Predicting acupuncture efficacy for neck pain based on functional connectivity features: a machine learning study

Abstract

Objective

To explore the mechanisms of acupuncture-induced neck pain relief and identify appropriate candidates using neuroimaging and machine learning techniques.

Methods

Eighty neck pain patients were included, with clinical data and functional magnetic resonance imaging scans collected pre- and post-treatment. A support vector machine (SVM) model was built using pre-treatment brain functional connectivity to predict acupuncture responsiveness, identifying key features as potential biomarkers for effectiveness. Longitudinal analysis of these features was conducted in responders and non-responders.

Results

This study enrolled 80 neck pain patients (48 acupuncture responders and 32 non-responders) for SVM model construction and longitudinal analysis of predictive features pre-/post-treatment. The SVM model achieved an accuracy of 0.85 in distinguishing the two groups. A total of 117 functional connectivity edges were identified as predictive features, potential biomarkers for acupuncture responses. Longitudinal analysis showed 6 predictive features altered post-treatment in responders versus 44 in non-responders. After FDR correction, only 3 functional connectivity features in responders negatively correlated with pain VAS scores (p < 0.05). These findings indicate more targeted changes in predictive features among responders compared to non-responders.

Conclusion

Using pre-treatment neuroimaging features to predict acupuncture effectiveness for neck pain shows promise. This approach could aid in developing personalized acupuncture strategies by identifying likely beneficiaries, guiding alternative interventions for non-responders.

Trial registration

International Traditional Medicine Clinical Trial Registry (registration number: ITMCTR2023000001, protocol version number: V1.0)

Background

Neck pain (NP), as a common musculoskeletal disorder, has become an important healthcare and social problem due to its high prevalence [1], heavy economic burden [2], and serious reduction in quality of life [3]. Traditional treatments for NP often fall short of expectations. While medications such as muscle relaxants and non-steroidal anti-inflammatory drugs provide relief for acute pain, their potential side effects, including hepatotoxicity, cannot be overlooked [4]. Consequently, both patients and healthcare providers are increasingly interested in effective alternative therapies. These include chiropractic care, which can offer mechanical adjustments to alleviate pain [5], and physiotherapy, which uses exercises and manual therapy to enhance mobility and strength [6].

Acupuncture, as a component of traditional Chinese medicine, has been extensively used in the treatment of pain for more than 3000 years. Numerous studies have shown that acupuncture can effectively reduce pain and disability, and improve quality of life in patients with neck pain [1,7,8]. However, owing to individual variations such as physiological conditions, the effectiveness of acupuncture for neck pain can differ considerably across patients. Being able to predict a patient’s individual response to acupuncture before treatment would aid in tailoring more personalized treatment plans for those suffering from chronic pain.

Several studies have attempted to predict the clinical efficacy of physical therapies such as acupuncture for neck pain based on clinical characteristics. For example, Zhang et al. constructed an effective prediction model to assess the effectiveness of acupuncture in treating different patterns of neck pain in traditional Chinese medicine and explored the potential differences between different patterns [9]. Moustafa et al. found through a linear regression model that cervical lordosis angle is a key predictor of the prognosis of patients with neck pain after cervical traction [10]. These earlier studies offer promising initial evidence and confirm the practicality of forecasting the outcomes of acupuncture treatments based on patients’ initial health status. Nevertheless, clinical data collection is frequently skewed by the personal biases of healthcare providers, which can hinder a precise and unbiased representation of patients’ conditions. Consequently, there is a pressing demand for more objective and effective biomarkers that can accurately indicate patients’ health status and predict the benefits of acupuncture for neck pain sufferers.

Brain imaging shows potential as an objective tool for identifying chronic pain, particularly since a 2015 ruling by a U.S. court allowed fMRI evidence to support claims of chronic pain [11,12]. Functional magnetic resonance imaging (fMRI) has been extensively utilized in studying the central processes underlying chronic neck pain, highlighting that changes in brain connectivity are key neuropathological indicators of neck pain [13–15]. Moreover, acupuncture has been shown to significantly adjust abnormal resting-state functional connectivity (rsFC) patterns in individuals with neck pain [16]. Based on these observations, we propose that brain functional connectivity may forecast the therapeutic outcomes of acupuncture for neck pain sufferers.

