Abstract
Inguinal hernia represents a clinically significant yet underreported complication of robot-assisted radical prostatectomy (RARP) for localized prostate cancer, with a notably high incidence within the first postoperative year. Despite its adverse impact on quality of life and potential for severe sequelae, predictive tools for this outcome remain limited. To develop and validate the first machine learning (ML)-based clinical prediction model for inguinal hernia within 1 year after RARP, leveraging explainable artificial intelligence (AI) techniques for clinical interpretability. This retrospective study analyzed localized prostate cancer patients who underwent RARP between June 1, 2021 and May 1, 2023 at our center. Least absolute shrinkage and selection operator (LASSO) regression identified five key predictors from multiple clinical parameters. Five ML algorithms were developed and evaluated on a 70:30 training–test split. Model performance was assessed via area under the curve (AUC), accuracy, specificity, and decision curve analysis (DCA). SHapley Additive exPlanations (SHAP) methodology provided interpretable feature attribution. The final analysis included 652 eligible patients. Extreme gradient boosting (XGBoost) demonstrated superior discriminative ability, with an AUC of 0.833 (95% CI: 0.770–0.895) in the validation set and 0.791 (95% CI: 0.734–0.848) in the test set. SHAP analysis identified five critical predictors ranked by impact: age, body mass index (BMI), preoperative albumin level, T stage, and history of abdominal surgery. This study established the first ML-driven predictive model for post-RARP inguinal hernia, with XGBoost demonstrating optimal performance. High-risk patients identified by the model warrant personalized proactive interventions.
Keywords: Robot-assisted radical prostatectomy, Inguinal hernia, Machine learning, Predictive model, SHAP
Introduction
Prostate cancer is the most prevalent malignant tumor of the male genitourinary system [1]. Radical prostatectomy, the most effective treatment for localized prostate cancer, has evolved from traditional open radical prostatectomy (ORP) to laparoscopic radical prostatectomy (LRP) and subsequently to robot-assisted radical prostatectomy (RARP). Numerous studies have revealed that RARP results in less surgical trauma and fewer postoperative complications than does ORP [2]. However, various complications may ensue following RARP, including urinary incontinence, sexual dysfunction, inguinal hernia, and penile shortening, among other sequelae [3, 4]. An inguinal hernia refers to a pathological condition wherein the abdominal contents protrude through the inguinal canal or adjacent fascial weaknesses into the groin region. This condition represents one of the most prevalent hernia types, exhibiting a substantially higher incidence in males than in females, with peak occurrence observed in children and adults over 50 years of age. Clinically, it typically manifests as a palpable groin mass, often accompanied by localized pain or discomfort [5].
Regan et al. [6] were the first to describe inguinal hernia as a complication of ORP, reporting an incidence of 12% at six months postoperatively. Recent meta-analyses [7] indicate that the incidence of inguinal hernia after LRP is approximately 7.5% (95% CI: 5.2–9.8%), and after RARP, it is approximately 7.9% (95% CI: 5.0–10.9%). Indirect inguinal hernias constitute most cases (81.9%, 95% CI: 75.3–88.4%) and predominantly occur within the first two years after surgery, with a higher prevalence on the right side. The current literature on the incidence of inguinal hernia following RARP is inconsistent, with large-scale studies being notably scarce. Inguinal hernias can lead to pain and intestinal dysfunction and may necessitate emergency surgery due to the risk of bowel obstruction [8, 9]. Although urinary incontinence and sexual dysfunction have been extensively documented by urologists as prevalent complications following RARP [10, 11], inguinal hernia remains relatively underreported in the literature.
Postoperative inguinal hernia is a frequent complication following RARP and significantly impacts patients ’ quality of life. To enable early intervention, there is a critical clinical need for predictive models targeting this specific outcome. However, such models are currently scarce. Recent advances in machine learning (ML) and artificial intelligence (AI) offer promising avenues for developing clinical prediction models. Furthermore, techniques such as SHapley Additive exPlanations (SHAP) enhance model interpretability, making ML approaches clinically valuable. Therefore, this study aimed to construct a clinical prediction model for inguinal hernia after RARP via ML combined with SHAP analysis. The goal is to improve clinical decision-making and ultimately enhance patient outcomes. To our knowledge, this represents the first attempt to develop an ML-based predictive model for inguinal hernia following RARP.
Materials and methods
Patient selection
This single-institution retrospective cohort study was conducted at a 6038-bed tertiary care teaching hospital in Eastern China. From June 1, 2021 to May 1, 2023, clinical data from prostate cancer patients who underwent RARP at our center were analyzed retrospectively. The study protocol received full ethical approval from our Institutional Review Board (Approval No. PJ2024-12–93). All procedures were conducted in accordance with local legislation and institutional requirements.
Inclusion criteria: 1. Histologically confirmed localized prostate cancer (diagnosed via preoperative biopsy or postoperative pathology). 2. Complete or accessible baseline clinical data. 3. Regular postoperative follow-up ≥ 1 year with inguinal region assessment.
Exclusion criteria: 1. Preexisting inguinal hernia before RARP (unilateral/bilateral, regardless of repair status). 2. Severe, poorly controlled comorbidities complicating the postoperative hernia follow-up assessment. 3. Abdominopelvic or inguinal region surgery with the potential to alter abdominal wall integrity following RARP. 4. Hernia prevention strategies during RARP. 5. Prolonged occupational engagement in heavy manual labor after RARP. 6. Inadequate follow-up without documented inguinal examination (lost to follow-up).
