Abstract
With approximately 75% of smokers resuming cigarette smoking after using the Gold Standard Programme for smoking cessation, investigation into novel therapeutic approaches is warranted. Typically, smoking cue reactivity is crucial for smoking behaviour. Here we developed a novel closed-loop, smoking cue reactivity patterns EEG-based neurofeedback protocol and evaluated its therapeutic efficacy on nicotine addiction. During an evoked smoking cue reactivity task participants’ brain activity patterns corresponding to smoking cues were obtained with multivariate pattern analysis of all EEG channels data, then during neurofeedback the EEG activity patterns of smoking cue reactivity were continuously deactivated with adaptive closed-loop training. In a double-blind, placebo-controlled, randomized clinical trial, 60 nicotine-dependent participants were assigned to receive two neurofeedback training sessions (∼1 h/session) either from their own brain (n = 30, real-feedback group) or from the brain activity pattern of a matched participant (n = 30, yoked-feedback group). Cigarette craving and craving-related P300 were assessed at pre-neurofeedback and post-neurofeedback. The number of cigarettes smoked per day was assessed at baseline, 1 week, 1 month, and 4 months following the final neurofeedback visit. In the real-feedback group, participants successfully deactivated EEG activity patterns of smoking cue reactivity. The real-feedback group showed significant decrease in cigarette craving and craving-related P300 amplitudes compared with the yoked-feedback group. The rates of cigarettes smoked per day at 1 week, 1 month and 4 months follow-up decreased 30.6%, 38.2%, and 27.4% relative to baseline in the real-feedback group, compared to decreases of 14.0%, 13.7%, and 5.9% in the yoked-feedback group. The neurofeedback effects on craving change and smoking amount at the 4-month follow-up were further predicted by neural markers at pre-neurofeedback. This novel neurofeedback training approach produced significant short-term and long-term effects on cigarette craving and smoking behaviour, suggesting the neurofeedback protocol described herein is a promising brain-based tool for treating addiction.
Introduction
Nicotine addiction is the leading preventable cause of disease and death worldwide. With ∼75% of patients with nicotine dependence not responding fully to the Gold Standard Programme (a comprehensive intervention consisting of manual-based teaching sessions together with nicotine replacement therapy) for smoking cessation interventions (Rasmussen et al., 2017), high relapse rates during long-term follow-up periods remain a core feature of nicotine addiction. Therefore, there is an urgent need to develop novel therapeutic approaches for nicotine addiction.
Neurofeedback, a psychophysiological procedure that helps participants to self-regulate their brain activity, has been of growing interest among basic and clinical neuroscientists (Sitaram et al., 2017). Clinically, neurofeedback has been used in many psychiatric disorders, including attention deficit hyperactivity disorder (Arns et al., 2009), depression (Young et al., 2017b), anxiety (Scheinost et al., 2013) and drug addiction (Sokhadze et al., 2008). Further, recent functional MRI-based neurofeedback studies indicate preliminary efficacy in reducing cigarette craving in smokers (Li et al., 2013; Hartwell et al., 2016). However, the feasibility of turning functional MRI neurofeedback into a widely available clinical intervention is questionable. In contrast, EEG is a relatively inexpensive and portable brain imaging technique that can be easily implemented at any location and has more potential for widespread clinical implementation than functional MRI neurofeedback. Previous EEG-based neurofeedback protocols, including alpha training, alpha/theta training, and SMR (sensorimotor rhythm)/beta training, have been used in drug addiction treatment for more than four decades (Sokhadze et al., 2008). Using these training protocols, drug-dependent patients received the power of a single and fixed EEG frequency and self-regulated that signal (Schmidt et al., 2017). Most of these studies focused on facilitating relaxation and reducing anxiety. However, the efficacy on drug addiction has only been classified as ‘probably efficacious’ according to the report of the Association for Applied Psychophysiology and Biofeedback and the International Society for Neurofeedback and Research (Sokhadze et al., 2008; Schmidt et al., 2017). Additionally, recent studies question the clinical efficacy of previous EEG neurofeedback protocols (Thibault and Raz, 2016; Fovet et al., 2017; Schabus et al., 2017). Hence, a new direction for EEG neurofeedback in treating drug addiction is warranted.
