Effective Prediction of Neurofeedback based on Functional

Connection Characteristics of Brain Network in Insomnia

Kai Li

, Zhi Zou

, Huan Zhang

, Linyuan Wang

, Ying Zeng

, Fei Qi

and Chi Zhang

1,*

Henan Key Laboratory of Imaging and Intelligent Processing, PLA Strategic Support Force Information Engineering

University, Zhengzhou, Henan, China

Henan Provincial People’s Hospital, Zhengzhou, Henan, China

Center for Healthcare Data Science, Zhejiang Lab Zhongtai Street, Yuhang District, Hangzhou, 311100, Zhejiang, China

Keywords: Insomnia, Neurofeedback, Brain Network Connection.

Abstract: Real-time functional magnetic resonance imaging neurofeedback (rt-fMRI-nf) is a new means of emotion

regulation in insomnia, however, due to personal physiological and psychological differences, the effect of

neurofeedback training on different patients is significantly different. Using brain imaging data to predict the

curative effect is of great significance to improve the individual adaptability of clinical application of

neurofeedback training, reduce the treatment cost and reduce the burden of patients. In this article, we raise a

neurofeedback training effectiveness prediction method based on brain network functional connection. In this

method, network connection matrices of the default mode network (DMN), salience network (SAN),

executive control network (ECN), basal ganglia (BG), sensorimotor (SM) et.al. in insomnia are used as

features to construct the prediction model, so as to predict the training effect of patients' neurofeedback by

using machine learning method. The experimental results through the cross validation of CatBoost model

with leave one method show that, the prediction accuracy of whether an insomnia patient can benefit from

emotion regulation method produced by rt-fMRI-nf is 75%. This method can initially provide a reference

basis for insomnia patients to choose treatment methods.

1 INTRODUCTION

The drug treatment effect of patients with mental

diseases such as insomnia, depressive disorder and

anxiety shows that more than 50% of patients have

not been relieved after treatment (Li 2007). In recent

years, rt-fMRI-nf has become a conventional method

for the treatment of mental diseases. Due to the large

differences in individual treatment success rates,

some participants did not learn to control their brain

responses, resulting in "ineffective neurofeedback".

Ineffective regulation reduces the overall efficiency

of neurofeedback training and hinders its

transformation into clinical interventions (Haugg

2020). In neurofeedback studies, these participants

are often referred to as "non-responders", accounting

for 30% to 50% of the total population in study

(Alkoby 2017). For example, in 2021, Direito

(

Ramos 2019

) et al. explored whether a personalized

fMRI neurofeedback framework will have a positive

impact on the success of neurofeedback, the

adjustment threshold is defined for every subject

according to the maximum change of blood oxygen

level dependency (BOLD) area during positioning

operation in visual motion, it is found that 40% of the

subjects can successfully adjust the activation of

visual motor area, while 60% of the subjects do not.

Therefore, exploring the prediction model of the

effectiveness of neurofeedback therapy is of great

significance to guide individual clinical treatment

decisions.

In recent years, brain imaging indexes, combined

with traditional machine learning approach such as

support vector machine and random forest, have been

widely used to predict the prognosis of diseases. In

2017, Kesler et al. (Kesler 2017) used rt-fMRI data

and random forests to predict the long-term cognitive

ability of breast cancer patients after drug therapy. It

was found that the prediction accuracy of the training

model could reach 100% in the network and attention

network. In 2019, Zhutovsky et al. (Zhutovsky 2019)

used the functional connectivity of resting state brain

network components and Gaussian process classifier

to predict the treatment response of psychotherapy to

patients with post-traumatic stress disorder. It was

found that the network centered on auxiliary motor

Li, K., Zou, Z., Zhang, H., Wang, L., Zeng, Y., Qi, F. and Zhang, C.

Effective Prediction of Neurofeedback based on Functional Connection Characteristics of Brain Network in Insomnia.

DOI: 10.5220/0011235400003438

In Proceedings of the 1st International Conference on Health Big Data and Intelligent Healthcare (ICHIH 2022), pages 151-156

ISBN: 978-989-758-596-8

151

area before treatment contributed the most to the

classification of non-responders and responders.