Data collection
Age, height, weight, prostate-specific antigen (PSA) levels, past medical history, Gleason score, ISUP (International Society of Urological Pathology) grade, postoperative pathological stage, and other relevant parameters were retrospectively collected for all patients. The PSA value was based on the first instance of abnormal elevation, ensuring that this measurement was not influenced by any prior endocrine therapy. Fasting morning blood samples collected within two weeks prior to surgery were used to measure preoperative hemoglobin, albumin, and fasting blood glucose levels. On the first postoperative morning, blood samples were obtained to assess the postoperative levels of these parameters. The change in hemoglobin was calculated by subtracting the postoperative hemoglobin level from the preoperative value. Prostate dimensions (anterior–posterior, craniocaudal, and transverse diameters) were measured on preoperative MR images by professional radiology technicians. The maximum anterior‒posterior and craniocaudal diameters were measured on mid-sagittal T2W MR images, whereas the maximum transverse diameter was measured on axial T2W MR images. The prostate volume was calculated via the following formula: (anterior–posterior diameter × craniocaudal diameter × transverse diameter) × 0.52. The operation duration was defined as the time from the initial incision to skin closure. Pelvic lymph node dissection (PLND) was performed in patients with a high likelihood of lymph node involvement, in accordance with the European Association of Urology (EAU) guidelines [12]. Staging was conducted according to the American Joint Committee on Cancer (AJCC) TNM staging system [13].
Surgical technique
All surgical procedures were performed by a dedicated robotic surgery team comprising five board-certified urologists credentialed for robot-assisted laparoscopy. Each surgeon had completed ≥ 200 robotic radical prostatectomies as the primary operator by the study commencement date. The operations adhered to standardized procedural protocols, with senior attending urologists (≥ 5 years postcertification) serving as first assistants.
The standardized RARP protocol implemented at our tertiary referral center comprises the following critical steps: Following general anesthesia, patients are positioned supine with a sacral cushion. After standard prepping and urinary catheterization, pneumoperitoneum is established via a Veress needle at the supraumbilical margin. An 8-mm robotic camera trocar is inserted 2 cm above the umbilicus, followed by the placement of two 12-mm assistant trocars bilaterally at the rectus border and two 8-mm robotic trocars (left/right paraumbilical) for robotic arms 1 and 3. The patient is then placed in the Trendelenburg position, and the robotic cart is docked. Instrumentation includes a 30° endoscope, bipolar forceps (Arm 1), and monopolar scissors (Arm 3). The procedure commences with adhesiolysis (if present), followed by development of the Retzius space through sharp/blunt dissection between the umbilical arteries and pelvic sidewalls, identification of the pubic arches and transection of the vas deferens. After the endopelvic fascia is exposed, the dorsal vascular complex (DVC) is suture-ligated with 2–0 absorbable sutures. Bladder neck dissection is performed while preserving the urethral length. Seminal vesicles and vas deferens are dissected bilaterally via a combination of sharp and cold techniques. Posterior dissection is performed along Denonvilliers ’ fascia to the prostatic apex with careful preservation of neurovascular bundles via athermal techniques. Apical dissection maximizes urethral length preservation before cold transection. PLND is performed in patients with a high likelihood of lymph node involvement. The surgeon dissects along the iliac bifurcation inferiorly between the external iliac artery (lateral) and genitofemoral nerve/spermatic vessels (medial), extending to Cloquet’s node at the femoral canal. Ligate the encountered lymphatics with Hem-o-lock clips before division. This process is repeated contralaterally. A continuous 3–0 barbed suture anastomosis (10–12 stitches) is used to reconstruct the urethrovesical junction over a 20Fr catheter, with anterior reconstruction reinforcing the DVC complex. The pelvic peritoneum is reapproximated with barbed sutures. Hemostasis is verified via digital rectal examination and irrigation tests. A drain is placed before specimen retrieval in an Endobag and fascial closure.
Postoperative follow-up protocol
Patients were evaluated at 1, 3, 6, and 12 months postoperatively at our institution. The clinical staff documented the recovery status and performed standardized physical examinations of the inguinal regions. For patients with clinically suspected masses, confirmatory ultrasonography of the inguinal area was performed. Postoperative inguinal hernia diagnoses were confirmed through either physical examination during follow-up visits or ultrasonographic findings.
Machine learning model development and evaluation
The datasets were randomly partitioned into training (70%) and testing (30%) sets. The training subset underwent further 7:3 splitting for model development (new training) and validation. We implemented tenfold cross-validated least absolute shrinkage and selection operator (LASSO) regression to identify predictors for model inclusion.
Following feature selection, clinical prediction models for 1-year post-RARP inguinal hernia were developed via five ML algorithms: logistic regression, decision tree, random forest, extreme gradient boosting (XGBoost), and the light gradient boosting machine (LightGBM). Hyperparameter optimization was conducted on the validation set before the final evaluation was performed on the test set. Grid search optimization determines the optimal hyperparameters for each algorithm, with receiver operating characteristic area under the curve (ROC-AUC) maximization serving as the primary selection criterion.
Model performance was comprehensively assessed via the ROC–AUC, sensitivity, specificity, precision (positive predictive value), F1 score, and accuracy. Decision curve analysis (DCA) was used to quantify the net clinical benefit across probability thresholds.
SHAP interpretability analysis
To enhance model interpretability, we implemented the SHAP methodology, which is grounded in cooperative game theory’s Shapley values and quantifies feature contributions through additive attribution mechanisms. This approach employs three complementary visualization techniques: histograms for feature value distribution analysis, summary (bee swarm) plots for global feature influence, and waterfall plots for individual prediction deconstruction. Crucially, a feature’s predictive impact correlates directly with its absolute SHAP value magnitude, where larger values signify stronger model influence.
Statistical methods
Continuous data are expressed as the means ± standard deviations (SDs) for normally distributed variables or medians with interquartile ranges (IQRs) for nonnormally distributed variables. Categorical data are presented as frequencies and percentages. Group comparisons were performed via independent-samples t tests for normally distributed continuous variables, chi-square (χ2) tests for categorical variables, and Mann–Whitney U tests for nonnormally distributed continuous variables. The threshold for statistical significance is set at a two-sided P < 0.05. All analyses are conducted via R (version 4.4.3) and Python (version 3.10.4).
