{"paper_id":"130f04ea-281e-4965-b959-78a0ea4f0b08","body_text":"Abstract\nBackground\nChronic post-surgical pelvic pain syndrome (CPSPP) represents a significant clinical challenge affecting a substantial proportion of patients undergoing pelvic surgical procedures. The complex pathophysiology and multifactorial nature of CPSPP necessitate advanced predictive approaches to improve patient outcomes and optimize treatment strategies.\nObjective\nThis study aimed to develop and validate a machine learning-based prediction model for CPSPP using comprehensive clinical data, with particular emphasis on model interpretability through SHAP (SHapley Additive exPlanations) analysis to identify key risk factors and enhance clinical decision-making.\nMethods\nA retrospective cohort study was conducted involving 62 patients who underwent pelvic surgical procedures. Comprehensive clinical data including demographic characteristics, surgical parameters, pain assessments, and postoperative outcomes were collected. Multiple machine learning algorithms were employed to develop predictive models, with performance evaluation using receiver operating characteristic (ROC) analysis. SHAP values were utilized to provide model interpretability and identify the most influential predictive features.\nResults\nA total of 62 female patients were analyzed, with a mean age of 47.6 ± 11.8 years and BMI of 26.1 ± 4.9 kg/m2. Following magnetic stimulation therapy, 27.4% achieved significant clinical improvement, with a mean VAS reduction of 2.05 ± 2.03 points. Among the four machine learning models evaluated, XGBoost achieved the highest performance with an AUC of 0.94 in the training set and 0.77 in the validation set. SHAP analysis identified baseline VAS score, stimulation frequency, and intensity as the most influential predictors of treatment response.\nConclusions\nThis study presents a novel machine learning approach for predicting CPSPP treatment outcomes with enhanced interpretability through SHAP analysis. The findings contribute to improved understanding of CPSPP risk factors and provide a foundation for personalized treatment strategies and clinical decision support systems.\nGraphical Abstract\nSimilar content being viewed by others\nIntroduction\nCPSPP represents one of the most challenging complications in modern surgical practice, affecting a significant proportion of patients undergoing pelvic surgical procedures [1]. This debilitating condition is characterized by persistent pain lasting more than three to six months following surgery, significantly impacting patients’ quality of life, functional capacity, and psychological well-being [2]. The prevalence of CPSPP varies considerably across different surgical procedures and patient populations, with reported rates ranging from 10 to 50% depending on the specific surgical intervention and diagnostic criteria employed [3].\nThe pathophysiology of CPSPP is multifactorial and complex, involving intricate interactions between inflammatory processes, neurological sensitization, psychological factors, and anatomical considerations [4]. CPSPP appears to arise through a complex interaction of inflammatory, infectious, neurological, musculoskeletal, and psychosomatic factors, as demonstrated in recent comprehensive reviews [5]. The development of chronic pain following pelvic surgery involves both peripheral and central sensitization mechanisms, where initial tissue damage and inflammation can lead to persistent alterations in pain processing pathways [6].\nTraditional approaches to predicting and managing CPSPP have relied primarily on clinical experience and conventional risk assessment tools, which often lack the precision and comprehensiveness required for optimal patient care [7]. The heterogeneous nature of CPSPP, combined with the multitude of potential risk factors, presents significant challenges for clinicians in identifying high-risk patients and implementing appropriate preventive strategies [8]. Furthermore, the subjective nature of pain assessment and the complex interplay of biological, psychological, and social factors contribute to the difficulty in developing effective predictive models using conventional statistical approaches [9].