The efficacy of these traditional EEG neurofeedback approaches in the treatment of addiction remains dubious, in part, because addiction process involves many complex cognitive models [e.g. cue reactivity model (Chiamulera, 2005), negative reinforcement model (Koob, 2013)] but previous neurofeedback studies mainly targeted arousal and/or anxiety. Instead, drug cue reactivity can evoke the impulse for drug-seeking behaviour in addiction (Weiss et al., 2001). Previous studies from our group and others indicate that smoking cue reactivity is a central characteristic of nicotine addiction (Zhang et al., 2009; Engelmann et al., 2012) and can predict relapse vulnerability (Janes et al., 2010); thus, there is evidential support for a causal relationship between cue reactivity and relapse (Parvaz et al., 2011). Therefore, reducing brain reactivity to smoking cues has the potential to improve smoking cessation outcomes.
Recent EEG studies have reported that smoking cue reactivity is a complex brain activity pattern that involves multiple EEG features, including both time domain (e.g. P300, slow positive wave) and frequency domain (e.g. alpha oscillation) features (Littel and Franken, 2007; Littel et al., 2012; Cui et al., 2013). Typically, multivariate pattern analysis (MVPA) can enhance sensitivity of detecting a particular brain activity pattern by using multi-feature combinations for input to multivariate patterns (Haynes and Rees, 2006). A number of neurofeedback studies combined with MVPA have been impressively successful at improving attention and perceptual learning after only a few sessions (Shibata et al., 2011; deBettencourt et al., 2015), whereas some traditional EEG-based neurofeedback studies require dozens of sessions for any effects to be detected (Sokhadze et al., 2008).
In the current study, we evaluated a novel EEG neurofeedback paradigm, in which brain activity patterns corresponding to smoking cue reactivity were repeatedly deactivated. In this double-blind, randomized, placebo-controlled study, we first trained a personalized MVPA classifier to identify an EEG activity pattern associated with smoking cue reactivity, and subsequently trained participants to deactivate this pattern. To improve participants’ vigilance during neurofeedback, they were trained using an adaptive closed-loop approach, in which the neurofeedback stimulus and decoded brain state can influence each other in real-time (deBettencourt et al., 2015), e.g. when participants were more successful in deactivating the patterns for smoking cue reactivity, a less provocative image was shown, providing positive reinforcement of success. Second, cigarette craving is regarded as a core symptom of nicotine addiction. The P300 component (∼300–550 ms) of the event-related potential (ERP) evoked by substance-related cues is characteristic of substance use disorders in general and P300 amplitudes have been found to correlate with cigarette craving in the majority of studies (Knott et al., 2008; Field et al., 2009; Littel et al., 2012; Campanella et al., 2014). Therefore, the short-term effects on cigarette craving and craving-related P300 were compared pre-neurofeedback and post-neurofeedback across two sessions of neurofeedback training. Cigarette consumption is also a crucial problem in nicotine addiction. The long-term effects on daily cigarette consumption reflecting smoking behaviour were examined at three follow-up visits up to 4 months post-training. Third, neuroimaging biomarkers often provide better predictions on treatment response than behavioural scale measures (Gabrieli et al., 2015). Emerging evidence suggests that early neurophysiological characteristics could be particularly useful in predicting neurofeedback effects, including patients’ initial ability to regulate the neural activity (e.g. the average performance in first 30 trials) (Weber et al., 2011) and the ease of generating neural activity patterns (e.g. the classification accuracy of neural activity patterns) (Sadtler et al., 2014). Therefore, we additionally examine whether these short-term and long-term behavioural effects could be predicted by the classification accuracy of neural activity patterns at pre-neurofeedback and deactivated performance during the first cycle of neurofeedback, respectively.
