-Gleason score, PSM and PNI (all P < 0.05), with no statistical differences in other indicators (P > 0.05). For preoperative LDH, elevated levels (≥ 182.61) were correlated with significantly higher preoperative highest PSA and a larger proportion of high D’Amico risk classification (P < 0.05). No statistically significant differences were observed in age, BMI, other PSA-related indices and pathological features between the two subgroups (P > 0.05) (Tables 4 and 5). Collectively, low HALP and high LDH were closely associated with unfavorable clinicopathological characteristics in the study cohort. Table 4Correlation analysis of patients with HALP < 41.31 and HALP ≥ 41.31 with clinicopathological featuresCharacteristicsHALP<41.31 (n = 120)HALP ≥ 41.31(n = 200)t/Z/χ2P valueAge70.25 ± 6.5468.07 ± 6.632.8610.004*BMI [M(Q1, Q3)], kg/m225.25(23.45, 27.03)25.28(23.61, 27.32)0.8780.380Preoperative highest PSA [M(Q1, Q3)], ng/mL15.71(9.28, 25.47)12.23(8.38, 19.07)2.5550.011*f/t PSA [M(Q1, Q3)]12.86(9.18, 17.45)11.22(8.54, 16.40)1.4360.151PV [M(Q1, Q3)], cm335.07(24.96, 46.54)34.32(24.96, 48.75)0.2780.781PSAD [M(Q1, Q3)], ng/mL20.45(0.27, 0.90)0.34(0.23, 0.63)2.7080.007*BPC ratio [M(Q1, Q3)]51.92(25.00, 76.44)46.15(23.08, 66.67)1.6170.106D’Amico risk classification (n, %)3.426< 0.001* Low risk14(11.67)35(17.50) Intermediate risk29(24.17)79(39.50) High risk77(64.16)86(43.00)post-Gleason score (n, %)2.4510.014* ≤ 68(6.67)26(13.00) 760(50.00)111(55.50) ≥ 852(43.33)63(31.50)Surgical margin (n, %)31.427< 0.001* Negative27(22.50)109(54.50) Positive93(77.50)91(45.50)PNI (n, %)5.3170.021* No17(14.17)50(25.00) Yes103(85.83)150(75.00)T stage (n, %)0.5970.440 T271(59.17)127(63.50) T3 ~ T449(40.83)73(36.50)*means the p-value < 0.05 is considered statistically significantBMI Body mass index, PSA Prostate specific antigen, PV Prostate volume, PSAD Prostate specific antigen density, BPC Biopsy positive cores, post-Gleason score postoperative Gleason score, PNI Perineural invasion Correlation analysis of patients with HALP < 41.31 and HALP ≥ 41.31 with clinicopathological features *means the p-value < 0.05 is considered statistically significant BMI Body mass index, PSA Prostate specific antigen, PV Prostate volume, PSAD Prostate specific antigen density, BPC Biopsy positive cores, post-Gleason score postoperative Gleason score, PNI Perineural invasion Table 5Correlation analysis of patients with LDH<182.61 and LDH ≥ 182.61 with clinicopathological featuresCharacteristicsLDH<182.61 (n = 212)LDH ≥ 182.61(n = 108)t/Z/χ2P valueAge68.81 ± 6.8469.04 ± 6.350.2860.775BMI [M(Q1, Q3)], kg/m225.25(23.35, 26.96)25.34(23.76, 27.97)1.3490.177Preoperative highest PSA [M(Q1, Q3)], ng/mL11.79(7.97, 19.50)14.90(11.26, 25.32)2.9910.003*f/t PSA [M(Q1, Q3)]11.70(8.98, 16.49)11.68(8.54, 17.70)0.1940.847PV [M(Q1, Q3)], cm333.54(24.57, 46.80)34.71(27.10, 46.48)0.7150.475PSAD [M(Q1, Q3)], ng/mL20.35(0.23, 0.69)0.43(0.27, 0.76)1.7720.076BPC ratio [M(Q1, Q3)]48.08(25.00, 69.23)51.92(23.08, 76.44)0.4430.658D’Amico risk classification (n, %)2.1030.035* Low risk36(16.98)13(12.04) Intermediate risk77(36.32)31(28.70) High risk99(46.70)64(59.26)post-Gleason score (n, %)1.8890.059 ≤ 624(11.32)10(9.26) 7120(56.60)51(47.22) ≥ 868(32.08)47(43.52)Surgical margin (n, %)0.8070.351 Negative94(44.34)42(38.89) Positive118(55.66)66(61.11)PNI (n, %)0.2200.639 No46(21.70)21(19.44) Yes166(78.30)87(80.56)cT stage (n, %)2.7600.097 T2138(65.09)60(55.56) T3 ~ T474(34.91)48(44.44)*means the p-value < 0.05 is considered statistically significantBMI Body mass index, PSA Prostate specific antigen, PV Prostate volume, PSAD Prostate specific antigen density, BPC Biopsy positive cores, PNI Perineural invasion, cT stage clinical T stage Correlation analysis of patients with LDH<182.61 and LDH ≥ 182.61 with clinicopathological features *means the p-value < 0.05 is considered statistically significant BMI Body mass index, PSA Prostate specific antigen, PV Prostate volume, PSAD Prostate specific antigen density, BPC Biopsy positive cores, PNI Perineural invasion, cT stage clinical T stage Standardized risk stratification for recurrence risk assessment Based on the cutoff values and five key clinical features, we developed a standardized three-tier risk stratification