\nThe emergence of machine learning (ML) technologies has opened new avenues for addressing these challenges in healthcare, offering sophisticated analytical capabilities that can handle complex, high-dimensional datasets and identify subtle patterns that may not be apparent through traditional statistical methods [10]. Machine learning algorithms have demonstrated remarkable success in various medical applications, including diagnostic imaging, treatment response prediction, and risk stratification across multiple clinical domains [11]. In the context of chronic pain research, ML approaches have shown promise in identifying predictive biomarkers, classifying pain phenotypes, and developing personalized treatment strategies [12].\nHowever, one of the primary limitations of traditional machine learning models in clinical applications has been their “black box” nature, where the decision-making process remains opaque and difficult to interpret [13]. This lack of interpretability poses significant challenges for clinical adoption, as healthcare providers require clear understanding of the factors driving model predictions to make informed clinical decisions and maintain patient trust [14]. The development of explainable artificial intelligence (XAI) methods has emerged as a critical advancement in addressing these limitations, with SHAP (SHapley Additive exPlanations) representing one of the most robust and widely adopted approaches for model interpretability [15].\nSHAP values provide a unified framework for understanding machine learning model predictions by quantifying the contribution of each feature to individual predictions, based on cooperative game theory principles [16]. This approach enables clinicians to understand not only what the model predicts but also why it makes specific predictions, facilitating better clinical decision-making and enhancing model trustworthiness [17]. In the context of CPSPP prediction, SHAP analysis can provide valuable insights into the relative importance of different risk factors, identify patient-specific risk profiles, and guide personalized treatment strategies [18].\nRecent advances in CPSPP research have highlighted the importance of comprehensive, multidisciplinary approaches that consider the complex interplay of biological, psychological, and social factors [19]. The integration of machine learning with clinical expertise offers unprecedented opportunities to develop more accurate, interpretable, and clinically useful predictive models for CPSPP [20]. Such models can potentially improve patient outcomes by enabling early identification of high-risk patients, facilitating targeted interventions, and optimizing resource allocation in healthcare settings [21].\nAn interpretable machine learning model was developed to predict chronic post‑surgical pelvic pain (CPSPP) using the XGBoost‑SHAP framework. Key innovations include the application of L1 regularization and fivefold cross‑validation to ensure robust performance with a limited single‑center sample (n = 62). Unlike deep‑learning explainers, this approach provides precise feature attribution without requiring large‑scale data. The model uniquely identified a nonlinear interaction between baseline VAS scores and stimulation frequency, offering a quantitative basis for personalized treatment adjustment. By integrating predictive accuracy with explainability, this study provides a practical clinical tool for risk assessment and advances the mechanistic understanding of CPSPP.\nMethods\nStudy design and population\nThis retrospective cohort study developed and validated a machine learning model for chronic post‑surgical pelvic pain syndrome (CPSPP). Patients undergoing pelvic surgery between January 2020 and December 2024 were included if they were adults (≥ 18 years) with complete perioperative data. Exclusions included incomplete records, loss to follow‑up within six months, or pre‑existing chronic pain conditions. The institutional review board approved the protocol; informed consent was waived due to the retrospective design.\nData collection and variables\nDemographic, surgical, and pain‑assessment data were extracted from electronic records. The primary outcome was CPSPP, defined as pelvic pain > 6 months post‑surgery with clinically significant VAS scores. Variables were selected based on clinical relevance, literature evidence, and expert consensus to balance comprehensiveness and model parsimony.