Prediction of myofascial pelvic pain syndrome based on random forest model.

In this study, we gather case data from the study participants, encompassing age, height, weight, body mass index (BMI), pregnancy history, and childbirth history [ 8 ]. Participants included in the study underwent pelvic floor pressure assessment administered by trained therapists specializing in pelvic floor care. The assessment utilized the Myotrac biofeedback device manufactured by Nanjing Vishee Medical Technology Co., Ltd. Various parameters are automatically recorded, including the average value and the coefficient of variation during the pre-resting phase, the maximum value and relaxation time during the fast contraction phase, the average value, coefficient of variation, and relaxation time during the tension contraction phase, and the average value and coefficient of variation are recorded during both the endurance contraction phase and the post-resting phase. The modified Oxford grading scale is used for pelvic floor muscle assessment. The evaluation criteria for pelvic floor muscle strength assessment [ 9 ] are as follows: Grade I indicates a slight contraction of the vaginal muscles during testing. Grade II signifies that the vaginal muscles can sustain a contraction for 2 s and repeat it twice. Grade III denotes that the vaginal muscles can sustain a contraction for 3 s and repeat it three times. Grade IV indicates that the vaginal muscles can sustain a contraction for 4 s and repeat it four times. Grade V reflects significant contraction of the vaginal muscles, with the ability to sustain a contraction for 5 s or longer and repeat it five times or more. These assessments are conducted to gather comprehensive data on pelvic floor muscle pressure and strength in the study participants.

Hang Yu: Writing – original draft, Methodology, Formal analysis, Data curation, Conceptualization. Hongguo Zhao: Investigation, Data curation, Conceptualization. Dongxia Liu: Data curation, Conceptualization. Yanhua Dong: Formal analysis. Manman Nai: Validation. Yikun Song: Validation. Jiaxi Liu: Validation. Luwen Wang: Writing – review & editing. Lei Li: Funding acquisition. Xinbin Li: Validation, Software, Methodology, Investigation, Formal analysis, Conceptualization.

2023-206-01.

This study is approved by the Ethics Committee of The Third Affiliated Hospital of Zhengzhou University (Ethics Approval Number: 2023-206-01, Approval Data: August 23, 2023). The patients whose data are included in this manuscription have consented for all clinical data and other data included in the manuscript to be published.

Foundation of He'nan Educational Committee (24A320036).

