Participants
The study protocol was approved by the Institutional Review Board of the First Affiliated Hospital of the Medical College of Xi’an Jiaotong University and was conducted according to the Declaration of Helsinki. All participants provided informed written consent before participation.
PDM subjects were screened and the diagnosis was confirmed by a gynecologist from the Department of Obstetrics and Gynecology at the First Affiliated Hospital of the Medical College of Xi’an Jiaotong University. The inclusion criteria for women with PDM included: (1) those satisfying the American College of Obstetricians and Gynecologists' diagnostic criteria for PDM; (2) those who have a regular menstrual cycle around 27–32 days; (3) those experiencing sharp abdominal pains during menstruation but no obvious abnormalities are found in the pelvic organs; and (4) those whose average menstrual pain rating is at least higher than 4 (0 = not at all, 10 = the worst imaginable pain) in the last 6 months. The exclusion criteria for all subjects were as follows: (1) individuals with organic pelvic disease; (2) those with psychiatric or neurological disorders; (3) those with brain trauma or brain surgery; (4) those with other comorbid chronic pain conditions (e.g., headache, fibromyalgia, irritable bowel syndrome, etc.); (5) those with immediate plans for pregnancy or a positive pregnancy test; (6) those using oral contraceptives and analgesics/antidepressants within the last 6 months; (7) those with a history of childbirth; and (8) those with any contraindications to MRI scans. The inclusion criteria for HCs included: (1) those who have a regular menstrual cycle around 27–32 days; and (2) those who are without pain during menses.
Clinical measurements, pain empathy evaluation, and MRI scans
For all subjects, basic information about the menstrual cycle and age at menarche was recorded. For each woman with PDM, the history and duration of menstrual pain were also recorded. Using a numerical rating scale (NRS) ranging from 0 (no pain) to 10 (the worst imaginable pain), the menstrual pain intensity was evaluated three times for three menstrual phases (1–2 days of the menstrual cycle). The trait pain was defined as the average NRS score of these three menstruation phases, and the state pain was defined as the NRS score during the 3rd menstrual phase. During the 3rd menstrual phase, an MRI scan was performed and pain empathy was evaluated using the picture stimulation paradigm (Fig. 1a).
For each HC, they underwent two MRI scans during the menstrual cycle, the first in the menstruation phase (1–2 days of the menstrual cycle) and the second in the luteal phase (7–8 days after ovulation). Each subject used an ovulation kit to verify the exact timing of ovulation. To further confirm whether the subjects were in the luteal phase, we measured progesterone and estrogen levels before the second scan. During the menstrual phase, state pain and pain empathy were evaluated (Fig. 1b).
Fifty-seven PDM subjects and fifty-three age-matched asymptomatic female controls from the local university were recruited in this study. Six individuals with PDM and eight individuals with HC were excluded from data analysis due to failure to acquire high-quality scanning images or complete behavioral and neuroimaging data. In addition, six and four individuals were excluded from the PDM and HC groups due to technical failures during the visual stimuli experiment, respectively. Forty-five PDMs and forty-one HCs were used in subsequent analyses.
Visual stimuli
To evoke the empathic pain of participants, 48 color pictures of the hands or feet of individuals in painful or nonpainful situations were presented as visual stimuli (Jackson et al., 2005). Jackson et al. had previously developed and validated the picture stimuli used in this study (Jackson et al., 2005). These pictures depicted daily life scenarios and were smoothed to avoid biases in judgment caused by age and gender. The details of the stimulation paradigm are as follows: participants received 24 trials divided into 4 blocks (each block consisted of 6 trials). Each trial consisted of the presentation of a picture (pain or neutral) for 4 s, followed by a fixation cross (1 s) and the pain rating (6 s). The interval between the trials (a black fixation cross on a white background) was randomly varied between 6 and 10 s (mean duration = 8 s). After each picture was presented, participants were asked to consider how the subject felt (0 = no pain, 10 = the worst pain). In our study, pain empathy was assessed during the menstruation phase in both HCs and women with PDM. For each subject, the pain empathy score was calculated as the difference between the sum of pain ratings for each pain picture and the sum of pain ratings for each no pain picture.