To examine our hypothesis, we constructed a predictive model using brain functional connectivity data. This model aims to determine the extent to which baseline rsFC characteristics can identify individual differences and forecast how patients with neck pain will respond to acupuncture. Additionally, the study examines the longitudinal changes in these predictive features before and after acupuncture treatment, comparing responders to non-responders among neck pain patients. This comparison is intended to further investigate the connection between these predictive features and the outcomes of acupuncture treatment. The findings of this research could lead to tailored treatment approaches for individuals with chronic pain, enhancing the efficiency of medical resource allocation, increasing the likelihood of successful treatments, and minimizing adverse effects. Furthermore, the results may contribute to a better understanding of the workings of acupuncture, scientific inquiry and interdisciplinary collaboration in acupuncture research, and provide a basis for improving patients’ quality of life and advancing the field of precision medicine.

Methods

Participants

Study protocols were approved by the Institutional Review Board of Shanxi Acupuncture Hospital (approved number: 2023-007) and registered at International Traditional Medicine Clinical Trial Registry (registration number: ITMCTR2023000001, protocol version number: V1.0). All participants provided voluntary written consent following ethical principles outlined in the Declaration of Helsinki before study initiation. Methodology and reporting were aligned with CONSORT recommendations.

Patients with neck pain were recruited from the campus of Shanxi University of Chinese Medicine and the Shanxi Provincial Hospital of Acupuncture between March 2023 and May 2024. To be included in the study, participants had to meet the criteria for NP as outlined by the American Physical Therapy Association’s Orthopedic Section [17]. Eligible participants were those who: 1) primarily complained of NP with a Visual Analog Scale (VAS) pain severity score of more than 3 points (out of a possible 10), 2) had experienced symptoms for no more than 3 months, 3) were between 18 and 60 years old and right-handed, and 4) were willing to provide informed consent. Those who met any of the following criteria were excluded: 1) pregnant or breastfeeding women, 2) individuals with serious primary health conditions or psychiatric/neurological disorders, 3) those who had MRI contraindications, and 4) participants enrolled in other clinical trials.

Intervention

The acupuncture treatment group had four Ashi points selected in the neck area. These Ashi points were identified using a handheld pressure pain tester (Force One FDIX, Wagner Instruments, Greenwich, CT, USA).

For acupuncture, single-use sterile needles (0.30 × 40 mm, Huatuo Medical Instrument Co., Ltd., China) were employed. The insertion depth and direction were determined based on the patient’s treatment site, with the aim of achieving the Deqi sensation. The needles were then retained for 30 min.

Participants received three acupuncture sessions weekly, totaling six treatments over a two-week period.

Outcome measures

All patients assessed their neck pain symptoms using the VAS at baseline and after treatment (on a scale of 0-10, where 0 indicates no pain and 10 indicates intolerable pain). Following the criteria set by Romy Lauche et al. in a randomized controlled study on needle cupping for neck pain [18], patients with a VAS score improvement of ≥21% after acupuncture treatment were classified as “acupuncture responders”, while those with an improvement of <21% were classified as “acupuncture non-responders”. “Acupuncture responders” refer to patients who show a significant positive response to acupuncture treatment, benefiting greatly from it, with notable symptom improvement after treatment. In contrast, “acupuncture non-responders” indicate patients with a minimal response to acupuncture, exhibiting little relief in symptoms post-treatment.

MRI data acquisition

Resting-state fMRI acquisitions were conducted on a Siemens Magnetom Trio 3 T system (Erlangen, Germany) at Shanxi Provincial People’s Hospital, following a standardized protocol [19,20]: 1) Anatomical localization, 2) 3D-T1WI utilizing magnetization-prepared rapid gradient-echo (MPRAGE) sequence: repetition time (TR) = 1900 ms, echo time (TE) = 2.26 ms, 176 sagittal slices, 256 × 256mm² field of view (FOV), 1 mm isotropic voxels; 3) BOLD-fMRI employing gradient-echo planar imaging: 2000 ms TR, 30 ms TE, 240 × 240mm² FOV, 128 × 128 acquisition matrix, 3.5 × 3.5 × 5 mm voxel size, 31 transverse slices (5 mm thickness), 90° flip angle, totaling 240 functional volumes.

fMRI preprocessing

fMRI data preprocessing was performed via DPARSF toolbox (http://www.fil.ion.ucl.ac.uk/spm/) in MATLAB 2022b [20]. The initial ten temporal volumes were discarded to remove signal instability. Slice timing correction resolved inter-slice acquisition delays, while head motion correction aligned volumes through rigid-body registration. Anatomical normalization reoriented individual data to Montreal Neurological Institute (MNI) template coordinates. Spatial smoothing utilized a 4-mm full-width-at-half-maximum Gaussian kernel to reduce noise. Datasets exhibiting head motion exceeding 2 mm translation or 2° rotation in any axis were excluded to mitigate motion-induced artifacts, enhancing analytical reliability.