Results
Participant enrollment
From June 1, 2021 to May 1, 2023, 823 consecutive patients underwent RARP for prostate cancer at our institution. After excluding 115 patients lost to follow-up, 708 individuals completed the retrospective study protocol. Following the exclusion of 56 patients who did not meet the predefined inclusion criteria, the final analytical cohort comprised 652 patients. Reasons for exclusion included preexisting inguinal hernia before RARP (N = 31), severe and poorly controlled comorbidities that complicated postoperative hernia assessment (N = 10), abdominopelvic or inguinal region surgery with potential to alter abdominal wall integrity following RARP (N = 8), and prolonged occupational engagement in heavy manual labor after RARP (N = 7). Among these patients, 79 patients (12.1%) developed inguinal hernias within the first postoperative year, whereas the remaining 573 patients (87.9%) remained complication-free throughout the 12-month surveillance period. Among 79 patients diagnosed with inguinal hernia, 52 (65.8%) presented right-sided hernias, 13 (16.5%) left-sided, and 14 (17.7%) bilateral cases. Figure 1 graphically displays the participant flow diagram.
Fig. 1.
Flow chart of study population inclusion and exclusion
Basic characteristics
Table 1 compares the baseline characteristics between the training (n = 456) and test (n = 196) sets. The overall cohort (N = 652) had a mean age of 68.5 years (± 6.9 SD), body mass index (BMI) of 23.4 (± 3.1), and an incidence of inguinal hernia of 12.1%. Common comorbidities included hypertension (46.5%), smoking (15.2%), diabetes (12.9%), and prior abdominal surgery (14.3%). Among 93 patients with a history of abdominal surgery, appendectomy was the most common prior procedure (n = 56, 60.2%). Oncologically, 77.5% were stage T1–T2, and 49.2% had a Gleason score of 7. Surgical margins were largely negative (Apical: 87.1%; Basal: 91.7%). Postoperative urinary continence rates gradually improved: 27.6% at 1 month, 65.6% at 3 months, and 77.8% at 6 months. Mean postoperative hospital stay was 6.1 days (SD 2.1). All p values for differences between the training and test sets exceeded 0.05 (range 0.131–1), indicating that there were no statistically significant differences. The distributions of key variables, such as age, BMI, inguinal hernia incidence, comorbidities (hypertension, respiratory disease, diabetes), and tumor characteristics (T stage, ISUP grade, Gleason score), were similar. Continuous measures (e.g., prostate volume, albumin levels) also demonstrated comparable means and SDs across sets. This homogeneity confirms successful randomization for cohort splitting and supports fair comparison for subsequent model training and validation phases.
Table 1.
Basic characteristics
| N | 652 | 456 | 196 | ||
|---|---|---|---|---|---|
| Inguinal hernia, n (%) | No | 573 (87.9) | 401 (87.9) | 172 (87.8) | 1.000 |
| Yes | 79 (12.1) | 55 (12.1) | 24 (12.2) | ||
| Age, mean (SD) | 68.5 (6.9) | 68.3 (7.1) | 69.1 (6.3) | 0.148 | |
| BMI, mean (SD) | 23.4 (3.1) | 23.4 (3.1) | 23.5 (3.1) | 0.675 | |
| Previous abdominal surgery, n (%) | No | 559 (85.7) | 393 (86.2) | 166 (84.7) | 0.706 |
| Yes | 93 (14.3) | 63 (13.8) | 30 (15.3) | ||
| History of TURP, n (%) | No | 592 (90.8) | 413 (90.6) | 179 (91.3) | 0.874 |
| Yes | 60 (9.2) | 43 (9.4) | 17 (8.7) | ||
| Respiratory disease, n (%) | No | 632 (96.9) | 444 (97.4) | 188 (95.9) | 0.461 |
| Yes | 20 (3.1) | 12 (2.6) | 8 (4.1) | ||
| Hypertension, n (%) | No | 349 (53.5) | 249 (54.6) | 100 (51.0) | 0.450 |
| Yes | 303 (46.5) | 207 (45.4) | 96 (49.0) | ||
| Drink, n (%) | No | 573 (87.9) | 405 (88.8) | 168 (85.7) | 0.326 |
| Yes | 79 (12.1) | 51 (11.2) | 28 (14.3) | ||
| Smoke, n (%) | No | 553 (84.8) | 392 (86.0) | 161 (82.1) | 0.259 |
| Yes | 99 (15.2) | 64 (14.0) | 35 (17.9) | ||
| Diabetes, n (%) | No | 568 (87.1) | 395 (86.6) | 173 (88.3) | 0.655 |
| Yes | 84 (12.9) | 61 (13.4) | 23 (11.7) | ||
| Operation duration, mean (SD) | 162.0 (53.6) | 160.5 (53.9) | 165.2 (52.8) | 0.300 | |
| PLND, n (%) | No | 376 (57.7) | 268 (58.8) | 108 (55.1) | 0.434 |
| Yes | 276 (42.3) | 188 (41.2) | 88 (44.9) | ||
| PSA, n (%) | ≤ 20 ng/ml | 402 (61.7) | 286 (62.7) | 116 (59.2) | 0.445 |
| > 20 ng/ml | 250 (38.3) | 170 (37.3) | 80 (40.8) | ||
| T stage, n (%) | T1-2 | 505 (77.5) | 357 (78.3) | 148 (75.5) | 0.499 |