Image acquisition
MRI scans were carried out in a 3.0 Tesla Signa GE scanner with an 8-channel phase array head coil. For each subject, we acquired: (1) a T1-weighted anatomical scan (3D IR-FSPGR sequence, voxel size = 1 × 1 × 1 mm3, data matrix = 256 × 256, repetition time (TR) = 1900 ms, echo time (TE) = 2.6 ms, field of view (FOV) = 256 × 256 mm2), and (2) a T2*-weighted resting-state fMRI scan (echo-planar imaging sequence, voxel size = 3.75 × 3.75 × 4 mm3, data matrix = 64 × 64, TR = 2000 ms, TE = 30 ms, FOV = 240 × 240 mm2, flip angle = 90º, slices = 35 with no gap).
Data preprocessing
Functional MRI data were performed using the DPARSF toolbox (A Toolbox for Data Processing & Analysis for Brain Imaging, http://www.rfmri.org/dpabi) in MATLAB. All functional images were accepted to slice timing, co-registration to the individual’s anatomical image, normalization to the Montreal Neurological Institute (MNI-152) template, outlier detection, and smoothing with a 6 mm full-width at half-maximum Gaussian kernel.
Intrinsic connectivity network definition and FNC analysis
In this study, the brain networks used for sFNC and dFNC analysis were defined using group independent component analysis (GICA) with the GIFT toolbox (https://trendscenter.org/software/gift/) (Calhoun et al., 2001). Specifically, we performed GICA with the GIFT toolbox to parcellate the fMRI data into multiple independent components (ICs), and then identified 53 reliable intrinsic connectivity networks (ICNs) from 100 ICs according to their spatial maps (Du et al., 2020). Considering prior knowledge of their functional meanings and the anatomical information, 53 ICNs were categorized into 7 brain functional networks, including: the subcortical network (SCN), auditory network (AUD), sensorimotor network (SMN), visual network (VSN), cognitive control network (CCN), default mode network (DMN), and cerebellar network (CBN) (Du et al., 2020) (see details in Supplementary Table 1).
Additional post-processing steps were performed on the time courses of ICNs, which included: 1) detrending linear, quadratic, and cubic trends; 2) conducting multiple regressions of the 24 realignment parameters and their temporal derivatives; 3) de-spiking detected outliers; and 4) low-pass filtering with a cut-off frequency of 0.15 Hz.
Based on the processed time courses of ICNs, we calculate both static and dynamic FNC. For sFNC measurement, the Pearson correlation coefficient of the whole time series between different ICNs is adopted. Specifically, 53×53 sFNC matrices containing 1,378 unique connectivities were constructed for each subject and then transformed to the z-values using Fisher’s r-to-z transformation. For dFNC measurement, a sliding window approach was used (Allen et al., 2014). The window length is 22 TRs (44 s), sliding along 175 TRs (250 s) in steps of 1 TR (2 s) and yielding a total of 153 continuous time windows. For each time window, the covariance matrix was calculated by using the windowed time series of ICNs to measure the dFNC between the 53 ICNs. Thus, each subject has 153 dFNC matrices with a size of 53×53. To quantify the variability of dFNC over time, we computed the standard deviation of the dFNC value across time windows. Thus, each subject obtained a 53×53 dFNC variability matrix (dFNC-var).
In addition, studies have demonstrated that resting state FNC brain networks show recurring spatial FNC states that are reproducible over time and across individuals (Abrol et al., 2017). To identify recurring dFNC states over time, we applied a k-means clustering algorithm on the windowed covariance matrices of all subjects. A cluster validity analysis using a silhouette method was used to estimate the optimal number of clusters, and it was determined to be 2 (state 1 and state 2) (Kim et al., 2017). The windowed covariance matrices were then averaged within each state. Thus, each subject had a 53 × 53 dFNC matrix corresponding to each state (dFNC-state1, dFNC-state2).
We also investigated the temporal characteristics of dFNC states, as expressed by the mean dwell time. The mean dwell time was calculated by averaging the number of consecutive windows belonging to the same state.