Functional connectivity construction

Using the GRETNA toolbox within MATLAB 2022b, a functional connectivity matrix was created using the AAL 116 template. The preprocessed fMRI data were first loaded into GRETNA, and the AAL 116 template was applied to segment the brain into 116 regions. The time series for these 116 regions were then extracted from each participant’s fMRI data, and their Pearson correlation coefficients were computed to form a 116 × 116 functional connectivity matrix. This matrix provides a clear depiction of the functional connectivity strengths between various brain regions.

Machine learning algorithm

The LIBSVM toolbox in MATLAB 2022b was applied to forecast the outcomes of acupuncture therapy. A Support Vector Machine (SVM) algorithm with a linear kernel was used to develop a prediction model that distinguishes between patients who respond to acupuncture and those who do not.

Label

In the SVM model, the labels were designated as “1” and “-1”, with “1” indicating patients who respond to acupuncture and “-1” indicating those who do not respond.

SVM prediction models construction and performance evaluation

Using a leave-one-out cross-validation (LOOCV) framework with grid search for hyperparameter tuning, each sample was iteratively designated as the test set while the remaining samples formed the training set, enabling an unbiased estimate of the model’s generalizability to unseen data.

Due to the high-dimensional nature of brain functional connectivity data, feature selection was conducted using the F-score method, which evaluates the discriminative power of each functional connectivity feature by calculating the ratio of between-class variance to within-class variance. Features were ranked according to their F-scores and incrementally added to assess model performance through LOOCV, determining the optimal subset. This process aims to enhance the accuracy and generalizability of the model while reducing the risk of overfitting.

The model’s performance was assessed using metrics such as accuracy, sensitivity, specificity, and the Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC) curve. The significance of the prediction outcomes was determined through permutation testing, conducted with 1000 iterations (statistical significance level = 0.001).

Predictive features identification and weight calculation

Through multiple rounds of leave-one-out cross-validation, features that consistently emerge as the best choices are known as consistent features. These features demonstrate robustness and uniformity across various training and testing combinations, indicating their frequent importance. The higher the weight of a consistent feature, the more pronounced its impact on differentiating between classes.

The changes of the predictive features after acupuncture treatment

To delve deeper into the link between predictive features and acupuncture outcomes, this study examined the alterations in functional connectivity of the identified predictive features both before and after acupuncture treatment. The analysis involved: 1) observing changes in predictive features for patients who responded to acupuncture before and after treatment, 2) observing changes in predictive features for patients who did not respond to acupuncture before and after treatment, and 3) analyzing the correlation between changes in functional connectivity and clinical VAS scores for all patients. A significance threshold of p < 0.05 was used for these analyses.

Statistical analysis

Statistical analyses for all demographic data and VAS, SAS, SDS scores were conducted using the statistical program SPSS 26.0. The normality of all continuous variables was assessed using the Shapiro-Wilk test (α = 0.05). For intergroup comparisons of baseline characteristics, independent samples t-tests were used for normally distributed continuous variables, and the Mann-Whitney U test was used for non-normal data; categorical variables were analyzed using the chi-square test. Intragroup comparisons before and after treatment were performed using paired samples t-tests or the Wilcoxon signed-rank test. A p-value of less than 0.05 was considered statistically significant.

To address the issue of multiple comparisons in the functional connectivity analysis (including longitudinal changes before and after treatment and correlations with VAS scores), we applied the Benjamini-Hochberg FDR method for correction, with a significance threshold set at α = 0.05.

For the machine learning component, to validate the statistical significance of the SVM model’s performance, we conducted a permutation test: patient labels (responders/non-responders) were randomly shuffled 1000 times, and the model was retrained and the accuracy rate calculated each time. The P-value for the original model’s accuracy was determined by comparing it to the distribution of random accuracy rates in the permutation distribution (one-tailed test), with a significance threshold set at p < 0.001.