| T3 | 147 (22.5) | 99 (21.7) | 48 (24.5) | ||
| ISUP grade, n (%) | 1 | 101 (15.5) | 64 (14.0) | 37 (18.9) | 0.457 |
| 2 | 177 (27.1) | 130 (28.5) | 47 (24.0) | ||
| 3 | 144 (22.1) | 99 (21.7) | 45 (23.0) | ||
| 4 | 103 (15.8) | 75 (16.4) | 28 (14.3) | ||
| 5 | 127 (19.5) | 88 (19.3) | 39 (19.9) | ||
| Gleason score, n (%) | 6 | 102 (15.6) | 65 (14.3) | 37 (18.9) | 0.287 |
| 7 | 321 (49.2) | 229 (50.2) | 92 (46.9) | ||
| 8 | 103 (15.8) | 75 (16.4) | 28 (14.3) | ||
| 9 | 121 (18.6) | 82 (18.0) | 39 (19.9) | ||
| 10 | 5 (0.8) | 5 (1.1) | 0 (0.0) | ||
| Preoperative albumin, mean (SD) | 43.0 (3.8) | 43.1 (3.9) | 42.8 (3.6) | 0.423 | |
| Preoperative FBG, mean (SD) | 5.9 (1.4) | 5.8 (1.4) | 5.9 (1.3) | 0.408 | |
| Postoperative albumin, mean (SD) | 33.7 (3.0) | 33.6 (3.2) | 33.7 (2.7) | 0.755 | |
| Postoperative FBG, mean (SD) | 6.8 (2.7) | 6.9 (2.8) | 6.5 (2.5) | 0.131 | |
| Hb change, mean (SD) | 18.9 (9.6) | 19.1 (9.4) | 18.4 (10.1) | 0.438 | |
| Transverse diameter, mean (SD) | 39.9 (7.5) | 39.9 (7.6) | 40.1 (7.4) | 0.738 | |
| Anteroposterior diameter, mean (SD) | 45.5 (7.8) | 45.3 (7.9) | 45.8 (7.5) | 0.438 | |
| Craniocaudal diameter, mean (SD) | 39.1 (7.1) | 39.2 (7.1) | 38.7 (7.0) | 0.387 | |
| Prostate volume, mean (SD) | 38,516.9 (18,707.1) | 38,463.2 (18,633.3) | 38,641.8 (18,925.1) | 0.912 | |
| Surgical margin | |||||
| Apical surgical margin, n (%) | Negative | 568 (87.1) | 398 (87.3) | 170 (86.7) | 0.949 |
| Positive | 84 (12.9) | 58 (12.7) | 26 (13.3) | ||
| Basal surgical margin, n (%) | Negative | 598 (91.7) | 418 (91.7) | 180 (91.8) | 1.000 |
| Positive | 54 (8.3) | 38 (8.3) | 16 (8.2) | ||
| Postoperative hospital stay (Days), mean (SD) | 6.1 (2.1) | 6.1 (2.1) | 6.3 (2.1) | 0.120 | |
| Postoperative recovery of urinary continence | |||||
| Urinary continence at 1 month, n(%) | Continence | 180 (27.6) | 127 (27.9) | 53 (27.0) | 0.907 |
| Incontinence | 472 (72.4) | 329 (72.1) | 143 (73.0) | ||
| Urinary continence at 3 months, n(%) | Continence | 428 (65.6) | 303 (66.4) | 125 (63.8) | 0.569 |
| Incontinence | 224 (34.4) | 153 (33.6) | 71 (36.2) | ||
| Urinary continence at 6 months, n(%) | Continence | 507 (77.8) | 355 (77.9) | 152 (77.6) | 1.000 |
| Incontinence | 145 (22.2) | 101 (22.1) | 44 (22.4) |
SD standard deviation, BMI body mass index, TURP transurethral resection of the prostate, PLND pelvic lymph node dissection, PSA prostate-specific antigen, ISUP International Society of Urological Pathology, FBG fasting blood glucose
Predictive variable screening
Through LASSO regression cross-validation, two regularization parameters were identified: λmin (0.0104394) and λ1 se (0.0318806). To enhance model generalizability and mitigate overfitting, λ1 se (0.0318806) was selected as the optimal regularization parameter. This yielded five significant predictors: age, BMI, history of abdominal surgery, preoperative albumin level, and T stage. The complete variable selection process is illustrated in Fig. 2.
Fig. 2.
Selection of Clinical Features via least absolute shrinkage and selection operator (LASSO) regression. A Variable selection trace plot across multiple candidate clinical features. B Binomial deviance curve versus log(lambda). Dashed vertical lines indicate optimal λ values selected by minimum criteria (left) and one-standard-error rule (right)
Machine learning models and evaluation
By employing LASSO regression for feature selection, we identified five significant predictors associated with 1-year postoperative inguinal hernia following RARP. Utilizing these key variables, we then constructed and evaluated five distinct ML prediction models.
ROC analysis demonstrated robust discriminative ability for all the models in both the validation and test sets (Fig. 3A, B). XGBoost achieved the highest performance, with AUCs of 0.833 in the validation cohort and 0.791 in the test cohort, significantly outperforming logistic regression and other ML models. Consistent performance on unseen data underscores the model’s clinical generalizability. DCA revealed that the XGBoost model consistently achieved the highest net benefit across clinically relevant threshold probabilities in both the validation and test sets (Fig. 3C, D). Its clinical utility surpassed that of the "treat-all" and "treat-none" strategies, demonstrating robust potential for clinical decision support.
Fig. 3.