The results of the chi-square test and Mann-Whitney U test revealed statistically significant differences ( P  < 0.05) between the two groups in the following variables: parity, gravidity, number of vaginal deliveries, perineal laceration, average value and the coefficient of variation during the pre-resting phase, the maximum value and relaxation time during the fast contraction phase, the average value and coefficient of variation during the tension contraction phase, the average value and coefficient of variation during the endurance contraction phase, and the coefficient of variation during the post-resting phase. Please refer to Table 1 and Table 2 for detailed analysis results. Table 1 General demographic characteristics of MPPS and non-MPPS groups (Example (%)). Table 1 Feature Total number of people(n = 1245) The non-MPPS group(n = 702) The MPPS group (n = 543) χ 2 P Age (years) 1.134 0.567  40 175(14.06 %) 629(50.52 %) 441(35.42 %) 88(12.54 %) 338(48.15 %) 276(39.32 %) 87(16.02 %) 291(53.59 %) 165(30.39 %) BMI(Kg/m 2 )  <18.5 1.015 0.602 73(5.86 %) 38(5.41 %) 35(6.45 %)  18.5∼<24 770(61.85 %) 431(61.40 %) 339(62.43 %)  ≥24 402(32.29 %) 233(33.19 %) 169(31.12 %) Number of pregnancies (times) 78.994 2 764(61.37 %) 373(53.13 %) 391(72.01 %) Number of production (times) 37.939 2 157(12.61 %) 89(12.67 %) 68(12.52 %) Number of vaginal births (times) 18.846 2 117(9.40 %) 68(9.69 %) 49(9.02 %) Number of cesarean sections (times) 1.017 0.601  0 871(69.96 %) 495(70.51 %) 376(69.24 %)  1∼2 363(29.16 %) 200(28.49 %) 163(30.02 %)  >2 11(0.88 %) 7(1.00 %) 4(0.74 %) History of perineal laceration 7.027 0.008  Yes 54(4.34 %) 21(2.99 %) 33(6.08 %)  No 1191(95.66 %) 681(97.01 %) 510(93.92 %) History of hysterectomy 0.203 0.653  Yes 47(3.78 %) 25(3.56 %) 22(4.05 %)  No 1198(96.22 %) 677(96.44 %) 521(95.95 %) Table 2 Univariate analysis of height, weight, pelvic floor pressure assessment, and modified Oxford muscle strength assessment between MPPS and non-MPPS groups. Table 2 Feature The non-MPPS group(n = 702) The MPPS group (n = 543) Z P Height(cm) 161(158, 165) 161(158,165) −0.491 0.623 Weight (Kg) 60(54, 65) 59(53, 64) −1.900 0.057 The pre-resting stage Average value (mmHg) 3.55(2.97, 4.53) 3.87(3.34,4.53) −4.095 <0.001 Variable coefficient 0.15(0.11, 0.21) 0.17(0.13,0.22) −3.302 0.001 Rapid contraction phase Maximum value (mmHg) 13.57(9.62, 19.11) 12.24(7.89, 18.21) −3.713 0.001 Relaxation time(s) 0.92(0.14, 1.61) 0.54(0.42, 1.90) −3.011 0.003 the tension contraction phase Average value (mmHg) 11.77(9.39,14.14) 9.84(7.44, 13.03) −7.757 <0.001 Coefficient of variation 0.17(0.11, 0.25) 0.18(0.12, 0.26) −2.668 0.008 Relaxation time(s) 0.82(0.09, 1.35) 0.22(0.07, 2.03) −0.955 0.339 the endurance contraction phase Average value (mmHg) 9.35(7.59, 11.67) 7.41(5.41, 9.93) −9.670 <0.001 Coefficient of variation 0.18(0.14, 0.23) 0.21(0.16, 0.26) −5.987 <0.001 the post-resting phase Average value (mmHg) 3.99(2.57, 5.50) 4.23(2.82, 5.73) −1.731 0.083 Coefficient of variation 0.19(0.17, 0.20) 0.21(0.14, 0.27) −1.999 0.046 Deep class I muscle strength 2(1, 3) 2(1, 3) −1.079 0.28 Superficial class I muscle strength 2(1, 3) 2(1, 3) −1.359 0.174 General demographic characteristics of MPPS and non-MPPS groups (Example (%)). Univariate analysis of height, weight, pelvic floor pressure assessment, and modified Oxford muscle strength assessment between MPPS and non-MPPS groups. The results of the univariate logistic regression analysis indicates significant differences ( P  < 0.05) between the two groups in various variables, including parity, gravidity, number of vaginal deliveries, perineal laceration, maximum value and relaxation time in the fast contraction phase, the average value and coefficient of variation in the tension contraction phase, the average value and coefficient of variation in the endurance contraction phase, and the coefficient of variation in the post-resting phase. These results suggest a significant association between these factors and MPPS. For detailed analysis results, please refer to Table 3 . Table 3 Univariate Logistic regression analysis of MPPS. Table 3 Feature index Feature b SE Wald χ2 P OR 1 Age −0.008 0.006 2.055 0.152 0.992 2 Height −0.007 0.009 0.54 0.462 0.993 3 Weight −0.01 0.007 2.455 0.117 0.99 4 BMI −0.002 0.013 0.02 0.889 0.998 5 Number of pregnancies 0.381 0.038 101.611 <0.001 1.463 6 Number of production 0.332 0.062 28.899 <0.001 1.394 7 Number of vaginal births 0.217 0.053 16.794 <0.001 1.242 8 Number of cesarean sections 0.075 0.078 0.916 0.339 1.078 9 History of perineal laceration 0.741 0.285 6.752 0.009 2.098 10 History of hysterectomy 0.134 0.298 0.202 0.653 1.143 11 Average value during the pre-resting phase 0.089 0.046 3.788 0.052 1.093 12 Coefficient of variation during the pre-resting phase 0.493 0.329 2.245 0.134 1.638 13 Maximum value during the fast contraction phase −0.021 0.007 7.933 0.005 0.979 14 Relaxation time during the fast contraction phase 0.09 0.035 6.719 0.01 1.094 15 Average value during the tension contraction phase −0.081 0.015 30.956 <0.001 0.922 16 Coefficient of variation during the tension contraction phase 1.243 0.561 4.914 0.027 3.465 17 Relaxation time during the tension contraction phase 0.034 0.041 0.685 0.408 1.035 18 Average value during the endurance contraction phase −0.137 0.018 57.535 <0.001 0.872 19 Coefficient of variation during the endurance contraction phase 2.256 0.684 10.878 0.001 9.542 20 Average value during the post-resting phase 0.043 0.027 2.575 0.109 1.044 21 Coefficient of variation during the post-resting phase 3.495 0.806 18.784 <0.001 32.958 22 Deep class I muscle strength −0.089 0.059 2.231 0.135 0.915 23 Deep class II muscle strength −0.024 0.054 0.205 0.651 0.976 24 Superficial class I muscle strength −0.081 0.062 1.713 0.191 0.922 25 Superficial class II muscle strength −0.051 0.054 0.897 0.344 0.95 Univariate Logistic regression analysis of MPPS. In the multivariate logistic regression analysis, the factors that demonstrated a significant association ( P  < 0.05) with MPPS in the chi-square test, Mann-Whitney U test, and univariate logistic regression analysis are included. The results indicates that the following factors independently influence MPPS: parity (OR = 1.475, P  < 0.05), average value in the pre-resting phase (OR = 1.347, P  < 0.05), coefficient of variation in the tension contraction