Materials and methods
Participants
Sixty participants who met the following criteria participated in the experiment: smoking 10 or more cigarettes per day for 2 years or more; right-handedness; between 18 and 40 years of age; having normal or corrected to normal vision; and in good mental and physical health assessed by the Mini-International Neuropsychiatric Interview (Sheehan et al., 1998). The exclusion criteria included chronic neurological, psychiatric, or medical conditions; treatment with any drugs during the previous 3 months; and being generally ill-suited to perform EEG. Finally, only males were eligible to participate in this study because of very low prevalence (2.7%) of female Chinese smokers and the possible influence of the menstrual cycle phase on smoking cue reactivity and cigarette craving (Franklin et al., 2015). All participants were recruited through online advertisements and posters, and they did not engage with any other smoking cessation plan simultaneously.
The study protocol was approved by the Human Ethics Committee of the University of Science and Technology of China and registered on the Chinese Clinical Trial Registry (ChiCTR-IPR-17011710, http://apps.who.int/trialsearch/Trial2.aspx?TrialID=ChiCTR-IPR-17011710). All participants provided written informed consent prior to the study and were paid 600 RMB (∼USD $95) after completing four experimental visits. Participants were randomly assigned to the real-feedback group (n = 30) or the yoked-feedback group (n = 30). The real-feedback group regulated their own online brain patterns. The yoked-feedback group regulated the brain activity pattern of a matched participant in the real-feedback group (deBettencourt et al., 2015). For more detailed information, refer the the trial profile (Fig. 2) and the online Supplementary material.
Experimental procedure
The study was a double-blind, randomized, placebo-controlled design. The experimental procedure consisted of four stages (Fig. 1A): (i) baseline session (Visit 1); (ii) two neurofeedback training sessions (Visits 2 and 3); (iii) post-training behavioural session (Visit 4); and (iv) follow-up session (Visit 5). For more details, see the Supplementary material. Participants were required to be abstinent from smoking cigarette for 2 h prior to every visit, which ensured participants’ some degree of craving and responsiveness to the cues without the potential confound of a ceiling effect from prolonged abstinence. No participant reported uncomfortable feelings after neurofeedback training.
Experimental procedure and neurofeedback paradigm. (A) Experimental procedure. (B) Paradigm of neurofeedback training. The neurofeedback procedure consisted of two parts: offline classifier construction and real-time feedback training. (C) A screenshot showing the feedback within a neurofeedback training cycle. Each cycle consisted of 40 training trials. Each trial was updated every 2 s including 1-s acquisition and 1-s computing. Using the adaptive closed-loop method, the curve of probabilistic score at the bottom of neurofeedback display indicated the extent to which the current brain activity pattern matched the pattern for reactivity to the smoking cue. The picture at the top of neurofeedback display was converted by the probabilistic score: lower probabilistic score resulted in a picture with a lower craving score. The red region at the bottom of neurofeedback display indicated the probabilistic score ranging from 0.5 to 1 and the blue region indicated the score ranging from 0 to 0.5. NF = neurofeedback; SVM = support vector machine; TCQ = Tobacco Craving Questionnaire.
Trial profile..
Primary outcome
Cigarette craving was measured by the Tobacco Craving Questionnaire (TCQ) at pre-neurofeedback (before the first neurofeedback training session) and post-neurofeedback (after the second neurofeedback training session). The P300 of the ERP evoked by smoking-related cues was assessed by the smoking cue reactivity task at pre-neurofeedback and post-neurofeedback. The number of cigarettes smoked per day was assessed at 1 week, 1 month, and 4 months after neurofeedback training by self-reporting. The number of cigarettes smoked per day refers to the average number of cigarettes smoked per day from the last visit/telephone interview to the current one.