Comparative performance evaluation of five machine learning models. A Receiver operating characteristic (ROC) curves in validation cohort; B ROC curves in independent test cohort; C Decision curve analysis (DCA) curves assessing clinical utility (validation cohort); D DCA curves assessing clinical utility (test cohort)
On the basis of the results of DCA, the probability threshold for predictive classification was set at 0.15 to maximize the clinical net benefit. Model performance metrics across both the validation and independent test sets are detailed in Table 2. Crucially, model utility for clinical application was evaluated primarily on the basis of generalization performance within the test set (i.e., unseen data). In this critical assessment, the XGBoost model demonstrates compelling superiority: it achieved the highest accuracy (0.832), signifying optimal overall classification capability, along with the highest specificity (0.884), indicating exceptional reliability in identifying negative cases (e.g., ruling out nonevents), which is a critical advantage for minimizing false positives. Notably, XGBoost also attained the top position in the core discriminative metric (AUC: 0.791, 95% CI: 0.734–0.848), outperforming other models. While all the models exhibited the expected generalization decay in the test set (e.g., the XGBoost AUC decreased by 0.042), XGBoost maintained superior performance across key test set metrics compared with alternative models (logistic regression, decision tree, random forest, LightGBM). Consequently, XGBoost is definitively established as the optimal model for deployment on unseen clinical data.
Table 2.
Evaluation of model performance in the validation set and test set
| Data set | Model | AUC | AUC 95% CI Lower | AUC 95% CI Upper | Accuracy | Precision | Sensitivity | Specificity | F1 Score |
|---|---|---|---|---|---|---|---|---|---|
| Validation set | Logistic | 0.843 | 0.810 | 0.877 | 0.805 | 0.351 | 0.727 | 0.815 | 0.473 |
| Decision Tree | 0.759 | 0.687 | 0.831 | 0.839 | 0.391 | 0.529 | 0.883 | 0.450 | |
| Random Forest | 0.742 | 0.669 | 0.816 | 0.759 | 0.289 | 0.647 | 0.775 | 0.400 | |
| XGBoost | 0.833 | 0.770 | 0.895 | 0.854 | 0.429 | 0.529 | 0.900 | 0.474 | |
| LightGBM | 0.793 | 0.725 | 0.861 | 0.818 | 0.278 | 0.294 | 0.892 | 0.286 | |
| Test set | Logistic | 0.781 | 0.723 | 0.839 | 0.781 | 0.306 | 0.625 | 0.802 | 0.411 |
| Decision Tree | 0.702 | 0.638 | 0.766 | 0.791 | 0.282 | 0.458 | 0.837 | 0.349 | |
| Random Forest | 0.760 | 0.700 | 0.819 | 0.740 | 0.245 | 0.542 | 0.767 | 0.338 | |
| XGBoost | 0.791 | 0.734 | 0.848 | 0.832 | 0.355 | 0.458 | 0.884 | 0.400 | |
| LightGBM | 0.769 | 0.710 | 0.828 | 0.811 | 0.341 | 0.583 | 0.843 | 0.431 |
AUC area under the curve, XGBoost extreme gradient boosting, LightGBM light gradient boosting machine
SHAP analysis and model interpretability
Figure 4A displays the SHAP-derived feature importance for the XGBoost model, which ranks the predictors by the mean absolute SHAP values (descending impact order). The five most influential features were age, BMI, preoperative albumin level, T stage, and history of abdominal surgery. The complementary analysis in Fig. 4B illustrates directional feature effects through a SHAP summary plot. Higher positive SHAP values were correlated with increased post-RARP inguinal hernia probability. Specifically, older age, advanced T stage, and prior abdominal surgery consistently resulted in positive SHAP values, indicating elevated hernia risk. Conversely, negative SHAP values for BMI and preoperative albumin suggested protective effects against hernia development.
Fig. 4.
SHapley Additive exPlanations (SHAP) analysis for the extreme gradient boosting (XGBoost) model. A Feature importance ranked by mean absolute SHAP value. B Summary plot displaying SHAP value distributions across features (dots = per-patient observations; color gradient: red = high feature value, blue = low). C, D Waterfall plots of representative cases: (C) true-positive patient and (D) true-negative patient. For both plots: red/blue bars indicate positive/negative feature contributions respectively; bar length denotes contribution magnitude
We employed the SHAP methodology to deconstruct individual predictions from our optimal model via waterfall plots (Fig. 4C, D). These visualizations encode feature contributions through horizontal bar length (magnitude) and color directionality (red = risk-enhancing, blue = protective). Figure 4C shows risk accumulation in a patient who developed post-RARP inguinal hernia, with significantly positive SHAP contributions from age (+ 0.19), T stage (+ 0.08), BMI (+ 0.06), and postoperative albumin (+ 0.05), yielding an elevated predicted risk of 0.495. Conversely, Fig. 4D depicts a hernia-free case exhibiting protective influences: postoperative albumin level (-0.02), BMI (-0.02), age (-0.01), and T stage (-0.01), resulting in a minimal predicted risk of 0.051.
Discussion
To our knowledge, this study is the first to apply ML and SHAP methodology to develop a clinical predictive model for post-RARP inguinal hernia. We developed a predictive model using five ML algorithms, among which the XGBoost model demonstrated the best prediction performance. Through SHAP analysis, we identified five key predictors: age, BMI, preoperative albumin level, T stage, and previous abdominal surgery. Furthermore, interpretability analysis of the XGBoost model revealed the specific contributions of each predictor to the model’s decisions, providing an in-depth understanding of the main factors affecting the incidence of post-RARP inguinal hernia. These results offer a clear direction for clinical intervention.
The incidence of inguinal hernia following RARP is a significant clinical issue. A meta-analysis indicated that RARP has a lower inguinal hernia rate (7.9%; 95% CI 5.0–10.9) than does ORP (13.7%; 95% CI 12.0–15.4). However, patients who undergo any prostatectomy have significantly higher hernia rates than untreated controls do [7]. Another study reported RARP-associated hernia incidences of 11.3%, 14.0%, and 15.4% at 1, 2, and 3 years postoperatively, respectively [14]. Toide et al. utilized a multicenter database of 3,195 Japanese patients undergoing RARP and reported cumulative incidences of postoperative inguinal hernia of 5.7% at 1 year, 8.3% at 2 years, and 9.5% at 3 years [15]. Our results revealed a 12.1% cumulative incidence of inguinal hernia within 1 year after RARP, which aligns with the established literature.