Results

Demographic and clinical characteristics

A total of 90 patients were enrolled for treatment of neck pain, with 83 completing a two-week acupuncture program and a subsequent fMRI scan. Three patients were removed due to excessive head movement. Consequently, 80 patients with neck pain were included in the predictive analysis and longitudinal fMRI data analysis. Based on the established criteria, 48 patients were classified as acupuncture responders, while 32 were categorized as non-responders. Their demographic and baseline clinical data are detailed in Table 1.

Table 1.

The demographic and clinical characteristics of neck pain patients.

	Acupuncture responders	Acupuncture non-responders	Statistics	P-value
Gender (Male/Female)	16/32	10/22	Χ2 = 0.038	P = 0.845
Age (Year)	23(21-25)	23(21-26)	Z = −0.124	P = 0.902
Height (cm)	165.15 ± 7.93	164.59 ± 7.74	t = −0.308	P = 0.759
Weight (kg)	55.00(45.25-71.50)	56.50(47.25-73.75)	Z = −0.034	P = 0.973
educational attainment (Year)	15(14-16)	15(14-17)	Z = −0.871	P = 0.384
Duration (Month)	14.00(7.00-30.00)	18.00(14.00-22.75)	Z = 1.090	P = 0.276
Baseline VAS	6(5-6)	5(5-6)	Z = −1.740	P = 0.082
After VAS	3.00(2.00-3.58)	5.00(4.00-5.00)	Z = −6.373	P = 0.001*
Baseline SAS	42.50(39.06-48.44)	45.03(40.31-51.88)	Z = −0.955	P = 0.340
After SAS	37.50(31.25-42.19)	45.00(37.50-48.75)	Z = −2.896	P = 0.004*
Baseline SDS	0.43(0.36-0.51)	0.48(0.37-0.55)	Z = −1.268	P = 0.205
After SDS	0.38(0.30-0.44)	0.44(0.36-0.50)	t = −2.113	P = 0.035*

Note: Depending on the data distribution, present values as mean ± standard deviation or median (interquartile range).

Treatment effects on NP patients

The analysis within each group showed significant changes in VAS, SAS, and SDS scores for acupuncture responders before and after treatment. In contrast, for acupuncture non-responders, there was no significant change in SAS scores, but significant differences were found in VAS and SDS scores before and after treatment.

The comparison between groups indicated that following treatment, there were noticeable disparities in VAS, SAS, and SDS scores between those who responded to acupuncture and those who did not, as reflected in Table 1.

Optimal number of connections

Given the extensive number of connections ((116×116-116)/2), the SVM training and prediction utilized the feature subset with the maximum F-score to determine the number of features, ranging from 0 to 350 with an increment of 1. The final choice was the number of connections with the highest accuracy, which was 217. Figure 1 depicts the relationship between the number of features and accuracy. As depicted in Figure 1, the model’s performance initially increased rapidly with the number of features, peaked at 217, and then began to decline. Consequently, the number of features was fixed at 217.

Figure 1.

Optimal number of connections.

It shows the optimal number of features, with the model’s performance initially increasing rapidly with the number of features, peaking at 217, and then beginning to decline.

Performance of the SVC prediction model

We employed a feature selection technique derived from the F-score to enhance our SVM model. The features used were functional connectivity edges derived from the AAL116 atlas, computed as Pearson correlation coefficients between the time series of each pair of the 116 brain regions. Following 80 cycles of iteration, the resulting Support Vector Classification (SVC) model demonstrated an accuracy rate of 0.85 for distinguishing between acupuncture responders and non-responders. It exhibited a specificity of 0.83, a sensitivity of 0.91, and an AUC of 0.91 (refer to Figure 2). The accuracy was found to be statistically significant according to permutation testing (P_accuracy < 0.001).

Figure 2.

Receiver operating characteristic curve.

Predictive features identification and weight calculation

We pinpointed 117 edges as predictive consistency features, which could serve as potential biomarkers for forecasting how patients with neck pain will respond to acupuncture therapy (refer to Figure 3a and 3b). The five predictive features with the greatest significance are the left anterior cingulate and paracingulate gyri - vermis_10 (with a weight of 0.803), the right superior frontal gyrus (dorsolateral) - vermis_1_2 (with a weight of 0.791), the right posterior cingulate gyrus - right middle temporal gyrus (with a weight of 0.773), the right middle occipital gyrus - left angular gyrus (with a weight of 0.706), and the left cerebellum_crus1 - left cerebellum_10 (with a weight of 0.686). A comprehensive breakdown of these predictive features and their respective weights can be found in Supplementary Material 1.

Figure 3.