Smoking cue reactivity task
The smoking cue reactivity task was adapted from a previous study (Due et al., 2002). Within this task, participants were instructed to view a picture displayed on the screen. To make them focus on every picture, participants pressed the button as soon as possible when an animal picture was shown. There were 330 pictures [150 smoking-related (e.g. a cigarette in the hand), 150 neutral (e.g. a pencil in the hand), and 30 animal-related (e.g. a kangaroo) cues] selected from our previous studies (Zhang et al., 2009). These pictures were divided into six blocks, including three smoking blocks and three neutral blocks. The block design helped improve the signal-to-noise ratio of EEG smoking cue reactivity and was consistent with the later neurofeedback training design. The order of six blocks was made random across participants. Within a block, each trial contained a picture presented for 1.5 s and a fixation (+) was presented for ∼1–1.5 s. Animal pictures were shown randomly during all blocks. After completing a block, participants had a 90-s rest. We found all participants could press the button correctly when the animal picture appeared (100% accuracy).
Neurofeedback paradigm
The neurofeedback training paradigm consisted of two parts (Fig. 1B). First, we trained a personalized classifier to distinguish the EEG activity patterns corresponding to smoking and neutral cue reactivity using the smoking cue reactivity task. Based on the acquisition EEG signals within the smoking cue reactivity task, raw signals were preprocessed offline in EEGLAB (https://sccn.ucsd.edu/eeglab). Preprocessing included the following steps: high-pass filter (0.5 Hz), epoch (∼0–1000 ms), and blink artefacts correction (using a conventional recursive least squares algorithm). Afterwards, time domain (EEG potential) and time-frequency domain (EEG power) features were calculated by contrasting smoking with neutral cues using a permutation test. Time-frequency domain features were calculated by wavelet analysis and divided into five common bands (Supplementary Fig. 2). Under the threshold (P < 0.05) of the permutation test, the EEG potential and EEG power features surviving this threshold were separately formed into temporal-spatial clusters by grouping them at adjacent time points and electrodes using a cluster-based statistic (Groppe et al., 2011) (Supplementary Fig. 1). Once the temporal-spatial clusters were identified, the EEG features for constructing the pattern classifier were extracted from these clusters (Supplementary Fig. 2). The mean values (potential and power) from each temporal-spatial cluster were calculated and combined into a linear support vector machine (SVM, penalty parameter C = 1) classifier using the MATLAB (https://www.mathworks.com) function fitcsvm. The SVM classifier was selected as it often outperforms other classifiers for neurofeedback (Lotte et al., 2018). To evaluate the predictive power of the personalized classifier, we assessed the classification accuracy for the classifier by 5-fold cross-validation. In each cross-validation cycle, 80% of the trials were used as the training samples to extract features and train the classifier, and the remaining 20% of the trials were used as the testing samples to compute the classification accuracy. This procedure was repeated five times for each participant. The classifier for subsequent neurofeedback was obtained using the entire sample data in the smoking cue reactivity task with a mean classification accuracy of 70% (range: 55–85%) (Supplementary Fig. 3).
Next, during neurofeedback training, participants were asked to repeatedly and continuously deactivate their real-time EEG activity patterns of smoking cue reactivity calculated using a previously constructed classifier. For each real-time raw EEG signal lasting 1 s, the preprocessing included high-pass filter and blink correction using the same algorithm as in the previous EEG preprocessing. The time domain and time-frequency domain features were then extracted from the previous temporal-spatial clusters by calculating the mean values (EEG potential and EEG power), and then input into the personalized classifier. The classifier estimated the probabilistic score in real time reflecting the extent to which the brain activity pattern matched the pattern for reactivity to the smoking cue. The probabilistic scores (from 0 to 1) for the classifier were conducted using the MATLAB function fitSVMPosterior. As a result, when a participant successfully deactivated the smoking cue pattern, the probabilistic score decreased. That is, when the current activity patterns were more similar to neutral cue activity patterns, the score decreased, and when the current activity patterns were more similar to the smoking cue activity patterns, the score increased.