phase (OR = 7.043, P  < 0.05), average value in the endurance contraction phase (OR = 0.83, P  < 0.05), coefficient of variation in the endurance contraction phase (OR = 4.919, P  < 0.05), and coefficient of variation in the post-resting phase (OR = 14.569, P  < 0.05). These results indicate that these factors play an independent role in the development of MPPS. For detailed analysis results, please refer to Table 4 . Table 4 Multivariate Logistic regression analysis of MPPS. Table 4 Feature b SE Wald χ2 P OR Number of pregnancies 0.389 0.05 61.66 <0.001 1.475 Number of production −0.008 0.112 0.005 0.943 0.992 Number of vaginal births −0.084 0.089 0.895 0.344 0.919 History of perineal laceration 0.42 0.328 1.645 0.2 1.523 average value during the pre-resting phase 0.298 0.062 23.219 <0.001 1.347 Coefficient of variation during the pre-resting phase 0.711 0.414 2.955 0.086 2.037 Maximum value during the fast contraction phase 0.007 0.017 0.156 0.693 1.007 Relaxation time during the fast contraction phase 0.084 0.044 3.615 0.057 1.088 Average value during the tension contraction phase 0.007 0.04 0.027 0.871 1.007 Coefficient of variation during the tension contraction phase 1.952 0.812 5.778 0.016 7.043 Average value during the endurance contraction phase −0.186 0.035 27.709 <0.001 0.83 Coefficient of variation during the endurance contraction phase 1.593 0.811 3.86 0.049 4.919 Coefficient of variation during the post-resting phase 2.679 0.884 9.179 0.002 14.569 constant −2.614 0.417 39.345 <0.001 0.073 Multivariate Logistic regression analysis of MPPS. In this study, the prediction model is developed using MATLAB (version 2020b). A total of 1245 cases are collected, comprising 543 MPPS-positive and 702 MPPS-negative cases. To enhance result accuracy when employing the RF method for binary classification, maintaining a 1:1 ratio between negative and positive samples is recommended. Consequently, 543 individuals from the negative cases were randomly chosen to create a balanced dataset of 1086 individuals, combined with the positive cases. Subsequently, the dataset was split into a 7:3 ratio, with 760 cases assigned to the training set and 326 cases to the cross-validation set. The modeling process of the random forest model is depicted in Fig. 1 . To assess the classification model's reliability, various metrics including the confusion matrix, accuracy, sensitivity, specificity, precision, receiver operating characteristic (ROC) curve, and area under the curve (AUC) value are analyzed. The out-of-bag (OOB) classification error of the classification model is illustrated in Fig. 2 . It is observed that the OOB error stabilizes at a value below 0.15 when the number of trees in the RF algorithm exceeds 100. This suggests that the classification model in this study exhibits strong performance. Fig. 2 The OOB classification error of the classification model. Fig. 2 The OOB classification error of the classification model. To examine the correlation between different clinical features and the incidence rate of MPPS, the significance of each clinical feature is assessed, as depicted in Fig. 3 . The feature indices and their definitions are provided in Table 3 . The top five significant features, listed in descending order of importance, include the coefficient of variation during the resting phase, relaxation time during rapid contraction, average value during endurance contraction, average value during sustained contraction, and maximum value during rapid contraction. Fig. 3 Importance of the analyzed clinical features. Fig. 3 Importance of the analyzed clinical features. The accuracy, sensitivity, specificity and precision of the classification model are usually used as the evaluation metrics of the prediction model. The definitions of the above indicator are [ 10 , 11 ]. (1) Accuracy = TP + TN TP + TN + FP + FN (2) Sensitivity = TP TP + FN (3) Specificity = TN TN + FP (4) Precision = TP TP + FP where, TP, TN, FP, and FN are the numbers of true positive, true negative, false positive, and false negative, respectively. The comparisons of accuracy, sensitivity, specificity, and precision between the random forest and logistic regression models are presented in Fig. 4 . The accuracy, sensitivity, specificity, and precision of the random forest model are much higher than those of logistic regression model. The differences are 0.1715, 0.332, 0.1048, and 0.1494, respectively. These results indicate that the random forest model exhibits significantly superior predictive performance compared to the logistic regression model. Fig. 4 Four evaluation metrics of the random forest and logistic regression models. Fig. 4 Four evaluation metrics of the random forest and logistic regression models. The ROC curves for the logistic regression and random forest models are provided in Fig. 5 (a) and (b), respectively, with their corresponding evaluation metric comparisons presented in Table 5 . The logistic regression analysis yielded a sensitivity of 57.46 %, specificity of 79.63 %, and an AUC of 0.7550 for the MPPS model. In contrast, the random forest model achieved a sensitivity of 90.66 %, specificity of 90.91 %, and an AUC of 0.9426 for MPPS. Evidently, the random forest model demonstrates superior predictive performance in comparison to the logistic regression model. Fig. 5 ROC curve of (a) RF model (AUC = 0.9426) and (b) logistic model (AUC = 0.755). Fig. 5 Table 5 Evaluation metrics of the prediction models. Table 5 Accuracy Sensitivity Specificity Precision RF model 0.8711 0.9066 0.9091 0.8351 Logistic regression model 0.6996 0.5746 0.7963 0.6857 ROC curve of (a) RF model (AUC = 0.9426) and (b) logistic model (AUC = 0.755). Evaluation metrics of the prediction models. The box plot illustrating the random forest and logistic regression models is displayed in Fig. 6 . The median, 25th percentile, 75th percentile, maximum value, and minimum value of the random forest model are 0.482, 0.423, 0.5468, 0.726, and 0.252, respectively. Corresponding values for the logistic regression model are 0.3982, 0.164, 0.6862, 1, and 0.00889, respectively. The random forest model demonstrates a distribution closer to normal, suggesting its superiority over the logistic regression model. Fig. 6 Box plot of the random forest and logistic regression models. Fig. 6 Box plot of the random forest and logistic regression models.