Inguinal hernias rank among the most prevalent surgical conditions, with serious implications. They may lead to acute complications, such as bowel obstruction and intestinal necrosis, requiring emergency intervention to prevent life-threatening outcomes [8, 9]. While hernia repair is common, it carries substantial risks. Postoperative complications include recurrence and chronic pain, both of which significantly impair quality of life [16]. Recurrence rates deserve particular attention, as studies have demonstrated that repeat repairs markedly increase the likelihood of complications, including persistent pain and reherniation [17]. In summary, inguinal hernia after RARP not only compromises patients ’ quality of life but also may lead to serious clinical complications. Therefore, there is a critical clinical need for predictive models to stratify high-risk patients early, enabling personalized interventions to optimize outcomes.
The risk factors for inguinal hernia after radical prostatectomy can be categorized into three main aspects: preoperative, intraoperative, and postoperative. Preoperative factors included advanced age, lower BMI, a history of previous abdominal surgeries, a history of previous inguinal hernia repair, and the preoperative International Prostate Symptom Score (IPSS). Intraoperative factors include the presence of a patent processus vaginalis (PPV) during surgery; surgical methods (ORP, LRP, or RARP); and surgical approaches (transabdominal or extraperitoneal). Postoperative factors included urinary continence recovery and anastomotic stricture development following radical prostatectomy [18, 19].
SHAP bar plot analysis (Fig. 4) revealed that age was associated with the highest mean absolute SHAP value, indicating that advanced age was the most significant risk factor for inguinal hernia following RARP. This finding aligns with previous reports in the literature [20, 21]. Patients of advanced age may have weak supporting connective tissue around the internal inguinal ring and weak abdominal muscles. Additionally, our study confirms that a low BMI is a significant risk factor for inguinal hernia development after RARP, which is consistent with the findings of prior studies. Notably, individuals with a lower BMI are more likely to develop inguinal hernias, whereas individuals with a higher BMI may be somewhat protected, as body fat may act as a "plug", preventing abdominal contents from entering the inguinal canal [22, 23].
SHAP analysis (Fig. 4) indicated that patients with low preoperative albumin levels are more prone to developing inguinal hernias following RARP. Studies have confirmed that patients with hypoalbuminemia upon hospital admission have significantly increased short- and long-term mortality rates, whereas those whose albumin levels normalize before discharge have a significantly lower mortality risk [24]. Furthermore, hypoalbuminemia is prevalent among elderly inpatients and is correlated with malnutrition and prolonged hospitalization. Consequently, monitoring albumin levels during hospitalization is recommended to assess the risk of malnutrition and other complications [25]. Research indicates that preoperative albumin assessment in complex abdominal wall reconstruction surgeries aids in predicting postoperative complication risks. Specifically, subnormal albumin levels predispose patients to delayed wound healing, skin necrosis, and other factors contributing to hernia formation [26]. We postulate that diminished serum albumin levels correlate with suboptimal nutritional status and impaired postoperative recovery, concurrently exacerbating abdominal wall structural weakness. This multifactorial predisposition substantially elevates hernia risk following urological procedures. In addition, our analysis revealed that pathological T3-stage prostate cancer patients are significantly more susceptible to post-RARP inguinal hernia than their T1/T2-stage counterparts are. We hypothesize that this elevated risk stems from extracapsular tumor extension, which compromises endopelvic fascia integrity through collagen disruption, thereby weakening this critical support structure of the posterior inguinal canal and diminishing abdominal pressure resistance. Compounding factors include advanced age and poorer nutritional status, which are frequently observed in patients with higher T stages. Concurrently, patients with prior abdominal surgery exhibit increased post-RARP hernia susceptibility, with appendectomy representing the most prevalent surgical history. This observation aligns with previous research. In a cohort of 205 prostate cancer patients who underwent RARP, Lee et al. [27] identified prior abdominal surgery as an independent risk factor for postoperative inguinal hernia development. We propose that prior abdominal operations may disrupt the structural integrity of the abdominal wall, rendering patients more vulnerable to developing inguinal hernias following RARP.
This model identifies patients at high risk for inguinal hernia following RARP, guiding informed consent and preventive strategies implemented across three phases. Preoperatively, patients and families should be counseled regarding the potential risk and available preventive options. Intraoperatively, for patients with financial feasibility, prophylactic mesh placement is recommended as a safe, effective preventive measure during RARP [28, 29]. Alternative surgical approaches like extraperitoneal approach RARP (EP-RARP) [30] or Retzius-sparing RARP (RS-RARP) [31, 32] should also be considered for eligible patients with low-risk prostate cancer and favorable surgical anatomy. Additionally, modifying dissection techniques around the internal inguinal ring provides a simple intraoperative preventive step [33, 34]. Postoperatively, patients require structured nutritional support and scheduled inguinal surveillance. For instance, in a patient with a 50% predicted risk by the model, preoperative counseling would address risk disclosure and prevention alternatives; the primary intraoperative strategy would prioritize mesh placement if financially viable, with alternative surgical approaches considered for eligible patients or modified dissection techniques employed if others are contraindicated; followed by postoperative nutritional optimization and regular inguinal examinations.