Using the characteristics of resting state brain activity

to establish machine learning prediction model plays

an important role in improving the efficiency of

treatment and reducing the cost of treatment. It can be

seen that the resting state brain activity characteristics

of functional magnetic resonance imaging can

effectively predict disease treatment.

Recent studies have shown that the DMN, ECN,

SAN and other connection modes of resting state are

abnormal in insomnia, depressive disorder and other

diseases (Ma 2018, Albert 2018). These networks are

closely related to emotion processing, executive

function and attention, and the regulation between

brain networks affects the behavior of patients. For

insomnia patients, not only the weak connection

between the SAN responsible for discovering the

surrounding sensory and emotional stimuli and the

DMN related to self-reference and meditation may

lead to the over processing of negative information,

but also the enhanced connection between the ECN

responsible for attention control and advanced

cognitive control tasks and the DMN may be related

to the large cognitive load (Zhang 2021). Research

has pointed previously that the internal brain

functional network connection based on fMRI can be

used to predict the performance of remission of major

depressive disorder after drug treatment (Korgaonkar

2020). Researchers used a fully connected group

method to explore the internal brain function network

that can predict the treatment outcomes of

antidepressants in patients with depression before the

treatment, which found that patients with high

connectivity among DMN, frontal parietal and motor

network were most likely to benefit from

antidepressant treatment. Chin et al. (Fatt 2019) found

that depressive disorder patients with high

connectivity between DMN and ECN support the

treatment of antidepressants. Therefore, the brain

network connection model before neurofeedback

training is expected to become a potential feature to

predict the training effect, and plays an important role

in predicting which patients may be effective for

neurofeedback training in insomnia.

Aiming at the rt-fMRI neurofeedback emotion

regulation training in insomnia, this paper constructs

a resting state brain network functional connectivity

feature set based on reflecting emotion correlation,

and proposes a neurofeedback training effectiveness

prediction method based on brain network

connection. The results show that this method can

accurately predict the training effect of insomnia

patients, and provide an important basis for patients'

treatment decision-making.

2 MATERIALS AND METHODS

2.1 Participants

The subjects of this experiment were recruited by

Henan Provincial People's hospital through outpatient

and advertising. A total of 24 patients with right-

handed insomnia (average age 47.13 ± 12.76 years, 5

males and 19 females) were enrolled, which met the

criteria of DSM (American 1994). The degree of

insomnia of all patients was consistent with the

Pittsburgh sleep quality index (PSQI) total score of

more than 10 or the insomnia severity index (ISI)

score of more than 8 (PSQI > 10 and ISI > 8 showed

insomnia disorder. The higher the score, the more

serious insomnia). They subjectively showed

symptoms such as difficulty in getting in and out of

sleep, dreaminess and easy to wake up, depression

and emotional abnormalities. The exclusion criteria

included serious suicidal ideation, history of mental

or cardiovascular diseases, drinking or taking drugs

affecting brain function during the experiment, and

MRI contraindications.

2.2 Experimental Paradigm

The neurofeedback training of insomnia disorder

requires subjects to complete six stages of

experiments, once a week, and fixed on weekends. At

visit1, insomnia patients asked to fill in a

demographic scale and took an overnight

Polysomnography test (PSG) measurement. At visit2,

subjects underwent baseline scans, including T1

structural images and a resting state scan of 6min20s.

During vivst2 to vivst6, subjects were asked to fill in

six scales to evaluate sleep, depression and emotional

status before and after the experiment, including

Pittsburgh sleep quality index (PSQI), insomnia

severity index (ISI), Hamilton Depression Scale

(HAMD), Hamilton Anxiety Scale (HAMA), positive

and Negative Emotion Scale (PANAS) and Baker

Depression Scale (BDI). From vivst3 to vivst5, each

subject underwent three neurofeedback training runs,

each session lasted about 50 minutes, including the

formal experimental run and the rest time between

each run.