To be eligible for participation, individuals have to meet specific criteria: (1) Exhibit a history of chronic pelvic pain lasting a minimum of 3–6 months, diagnosed as Myofascial Pelvic Pain Syndrome (MPPS) [ 7 ]. (2) Not have undergone any other treatment pertaining to this condition in the preceding month. (3) Be aged 18 years or above. (4) Patients with a history of sexual activity who can undergo vaginal examinations. Patients meeting any of the following criteria are excluded from the study: (1) Patients with chronic pelvic pain (CPP) attributed to known factors like infection, adenomyosis, uterine fibroids, etc. (2) Patients who have undergone gynecological examination, gynecological and urological ultrasound, examination of vaginal secretions and urine pathogens, and have been excluded due to acute infectious diseases or other organic diseases. (3) Pregnant patients or those within 3 months postpartum are excluded. (4) Patients with implanted cardiac pacemakers, metal intrauterine devices, or allergies to electrical stimulation are excluded. (5) Patients with active vaginal bleeding, local pelvic skin/mucosal damage, or infection are excluded. (6) Patients with other pelvic floor dysfunction diseases like urinary incontinence or pelvic organ prolapse are excluded. (7) Patients with neurological or psychiatric disorders or other severe illnesses that could impede cooperation are excluded. (8) Patients with a history of spinal disc disease, sciatica, or other neurological disorders are excluded. (9) Patients with incomplete clinical baseline data are excluded.