system for predicting postoperative recurrence risk. Patients were categorized into low-risk, intermediate-risk, and high-risk groups according to the presence of adverse factors, which included PSM, postoperative Gleason score ≥ 8, preoperative HALP score < 41.31, preoperative maximum PSA ≥ 31.61 ng/mL, and preoperative LDH ≥ 182.61 U/L. Specifically, the classification was as follows: (1) low-risk group: no adverse prognostic factors present; (2) intermediate-risk group: 1 to 3 adverse prognostic factors; (3) high-risk group: 4 to 5 adverse prognostic factors. This approach enables effective risk assessment and supports personalized postoperative management in clinical settings. Discussion In this study, a total of 320 PCa patients were enrolled and stratified into the recurrence group (n = 125) and non-recurrence group (n = 195) according to disease recurrence status. Significant differences in clinical pathological features and related preoperative blood indicators were observed between the two groups. These differences suggested that tumor invasive behavior, pathological features, and preoperative nutritional-inflammatory status were closely linked to PCa postoperative recurrence. Patients in the recurrence group exhibited significantly higher preoperative and postoperative Gleason scores compared to those in the non-recurrence group (P < 0.001). Additionally, the recurrence group showed a significantly increased proportion of PSM, lymphovascular invasion, BPC and PNI, consistent with previous reports [14–16]. A higher Gleason score indicates greater heterogeneity and malignancy of tumor cells. Meanwhile, PSM, PNI, and lymphovascular invasion directly reflect the tumor’s potential to breach local anatomical barriers, forming microinvasion and further distant metastasis [17, 18]. The presence of these pathological features significantly increases the risk of tumor residuals and postoperative recurrence. HALP is an immunonutritional biomarker integrating hemoglobin, albumin, lymphocytes, and platelets, which has been widely used to predict the prognosis of patients with various malignant tumors [19]. Combining these four routine laboratory indicators enables HALP to outperform single biomarkers alone. Hemoglobin and albumin mainly reflect nutritional status, while lymphocytes and platelets are closely linked to tumor progression and clinical prognosis [20–22]. A reduced HALP score usually suggests malnutrition, weakened lymphocyte-mediated antitumor immunity, and a tumor-promoting inflammatory microenvironment. These factors together promote tumor cell survival, invasion, and an increased risk of early recurrence. Clinical studies have validated the good prognostic performance of HALP in various cancers, which helps clinicians identify high-risk patients and make personalized therapeutic decisions [23–26]. These studies have shown that a high HALP score is correlated with favorable survival outcomes in cancer patients. Taha H F et al. demonstrated that in patients with metastatic hormone-sensitive PCa (mHSPC) treated with docetaxel combined with ADT, a higher rate of complete response occurred in those with a high HALP score than in those with a low HALP score (53.8% vs. 5.4%, respectively; P = 0.001) [27]. In addition, several studies have confirmed the HALP score as an independent prognostic factor for PSA-progression free survival (PFS) in patients with metastatic PCa after cytoreductive RP [28]. Low preoperative HALP value (< 32.4) was significantly associated with a decreased PSA-PFS. Similarly, preoperative HALP was identified as an independent risk factor for recurrence following LRP in this study, exhibiting the optimal predictive performance with an AUC of 0.843. Patients with a low HALP score presented adverse clinicopathological features, such as older age, advanced tumor stage, and high Gleason score. Meanwhile, Kaplan-Meier analysis demonstrated that PCa patients with a preoperative HALP value < 41.31 had shorter PFS (P < 0.0001). Compared to previous studies, the HALP cutoff identified in this research is slightly higher than the values reported for other solid tumors. Such differences are probably due to distinct patient populations and therapeutic strategies