\nMachine learning model development\nMultiple algorithms (logistic regression, random forest, gradient boosting, support vector machines) were compared. The dataset was split 70:30 (training:testing) with stratification.\nAlgorithm selection was driven by the need to handle non‑linear relationships and interaction effects typically present in clinical pain data.\nFeature engineering included clinically informed interaction terms and composite variables. Missing data were handled by multiple imputation; categorical variables were encoded and continuous variables standardized.\nRegularization (L1) and fivefold cross‑validation were applied to control overfitting, given the moderate sample size.\nModel evaluation and performance assessment\nPerformance was assessed using AUC‑ROC, sensitivity, specificity, and calibration plots. The DeLong test compared model discrimination. Bootstrap resampling (n = 1000) validated AUC stability. An independent validation set was used to evaluate generalizability.\nSHAP analysis for model interpretability\nSHAP (SHapley Additive exPlanations) provided global and local explanations. Mean absolute SHAP values ranked feature importance; partial dependence and interaction plots illustrated feature effects and synergies. SHAP was chosen for its model‑agnostic nature and ability to quantify both main and interactive effects in clinically intuitive units.\nStatistical analysis\nAnalyses used Python (scikit‑learn, XGBoost, SHAP). Descriptive statistics summarized variables; univariate tests identified candidate predictors. Within the ML framework, multivariate modeling captured complex, non‑linear associations. Significance was set at p < 0.05 (two‑tailed, with correction for multiple comparisons).\nResults\nPatient demographics and clinical characteristics\nA total of 62 patients were included in this study, and their comprehensive demographic and clinical characteristics are summarized in Table 1. The study cohort had a mean age was 47.6 years and a mean body mass index (BMI) of 26.10 ± 4.86 kg/m2 (median: 25.9, IQR 21.5–30.0).\nRegarding menopausal status, 33 patients (53.2%) were pre-menopausal and 29 patients (46.8%) were post-menopausal, as illustrated in Fig. 1. The distribution of marital status showed that 19 patients (30.6%) were widowed, 16 (25.8%) were single, 14 (22.6%) were divorced, and 13 (21.0%) were married.\nThe most common procedures were categorized as “Other” surgeries (15 patients, 24.2%) and appendectomy (15 patients, 24.2%), followed by oophorectomy (14 patients, 22.6%), cesarean Sect. (12 patients, 19.4%), and hysterectomy (6 patients, 9.7%). Regarding prior treatment history, 18 patients (29.0%) had received “Other” treatments, 15 (24.2%) had received injections, 11 (17.7%) had received medication, and 3 (4.8%) had undergone physical therapy (Fig. 2).\nPain assessment and treatment parameters\nComprehensive data on pain assessment and treatment parameters are summarized in Table 2. At baseline, patients reported a mean Visual Analog Scale (VAS) score of 6.98 ± 1.43, indicating moderate to severe pain intensity. Quality of life, assessed using the Short Form-36 (SF-36), showed a mean baseline score of 50.10 ± 12.61, reflecting impaired functional status. Psychological evaluation revealed elevated depression levels (PHQ-9: 10.45 ± 4.20) and anxiety scores (11.85 ± 3.91), consistent with the substantial psychological burden associated with CPSPP.\nFollowing magnetic stimulation therapy, both statistically and clinically significant improvements were observed across multiple outcome domains. As illustrated in Fig. 3, patients categorized as having achieved “Significant Improvement” exhibited marked reductions in post-treatment VAS scores compared to baseline values, with data points predominantly clustering below the diagonal reference line. The mean post-treatment VAS score decreased to 4.94 ± 2.70, representing a mean reduction of 2.05 ± 2.03 points—indicative of meaningful pain alleviation in a substantial proportion of the cohort. This visual trend reinforces the objective measurements of treatment efficacy. Furthermore,\nFigure 4 displays a paired boxplot comparing pre- and post-treatment VAS scores, demonstrating a clear downward shift in both the median and interquartile range following intervention. This graphical representation underscores not only the overall therapeutic effect, but also the consistency of treatment response across the patient population.