The random forest model has exhibited exceptional predictive performance in the initial screening evaluation of MPPS disease. The development of this predictive framework holds significant importance in refining the precision of MPPS prediction within clinical environments and elevating treatment outcomes. This research carries profound global implications, given the potentially elevated misdiagnosis rates and delayed diagnosis proportions of MPPS on a worldwide scale, coupled with a potential scarcity of seasoned healthcare providers. Moving forward, continual refinement and validation of the model will be imperative to further augment the precision of MPPS risk assessment, thereby furnishing clinicians with more dependable decision-making support in clinical practice.

This study is grounded in the clinical variables of the patients enrolled in the research, in conjunction with assessments of pelvic floor pressure and data from the modified Oxford muscle strength evaluation. The study analyzes the pertinent risk factors and constructs a random forest model. This model introduces a preliminary risk assessment approach for MPPS in patients. A comparison between the performance of the random forest model and the logistic regression prediction model illustrates the superior predictive performance of the random forest model. In descending order of importance, the top five factors are identified as the coefficient of variation in the late resting phase, relaxation time in the fast contraction phase, mean value of the endurance contraction phase, mean value of the tension contraction phase, and maximum value in the fast contraction phase. Consequently, these indicators merit significant attention during the clinical screening process. MPPS is a non-inflammatory disease that occurs in the pelvic floor muscles and fascia. It presents with local adhesions and spasms resulting from pelvic floor muscles and fascia damage, leading to persistent chronic pain [ 12 ]. The current diagnosis of MPPS continues to depend on clinical assessment, guided by existing research and clinical guidelines [ 13 , 14 ]. Within the clinical diagnostic process, the diagnosis of MPPS lacks specificity, and a single clinical feature is insufficient for a definitive diagnosis. A comprehensive analysis of multiple clinical factors is necessary, posing a significant challenge to the clinical experience of healthcare providers. These factors can influence treatment outcomes, patient quality of life, and even mental health implications [ 15 , 16 ]. While depending on experienced healthcare providers can enhance diagnostic accuracy for MPPS, the complexity of the condition and the scarcity of experienced providers render this approach impractical. Therefore, the development and implementation of reliable predictive models are crucial for timely identification of these patients and providing effective interventions to improve their prognosis [ 17 ]. Currently, machine learning models are widely used in the diagnosis and treatment of various clinical conditions [ [18] , [19] , [20] , [21] , [22] ], and research by Wenhui Jiang and colleagues has shown that the random forest model demonstrates good predictive performance in clinical practice [ 18 ]. Based on a dataset of 1245 cases, we implement the random forest algorithm for the first time in the screening of MPPS. In this investigation, the Logistic regression model forecasts the accuracy, sensitivity, specificity, and precision of MPPS at 69.96 %, 57.46 %, 79.63 %, and 68.57 % respectively, with an AUC of the ROC curve at 0.755. Conversely, the random forest prediction model exhibits accuracy, sensitivity, specificity, and precision rates of 87.11 %, 90.66 %, 90.91 %, and 83.51 % respectively, with an AUC of the ROC curve at 0.942. The random forest model showcases exceptional predictive performance during the initial screening of MPPS. The development of this predictive model holds significant importance in enhancing the predictive accuracy and treatment outcomes in clinical MPPS. Research findings highlight that the random forest model displays superior clinical predictive performance in identifying risk factors among MPPS patients, offering vital support and guidance for bolstering early identification and personalized treatment for MPPS. In the early stages of MPPS, treatment costs are minimal, necessitating solely rehabilitation therapy without clinical intervention. Patients in this phase typically exhibit subtle symptoms, presenting a notable diagnostic challenge for healthcare professionals. Through the predictive model proposed in this study, timely diagnosis in the early stages of MPPS becomes achievable, arresting the progression of the condition and eliminating the need for clinical intervention. This model aids in reducing economic burdens, enhancing patients' quality of life, and is pivotal in boosting the efficiency and quality of healthcare services. By enabling accurate prediction and timely intervention, it becomes feasible to significantly enhance patients' treatment outcomes and quality of life. The factors contributing to the superior predictive performance of the random forest model over the logistic regression model are multifaceted. Firstly, the model utilize in this study encompasses a multitude of features, posing challenges for the logistic regression model in managing a problem of such complexity. In contrast, the Random Forest model remains unaffected by the abundance of features, rendering it well-suited for addressing multi-feature complexities. Secondly, the logistic regression model operates under the assumption of relatively independent features, adhering to a linear framework. However, in complex scenarios such as predicting MPPS, where features demonstrate intricate interdependencies and significant non-linearity, the predictive performance of the logistic regression model is compromised. The Random Forest model comprises numerous independent decision trees. During its construction phase, the model randomly selects subsets of samples and features. This capability enables the model to identify interactions among features, thereby improving classification accuracy and generalization abilities [ 23 ]. Finally, it is imperative to acknowledge that errors may arise during the screening process conducted by healthcare professionals, potentially introducing biases into the dataset. Such biases can significantly impact the performance of the logistic regression model. Nonetheless, as previously noted, the Random Forest model comprises numerous independent decision trees, endowing it with a degree of resilience against external influences [ 24 ]. These factors contribute to the superior predictive performance exhibited by the Random Forest model in MPPS prediction. Although the model in this study demonstrated excellent performance, it is imperative to acknowledge its inherent limitations. The restricted sample size originating solely from a singular research center may introduce selection bias, thereby potentially compromising the model's generalizability and applicability. Future endeavors will necessitate a more extensive dataset for external validation to substantiate the model's effectiveness. Our forthcoming objective is to augment data collection efforts and diversify the array of predictive variables to enhance the model's comprehensiveness and predictive prowess. The continuous integration of updates and additional sample data is poised to bolster the model's stability and accuracy. Furthermore, external validation of this predictive framework will be systematically conducted to fortify the reliability of our research findings. Through the ongoing processes of refinement and validation, we aim to refine the precision of MPPS risk assessment, furnishing clinicians with dependable decision-making support in clinical settings.