Study 1: Identifying and validating the brain FNC-based features for predicting pain empathy in HCs
To explore the association between brain FNC and pain empathy scores during the menstruation phase in HCs, we applied PLSR analysis to predict pain empathy scores by using features of sFNC, dFNC-var, dFNC-state1 and dFNC- state2, respectively (see Fig. 2 for methods schematic). We conducted PLSR within a nested-cross validation framework to evaluate the generalizability of the outputted brain model for pain empathy scores. There are 6 steps involved in the analysis pipeline. Step 1: We used the following data as input matrices: X is an n×m data matrix, Y is a factor vector of length n, where n is the number of subjects and m is the number of features. Step 2: The total dataset was split into 5 sets, one set was selected as the validation set, and the other 4 sets were combined into the training set for each outer-fold. Step 3: Within the inner-loop, we performed 100 times bootstrap to the sample training set and split it into a training set (80%) and a testing set (20%) (Xu & Goodacre, 2018). During inner-fold cross-validation, competitive adaptive reweighted sampling (CARS) coupled with PLSR was performed to select a combination of key brain features that had the capacity to predict pain empathy (Li et al., 2009). Step 4: Then, frequency-based feature selection was performed to select the stable features whose frequencies were greater than 50 in the 100 iterative subsets of the features. Step 5: Based on these stable brain features, PLSR was performed on the entire training set of the outer-loop to derive the multivariate brain model of pain empathy for each fold. Step 6: The final feature set was the combination of features from all folds, and its coefficients were averaged across all folds. To evaluate the performance of this model, we calculated the coefficient of determination (R²) between the estimated values of pain empathy in all folds of the outer-loop and the actual values.
After the above analysis, we obtained a prediction model based on each FNC matrix (sFNC, dFNC-var, dFNC-state1, and dFNC- state2). The model corresponding to the maximum R² was considered as the final prediction model in HCs.
We characterized the prediction model from three perspectives, including the best predictive FNC features, network degree, and weight pattern of the network. The network degree was calculated as the number of predictive FNCs passing through the network divided by the number of ICNs contained in this network, and the network weight pattern was estimated by adding the absolute weights of predictive FNCs passing through the network.
Test-retest reliability of the brain FNC-based features
To test-retest reliability of the identified brain FNC-based features, we predicted pain empathy scores in HCs by using the new imaging data from the luteal phase. The same data analysis pipeline was performed. Then, we compared whether the prediction model from the luteal phase was similar to that from the menstrual phase.
Study 2: Investigating the specificity of the identified brain FNC-based features for predicting pain empathy in women with PDM
To explore the association between brain FNC and pain empathy scores during the menstruation phase in women with PDM, we also used PLSR to predict pain empathy scores based on FNC features (sFNC, dFNC-var, dFNC-state1, and dFNC-state2) in women with PDM. The same data analysis pipeline was performed.
We tested whether the prediction model from HCs can be applied to women with PDM; and we also tested whether the prediction model from women with PDM can accurately predict the pain empathy of HCs. We assessed the relationship between the pain scores of women with PDM (state and trait pain) and predictive features of pain empathy using regression analysis.
Study 3: Identifying the abnormal brain FNC pattern of PDM
To identify the abnormal brain FNC pattern in women with PDM as compared with HCs, we first performed statistical comparisons for the temporal characteristics of dFNC states between different groups. Then we used logistic regression to construct the brain classification model by using FNC features of sFNC, dFNC-var, dFNC-state1, and dFNC-state2, respectively. Specifically, for each FNC feature, we first used the ReliefF algorithm (Robnik-Šikonja & Kononenko, 2003) to reduce the dimensionality of the initial feature. Then, feature selection was carried out based on the Gain algorithm, and a feature combination consisting of 25 high-correlation but low-redundancy measures was found from 100 bootstrap training samples (Vallières et al., 2015). After feature selection, we performed step-wise forward model construction based on the maximum area under the curve (AUC) in 1000 bootstrap training and testing samples, and further constructed logistic regression models combining 1 to 10 measures (Vallières et al., 2015). Finally, the best brain classification model among the 10 models was selected according to the maximum AUC, and the final model of each FNC feature was obtained. The model corresponding to the maximum AUC was considered the final classification model for women with PDM. The relationship between the most discriminative features and trait/state pain intensity was also investigated. Then, we assessed whether the abnormal brain pattern and pain empathy model in women with PDM were similar or different across the most important features of the respective models.