Predictive features of acupuncture responsiveness.

Figure 3a depicts the features used for predicting the response to acupuncture, and Figure 3b shows the importance or weight of each of these features.

Changes of predictive features before and after treatment

As an exploratory analysis, we investigated the changes in predictive consistency features before and after acupuncture treatment. In those who responded to acupuncture, we observed significant changes in six predictive features after two weeks of treatment. Notably, the connectivity between the right posterior cingulate gyrus and the right middle temporal gyrus saw a marked reduction following treatment (p < 0.05, uncorrected). Conversely, there were notable increases in the following features post-treatment: left rectus gyrus - vermis_9, left cerebellum_crus1 - left cerebellum_3, left anterior cingulate and paracingulate gyri - right cuneus, left anterior cingulate and paracingulate gyri - right calcarine fissure and surrounding cortex, and left anterior cingulate and paracingulate gyri - right posterior cingulate gyrus (all with p < 0.05, uncorrected) (refer to Figure 4a and Table 2). Among those who did not respond to acupuncture, a total of 44 predictive features, including the right anterior cingulate and paracingulate gyri - vermis_10, experienced significant decreases after treatment (p < 0.05, uncorrected) (refer to Figure 4b and Table 3). Despite these changes, they did not remain statistically significant following FDR correction.

Figure 4.

Changes in brain connectivity characteristics before and after treatment in acupuncture responders and non-responders.

Figure 4a illustrates the alterations in predictive features of acupuncture responders before and after treatment, while Figure 4b portrays the changes in predictive features of non-acupuncture responders. The color red indicates an increase from before to after treatment, while blue indicates a decrease.

Table 2.

Changes in brain connectivity characteristics before and after treatment among acupuncture responders.

Number	Functional Connectivity Edge	Statistic	P-value
1	Posterior cingulate gyrus Right - Middle temporal gyrus Right	2.206	0.032
2	Rectus gyrus Left - Vermis_9	−2.297	0.026
3	Cerebellum_Crus1_Left - Cerebellum_3_Left	−2.020	0.049
4	Anterior cingulate and paracingulate gyri Left - Cuneus Right	−2.618	0.012
5	Anterior cingulate and paracingulate gyri Left - Calcarine fissure and surrounding cortex Right	−2.242	0.030
6	Anterior cingulate and paracingulate gyri Left - Posterior cingulate gyrus Right	−2.340	0.024

Table 3.

Changes in brain connectivity characteristics before and after treatment among acupuncture non responders.

