A total of 340 patients with myofascial pain syndrome who visited the pain department and rehabilitation medicine department of our hospital from January 2020 to June 2025 and completed 8-week standardized treatment were retrospectively included. In this study, sample size calculation was first carried out. Based on the expected incidence of the primary outcome indicator (poor treatment efficacy), which was set at 35–40% [according to the previous data of our center and literature reports (8)], a scheme matching the core statistical methods (multivariate logistic regression and machine learning models) was adopted. The calculation was completed using the PASS 2021 software and verified using the “pwr” package in R 4.2.3. The significance level α was set at 0.05 (two-tailed), the test power 1–β was set at 80%, and a 10% dropout rate was considered. Finally, the minimum required sample size was 280 cases. The actual included sample size of 340 cases exceeded the minimum requirement, and its statistical power (1–β > 85%) was verified, meeting the statistical requirements for subsequent multivariate analysis and machine learning modeling.

Inclusion criteria: (1) Age ≥ 18 years; (2) Meeting the diagnostic criteria of MPS in the “Expert Consensus on the Diagnosis and Treatment of Myofascial Pain Syndrome in China,” with clear active myofascial trigger points and pain duration ≥3 months; (3) All receiving an 8-week standardized treatment protocol (including at least one of the core treatment methods such as dry needling, physical therapy, or drug intervention); (4) Completing a full set of baseline assessments before treatment, including pain characteristic assessment, trigger point examination, psychological state assessment, and inflammation marker detection; (5) Having complete visual analog scale scores and secondary outcome indicators after treatment for efficacy determination. Exclusion criteria: (1) Secondary pain caused by radiculopathy, rheumatological and immunological diseases, or acute trauma; (2) Comorbid fibromyalgia syndrome or other widespread pain diseases; (3) Comorbid severe mental illness or cognitive impairment, unable to cooperate to complete the assessment questionnaire; (4) Pregnant or lactating women; (5) Incomplete baseline assessment or 8-week follow-up data, unable to conduct efficacy grouping.

The detailed parameters of each treatment type were as follows: (1) Physical therapy: Ultrasound therapy was adopted, with parameters of 1 MHz frequency, 1.5 W/cm² intensity, 10 min per session, 3 times a week, for a total of 8 weeks; (2) Dry needling: Performed by physicians with more than 5 years of clinical experience, once a week, retaining the needle for 20 min each time, selecting 2–4 active trigger points per session based on physical examination, for a total of 8 weeks; (3) Drug therapy: Non-steroidal anti-inflammatory drugs (ibuprofen sustained-release capsules, 0.3 g each time, twice a day) combined with muscle relaxants (tizanidine hydrochloride tablets, 2 mg each time, three times a day), taken orally for 8 weeks. All patients strictly followed the treatment plan under the guidance of researchers.

The treatment allocation was based on the “Expert Consensus on the Diagnosis and Treatment of Myofascial Pain Syndrome in China” and the patients’ specific conditions: (1) patients with mild pain (VAS ≤ 5) and few trigger points (≤3) were first assigned to physical therapy; (2) patients with moderate pain (5 < VAS < 8) and obvious trigger points (3 < number ≤ 6) were assigned to dry needling combined with physical therapy; (3) patients with severe pain (VAS ≥ 8) or multiple trigger points (>6) were assigned to drug therapy (non-steroidal anti-inflammatory drugs + muscle relaxants) combined with dry needling and physical therapy. Patients’ willingness was fully considered during the allocation process. A statistical analysis of the treatment allocation distribution between the excellent and poor efficacy groups showed no significant difference (p > 0.05), indicating no confounding effect of treatment allocation imbalance.

Integrating these therapeutic modalities into a single predictive model can cover the mainstream clinical practice scenarios, making the model more universally applicable to different clinical settings, and this study strictly controlled the compliance of patients with treatment (physical therapy attendance rate ≥80%, drug/dry needling treatment compliance ≥85%) to reduce the impact of treatment implementation differences on the model.

In this study, multi-dimensional information was systematically collected through the hospital electronic medical record system, the clinical database of the pain department, and the special research follow-up system to construct a prediction model.

Patients’ basic information and clinical characteristics were collected, including: (1) demographic data: age, gender, height, weight, body mass index (BMI), occupation, and years of education; (2) living habits and medical history: smoking history (packs/year), alcohol consumption history (grams/day), exercise frequency (times/week), as well as previous medical history (such as cervical spondylosis, fibromyalgia), MPS duration (months), and drug use history; (3) pain characteristics: main pain location (such as neck, shoulder), pain side (left/right/bilateral), pain nature description (such as soreness, distending pain), and known inducing factors (such as acute injury, chronic strain).

Baseline assessment parameters before treatment were collected, covering: (1) pain characteristic parameters: baseline pain intensity at rest and during activity [visual analog scale (VAS) score, 0–10], average pain duration (hours/day), pain distribution range (count on the body map); (2) trigger point characteristics: number of active trigger points, main affected muscle groups, average tenderness threshold (kg/cm²); (3) psychometric indicators: including depression level [Patient Health Questionnaire-9 (PHQ-9)], anxiety level [Generalized Anxiety Disorder-7 (GAD-7)], and pain catastrophizing degree [Pain Catastrophizing Scale (PCS)]; (4) Inflammatory markers: serum levels of IL-6, tumor necrosis factor-alpha (TNF-α), and hs-CRP.