Before neurofeedback training, the subjects were

asked to write down three or more specific positive

autobiographical memories, and explain the specific

tasks of the experiment to the subjects. Each rt-fMRI

ICHIH 2022 - International Conference on Health Big Data and Intelligent Healthcare

152

neurofeedback experiment session included

functional localization before training and 7 min

resting state scanning. Before the formal

neurofeedback training, there will be a pre-training

lasting for 6 min and 30s and subjects will train

without feedback signal. Then, three formal

neurofeedback trainings composed of seven 30s

"rest" blocks and six 30s "emotion regulation" blocks

guided by the prompt "happy" were conducted, each

lasting for 6 min and 30s. In the "rest" block, the

patient is required to look at the cross on the screen to

calm their emotions. In the "emotion regulation"

block, the activity signal of the left amygdala of the

insomnia patient is fed back to the subjects in the

shape of a bar column, and the patient is instructed to

adjust the height of the bar column on the screen by

specifically recalling positive autobiographical

memory. Each repetition time of the feedback signal

(TR = 2S) update once. Then a transfer training run

without feedback signal but the same as the formal

feedback run experimental paradigm was carried out,

which also lasted for 6 min and 30s. Finally, a 7 min

resting state scan was performed. Visit 6 is the

follow-up period, during which the T1 structure

image and resting state of the subjects were scanned

again. The neurofeedback process is shown in Figure

Figure1 ： The process of neurofeedback training in

insomnia

2.3 Experimental Acquisition

All fMRI data in this experiment are from the medical

imaging center of Henan people's hospital. The data

collection is completed through their magnetic

machine (Siemens prism 3T). The head coil is 64

channels. Before the experiment, we first fixed the

subject's head with a sponge pad, and then pasted

medical tape laterally on the subject's forehead to

both sides of the coil for fixation to prevent excessive

head movement from affecting the experimental

results. This paper uses the resting state data of

insomnia patients before and after neurofeedback.

2.4 Data Processing

The DPABI is used to process resting state fMRI data,

including converting DICOM raw data to NII data

(deleting the first 10 time points), slice timing,

realignment, reorientation, coregister, brain

component segment, smooth, detrending, filter, etc.

This paper performed group ICA analysis by gift

v3.0b toolbox (Calhoun2010) on the preprocessed

data to gain the resting state brain network

components of the whole brain. The minimum

description length criterion (MDL) (Li 2007) was

used to calculate the optimal number of resting state

components in the whole brain, and the INFOMAX

algorithm was used to decompose the fMRI data. The

reliability and robustness of component analysis are

improved through repeated calculation for 20 times

by ICASSO (Himberg 2003). Then, the simplified

data of principal component analysis (PCA) is

inversely reconstructed and decomposed into a series

of spatially independent components and their

corresponding time processes. Normalization

converts each independent component to a z-value.

The brain network of interest is selected by the

method of maximum spatial correlation, and the brain

network template is used as the spatial template of the

component. We single out 10 brain networks

involved in emotional cognitive processing and

sensory motor in the resting brain network template

proposed by Stanford cognitive and System

Neuroscience Laboratory in 2012 (Shirer 2012),

anterior salience network (ASN), posterior salience

network (PSN), dorsal default mode network

(DDMN), ventral default mode network (VDMN),

left executive control network (LECN) and right

executive control network (RECN), basal

ganglia(BG), sensorimotor(SM), precuneus and

auditory. The Pearson correlation between each

network time series is calculated, and the correlation

coefficient is transformed by fisher-z transform.

Finally, the functional connection matrix of the brain

network before neurofeedback is used as the

characteristic input of the prediction model.

2.5 Machine Learning Model for

Treatment Effect Prediction

In recent years, people are increasingly interested in

the application of machine learning technology in the

diagnosis, classification and effect prediction of

clinical diseases such as depression, schizophrenia,

bipolar disorder and autism disorder. We use machine

learning method combined with brain image

characteristics to predict the effect of neurofeedback

Effective Prediction of Neurofeedback based on Functional Connection Characteristics of Brain Network in Insomnia

153

training for insomnia patients, further promote the

decision-making of individual treatment plan for

insomnia, and accelerate the clinical transformation

and application of rt-fMRI neurofeedback

technology.

In this paper, 10 functional brain networks and 68

network sub-components of 24 insomnia patients

were analyzed in time, and twenty-four 78× 78

functional connection matrix were generated within

and between brain networks. We select 3003 non-

repeated function connection values in the generated

function connection matrix as features. The machine

learning method is adopted, and CatBoost,

RandomForest, LightGBM, XGboost, ExtraTrees,

KNeighbors and other models are used for secondary

classification. Label 0 represents the ineffective

neurofeedback treatment, and label 1 represents the

effective neurofeedback treatment.