Number	Functional Connectivity Edge	Statistic	P-value
1	Anterior cingulate and paracingulate gyri Right - Vermis_10	2.111	0.043
2	Posterior cingulate gyrus Right - Middle temporal gyrus Right	2.544	0.016
3	Middle occipital gyrus Right - Angular gyrus Left	2.334	0.026
4	Cerebellum_Crus1_Left - Cerebellum_10_Left	2.945	0.006
5	Cerebellum_Crus1_Right - Cerebellum_10_Left	2.934	0.006
6	Superior frontal gyrus (dorsolateral) Right - Anterior cingulate and paracingulate gyri Left	3.668	0.001
7	Posterior cingulate gyrus Left - Cerebellum_3_Right	3.081	0.004
8	Cerebellum_6_Left - Cerebellum_10_Left	3.454	0.002
9	Anterior cingulate and paracingulate gyri Left - Vermis_8	2.533	0.017
10	Inferior occipital gyrus Right - Cerebellum_10_Left	2.958	0.006
11	Posterior cingulate gyrus Right - Cerebellum_3_Right	2.912	0.007
12	Anterior cingulate and paracingulate gyri Left - Vermis_4_5	2.521	0.017
13	Cerebellum_Crus1_Left - Cerebellum_4_5_Left	2.548	0.016
14	Cerebellum_9_Right - Cerebellum_10_Right	2.546	0.016
15	Middle frontal gyrus Left - Vermis_4_5	2.677	0.012
16	Calcarine fissure and surrounding cortex Left - Cerebellum_10_Right	3.778	0.001
17	Cerebellum_Crus1_Left - Cerebellum_3_Left	2.749	0.010
18	Paracentral lobule Right - Cerebellum_9_Right	2.295	0.029
19	Lingual gyrus Left - Cerebellum_10_Right	3.869	0.001
20	Cerebellum_6_Left - Cerebellum_10_Right	3.131	0.004
21	Inferior occipital gyrus Left - Cerebellum_10_Left	2.516	0.017
22	Cuneus Left - Temporal pole: superior temporal gyrus Right	2.190	0.036
23	Middle frontal gyrus (orbital part) Right - Inferior frontal gyrus (orbital part) Left	2.362	0.025
24	Anterior cingulate and paracingulate gyri Right - Vermis_4_5	2.342	0.026
25	Cerebellum_Crus2_Left - Cerebellum_4_5_Left	2.299	0.028
26	Superior frontal gyrus (medial) Right - Thalamus Right	2.108	0.043
27	Middle occipital gyrus Right - Cerebellum_3_Right	2.359	0.025
28	Calcarine fissure and surrounding cortex Right - Cerebellum_10_Right	2.619	0.014
29	Middle frontal gyrus (orbital part) Left - Parahippocampal gyrus Left	2.382	0.024
30	Temporal pole: superior temporal gyrus Left - Cerebellum_Crus1_Left	2.666	0.012
31	Superior frontal gyrus (dorsolateral) Right - Anterior cingulate and paracingulate gyri Right	3.664	0.001
32	Precuneus Left - Cerebellum_3_Right	2.217	0.034
33	Precuneus Left - Temporal pole: superior temporal gyrus Left	2.454	0.020
34	Superior frontal gyrus (medial) Left - Vermis_4_5	3.313	0.002
35	Calcarine fissure and surrounding cortex Left -Lenticular nucleus putamen Right	2.079	0.046
36	Superior frontal gyrus (dorsolateral) Left - Vermis_4_5	2.377	0.024
37	Superior frontal gyrus (dorsolateral) Left - Lenticular nucleus putamen Left	2.310	0.028
38	Inferior frontal gyrus (triangular part) Right - Cuneus Left	3.378	0.002
39	Middle frontal gyrus (orbital part) Right - Middle temporal gyrus Right	2.665	0.012
40	Parahippocampal gyrus Left - Inferior temporal gyrus Left	2.206	0.035
41	Superior frontal gyrus (medial orbital) Left - Cerebellum_7b_Right	2.087	0.045
42	Superior frontal gyrus (medial orbital) Left - Cerebellum_3_Right	4.427	0.001
43	Cerebellum_6_Right - Cerebellum_10_Left	3.322	0.002
44	Paracentral lobule Left - Cerebellum_Crus1_Right	2.370	0.024

Additionally, the analysis exploring the relationship between predictive consistency features and VAS scores before and after treatment in both groups found that among acupuncture responders, there was a negative correlation between the functional connectivity (FC) values of the right posterior cingulate gyrus-right middle temporal gyrus, left cerebellum_crus1 - left cerebellum_3, left anterior cingulate and paracingulate gyri - right cuneus, and left anterior cingulate and paracingulate gyri - right calcarine fissure and surrounding cortex, and VAS scores (p < 0.05). On the other hand, in non-responders to acupuncture, a positive correlation was observed between the FC values of left cerebellum_crus1 - left cerebellum_10 and right middle occipital gyrus – right cerebellum_3, and VAS scores (p < 0.05) (refer to Figure 5 and Table 4).

Figure 5.

Correlation analysis between predictive features and VAS scores before and after treatment in the two groups.

a, b, c, and d show the correlation between predictive features and VAS scores for individuals who respond to acupuncture, while e and f illustrate the correlation for those who do not respond to acupuncture.

Table 4.

Correlation analysis between predictive features and VAS scores before and after treatment in the two groups.

Group	Functional Connectivity Edge	Statistic	P-value
acupuncture responders	Posterior cingulate gyrus Right - Middle temporal gyrus Right	−0.312	0.031
	Cerebellum_Crus1_Left - Cerebellum_3_Left	−0.488	0.001
	Anterior cingulate and paracingulate gyri Left - Cuneus Right	−0.428	0.002
	Anterior cingulate and paracingulate gyri Left - Calcarine fissure and surrounding cortex Right	−0.473	0.001
acupuncture non responders	Cerebellum_Crus1_Left - Cerebellum_10_Left	0.401	0.023
	Middle occipital gyrus Right-Cerebellum_3_Right	0.355	0.046

Following False Discovery Rate (FDR) correction, the negative correlation between the FC values of left cerebellum_crus1 – left cerebellum_3, left anterior cingulate and paracingulate gyri - right cuneus, and left anterior cingulate and paracingulate gyri - right calcarine fissure and surrounding cortex, and VAS scores remained significant only for acupuncture responders (p < 0.05, FDR corrected) (refer to Table 5).