across studies. LDH is a key enzyme in anaerobic glycolysis. Higher LDH reflects increased tumor metabolic activity, higher tumor burden, and more aggressive tumor biology. Elevated LDH has been reported in a range of malignant tumors and shows prognostic value across multiple solid tumors [12, 29]. Meta-analysis results indicate that LDH levels > 193 U/L are associated with higher mortality and progression risks in mPCa patients, and those with elevated LDH have shorter OS than patients with normal levels [30]. Consistently, Qiu R et al. reported that rapidly progressive PCa was associated with higher initial PSA, LDH (> 191.5 U/L), and alkaline phosphatase levels [31]. Meanwhile, high LDH was an independent prognostic factor for poor PFS/OS and disease progression in castration-resistant PCa patients treated with cabazitaxel [32]. This was generally consistent with our findings. Preoperative LDH level was significantly elevated in the recurrence group (P = 0.009), and high LDH was closely associated with higher preoperative peak PSA and D’Amico risk classification. When preoperative LDH ≥ 182.61 U/L, patients exhibited poorer RFS (HR = 1.37, P = 0.083), suggesting LDH is a potential risk factor for recurrence. Although not statistically significant, it still has clinical significance. We hypothesized that the contribution of LDH might be temporally dynamic. By utilizing time-dependent explanations of the machine learning survival model, we found that LDH’s contribution to the model grew initially before it stabilized during the mid-to-long-term follow-up (Supplementary Fig. 2). ML and deep learning have become valuable tools in precision oncology, improving the diagnosis, risk stratification and clinical management of prostate cancer with better diagnostic accuracy than conventional approaches [5]. Among the ten evaluated ML models in this study, the GBM model exhibited the optimal discriminative performance for PCa recurrence prediction (AUC = 0.891). After removing less important features, we refined the model to the top 5 features, with an AUC of 0.912. The superior performance of the five-feature model relative to the all-feature model is likely due to noise reduction from removing less informative features. The best model outperformed traditional tools, including the UCSF-CAPRA score and the D’Amico risk classification. While XGBoost, RF, and LightGBM had a perfect AUC of 1.0 in the training set, their performance on the test set indicates a degree of overfitting. This could be due to the high learning capability of these models when applied to a relatively modest number of patients and features. This allowed the models to capture random patterns rather than useful clinical signals. To address this problem, we focused on test set performance and cross-validation for final model selection. GBM performed well because it is good at capturing complex nonlinear interactions among variables and assigning each feature the right weight. This allows it to outperform traditional risk systems that only use a limited number of clinical indicators [33, 34]. SHAP analysis showed that PSM, preoperative HALP score, postoperative Gleason score, preoperative maximum PSA, and preoperative LDH were the top five predictive features. HALP score was the strongest single predictor (AUC = 0.843), followed by PSM (AUC = 0.765). As an immunonutritional indicator, preoperative HALP is a more comprehensive marker than single blood tests [24, 35]. Additionally, the GBM model achieved significantly better predictive performance than any individual marker, consistent with previous studies [36, 37]. This confirms that combining different features compensate for the weaknesses of using just one. Overall, our results indicate that ML models using pathological, hematological and inflammatory factors can effectively identify high-risk PCa patients, supporting the clinical significance of ML tools for postoperative recurrence prediction. To facilitate clinical use, we developed a straightforward three-tier risk stratification system using the five key features. Patients were categorized into low-risk, intermediate-risk, and high-risk groups based on the number of the five adverse factors present. All five