This study acknowledges several inherent limitations. First, as a single-center retrospective investigation, the generalizability of our findings may be constrained by uncontrolled confounders characteristic of observational designs, while causal inferences remain inherently limited. Consequently, potentially relevant predictors—including intraoperative factors, such as patent processus vaginalis status and certain postoperative parameters—were unavailable for analysis, precluding their inclusion. Furthermore, due to resource considerations, ultrasonography was selectively performed only in patients with abnormal physical findings or symptoms, a pragmatic approach that risks under-detection of subclinical hernias and introduces potential detection bias. Additionally, external validation remains absent, necessitating future large-scale, multicenter prospective studies for robust verification. Subsequent investigations should incorporate expanded predictor variables, employ standardized diagnostic protocols, and implement advanced machine learning algorithms to optimize both predictive performance and clinical translational utility.
Conclusion
This study pioneers an interpretable ML model for predicting 1-year post-RARP inguinal hernia. The XGBoost-based tool demonstrated robust performance (AUC: 0.833 validation/0.791 test), outperforming other methods. SHAP analysis identified five key predictors: advanced age, lower BMI, low preoperative albumin level, higher T stage, and history of abdominal surgery. This model enables early identification of high-risk patients, guiding personalized informed consent and preventive strategies. Future multicenter validation is warranted to enhance its clinical adoption.
Author contributions
All authors contributed to the study conception and design. Material preparation, data collection, and analysis were performed by Weidong Yu, You Ma, Junchao Wu, Meng Zhang, and Cheng Yang. The first draft of the manuscript was written by Weidong Yu, and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.
Funding
National Natural Science Foundation of China, 82270818, Outstanding Scientific Research and Innovation Team for Male Genitourinary Diseases in Anhui Provincial Universities, 2022AH010071, Research Funds of the Center for Big Data and Population Health of IHM,JKS2022001.
Data availability
No datasets were generated or analyzed during the current study.
Declarations
Conflict of interest
The authors have no relevant financial or non-financial interests to disclose. The authors declare no competing interests.
Ethical approval
The study protocol received full ethical approval from the Institutional Review Board of Anhui Medical University (Approval No. PJ2024-12–93). All procedures were conducted in accordance with local legislation and institutional requirements.
Consent to participate
The ethics committee/institutional review board waived the requirement of written informed consent for participation from the participants because this study is a retrospective study.
Footnotes
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Weidong Yu, You Ma, and Junchao Wu contributed equally to this work.
References
- 1.Siegel RL, Giaquinto AN, Jemal A (2024) Cancer statistics, 2024. CA Cancer J Clin 74(1):12–49. 10.3322/caac.21820 [DOI] [PubMed] [Google Scholar]
- 2.Carlsson S, Nilsson AE, Schumacher MC et al (2010) Surgery-related complications in 1253 robot-assisted and 485 open retropubic radical prostatectomies at the Karolinska University Hospital, Sweden. Urology 75(5):1092–1097. 10.1016/j.urology.2009.09.075 [DOI] [PubMed] [Google Scholar]
- 3.Kadono Y, Nohara T, Kawaguchi S et al (2022) Impact of pelvic anatomical changes caused by radical prostatectomy. Cancers 14(13):3050. 10.3390/cancers14133050 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Novara G, Ficarra V, Rosen RC et al (2012) Systematic review and meta-analysis of perioperative outcomes and complications after robot-assisted radical prostatectomy. Eur Urol 62(3):431–452. 10.1016/j.eururo.2012.05.044 [DOI] [PubMed] [Google Scholar]
- 5.Abdulhai S, Glenn IC, Ponsky TA (2017) Inguinal hernia. Clin Perinatol 44(4):865–877. 10.1016/j.clp.2017.08.005 [DOI] [PubMed] [Google Scholar]
- 6.Regan TC, Mordkin RM, Constantinople NL, Spence IJ, Dejter SWJ (1996) Incidence of inguinal hernias following radical retropubic prostatectomy. Urology 47(4):536–537. 10.1016/S0090-4295(99)80491-9 [DOI] [PubMed] [Google Scholar]
- 7.Alder R, Zetner D, Rosenberg J (2020) Incidence of inguinal hernia after radical prostatectomy: a systematic review and meta-analysis. J Urol 203(2):265–274. 10.1097/JU.0000000000000313 [DOI] [PubMed] [Google Scholar]
- 8.Zhou J, Yuan X (2024) Establishment of a risk prediction model for bowel necrosis in patients with incarcerated inguinal hernia. BMC Med Inform Decis Mak 24(1):39. 10.1186/s12911-024-02440-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Ge BJ, Huang Q, Liu LM, Bian HP, Fan YZ (2010) Risk factors for bowel resection and outcome in patients with incarcerated groin hernias. Hernia 14(3):259–264. 10.1007/s10029-009-0602-2 [DOI] [PubMed] [Google Scholar]
- 10.Moretti TBC, Magna LA, Reis LO (2024) Continence criteria of 193 618 patients after open, laparoscopic, and robot-assisted radical prostatectomy. BJU Int 134(1):13–21. 10.1111/bju.16180 [DOI] [PubMed] [Google Scholar]
- 11.Moretti TBC, Magna LA, Reis LO (2024) Erectile dysfunction criteria of 131,350 patients after open, laparoscopic, and robotic radical prostatectomy. Andrology 12(8):1865–1871. 10.1111/andr.13634 [DOI] [PubMed] [Google Scholar]
- 12.European Association of Urology (EAU) guidelines on Prostate Cancer. https://uroweb.org/guidelines/prostate-cancer/chapter/treatment. Accessed Dec 2024.