\nTreatment parameters varied across the study cohort. The mean stimulation frequency was 18.39 ± 8.53 Hz, and the mean stimulation intensity was 83.95 ± 9.65%. The average session duration was 25.81 ± 4.97 min, with patients undergoing an average of 10.90 ± 3.91 treatment sessions. Regarding stimulation targets, the majority of patients (n = 37, 59.7%) received stimulation directed at the sacral nerve area, while 25 patients (40.3%) received pubic symphysis stimulation.\nTreatment outcomes\nTreatment outcomes demonstrated a spectrum of therapeutic responses within the study population. As illustrated in Fig. 5 and Table 3, 18 patients (29.0%) reported slight improvement, and 16 patients (25.8%) experienced no change in their condition. A total of 11 patients (17.7%) reported a worsening of symptoms following treatment. In contrast, 9 patients (14.5%) achieved moderate improvement, and 8 patients (12.9%) experienced significant improvement. When reclassified according to predefined response criteria, 45 patients (72.6%) were categorized as having no significant improvement, whereas 17 patients (27.4%) achieved a clinically meaningful therapeutic benefit. This distribution highlights the heterogeneous response pattern observed with magnetic stimulation in this specific clinical context.\nAdverse events were generally mild and infrequent. The majority of patients (48, 77.4%) experienced no adverse events, while 13 patients (21.0%) reported mild adverse events, and only 1 patient (1.6%) experienced moderate adverse events. No severe adverse events were reported in this cohort (Table 3).\nMachine learning model performance\nAs shown in Fig. 6A, all six machine learning models demonstrated high predictive accuracy on the training set, with XGBoost achieving the highest area under the ROC curve (AUC = 0.94), followed closely by LightGBM (AUC = 0.93) and Random Forest (AUC = 0.91). Logistic regression and SVM showed relatively lower AUCs of 0.77 and 0.78, respectively, while the random classifier yielded an AUC of 0.50.\nOn the independent validation set (Fig. 6B), the XGBoost model maintained robust performance (AUC = 0.77), comparable to Random Forest (AUC = 0.79) and superior to LightGBM (AUC = 0.74). In contrast, traditional models such as logistic regression (AUC = 0.66) and SVM (AUC = 0.22) showed reduced discriminative ability, indicating limited generalizability.\nModel robustness was validated through fivefold cross‑validation (mean AUC = 0.98 ± 0.03) (Supplementary Figure S1) and L1‑regularized XGBoost optimization (α = 0.01) (Supplementary Figure S2). Bootstrap resampling (n = 1000) confirmed stable AUC distribution (95% CI 0.82–0.91) (Supplementary Table S1).\nSupplementary validation confirms the consistently superior performance of XGBoost across key clinical subgroups, including patients who underwent laparoscopic surgery and those with a history of chronic pain (AUC improvement (XGBoost vs LR): 17.5%) (Supplementary Figure S2, S4). Therefore, the conclusion that “XGBoost outperforms other models in both the training and validation cohorts” remains valid.\nTaken together, XGBoost consistently outperformed other models across both training and validation cohorts. The XGBoost model (AUC = 0.94) significantly outperformed both the VAS‑PCI integrated model (AUC = 0.66) and the ICPC risk‑stratification scale (AUC = 0.71) (Supplementary Figure S3). Its ability to capture complex nonlinear interactions among preoperative laboratory features makes it a suitable and reliable model for predicting intrapartum analgesia efficacy. Thus, XGBoost was selected as the optimal algorithm for further clinical interpretation and individualized prediction modeling.\nSHAP analysis for model interpretability\nSHAP analysis provided critical insights into the factors driving the model’s predictions, as visualized in Fig. 7. The SHAP summary plot illustrates the global importance of each feature and its corresponding impact on model output. The analysis revealed that baseline VAS score, stimulation frequency (Hz), and stimulation intensity (%) were the most influential predictors of treatment response.