Review Article

Machine Learning in Neuroimaging: A New Approach to Understand Acupuncture for Neuroplasticity

Table 1

The detailed characteristics of the included studies.

Participants
Sample size, gender (M/F), age (Y)
InterventionModalityFeaturePurpose (C/R)MLFeature selectionValidationModel assessmentMVPA findingsUnivariate analysis resultsConclusion

López et al., 2013Migraine
Verum ACU: 18, 3/15, ; sham ACU: 18, 7/11,
One session of verum or sham ACU stimulationTask-SPECT (function)
Task: ACU when image data acquisition
Blood perfusionCLinear SVMFilter: discarding voxels with intensity values under 25% of the maximumLOOCVACC
SPE
SEN
The classifier performed better when the training data was extracted from the verum ACU group than from the sham ACU group.Verum ACU yielded greater changes in the perfusion patterns than sham ACU. Verum ACU produced a more significant decrease in blood perfusion.SVM can distinguish the SPECT images of pre- and post-ACU acquisitions. Changes in blood perfusion following verum ACU is greater than sham ACU.
Jung et al., 2019HS
14, 14/0,
ACU at L HT7 or L PC6 for 20 blocksTask-fMRI (function)
Task: block design (16 s rest+6 s ACU+4 s stimulation location+4 s intensity report) 20
BOLD signalCLinear SVMNo feature selection stepsLOOCVACCThe classifier got an accuracy of 58.6% for classifying HT7 and PC6 with the features extracted from SI, MI, paraCL, anterior and posterior insula, SMG, ACG, vmPFC, PPC, and IPL. Using signal of ROI as feature, the classifier got higher accuracy (MI, 65%; SMA, 64%; SMG 62%; SI, 62%; and dlPFC, 62%).No significant difference in BOLD signal alteration following HT7 and PC6 stimulationSpatial localization of pain perceptions to ACU needle can be predicted by the neural response patterns in the somatosensory areas and the frontoparietal areas.
Yu et al., 2019HS
TR: 30, 16/14, 23-27; LT: 30, 18/12, 23-28
One session of TR or LT manipulation at ST36Task-EEG (function)
Task: ACU manipulation (TR or LT) for 3 min
Graph theoryCDT
NB
SVM
KNN
LDA
BP
TSK
Selecting the features of interest6-fold
CV
ACC
AUC
The classifier got an accuracy of 92.14% and AUC of 0.9570 with all graph theory features as inputs. With the increase of filter number, the accuracy was gradually improved. The highest accuracy was 92.37% with 6 filters in the TSK model.PLV of TR was stronger than the baseline, while PLV of LT was weaker than the baseline. The value of all the six graph theory features of TR was significantly lower than that of LT.Different ACU manipulations have different effects on functional brain networks. Classification of different ACU manipulations based on EEG with network features is feasible.
Liu et al., 2018MWOA
Responder: 38, /, ; nonresponder: 56, /,
24 sessions of sham ACU at NAP in 8 weeksDTI (structure)TABACLinear SVMFilter+wrapper: traversing the values of the two-sample test from 0.01 to 1 with a 0.01 interval to find the best for classifierLOOCVACC
SPE
SEN
PPV
NPV
The single FA, MD, AD, and RD of the mPFC-amygdala fiber contributed to lackluster classification accuracy. The classifier got a higher accuracy with the combined features of FA, MD, and RD (in which ACC, SEN, SPE, PPV, and NPV were 84.0%, 90.2%, 76.7%, 82.1%, and 86.8%, respectively). The external capsule, ACG, and mPFC significantly contributed to the discrimination of responders and nonresponders.The increased FA, decreased MD, decreased AD, and decreased RD of the mPFC-amygdala fiber were detected in MWOA patients than HS.The variability of placebo treatment outcomes in migraineurs could be predicted from prior diffusion measures along the fiber pathways of the mPFC-amygdala.
Yang et al., 2020MWOA
Responder: 19, 6/13, ; nonresponder: 22, 2/20,
12 sessions of ACU at GV20, GV24, bil-GB13, bil-GB8, and bil-GB20 in 4 weeksT1 (structure)GMVCLinear SVMFilter+wrapper+embedded: traversing the values of the two-sample test from 0.0025 to 0.05 with a step of 0.0025 to select the best for classifier
LASSO
10-fold
CV
ACC
SPE
SEN
AUC
DSC
Using the clusters located at the frontal, temporal, parietal, precuneus, and cuneus gyri as features, the classifier got the SEN of 73%, SPE of 85%, ACC of 83%, and AUC of 0.7871.The baseline GMV in all predictive regions significantly differed between responders and nonresponders. Alterations of migraine days were correlated with the baseline GMV of L cuneus, R MiFG/IFG, L IPL, and SPL/IPL. The responders achieved an increase in GMV of the L cuneus after ACU.The pretreatment brain structure could be a novel predictor for ACU treatment of MWOA.
Tu et al., 2019cLBP
Real ACU: 24, 8/16, ; sham ACU: 26, 11/15,
6 sessions of ACU in 4 weeks, 8-12 effective acupoints were used in the real ACU group; 12 sham points were used in the sham ACU group.Resting-fMRI (function)ICA+rsFCRRBF SVRSelecting the features of interest5-fold
CV