Table 5.

Correlation analysis between predictive features and VAS scores before and after treatment in the two groups(FDR).

Group	Functional Connectivity Edge	Statistic	P-value
acupuncture responders	Cerebellum_Crus1_Left - Cerebellum_3_Left	−0.488	0.050
	Anterior cingulate and paracingulate gyri Left - Cuneus Right	−0.428	0.033
	Anterior cingulate and paracingulate gyri Left - Calcarine fissure and surrounding cortex Right	−0.473	0.025

Discussion

Utilizing a substantial sample size, this study investigated the predictive power of functional brain network characteristics for the outcomes of acupuncture treatment in individuals with neck pain. The findings reveal, for the first time, that the potential response of neck pain patients to acupuncture can be anticipated through the analysis of their functional brain network features prior to treatment. Notably, the alterations in resting-state functional connectivity, particularly involving the anterior cingulate cortex (ACC), were more significant in those who responded to treatment compared to those who did not. This suggests that the pre-treatment resting-state functional connectivity profiles are essential indicators of the effectiveness of acupuncture for neck pain and could potentially serve as valuable biomarkers to identify patients who would be more likely to benefit from acupuncture therapy for neck pain.

The support vector machine model based on functional connectivity features is a reliable tool for predicting the efficacy of acupuncture

Machine learning has shown potential in characterizing individual patient states and predicting treatment responses [21,22]. For example, Lu and colleagues established a predictive model using degree centrality as a feature, which accurately predicted the effectiveness of acupuncture in motor recovery following a stroke [23]. Tu and team’s study indicated that the responses of patients with chronic low back pain to both actual and placebo acupuncture could be anticipated through pre-treatment resting-state functional connectivity [24]. Yin and associates developed a support vector machine model that effectively forecasted the treatment response in patients with functional dyspepsia to acupuncture [25]. In light of previous research revealing abnormal functional brain network connections in patients with neck pain [13,15], and the fact that acupuncture can notably relieve neck pain symptoms while correcting these abnormal brain functional connections [16], this study chose whole-brain functional connectivity as a predictive element to create a model for predicting the outcomes of acupuncture treatment in individuals suffering from neck pain. In choosing a machine learning algorithm, we opted for linear SVM, which excels in the analysis of MRI data. The algorithm’s approach of identifying the maximum margin hyperplane enables it to efficiently manage high-dimensional data while preserving robust generalization performance. Furthermore, its efficient computation and the clarity of its results make it an excellent choice for feature extraction and classification tasks in the context of MRI data analysis [26,27].

Enhanced functional connectivity of the anterior cingulate cortex mediates the analgesic effect of acupuncture

Pain is a complex emotional experience involving multiple dimensions. In this study, we found that the ACC played a pivotal role among the predictive features, which aligns with its function in processing the emotional and cognitive dimensions of pain [28]. The ACC is not only involved in the processing of the unpleasantness of pain but also influences emotional, cognitive decision-making, and responses related to pain. Consequently, patients with neck pain often exhibit abnormal functional activity in the ACC due to prolonged exposure to repetitive noxious stimuli [29]. This study observed enhanced functional connectivity between the ACC and pain-related brain regions after acupuncture treatment, which may suggest that acupuncture plays a role in integrating the multidimensional experience of pain. This finding is consistent with previous research, confirming that acupuncture exerts its analgesic effects by regulating the functional connectivity network of the ACC [30]. Although these variations did not remain statistically significant after FDR correction, they still offer potential leads, suggesting that the variability in responses to acupuncture for chronic neck pain might be linked to alterations in ACC functional connectivity traits. It is noteworthy that vision has been confirmed to participate in the integration of brain function in pain through central nervous system cross-modal integration [31]. Further analysis in this study found that after acupuncture treatment, the functional connectivity between the ACC and the visual cortex (such as the calcarine fissure, cuneus) significantly increased in responders, and this change was negatively correlated with the degree of pain relief. This phenomenon may reveal the mechanism of acupuncture analgesia: by enhancing the functional coupling between the ACC and visual-related brain regions, thereby alleviating pain perception. These findings emphasize that regulating the functional activity of the ACC may provide new strategies for treating chronic neck pain and associated emotional disorders.