\nSupplementary Table S2 compares SHAP consistency across models. XGBoost showed significantly higher SHAP‑CI than random forest (0.89 vs 0.76, p < 0.02) and was associated with 81% clinical decision accuracy. These results indicate that XGBoost provides not only better prediction, but also more stable explanations, reducing decision bias and supporting its higher clinical reliability for individualized treatment.\nFigure 7 presents the SHAP summary plot, which illustrates the global importance of each feature and its impact on the model output. The analysis revealed that baseline VAS score, stimulation frequency, and stimulation intensity were the most influential predictors of treatment response.\nBaseline VAS score is the most significant negative predictor (mean |SHAP|= 0.24), indicating that higher initial pain intensity significantly reduces the likelihood of post‑treatment pain relief (OR = 0.72, 95% CI 0.63–0.82) (Supplementary Figure S5). Stimulation frequency demonstrates a significant non‑linear effect, with the 50–80 Hz interval associated with the highest treatment response rate, consistent with peripheral nerve conduction physiology [22].\nStimulation intensity showed a generally positive association with treatment success, indicating that higher-intensity stimulation may enhance the likelihood of significant clinical improvement. Supplementary analysis confirms a monotonic dose–response relationship, with optimal stimulation intensity at 2.71 mA. Response distribution shows 12.9% significant and 29% mild improvement rates (Supplementary Figure S6, S7).\nThe SHAP analysis allows clinicians to interpret not only the overall influence of each factor, but also the direction and magnitude of its effect on individual predictions. This level of interpretability is essential for clinical implementation, providing actionable insights to inform personalized treatment strategies and enhance shared decision-making.\nDiscussion\nThis study successfully developed and validated a machine learning-based predictive model for CPSPP treatment outcomes, integrating comprehensive clinical data with advanced interpretability techniques. Our findings demonstrate the exceptional potential of machine learning in identifying patients likely to achieve significant improvement with targeted interventions, thereby facilitating personalized treatment strategies and optimized resource allocation.\nThe XGBoost model demonstrated outstanding predictive performance (AUC‑ROC = 0.94), surpassing all other algorithms. Although this high accuracy raises potential overfitting concerns, model robustness is supported by two key factors: performance was maintained on a strictly independent test set, and high predictive consistency was observed across multiple algorithms (logistic regression, random forest, SVM), indicating that the identified features capture clinically meaningful patterns. These results suggest that XGBoost is particularly suitable for modeling treatment response in chronic post-surgical pelvic pain following magnetic stimulation therapy.\nThe exceptional model performance can be attributed to several factors inherent in the dataset and study design. The comprehensive feature set included detailed baseline assessments, treatment parameters, and demographic characteristics that collectively capture the multifactorial nature of treatment response in CPSPP. The relatively homogeneous patient population, all receiving similar interventions within a single institution, may have reduced confounding variables that typically complicate predictive modeling in heterogeneous clinical settings.\nIn this study, SHAP was applied to a traditional machine‑learning model (XGBoost), complementing the explanation logic of deep‑learning frameworks such as DeepXplainer [23]. Compared with deep‑learning interpretability frameworks like DiaXplain, the XGBoost‑SHAP approach offers higher computational efficiency and is better suited for small‑sample clinical datasets [24].\nStimulation parameters, particularly frequency and intensity, are important predictors of treatment outcomes. SHAP analysis provides quantitative guidance for adjusting these parameters to optimize treatment success at the individual level, supporting a shift toward personalized protocols. While IoT device data were not included, the analysis of clinical feature interactions establishes a robust interpretative basis for future multi-source data integration.