features can be easily obtained through clinicopathological examinations and routine blood tests. No special or invasive examinations are required, which greatly improves the clinical practicability of the model. This standardized stratification allows clinicians to quickly evaluate recurrence risk, guide postoperative surveillance and plan adjuvant treatment. The study has some limitations. First, this was a retrospective study, which may introduce selection bias. Second, the sample size was relatively small, and more large-scale, prospective, multicenter studies are needed to further validate the performance and generalizability of our model across diverse regions and populations in the future. Finally, the prognostic value of LDH was only marginally significant in our cohort, necessitating further validation in larger multicenter studies. Conclusion This study developed ML models for PCa recurrence prediction, with the GBM model showing optimal performance. A five-feature parsimonious model (PSM, preoperative HALP score, postoperative Gleason score, preoperative maximum PSA, preoperative LDH) was constructed, achieving an AUC of 0.912 in the internal test cohort and 0.895 in the external validation cohort. SHAP analysis clarified core predictive values and derived actionable cut-offs for prognostic stratification. A three-tier risk stratification system was further established to predict postoperative recurrence risk. This model can serve as a practical tool for individualized clinical decision-making in PCa management. Supplementary Information Supplementary Material 1. Supplementary Table 1. Comparison of clinicopathological characteristics between non-recurrence and recurrence group in the external validation cohort. Supplementary Figure 1. Sensitivity analysis on BCR. (A) Area under the curve, (B) calibration, and (C) DCA analysis with outcome restricted to BCR. Supplementary Figure 2. Time-dependent explanations of machine learning survival models. The importance of LDH increased with time before stabilized. Acknowledgements We would like to thank the technicians at Beijing Chao Yang Hospital, Capital Medical University, Beijing, China for technical support. Conflict of interest The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. Authors’ contributions Hao Wang, Xuemeng Qiu, and Zhen Li were responsible for project implementation, data analysis, and manuscript writing. Jiyue Wu and Lijian Gan were responsible for data management and data analysis. Dawei Xie and Yirui Wei were responsible for data collection. Wei Wang and Jianwen Wang were responsible for project design, manuscript writing, and approval of final version of manuscript. Funding This study was supported by The Capital Health Research and Development of Special Fund, Beijing, China (grant No. 2020-2-2033). Data availability The datasets and code used and/or analyzed during the current study are available from the corresponding author on reasonable request. Declarations Ethics approval and consent to participate The study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of Beijing Chaoyang Hospital (2020-science-299-1). Due to the retrospective nature of the study, Ethics Committee of Beijing Chaoyang Hospital waived the need of obtaining informed consent in the manuscript. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Hao Wang, Xuemeng Qiu and Zhen Li contributed equally to this work and considered as co-first authors. Contributor Information Jianwen Wang, Email: wjianw999@163.com. Wei Wang, Email: weiwang0920@163.com. References Associated Data Supplementary Materials Supplementary Material 1. Supplementary Table 1. Comparison of clinicopathological characteristics between non-recurrence and recurrence group in the external validation cohort. Supplementary Figure 1. Sensitivity analysis on BCR. (A) Area under the curve, (B) calibration, and (C) DCA analysis with outcome restricted to BCR. Supplementary Figure 2. Time-dependent explanations of machine learning survival models. The importance of LDH increased with time before stabilized. Data Availability Statement The datasets and code used and/or analyzed during the current study are available from the corresponding author on reasonable request.