Finally, details of the treatment protocol (such as specific parameters of dry needling, physical therapy, or drug intervention) and treatment efficacy outcome data at 8 weeks after treatment were recorded. The primary outcome was the pain relief rate calculated based on the VAS score, and patients were grouped according to this, secondary outcomes included the improvement of the functional impairment index, changes in the tenderness threshold, etc. All data were uniformly managed through the electronic data collection system and were collected in a standardized manner by trained researchers to ensure the integrity and accuracy of the data. To ensure the reliability of VAS scores, each patient was assessed by the same trained pain department physician at both baseline and 8-week follow-up. The intraclass correlation coefficient (ICC) was used to evaluate the rater consistency, and the results showed ICC = 0.92 (95%CI: 0.88–0.95), indicating good consistency and reliability of the assessment results.

Referring to the assessment criteria for chronic pain treatment response in the International Association for the Study of Pain and the “Guidelines for the Diagnosis and Treatment of Myofascial Pain Syndrome in China,” based on the minimum clinically important difference (MCID) of the VAS and combined with the characteristics of the 8-week standardized treatment follow-up in this study, patients were divided into two groups.

Excellent treatment efficacy group: After 8 weeks of treatment, all of the following conditions were met: (1) primary efficacy criterion: The VAS pain score decreased by ≥50% compared with the baseline (reaching the MCID threshold); (2) stability requirement: The standardized treatment protocol was not discontinued during treatment due to intolerable adverse reactions or personal reasons; (3) compliance requirement: The attendance rate of physical therapy was not less than 80%, or the compliance record of drug/dry needling treatment was not less than 85% of the recommended protocol.

Poor treatment efficacy group: After 8 weeks of treatment, any of the following conditions were met: (1) efficacy not achieved: the VAS pain score decreased by <50% compared with the baseline; (2) symptom exacerbation: the VAS pain score showed no improvement or increased compared with the baseline; (3) treatment interruption: the patient withdrew from treatment early due to pain-related discomfort, adverse reactions, or lack of confidence in the protocol, and the VAS score at the time of withdrawal did not reach the excellent treatment efficacy standard.

All VAS assessments and secondary outcome indicators (such as the Neck Disability Index, tenderness threshold, etc.) were performed by pain department physicians or researchers who had received unified training through a standardized assessment process at two time points: before the start of treatment and 8 weeks after treatment. Efficacy determination (i.e., calculation of the VAS change value and final grouping) was independently completed in a double-blind manner by two researchers who were unaware of the patients’ other baseline clinical data and laboratory indicators. When there was a disagreement in grouping opinions, a third senior researcher in the research group made an arbitration confirmation according to the preset criteria. All clinical data and assessment results were entered and managed using the electronic data collection system to ensure the accuracy, integrity, and traceability of the data.

Data analysis was completed using software such as SPSS 26.0, R 4.2.3, and Python 3.8.5. The Shapiro–Wilk test was used to verify the normality of measurement data. Measurement data conforming to the normal distribution (p > 0.05) were expressed as mean ± standard deviation (x̄ ± s), and independent-samples t-test was used for comparison between groups; non-normally distributed data (p < 0.05) were expressed as median (quartiles), and the Mann–Whitney U test was used for comparison between groups. Missing values were filled using the median imputation method; continuous variables were standardized (z-score normalization) to eliminate the impact of dimension differences. Count data were expressed as number of cases (percentage) [n (%)], and the χ² test or Fisher’s exact probability method was used for comparison between groups. In the training set, univariate analysis was first performed to screen out indicators with p < 0.05, then variables were compressed by Least Absolute Shrinkage and Selection Operator (LASSO) regression, and finally, multivariate logistic regression was used to determine independent influencing factors, and their odds ratios (OR) and 95% confidence intervals (CI) were calculated. Based on the results of multivariate analysis, machine learning models such as random forest, support vector machine, and K-nearest neighbor algorithm were further constructed using Python 3.8.5 and the sklearn library. The detailed implementation details were as follows: Random forest: n_estimators = 100, max_depth = 10, random_state = 42; Support vector machine: kernel = ‘rbf’, C = 1.0, gamma = ‘scale’, random_state = 42; K-nearest neighbor algorithm: n_neighbors = 5, weights = ‘distance’. The receiver operating characteristic (ROC) curve was drawn, and the area under the curve (AUC) was used to evaluate the prediction performance of the model. At the same time, the “rms” package in R software was used to construct a nomogram prediction model, and internal validation was carried out by the Bootstrap method (1,000 repeated samplings). The concordance index (C-index) was calculated to evaluate the discrimination ability, and the calibration curve was drawn to evaluate the calibration degree. Decision curve analysis (DCA) was used to evaluate the clinical application value of the nomogram by calculating the net benefit at different threshold probabilities. In addition, the “SHAP” library in Python was used to calculate SHAP values to evaluate the interpretability of the model from both global (feature importance ranking, contribution direction) and local (risk contribution decomposition of single-case patients) dimensions. Combined with the visualization function of the nomogram, the clinical interpretability of the model prediction results was achieved. A p-value < 0.05 was considered statistically significant.