MAE
The prediction model obtained an of and an MAE of between actual and predicted treatment responses for real ACU. mPFC FC (mPFC-insula, mPFC-putamen, mPFC-caudate, and mPFC-AG) and other FC (PCC-MiFG, insula-IFG, insula-SPL, and caudate-AG) significantly contributed to prediction. The prediction model got an of and an MAE of for sham ACU. Connections of mPFC-dACG, mPFC-SPL, mPFC-paraCL, SFG-PreCG, SFG-MiFG, and ACG-paraCL provided significant information for prediction.Changes of pain severity correlated with baseline mPFC-SN and mPFC-AG FC in the real ACU group. Baseline mPFC-dACG FC was correlated with changes in pain severity in the sham ACU group. Changes of FC between the mPFC and insula/AG were correlated with the relief of pain severity after real treatment, while changes of FC between the mPFC and paraCL/SPL were correlated with the relief of pain severity after sham ACU treatment.Pretreatment rsFC could predict symptom changes for real and sham treatment, and the rsFC characteristics that were significantly predictive for real and sham treatment differed.
Xue et al., 2011HS
12, 9/3, 21-26
ACU at GB40 or KI3 for 3 blocks, switching after a one-week intervalTask-fMRI (function)
Task: on/off block design (1 min rest+1 min ACU) 3
BOLD signalCLinear SVMSingular value decomposition/SDMThe performance of the classifier was not mentioned in this study. ACU stimulation at GB40 produced predominantly signal increases in the insula, red nucleus, thalamus, and amygdala. ACU at KI3 elicited more extensive decreased neural responses in the MFG, PCC, thalamus, and ACG.ACU at GB40 and KI3 can both evoke similar widespread signal decreases in the limbic and subcortical structures.Neural response patterns between ACU stimulation at GB40 and KI3 are distinct. Conventional GLM analysis is insensitive to detect neural activities evoked by ACU stimulation.
Yin et al., 2020FD
Responders: 21, 2/19, ; nonresponder: 12, 2/10,
20 sessions of ACU in 4 weeks. One or two acupoints among CV12, ST36, and BL21 were used.Resting-fMRI (function)rsFCCLinear SVMWrapper: recursive feature eliminationLOOCVACC
SPE
SEN
AUC
The classifier obtained an ACC of 84.9%, SEN of 78.6%, SPE of 89.5%, and AUC of 86.8%. The FC between R insula-L precuneus, L MiOFG-L thalamus, L insula-L ACG, R ACG-R temporal pole, R SOG-R cerebellum-3 contributed crucial information for prediction./The whole-brain resting-state functional brain network has good predicting potential for ACU treatment to FD patients.
Hao et al., 2008HS
60, /, 24-75
One session of electro-ACU at ST36Task-EEG/ECG (function)
Task: ACU when image data acquisition
BIS
TPI
LF/HF
HR
HRV
RFNNSelecting the features of interestValidation with an independent setAAEWith the FNN, the AAE of the estimation and true value is 10.2278./The alteration of β-endorphin following electro-ACU can be predicted by monitoring EEG and ECG signal parameters.
Li et al., 2010HS
11/11, ; GB37: 11, /, /; NAP: 11, /, /
ACU at GB37 or NAP for 2 blocksTask-fMRI (function)
Task: on/off block design (1 min rest+two 30 s ACU separated by a 50 s rest period+50 s rest)
BOLD signalCLinear SVMSearchlight+singular value decompositionLOOCVACCThe occipital cortex, limbic-cerebellar areas, and somatosensory cortex could help to differentiate the central neural response patterns induced by real or sham ACU stimulation with higher accuracy above the chance level.Compared with the sham group, the ACU group induced higher signal intensity at some major regions of limbic-cerebellar system and small regions of the primary somatosensory cortex and supplementary motor area.Neural response patterns of brain cortex to the ACU stimulation at GB37 and a nearby NAP could differ from each other effectively with the application of the MVPA approach.