\nThe multifactorial etiology of CPSPP, involving inflammatory, neurological, musculoskeletal, and psychological components, makes its treatment particularly challenging. Our model’s ability to integrate diverse clinical variables and identify complex interaction patterns demonstrates the value of machine learning approaches in addressing such multifaceted clinical problems. The insights gained from SHAP analysis not only validate existing clinical knowledge, but also reveal novel relationships that may guide future research and treatment development.\nRecent studies confirm SHAP’s consistent interpretability across disease predictions. For instance, SHAP has been used to identify key predictors in social anxiety (Chadaga et al., 2024) [25], reveal predictive features in gestational diabetes (Hassan et al., 2025) [26], and explain critical gene impacts in glioma survival (Vershinina et al., 2025) [27]. Similarly, it has clarified deep‑learning decisions in lung‑cancer detection (Wani et al.) [23] and has been integrated into IoMT‑based Alzheimer’s frameworks (Khan et al.)[28], underscoring its versatility across diverse medical domains.\nOur observed treatment response rate (27.4% with significant improvement) aligns with existing literature for similar CPSPP interventions. However, our model’s predictive accuracy represents an advance over conventional assessments. The mean VAS improvement of 2.05 points is clinically meaningful, and the low adverse event rate (22.6% total, 1.6% moderate) supports the intervention’s safety.\nThis single‑center retrospective study (n = 62) employed strict criteria to ensure homogeneity; its results require validation in larger, multicenter cohorts. The retrospective design may introduce selection bias and limits causal inference. Future prospective studies integrating multimodal data (e.g., genetic, imaging) are needed to enhance accuracy and generalizability.\nDespite these limitations, the study demonstrates the feasibility of an interpretable machine‑learning model for CPSPP, providing a foundation for precision medicine. The framework can optimize treatment selection and resource allocation. Future work should focus on prospective multicenter validation, integration of additional data modalities, and the development of real‑time clinical decision support systems.\nFinally, the exceptional predictive performance observed in this study raises important questions about the underlying mechanisms of treatment response in CPSPP. Future mechanistic studies should investigate the biological and psychological factors that contribute to the predictive features identified through SHAP analysis, potentially leading to the development of novel therapeutic targets and treatment approaches.\nConclusion\nIn conclusion, this study successfully developed and validated a machine learning-based predictive model for CPSPP syndrome treatment outcomes, achieving exceptional predictive performance through the integration of comprehensive clinical data and advanced interpretability techniques. The logistic regression model demonstrated perfect discrimination between patients likely to achieve significant improvement and those unlikely to benefit from treatment, with AUC-ROC, sensitivity, and specificity all reaching 1.000.\nThe SHAP analysis provided crucial insights into the factors driving treatment outcomes, identifying baseline VAS score, stimulation frequency, and stimulation intensity as the most influential predictors. These findings not only validate existing clinical knowledge, but also provide quantitative guidance for personalizing treatment protocols to maximize the likelihood of success for individual patients.\nThe study’s findings have significant implications for clinical practice, offering the potential to transform CPSPP treatment from empirical approaches to precision medicine strategies. The ability to accurately predict treatment outcomes can guide clinical decision-making, optimize resource allocation, and improve patient counseling by providing realistic expectations about treatment benefits.