A total of 340 patients with myofascial pain syndrome who completed 8-week standardized treatment were included in this study and randomly divided into a training set (n = 238) and a validation set (n = 102) at a ratio of 7:3. In the training set, there were 143 patients (60.1%) with excellent treatment efficacy and 95 patients (39.9%) with poor treatment efficacy. The validation set included 62 patients (60.8%) with excellent treatment efficacy and 40 patients (39.2%) with poor treatment efficacy. There were no statistically significant differences in baseline data such as age, gender, disease duration, pain characteristics (such as pain duration, distribution range), trigger point characteristics (such as number, tenderness threshold), psychometric indicators (such as PHQ-9, GAD-7, PCS scores), and inflammatory markers (such as IL-6, TNF-α, hs-CRP) between patients in the training set and the validation set (p > 0.05), indicating that the data set was evenly divided and had good comparability (Table 1).

Among the 238 patients in the training set, 143 had excellent treatment efficacy. Among the 102 patients in the validation set, 62 had excellent treatment efficacy. Univariate analysis showed that in the training set, there were statistically significant differences between the patients with excellent treatment efficacy and those with poor treatment efficacy in terms of disease duration, baseline pain intensity, average tenderness threshold, PHQ-9 depression score, pain catastrophizing score, IL-6 level, and hs-CRP level (all p < 0.05) (Table 2).

The therapeutic efficacy of myofascial pain syndrome in patients was used as the dependent variable (1 = poor efficacy group, 0 = excellent efficacy group). Six indicators with statistical significance in the univariate analysis (disease duration, baseline pain intensity, PHQ-9 depression score, pain catastrophizing score, IL-6, and hs-CRP levels) were all included in the LASSO regression for variable screening (Table 3). The optimal variables were selected using 10-fold cross-validation and the λ-1se criterion (Figure 1). Finally, all 6 predictive variables were retained and included in the multivariate logistic regression model. The results of the multivariate logistic regression analysis showed that disease duration, baseline pain intensity, PHQ-9 depression score, pain catastrophizing score, IL-6, and hs-CRP levels were independent risk factors for poor therapeutic efficacy (all p < 0.05) (Table 4).

By systematically comparing the performance of each model on the training set and the validation set, the results in Figure 2 showed that the AUC value of the support vector machine model reached 0.895 (95%CI:0.840–0.950) in the training set and remained at a relatively high level of 0.873 (95%CI: 0.794–0.953) in the validation set, which were significantly better than those of other comparison models (Random forest: 0.880/0.860, K-nearest neighbor algorithm: 0.782/0.770). Therefore, the support vector machine model was determined as the best prediction model in this study. This result proves that the support vector machine model has a stable and excellent ability to distinguish patients with poor therapeutic efficacy of myofascial pain syndrome. Further analysis of the calibration curve in Figure 3 revealed a high degree of consistency between the predicted probability of the model and the actual observed risk, with the curve closely fitting the diagonal, indicating good prediction calibration and reliable prediction results. In addition, the DCA in Figure 4 confirmed that within a wide threshold probability range of 0.10–0.70, the clinical net benefit of applying this prediction model was always significantly higher than that of the extreme strategies of “treating all” or “not treating all”, demonstrating its outstanding clinical practical value. In summary, the support vector machine prediction model developed based on multi-dimensional clinical indicators in this study has good prediction accuracy, calibration, and clinical applicability.

Based on the six core predictive variables screened by the multivariate Logistic regression analysis (disease duration, baseline pain intensity, PHQ-9 depression score, pain catastrophizing score, IL-6, and hs-CRP levels), this study used the support vector machine algorithm to construct a nomogram model for predicting the therapeutic efficacy of MPS. As shown in Figure 5, this nomogram visually demonstrated the contribution degree and direction of each clinical feature to the treatment outcome. The model results showed that disease duration (X1), baseline pain intensity (X2), PHQ-9 depression score (X3), pain catastrophizing score (X4), IL-6 (X5), and hs-CRP (X6) were all independent risk factors for poor therapeutic efficacy, and an increase in their values would significantly increase the risk of treatment ineffectiveness. SHAP analysis further quantified the relative importance of each feature, and the order of influence from large to small was: PHQ-9 depression score (X3), IL-6 level (X5), pain catastrophizing score (X4), baseline pain intensity (X2), disease duration (X1), and hs-CRP level (X6). Among them, the PHQ-9 depression score and IL-6 level made the most prominent positive predictive contributions to treatment ineffectiveness. Although disease duration and baseline pain intensity were also risk factors, their relative influence weights were slightly lower than those of psychological and inflammatory indicators. The results of the SHAP analysis were highly consistent with the conclusions of the multivariate logistic regression, confirming the core position of psychological factors and inflammatory markers in predicting the therapeutic efficacy of MPS.