ParticipantsA total of 131 subjects were included in this study, 6 of them were excluded by the image quality control. 71 subjects were in the HC group, 46 of them are cases in our hospital, and 25 of them were the normal health from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset. 54 subjects were in the CP group, 35 of them were fibromyalgia-ness chronic pain, and 19 of them were other over 3 months chronic pain which includes rheumatoid arthritis, chronic low back pain, and ankylosing spondylitis. All FMS-ness patients had been examined by a rheumatologist and a neurologist, and divided into the typical FMS group and the sub-clinical FMS group. The typical group fulfilled the diagnostic criteria for FMS according to the guidelines released by the American College of Rheumatology 2016 reversion22: (1) Presence of generalized pain (four quadrants and axial of five regions). (2) Symptoms have been present at a similar level for at least 3 months. (3) Widespread Pain Index (WPI) ≥ 7 and Symptom Severity (SS) ≥ 5, or WPI of 4–6 and SS ≥ 9. The sub-clinical groups exhibited similar fatigue or sleep disturbance for at least 3 months, with the presence of at least 3 regional pain points (WPI ≥ 3). Our hospital conducted comprehensive examinations on both groups to rule out other possible diagnoses. Additionally, both groups completed the self-rating anxiety scale (SAS) and self-rating depression scale (SDS). The Institutional Review Board (IRB) of the Zhejiang Provincial People’s Hospital (IRB No.2023KY053) reviewed and approved the study protocol. All patients gave their written informed consent, following the principles of the Declaration of Helsinki.Image acquisitionHigh-resolution 3D-FSPGR images were acquired by GE Discovery 750 3.0T with the following parameters: TE = 2.9 ms, TR = 6.7 ms, flip angle = 15 degree, FOV = 216 × 216 mm, matrix = 256 × 256, slice thickness = 1 mm, 192 scanning images. The date from ADNI were also acquired by the same machine and similar parameters: TE = 3.1 ms, TR = 7.4 ms, flip Angle = 11 degree, FOV = 216 × 216 mm, matrix = 256 × 256, slice thickness = 1 mm, 192 scanning images.Image preprocessing and segmentationData were preprocessed with the Computational Anatomy Toolbox (CAT12.8.2 r2130, http://www.neuro.uni-jena.de/cat/) ran under Statistical Parametric Mapping, Version 12 (SPM12, http://www.fil.ion.ucl.ac.uk/spm/software/spm12/)23. To ensure quality, we visually inspected all raw images for artifacts and statistically controlled all segmented images for inter-subject homogeneity by the ratio between weighted overall image quality and quartic mean Z-score. All the structural images were segmented into GM, WM, and cerebrospinal fluid (CSF), and were normalized to Montreal Neurological Institute (MNI) standard space by Geodesic Shooting templates. To preserve the volume of GM and WM, we applied “modulation” during the normalization step and then resampled the images to 1.5 × 1.5 × 1.5 mm3. Finally, we smoothed the modulated images with an isotropic 8 mm full-width half maximum Gaussian kernel. The graphical representation delineates the steps involved in the image processing pipeline, as illustrated in Fig. 1.Figure 1Radiomics signatures extractionRadiomics features were extracted using the PyRadiomics open-source Python package (version 2.1.0; https://pyradiomics.readthedocs.io/)24. The width of discretization bins for feature extraction was fixed to 25. We extracted the standard feature classes, which included shape, first-order, texture, wavelet, exponential, and square transform features25. ComBaTool and z-scores were then used to normalize the features. ComBaTool, a free online application (https://forlhac.shinyapps.io/Shiny_ComBat/)26, was used to pool features and minimize inter-scanner variability. Principal component analysis was utilized to visualize the effects of Combat on feature uniformization27. Finally, all radiomics features were standardized using z-scores. All features were extracted from transformed mwp0* images with the individual masks of GM and WM. Details can be found in the software package and the source code28.Feature selection and binary logistic regression (LR) model constructionThe patients were selected using random stratified sampling, with consistent control of the CP/HC ratio in the training and validation groups at 7:3. The model was constructed and features were selected from both GM and WM to distinguish between the two groups. The selected features were screened through three steps: (1) include features with significant differences using an independent-sample t-test or Mann-Whitney U test (p < 0.05). (2) Elimination of internal redundant features using Spearman’s rank correlation test (threshold of 0.8). (3) Reduction of inter-sequence redundancy and selection of the best predictive features using the least absolute shrinkage and selection operator (LASSO)- binary logistic regression model. The best lambda to determine non-zero regression coefficient variables was determined through a 10-fold cross-validation process. Important WM and GM features were used to construct respective models through logistic analysis. After eliminating a feature, logistic modeling analysis is conducted by merging the characteristics of both regions.Establishment and validation of the XGBoost modelTo classify CPs and HCs in the validation dataset, an XGBoost model was built using the training dataset. GridSearchCV was used to optimize the model parameters. The radiomics signatures were arranged and combined, and cross-validation was used to return evaluation index scores for all parameter combinations. The optimal value was selected as the parameter corresponding to the combination with the highest score, and the model prediction probability is taken as rad-score.To comprehend the reasons for the complexity of XGBoost models, Shapley Additive exPlanations (SHAP) is utilized for analyzing the correlation between features and outputs in the XGBoost “black box“29. With SHAP analysis, the contribution of each feature to changes in the model output is represented by its SHAP value. The prediction results are linearly decomposed into the influence of individual features, which enables the calculation of feature importance and visualization of the role of different features in the model based on their sensitivity to changes in output.Combined nomogram model buildingThe combined model was established using the rad-score of the radiomics model with the best prediction performance and the highly related clinical indicators of FMS by logistic regression analysis, which was presented in the form of a nomogram30. The accuracy of quantitative prediction was evaluated using the area under curve (AUC). The consistency between the predicted results and the actual results was evaluated using a calibration curve, and its clinical effectiveness was evaluated using a decision curve analysis (DCA).Statistical analysisThe study used two software programs, SPSS 26.0 and R software (4.1.3, http://www.r-project.org), to perform statistical analysis. A statistically significant difference was considered when P< 0.05. Clinical data were analyzed using the SPSS software. The chi-square test was used for classified variable analysis, the t-test for continuous variables of normal distribution, and the Mann–Whitney U test for abnormal or unknown distribution. Univariate and multivariate logistic analyses were performed to identify clinical indicators that have a high correlation with FMS. The R software was used to establish and evaluate the nomogram. The software packages “car”, “rms”, “pROC”, and “DecisionCurve” were used to analyze the nomogram, receiver operating characteristic (ROC) curve, calibration curve, and decision curve analysis (DCA)31.