M/F: male/female; Y: year; C/R: classification/regression; ML: machine learning; MVPA: multivariate pattern analysis; ACU: acupuncture; SPECT: single-photon emission computed tomography; SVM: support vector machine; LOOCV: leave-one-out-cross-validation; ACC: accuracy; SPE: specificity; SEN: sensitivity; HS: healthy subjects; fMRI: functional magnetic resonance imaging; BOLD: blood oxygenation level dependent; TR: twirling-rotating manipulation; LT: lifting-thrusting manipulation; EEG: electroencephalogram; DT: decision tree; NB: naïve Bayes; KNN: -nearest neighbor; LDA: linear discriminant analysis; BP: BP neural network; TSK: Takagi-Sugeno-Kang fuzzy system; CV: cross-validation; AUC: area under the ROC curve; PLV: phase locking value; MWOA: migraine without aura; NAP: nonacupoint; DTI: diffusion tensor image; TABA: tractography atlas-based analysis; PPV: positive predictive value; NPV: negative predictive value; FA: fractional anisotropy; MD: mean diffusion; AD: axial diffusivity; RD: radial diffusivity; GMV: gray matter volume; LASSO: least absolute shrinkage and selection operator; DSC: dice similarity coefficient; cLBP: chronic low back pain; rsFC: resting-state functional connectivity; RBF: radical basis function; SVR: support vector regression; MAE: mean absolute error; SDM: spatial discriminance map; GLM: general linear model; FD: functional dyspepsia; BIS: bispectral index; TPI: tip perfusion index: LF/HF: low/high-frequency ratio; HR: heart rate; HRV: heart rate variability; FNN: fuzzy neural network; AAE: absolute average error; L: left; R: right; SI: primary somatosensory cortex; MI: primary motor cortex; paraCL: paracentral lobe; SMG: supramarginal gyrus; ACG: anterior cingulate gyrus; vmPFC: ventromedial prefrontal cortex; PPC: posterior parietal cortex; IPL: inferior parietal lobe; dlPFC: dorsolateral prefrontal cortex; mPFC: medial prefrontal cortex; MiFG: middle frontal gyrus; IFG: inferior frontal gyrus; SPL: superior parietal lobe; AG: angular gyrus; dACG: dorsal ACG; PreCG: precentral gyrus; SFG: superior frontal gyrus; SN: subcortical network; MFG: medial frontal gyrus; ITG: inferior temporal gyrus; MiOFG: middle orbitofrontal gyrus; SOG: superior occipital gyrus.