Adaptive closed-loop design
To improve participants’ vigilance, and help them better self-monitor and evaluate their brain state during neurofeedback, we used an adaptive closed-loop method in which the neurofeedback stimulus and decoded brain state influenced each other in real-time (deBettencourt et al., 2015). Different craving level pictures evoked different degrees of smoking cue reactivity for smokers (Carter and Tiffany, 1999). In the current study, an approach of continually updating sensory stimuli (e.g. different craving level pictures) based on changing brain states (e.g. different degrees of smoking cue reactivity pattern) constituted a ‘closed-loop’ design. The logic of this closed-loop design is that, when a participant was unsuccessful in ‘deactivating’ the smoking cue pattern (i.e. the probabilistic score increased), a picture with a higher craving level was displayed to amplify and externalize the consequences of the participant’s smoking cue-related brain activity pattern (deBettencourt et al., 2015; deBettencourt and Norman, 2016). This made unsuccessful deactivation more salient and increased the self-monitoring demand of the task. In other words, we amplified the consequences of their cue pattern, rewarding successful deactivation by reducing difficulty with a lower craving level picture and punishing unsuccessful deactivation by increasing difficulty with a higher craving level picture. For more details, see the online Supplementary material.
The probabilistic score was presented at the bottom half of the screen and transferred into a picture presented on the top half of the screen at the same time (Fig. 1C). The association between the probabilistic score and the transferred picture was controlled by a linear positive correlation function. A lower probabilistic score resulted in a picture with lower craving level rated previously by a group of 20 independent smokers (Supplementary Fig. 4). To reduce fluctuations due to noise in the EEG signal, the probabilistic score value of each trial was calculated using a moving average of the current and two preceding values.
After neurofeedback practice, participants were required to identify 10 cognitive strategies that may be effective at deactivating the neurofeedback signal, but it was emphasized that they should adjust their strategies to find a method that works best for them during neurofeedback (Instruction: ‘Make the feedback curve move down and the picture induce less craving’). Each neurofeedback training session consisted of eight cycles, with 40 trials per cycle. Each trial was updated every 2 s including 1-s acquisition and 1-s computing. There was a 1-min rest between cycles. At the end of each cycle, the self-regulation performance during the previous cycle was presented. After each cycle, participants completed ratings on perceived control over the neurofeedback signal (Supplementary material). The final cumulative performance was translated into an additional money reward. Both groups received the same money after neurofeedback. After completing each training session, they were asked to practice the strategies that worked best for them during neurofeedback offline.
EEG acquisition
The Psychophysics Toolbox for MATLAB was used to run the experiment (http://psychtoolb ox.org/). EEG data were recorded using SynAmps RT amplifier (NeuroScan). Sixty-four Ag/AgCl electrodes were placed on the scalp at specific locations according to the extended international 10–20 system. In addition, the electrical activities were recorded over right and left mastoid. Vertical electrooculography (EOG) was recorded using bipolar channels placed above and below the left eye, and horizontal EOG was recorded using bipolar channels placed lateral to the outer canthi of both eyes. The reference electrode was attached to the tip of the nose and the ground electrode was attached to AFz. Impedance between the reference electrode and any recording electrode was kept under 5 kΩ. All signals were digitized at 500 Hz during data collection. The real-time EEG acquisition during neurofeedback was completed by NeuroScan Access SDK.
Offline data analysis
Offline ERP analysis in the smoking cue reactivity task was conducted using EEGLAB. Data preprocessing was identical to the steps in the neurofeedback paradigm. Each epoch was baseline corrected using the pre-stimulus interval from −200 ms to 0 ms. Epochs containing the ERP amplitude changes exceeding ±100 μv were rejected. The ERP were grand averaged based on each type of picture across participants. The statistical analysis was performed using MATLAB 2016a. The variance analysis on the measured variables was conducted by a two-way mixed-design repeated ANOVA. Student t -tests were performed to compare conditions, with the exception that the comparison of P300 amplitudes was performed with a permutation test. Correlation analysis between two variables was calculated using Pearson’s correlation. The mediation analysis was performed with bootstrapped methods. The effect size (d-value) was calculated online (https://www.psychometrica.de/effect_size.html). All reported P-values were two-tailed.