- 13.Chen N, Zhou Q (2016) The evolving gleason grading system. Chin J Cancer Res 28(1):58–64. 10.3978/j.issn.1000-9604.2016.02.04 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Yamada Y, Fujimura T, Fukuhara H et al (2017) Incidence and risk factors of inguinal hernia after robot-assisted radical prostatectomy. World J Surg Oncol 15(1):61. 10.1186/s12957-017-1126-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Toide M, Ebara S, Tatenuma T et al (2024) Incidence and risk factors of inguinal hernia after robot-assisted radical prostatectomy: a retrospective multicenter cohort study in Japan (the MSUG94 group). J Robot Surg. 10.1007/s11701-023-01761-1 [DOI] [PubMed] [Google Scholar]
- 16.Konschake M, Zwierzina M, Moriggl B et al (2020) The inguinal region revisited: the surgical point of view. Hernia 24(4):883–894. 10.1007/s10029-019-02070-z [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Köckerling F, Krüger C, Gagarkin I et al (2020) What is the outcome of re-recurrent vs recurrent inguinal hernia repairs? An analysis of 16,206 patients from the Herniamed Registry. Hernia 24(4):811–819. 10.1007/s10029-020-02138-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Stranne J, Lodding P, Sahlgrenska A et al (2011) Inguinal hernia after radical retropubic prostatectomy: risk factors and prevention. Nat Rev Urol 8(5):267–273. 10.1038/nrurol.2011.40 [DOI] [PubMed] [Google Scholar]
- 19.Fernando H, Garcia C, Hossack T et al (2019) Incidence, predictive factors and preventive measures for inguinal hernia following robotic and laparoscopic radical prostatectomy: a systematic review. J Urol 201(6):1072–1079. 10.1097/JU.0000000000000133 [DOI] [PubMed] [Google Scholar]
- 20.Pans A, Albert A, Lapiere CM, Nusgens B (2001) Biochemical study of collagen in adult groin hernias. J Surg Res 95(2):107–113. 10.1006/jsre.2000.6024 [DOI] [PubMed] [Google Scholar]
- 21.Wagh PV, Read RC (1972) Defective collagen synthesis in inguinal herniation. Am J Surg 124(6):819–822. 10.1016/0002-9610(72)90148-1 [DOI] [PubMed] [Google Scholar]
- 22.Rosemar A, Angeras U, Rosengren A (2008) Body mass index and groin hernia: a 34-year follow-up study in Swedish men. Ann Surg 247(6):1064–1068. 10.1097/SLA.0b013e31816b4399 [DOI] [PubMed] [Google Scholar]
- 23.Zendejas B, Hernandez-Irizarry R, Ramirez T et al (2014) Relationship between body mass index and the incidence of inguinal hernia repairs: a population-based study in Olmsted County, MN. Hernia 18(2):283–288. 10.1007/s10029-013-1185-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Akirov A, Masri-Iraqi H, Atamna A, Shimon I (2017) Low albumin levels are associated with mortality risk in hospitalized patients. Am J Med 130(12):1411–1465. 10.1016/j.amjmed.2017.07.020 [Google Scholar]
- 25.Brock F, Bettinelli LA, Dobner T et al (2016) Prevalence of hypoalbuminemia and nutritional issues in hospitalized elders. Rev Lat Am Enfermagem. 10.1590/1518-8345.0260.2736 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Ash M, Marxen T, Su S, Losken A (2025) The modified fragility index and perioperative albumin as predictors of complications in complex abdominal wall reconstruction. Ann Plast Surg 94(4):457–461. 10.1097/SAP.0000000000004316 [DOI] [PubMed] [Google Scholar]
- 27.Lee DH, Jung HB, Chung MS, Lee SH, Chung BH (2013) Patent processus vaginalis in adults who underwent robot-assisted laparoscopic radical prostatectomy: predictive signs of postoperative inguinal hernia in the internal inguinal floor. Int J Urol 20(2):177–182. 10.1111/j.1442-2042.2012.03118.x [DOI] [PubMed] [Google Scholar]
- 28.Li H, Ren L, Wang C et al (2025) A meta-analysis assessing the efficacy of concurrent robotic inguinal hernia repair (RIHR) during robotic-assisted radical prostatectomy (RARP). J Robot Surg 19(1):302. 10.1007/s11701-025-02437-8 [DOI] [PubMed] [Google Scholar]
- 29.Jaber AR, Moschovas MC, Rogers T et al (2023) Simultaneous hernia repair following robotic-assisted radical prostatectomy is safe with low rates of mesh-related complications. J Robot Surg 17(4):1653–1658. 10.1007/s11701-023-01574-2 [DOI] [PubMed] [Google Scholar]
- 30.Uy M, Cassim R, Kim J et al (2022) Extraperitoneal versus transperitoneal approach for robot-assisted radical prostatectomy: a contemporary systematic review and meta-analysis. J Robot Surg 16(2):257–264. 10.1007/s11701-021-01245-0 [DOI] [PubMed] [Google Scholar]
- 31.Chang KD, Abdel Raheem A, Santok GDR et al (2017) Anatomical Retzius-space preservation is associated with lower incidence of postoperative inguinal hernia development after robot-assisted radical prostatectomy. Hernia 21(4):555–561. 10.1007/s10029-017-1588-9 [DOI] [PubMed] [Google Scholar]
- 32.Kowalczyk KJ, Davis M, O Neill J et al (2020) Impact of Retzius-sparing versus standard robotic-assisted radical prostatectomy on penile shortening, Peyronie’s disease, and inguinal hernia sequelae. Eur Urol Open Sci 22:17–22. 10.1016/j.euros.2020.09.004 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Soma T, Fukuda S, Matsuyama Y et al (2023) Peritoneal closure and the processus vaginalis transection method to prevent inguinal hernia after robot-assisted radical prostatectomy. Int J Urol 30(6):532–538. 10.1111/iju.15170 [DOI] [PubMed] [Google Scholar]
- 34.Shimbo M, Endo F, Matsushita K et al (2017) Incidence, risk factors and a novel prevention technique for inguinal hernia after robot-assisted radical prostatectomy. Urol Int 98(1):54–60. 10.1159/000448339 [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
No datasets were generated or analyzed during the current study.