Additionally, the longitudinal analysis revealed that the changes in predictive features among acupuncture responders were more targeted and strongly associated with the reduction of symptoms in patients suffering from neck pain. This observation suggests that individuals who responded to acupuncture underwent favorable adaptive responses to the treatment, resulting in more pronounced alterations in specific brain areas before and after therapy. Based on these findings, we hypothesize that the observed changes in the rsFC pattern of the ACC may reflect the specific effects of acupuncture treatment and could play a critical role in predicting therapeutic outcomes and individual responses to acupuncture in patients with neck pain.

Default mode network connectivity patterns explain acupuncture response-related neuroplasticity differences

This study revealed that all patients with neck pain exhibited reduced resting-state functional connectivity (rsFC) in the default mode network (DMN), particularly centered on the posterior cingulate cortex (PCC), following acupuncture treatment. This suggests that acupuncture may alleviate pain perception by suppressing the hyperactivation of DMN during pain processing. These findings align with previous neuroimaging studies, which have identified DMN as a critical regulator in chronic musculoskeletal pain [32], and highlighted the modulation of DMN functional activity as a shared mechanism underlying acupuncture’s therapeutic effects [33].

Notably, by stratifying patients into responders and non-responders based on clinical outcomes, this study found that non-responders exhibited more extensive DMN reorganization compared to responders. This differential pattern of neuroplasticity may reflect distinct adaptive responses to acupuncture intervention: responders showed targeted connectivity changes associated with pain relief, whereas non-responders displayed broader, less specific neural adjustments that did not translate into significant clinical improvement. Specifically, responders showed significant changes in only a few predictive functional connectivity features after treatment, and these changes were correlated with pain scores. In contrast, while non-responders exhibited changes in more features, none of these changes had significant clinical relevance, suggesting that their widespread neural adjustments lack a clear association with treatment efficacy. This, to some extent, underscores the importance of using machine learning to distinguish treatment response patterns. By identifying specific DMN connectivity profiles, it may be possible to select patient subgroups more likely to benefit from acupuncture. However, the lack of a sham acupuncture control group in this study limits our ability to fully disentangle the contributions of specific versus non-specific factors to the observed DMN connectivity changes, which remains a key direction for future research.

Limitations

This study has several limitations that should be taken into account in future research. Firstly, as an exploratory investigation, while some notable differences were initially observed, these did not hold up after applying the FDR correction, indicating a potential for false positives. Further research is necessary to confirm these findings. Secondly, although the omission of Fisher Z transformation for functional connectivity data may be seen as deviating from conventional preprocessing procedures, the large sample size, the use of non-parametric statistical methods, and the distribution robustness of SVM in this study have effectively ensured the reliability and stability of the results. Thirdly, this study only included data from two timepoints (pre- and post-treatment), lacking multi-timepoint dynamic monitoring, which may limit insights into the long-term changes in acupuncture efficacy and the dynamic evolution of its neural mechanisms. Fourthly, although machine learning models based on rs-fMRI demonstrate potential in identifying acupuncture responders, their clinical translation still faces multiple challenges. Future research should focus on developing low-cost, highly accessible alternative biomarkers and conducting translational medical studies involving both clinicians and patients to evaluate the feasibility and acceptability of such predictive tools in real-world settings.

Conclusions

This study highlights the potential of an SVM model based on pre-treatment brain functional connectivity features in identifying neuroimaging biomarkers for acupuncture responders, laying the groundwork for the development of objective predictive tools. Furthermore, the resting-state functional connectivity of the ACC was identified as a critical predictor for accurate outcome forecasting. These insights enhance our understanding of acupuncture’s therapeutic mechanisms and set the stage for further research into predicting and personalizing acupuncture treatments for chronic neck pain.

Funding Statement

This work was financially supported by the National Natural Science Foundation of China (Grant No. 82074549), Shanxi Province Basic Research Program (Grant No. 202203021221195), and Scientific Research Fund for the Doctoral Scholars of Shanxi University of Chinese Medicine (Grant No. 2023BK17, 2023BKS12). Funders and sponsors have no role in the design of this study.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Ethical approval and consent to participate

This trial was approved by the Institutional Review Board of Shanxi Acupuncture Hospital (approved number: 2023-007) and registered at International Traditional Medicine Clinical Trial Registry (registration number: ITMCTR2023000001, protocol version number: V1.0). Only patients who had signed the informed consent form were included.

Data availability statement

The data that support the findings of this study are available from the corresponding author, Laixi Ji, upon reasonable request.