This study is a retrospective analysis. The clinical data of female patients who underwent pelvic floor screening due to chronic pelvic pain at the Pelvic Floor Rehabilitation Center of the Third Affiliated Hospital of Zhengzhou University from January 2021 to December 2023 is retrospectively analyzed. A total of 543 female patients meeting the study's inclusion and exclusion criteria are randomly selected from this dataset and allocated to the MPPS group. Furthermore, 702 healthy female patients who underwent pelvic floor screening during routine physical examinations within the same timeframe are randomly selected and assigned to the non-MPPS group. Ethical approval for this study was granted by the Ethics Committee of the Third Affiliated Hospital of Zhengzhou University (Ethics Approval Number: 2023-206-01, Approval Date: August 23, 2023).

Statistical analysis is conducted using SPSS 26.0 and MATLAB (version 2020b) software. The process modeling and analysis method is illustrated in Fig. 1 . Different methods are employed for various types of data: For continuous variables, data are presented as median ( P 25 , P 75 ), and between-group comparisons utilize the Mann-Whitney U test. For categorical variables, data are presented as counts and percentages (%), and between-group comparisons are conducted using the chi-square test. Logistic regression and random forest algorithms are utilized to develop a prediction model for Myofascial Pelvic Pain Syndrome (MPPS). The predictive accuracy, ROC curve, and AUC (Area Under the Curve) are compared for these models. A significance level of P < 0.05 is deemed as statistically significant. Fig. 1 Process of modeling and analysis method. Fig. 1 Process of modeling and analysis method.

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

The authors do not have permission to share data.