3 RESULTS

Referring to the common segmentation proportion of

machine learning, 24 patients are divided into 14

patients for training and 10 patients for testing. Using

automatic machine learning technology, the

recognition accuracy of 11 machine learning models

under this data set is investigated.

For the generated 10 test sets, the prediction of

each model is shown in Table 1.

Table 1: Prediction of 10 subjects under each model.

model Light-GBMLarge

Light-

GBMXT

XG-

Boost

Light-

GBM

Cat-

Boost

accurac

90% 70% 50% 70% 60%

Random-

ForestEnt

KNeigh-

orsUnif

KNeigh-

orsDist

Random-

ForestGini

Extra-

TreesEnt

Extra-

TreesGini

40% 30% 30% 30% 20% 20%

Considering that the sample size of this

experiment is relatively small, the left one method

(LOOCV) is used for cross validation, that is, the

functional connection matrix of each subject is

recycled as the test set, and the brain network

characteristics of the other 23 subjects are input into

the automatic machine learning model as the training

set for training, with a total of 24 cyclic evaluations.

The prediction results of each model are shown in

Table 2.

Table 2: The prediction of each model of 24 subjects under LOOCV.

model Cat-Boost

Random-

ForestEnt

Random-

ForestGini

Light-

GBMXT

Light-

GBM

accuracy 75% 71% 71% 63% 63%

Extra-

TreesEnt

Light-

GBMLarge

XG-Boost

KNeigh-

orsDist

KNeigh-

orsUnif

Extra-

TreesGini

63% 58% 54% 54% 54% 54%

From table2, CatBoost, RandomForestEntr and

RandomForestGini models have higher prediction

accuracy among all models.

Figure 2: ROC curve of CatBoost under LOOCV.

ICHIH 2022 - International Conference on Health Big Data and Intelligent Healthcare

154

As shown in Figure 1, the area is 0.84 under the

ROC curve of CatBoost, the accuracy of this model is

high. Compared with other models, CatBoost model

solves the problems of gradient deviation and

prediction offset, so the occurrence of over fitting is

reduced and the accuracy and generalization ability of

the algorithm are improved. Randomforestentr and

RandomForestGini belong to random forest models,

which can be over fitted by reducing randomness, so

their prediction accuracy is high. However, due to the

high randomness of their parameters, their accuracy

is lower than CatBoost. Xgboost model uses

approximate algorithm to improve operation speed.

Kneigborsdist and Kneigborsunif are too simple in

structure, resulting in insufficient fitting of data.

Extratreesgini model adopts randomization for model

structure and model parameters, resulting in too

strong randomization of model. Therefore, the

prediction accuracy of Xgboost, Kneigborsdist,

Kneigborsunif and Extratreesgini are lower.

4 DISCUSSION

The results of machine learning prediction model

based on the functional connection characteristics of

brain network show that the effectiveness of

neurofeedback in insomnia patients is significantly

higher than the random probability. The clinical

practical significance of the model can be explained

as whether an individual with insomnia can be

predicted to be suitable for neural feedback training

after resting state MRI scanning and brain network

data analysis, which provided a reliable basis for the

selection of treatment schemes for insomnia patients.

However, our study has some limitations. On the

one hand, the degree of each subject is different in

insomnia, and there are individual differences in the

treatment effect after neurofeedback training. For the

ineffective treatment of patients after neurofeedback,

whether it is caused by internal factors or external

factors such as errors in the experimental process,

these problems need to be further discussed. For

example, we can divide the subjects' disease degree

more carefully, explore the impact of different

severity of insomnia on the effect of neurofeedback

training, or adjust the neurofeedback experimental

design by changing the neurofeedback target area. On

the other hand, due to the limited experimental

conditions and the limited number of samples

included in the experiment, 24 subjects participated

in neurofeedback training. With the development of

more clinical experiments, this method is expected to

achieve more accurate prediction.

5 CONCLUSIONS

In the paper, we propose a neurofeedback

effectiveness prediction method based on the

functional connectivity of resting state brain

networks. The functional connections of brain

networks such as DMN, ECN, SAN and SM et.al.

before neural feedback in insomnia are extracted as

features, and a prediction model based on automatic

machine learning is constructed to predict the neural

feedback training effect of insomnia patients, so as to

realize the prediction of the effectiveness of neural

feedback training of insomnia patients based on

machine learning. The experimental results show that

the highest prediction accuracy of all machine

learning models reaches 75%, which provides an

important support for insomnia patients to make

decisions in the treatment plan.