Data availability
We have developed a MATLAB toolbox (named BRAINART1.0, software copyright: 2016SR202558) to support this novel neurofeedback method. The toolbox and the data that support the findings of this study are available from the corresponding author upon reasonable request.
Results
Demographic and clinical characteristics
To compare the baseline demographic and clinical characteristics before neurofeedback training between the two groups, independent sample t-tests were conducted. No significant group difference was found (Table 1).
Baseline demographic and clinical characteristics in two groups
| Characteristic | Real-feedback group (n = 28) | Yoked-feedback group (n = 25) | P-value |
|---|---|---|---|
| Age, years | 23.7 (3.8) | 23.4 (3.1) | 0.74 |
| Education, years | 14.8 (2.5) | 14.6 (2.4) | 0.83 |
| Cigarettes/day | 14.1 (4.5) | 15.0 (7.5) | 0.42 |
| Years of cigarette use | 7.1 (3.9) | 6.6 (2.8) | 0.57 |
| Score | |||
| FTND | 4.6 (1.9) | 4.8 (1.4) | 0.57 |
| MQS | 5.2 (2.6) | 5.0 (2.6) | 0.85 |
| BIS | 60.3 (5.4) | 59.2 (4.4) | 0.46 |
| SPSRQ | |||
| Punishment | 8.4 (4.1) | 7.6 (4.3) | 0.54 |
| Reward | 7.5 (2.4) | 6.7 (2.9) | 0.32 |
| ERQ | 45.9 (7.8) | 44.8 (6.6) | 0.56 |
| PANAS | |||
| Negative | 15.8 (6.7) | 17.1 (8.1) | 0.52 |
| Positive | 25.6 (7.8) | 27.4 (8.3) | 0.41 |
| BAI | 28.9 (9.3) | 30.5 (8.2) | 0.54 |
| BDI | 4.3 (4.9) | 3.3(2.9) | 0.38 |
| STAI | |||
| State | 41.6 (6.3) | 40.2 (6.4) | 0.42 |
| Trait | 41.0 (6.4) | 42.2 (5.7) | 0.49 |
| Characteristic | Real-feedback group (n = 28) | Yoked-feedback group (n = 25) | P-value |
|---|---|---|---|
| Age, years | 23.7 (3.8) | 23.4 (3.1) | 0.74 |
| Education, years | 14.8 (2.5) | 14.6 (2.4) | 0.83 |
| Cigarettes/day | 14.1 (4.5) | 15.0 (7.5) | 0.42 |
| Years of cigarette use | 7.1 (3.9) | 6.6 (2.8) | 0.57 |
| Score | |||
| FTND | 4.6 (1.9) | 4.8 (1.4) | 0.57 |
| MQS | 5.2 (2.6) | 5.0 (2.6) | 0.85 |
| BIS | 60.3 (5.4) | 59.2 (4.4) | 0.46 |
| SPSRQ | |||
| Punishment | 8.4 (4.1) | 7.6 (4.3) | 0.54 |
| Reward | 7.5 (2.4) | 6.7 (2.9) | 0.32 |
| ERQ | 45.9 (7.8) | 44.8 (6.6) | 0.56 |
| PANAS | |||
| Negative | 15.8 (6.7) | 17.1 (8.1) | 0.52 |
| Positive | 25.6 (7.8) | 27.4 (8.3) | 0.41 |
| BAI | 28.9 (9.3) | 30.5 (8.2) | 0.54 |
| BDI | 4.3 (4.9) | 3.3(2.9) | 0.38 |
| STAI | |||
| State | 41.6 (6.3) | 40.2 (6.4) | 0.42 |
| Trait | 41.0 (6.4) | 42.2 (5.7) | 0.49 |
Values are mean and values in parentheses are 1 standard deviation (SD).