\nWhile the exceptional model performance requires validation in larger, more diverse populations, this work establishes a strong foundation for the integration of machine learning and explainable AI in chronic pain management. The transparent, interpretable nature of the predictive model facilitates clinical adoption and provides actionable insights that align with clinical decision-making processes.\nFuture research should focus on prospective validation, integration of additional data modalities, and development of real-time clinical decision support systems to realize the full potential of these predictive models in improving patient outcomes and advancing personalized pain management strategies. The success of this approach in CPSPP treatment prediction suggests broader applications for machine learning-based outcome prediction across various chronic pain conditions, potentially revolutionizing pain management through data-driven, personalized treatment approaches.\nData availability\nThe datasets generated and analyzed during the current study are not publicly available due to institutional data policy and privacy concerns, but are available from the corresponding author on reasonable request.\nReferences\nWang H, Zhang J, Ma D, Zhao Z. The role of acupuncture and its related mechanism in treating chronic prostatitis/chronic pelvic pain syndrome. Int J Gen Med. 2023;16:4039–50. https://doi.org/10.2147/IJGM.S417066.\nZhao QX, Yang FY, Chen D, Xing NZ. Lycopene combined with quercetin and curcumin for chronic prostatitis/chronic pelvic pain syndrome in rats: effect and mechanism. Zhonghua Nan Ke Xue. 2021;27(2):99–105.\nFranz J, Kieselbach K, Lahmann C, Gratzke C, Miernik A. Chronic primary pelvic pain syndrome in men. Dtsch Arztebl Int. 2023;120(29–30):508–18. https://doi.org/10.3238/arztebl.m2023.0036.\nLamvu G, Carrillo J, Ouyang C, Rapkin A. Chronic pelvic pain in women: a review. JAMA. 2021;325(23):2381–91. https://doi.org/10.1001/jama.2021.2631.\nDydyk AM, Singh C, Gupta N. Chronic pelvic pain. In: StatPearls. Treasure Island (FL): StatPearls Publishing; 2025.\nAredo JV, Heyrana KJ, Karp BI, Shah JP, Stratton P. Relating chronic pelvic pain and endometriosis to signs of sensitization and myofascial pain and dysfunction. Semin Reprod Med. 2017;35(1):88–97. https://doi.org/10.1055/s-0036-1597123.\nGrinberg K, Sela Y, Nissanholtz-Gannot R. New insights about chronic pelvic pain syndrome (CPPS). Int J Environ Res Public Health. 2020;17(9):3005. https://doi.org/10.3390/ijerph17093005.\nPolackwich AS, Shoskes DA. Chronic prostatitis/chronic pelvic pain syndrome: a review of evaluation and therapy. Prostate Cancer Prostatic Dis. 2016;19(2):132–8. https://doi.org/10.1038/pcan.2016.8.\nDal Farra F, Aquino A, Tarantino AG, Origo D. Effectiveness of myofascial manual therapies in chronic pelvic pain syndrome: a systematic review and meta-analysis. Int Urogynecol J. 2022;33(11):2963–76. https://doi.org/10.1007/s00192-022-05173-x.\nLam JC, Stokes W. Acute and chronic prostatitis. Am Fam Physician. 2024;110(1):45–51.\nSpeer LM, Mushkbar S, Erbele T. Chronic pelvic pain in women. Am Fam Physician. 2016;93(5):380–7.\nNamazi G, Chauhan N, Handler S. Myofascial pelvic pain: the forgotten player in chronic pelvic pain. Curr Opin Obstet Gynecol. 2024;36(4):273–81. https://doi.org/10.1097/GCO.0000000000000966.\nLi R, Li Y, Wang H, et al. Effect and mechanism of aloin in ameliorating chronic prostatitis/chronic pelvic pain syndrome: network pharmacology and experimental verification. Drug Des Devel Ther. 2025;19:40110504.\nWang H, Chang H, Wang A, et al. Exploring the efficacy of acupuncture for tension-type headache: a literature review and insights from traditional Chinese medicine. J Oral Facial Pain Headache. 2024;38(4):11–23. https://doi.org/10.22514/jofph.2024.035.\nWang Z, Wu J, Deng C, et al. Efficacy and brain modulation mechanisms of acupuncture for chronic prostatitis/chronic pelvic pain syndrome revealed by structural MRI changes. Front Neurol. 2025;16:1579484. https://doi.org/10.3389/fneur.2025.1579484.\nJia Z, Lv D, Chen T, et al. Network pharmacology and in vivo experiment-based strategy for investigating the mechanism of chronic prostatitis/chronic pelvic pain syndrome in QianLieJinDan tablets. Heliyon. 2024;10(9):e29975. https://doi.org/10.1016/j.heliyon.2024.e29975.\nBałabuszek K, Toborek M, Pietura R. Comprehensive overview of the venous disorder known as pelvic congestion syndrome. Ann Med. 2022;54(1):22–36. https://doi.org/10.1080/07853890.2021.2014556. (PMID: 34935563).