ACKNOWLEDGEMENTS

This work was supported by the National Natural

Science Foundation of China [grant number

82071884]; and the Key Project of Medical Science

and Technology of Henan Province [grant number

LHGJ20200060].

REFERENCES

Alkoby, O., Abu-Rmileh, A., Shriki, O., & Todder, D..

(2017). Can we predict who will respond to

neurofeedback? a review of the inefficacy problem and

existing predictors for successful eeg neurofeedback

learning. Neuroscience, S0306452216307576.

Albert, K. M. , GG Potter, Boyd, B. D. , Kang, H. , &

Taylor, W. D. . (2018). Brain network functional

connectivity and cognitive performance in major

depressive disorder. Journal of Psychiatric Research.

American, P. . (1994). Diagnostic and statistical manual of

mental disorders. Encyclopedia of the Neurological

Sciences, 25(2), 4-8.

Calhoun, V. D. , Adali, T. , Pearlson, G. D. , & Pekar, J. J.

. (2010). A method for making group inferences from

functional mri data using independent component

analysis. Human Brain Mapping, 14(2), 140.

Fatt, C. R. C. , Jha, M. K. , Cooper, C. M. , Fonzo, G. , &

Trivedi, M. H. . (2019). Effect of intrinsic patterns of

functional brain connectivity in moderating

antidepressant treatment response in major depression.

American Journal of Psychiatry, 177(2),

appi.ajp.2019.1.

Haugg, A., Sladky, R., Skouras, S., Mcdonald, A., &

Scharnowski, F.. (2020). Can we predict real‐time fmri

Effective Prediction of Neurofeedback based on Functional Connection Characteristics of Brain Network in Insomnia

155

neurofeedback learning success from pretraining brain

activity?. Human Brain Mapping(28).

Himberg, J. , & Hyvarinen, A. . (2003). Icasso: software for

investigating the reliability of ICA estimates by

clustering and visualization. Neural Networks for

Signal Processing, 2003. NNSP'03. 2003 IEEE 13th

Workshop on. IEEE.

Kesler, S. R. , Rao, A. , Blayney, D. W. , Oakley-Girvan, I.

A. , Meghan, K. , & Oxana, P. . (2017). Predicting long-

term cognitive outcome following breast cancer with

pre-treatment resting state fmri and random forest

machine learning. Frontiers in Human Neuroscience,

11, 555.

Korgaonkar, M. S. , Goldstein-Piekarski, A. N. , Fornito, A.

, & Williams, L. M. . (2020). Intrinsic connectomes are

a predictive biomarker of remission in major depressive

disorder. NPG Open Access , 25(7).

Li, Y. O., Tülay Adal, & Calhoun, V. D. (2007). Estimating

the number of independent components for functional

magnetic resonance imaging data. Human Brain

Mapping, 28(11), 1251-1266.

Ma, X. , Jiang, G. , Fu, S. , Jin, F. , Wu, Y. , & Liu, M. , et

al. (2018). Enhanced network efficiency of functional

brain networks in primary insomnia patients. Frontiers

in Psychiatry, 9, 46.

Ramos, M. , Pereira, J. , Direito, B. , Sousa, T. , & Sayal,

A. . (2019). Directly exploring the neural correlates of

feedback-related reward saliency during fMRI-based

Neurofeedback.

W., R., Shirer, S., Ryali, & E., et al. (2012). Decoding

subject-driven cognitive states with whole-brain

connectivity patterns. Cerebral Cortex.

Zhutovsky, P. , Thomas, R. M. , Olff, M. , Rooij, S. ,

Kennis, M. , & Wingen, G. , et al. (2019). Individual

prediction of psychotherapy outcome in posttraumatic

stress disorder using neuroimaging data. Translational

Psychiatry, 9.

Zhang, Gao, et al. (2021). A research on resting-state

functional network connectivity after rt-fmri

neurofeedback in insomnia. Journal of Physics:

Conference Series, 1907(1), 012013 (7pp).

ICHIH 2022 - International Conference on Health Big Data and Intelligent Healthcare

156