\nClark MR, Taylor AC. Pelvic venous disorders: an update in terminology, diagnosis, and treatment. Semin Intervent Radiol. 2023;40(4):362–71. https://doi.org/10.1055/s-0043-1771041.\nYebes A, Toribio-Vazquez C, Martinez-Perez S, et al. Prostatitis: a review. Curr Urol Rep. 2023;24(5):241–51. https://doi.org/10.1007/s11934-023-01150-z.\nLi J, Yi X, Ai J. Broaden horizons: the advancement of interstitial cystitis/bladder pain syndrome. Int J Mol Sci. 2022;23(23):14594. https://doi.org/10.3390/ijms232314594.\nKarp BI, Stratton P. Chronic pelvic pain and botulinum toxin. Toxicon. 2025;258:108336. https://doi.org/10.1016/j.toxicon.2025.108336.\nLaporte Y. Continuous conduction of impulses in peripheral myelinated nerve fibers. J Gen Physiol. 1951;35(2):343–60. https://doi.org/10.1085/jgp.35.2.343.\nWani NA, Kumar R, Bedi J. DeepXplainer: an interpretable deep learning based approach for lung cancer detection using explainable artificial intelligence. Comput Methods Programs Biomed. 2024;243:107879. https://doi.org/10.1016/j.cmpb.2023.107879.\nMoulaei K, Afrash MR, Parvin M, Shadnia S, Rahimi M, Mostafazadeh B, et al. Explainable artificial intelligence (XAI) for predicting the need for intubation in methanol-poisoned patients: a study comparing deep and machine learning models. Sci Rep. 2024;14(1):15751. https://doi.org/10.1038/s41598-024-66481-4.\nChadaga K, Prabhu S, Sampathila N, Chadaga R, Bhat D, Sharma AK, et al. SADXAI: Predicting social anxiety disorder using multiple interpretable artificial intelligence techniques. SLAS Technol. 2024;29(2):100129. https://doi.org/10.1016/j.slast.2024.100129.\nHassan A, Ahmad SG, Iqbal T, et al. Enhanced model for gestational diabetes mellitus prediction using a fusion technique of multiple algorithms with explainability. Intern J Comput Intell Syst. 2025;18(1):1–33. https://doi.org/10.1007/s44196-025-00760-4.\nVershinina O, Turubanova V, Krivonosov M, Trukhanov A, Ivanchenko M. Explainable machine learning models for glioma subtype classification and survival prediction. Cancers (Basel). 2025;17(16):2614. https://doi.org/10.3390/cancers17162614.\nKhan AH, Ali D, Ahmed S, Alhumam A, Khan MF, Siddiqui SY. IoMT driven Alzheimer’s prediction model empowered with transfer learning and explainable AI approach in healthcare 5.0. Sci Rep. 2025;15(1):35382. https://doi.org/10.1038/s41598-025-19395-8.\nDeclaration of Generative AI and AI-assisted technologies in the writing process\nDuring the preparation of this work, the authors used ChatGPT (OpenAI, San Francisco, USA) to assist in improving the clarity and language of the manuscript. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.\nFunding\nThis research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.\nAuthor information\nAuthors and Affiliations\nContributions\n.J.H.X.and.Z.W. contributed equally to the study design, data collection, statistical analysis, and manuscript writing..Z.L.X. and Y.S. assisted with data collection and statistical analysis. Y. B. Z. were responsible for the study concept, supervision, and manuscript revision. All authors have read and approved the final manuscript.\nCorresponding author\nEthics declarations\nHuman ethics and consent to participate\nThis retrospective study was approved by the Institutional Review Board of the Second People’s Hospital of Hefei (Approval No. 2022-KY-067). All procedures were conducted in accordance with the Declaration of Helsinki. Given the retrospective nature of the study and the use of anonymized data, the requirement for informed consent was waived by the ethics committee.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nPublisher's Note\nSpringer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.\nSupplementary Information\nRights and permissions\nOpen Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.\nAbout this article\nCite this article\nXi, J., Wang, Z., Xu, Z. et al. Machine learning-based prediction model for chronic post-surgical pelvic pain syndrome: a comprehensive analysis using SHAP interpretability. Eur J Med Res 31, 323 (2026). https://doi.org/10.1186/s40001-026-03923-x\nReceived:\nAccepted:\nPublished:\nVersion of record:\nDOI: https://doi.org/10.1186/s40001-026-03923-x","source_license":"CC0","license_restricted":false}