BAI = Beck Anxiety Inventory; BDI = Beck Depression Inventory; BIS = Barratt Impulsivity Scale; ERQ = Emotion Regulation Questionnaire; FTND = Fagerström Test of Nicotine Dependence; MQS = Motivation of Quitting Smoking; PANAS = Positive and Negative Affect Scales; SPRSQ = The Sensitivity to Punishment and Sensitivity to Reward Questionnaire; STAI = State-Trait Anxiety Inventory.
Baseline demographic and clinical characteristics in two groups
| Characteristic | Real-feedback group (n = 28) | Yoked-feedback group (n = 25) | P-value |
|---|---|---|---|
| Age, years | 23.7 (3.8) | 23.4 (3.1) | 0.74 |
| Education, years | 14.8 (2.5) | 14.6 (2.4) | 0.83 |
| Cigarettes/day | 14.1 (4.5) | 15.0 (7.5) | 0.42 |
| Years of cigarette use | 7.1 (3.9) | 6.6 (2.8) | 0.57 |
| Score | |||
| FTND | 4.6 (1.9) | 4.8 (1.4) | 0.57 |
| MQS | 5.2 (2.6) | 5.0 (2.6) | 0.85 |
| BIS | 60.3 (5.4) | 59.2 (4.4) | 0.46 |
| SPSRQ | |||
| Punishment | 8.4 (4.1) | 7.6 (4.3) | 0.54 |
| Reward | 7.5 (2.4) | 6.7 (2.9) | 0.32 |
| ERQ | 45.9 (7.8) | 44.8 (6.6) | 0.56 |
| PANAS | |||
| Negative | 15.8 (6.7) | 17.1 (8.1) | 0.52 |
| Positive | 25.6 (7.8) | 27.4 (8.3) | 0.41 |
| BAI | 28.9 (9.3) | 30.5 (8.2) | 0.54 |
| BDI | 4.3 (4.9) | 3.3(2.9) | 0.38 |
| STAI | |||
| State | 41.6 (6.3) | 40.2 (6.4) | 0.42 |
| Trait | 41.0 (6.4) | 42.2 (5.7) | 0.49 |
| Characteristic | Real-feedback group (n = 28) | Yoked-feedback group (n = 25) | P-value |
|---|---|---|---|
| Age, years | 23.7 (3.8) | 23.4 (3.1) | 0.74 |
| Education, years | 14.8 (2.5) | 14.6 (2.4) | 0.83 |
| Cigarettes/day | 14.1 (4.5) | 15.0 (7.5) | 0.42 |
| Years of cigarette use | 7.1 (3.9) | 6.6 (2.8) | 0.57 |
| Score | |||
| FTND | 4.6 (1.9) | 4.8 (1.4) | 0.57 |
| MQS | 5.2 (2.6) | 5.0 (2.6) | 0.85 |
| BIS | 60.3 (5.4) | 59.2 (4.4) | 0.46 |
| SPSRQ | |||
| Punishment | 8.4 (4.1) | 7.6 (4.3) | 0.54 |
| Reward | 7.5 (2.4) | 6.7 (2.9) | 0.32 |
| ERQ | 45.9 (7.8) | 44.8 (6.6) | 0.56 |
| PANAS | |||
| Negative | 15.8 (6.7) | 17.1 (8.1) | 0.52 |
| Positive | 25.6 (7.8) | 27.4 (8.3) | 0.41 |
| BAI | 28.9 (9.3) | 30.5 (8.2) | 0.54 |
| BDI | 4.3 (4.9) | 3.3(2.9) | 0.38 |
| STAI | |||
| State | 41.6 (6.3) | 40.2 (6.4) | 0.42 |
| Trait | 41.0 (6.4) | 42.2 (5.7) | 0.49 |
Values are mean and values in parentheses are 1 standard deviation (SD).
BAI = Beck Anxiety Inventory; BDI = Beck Depression Inventory; BIS = Barratt Impulsivity Scale; ERQ = Emotion Regulation Questionnaire; FTND = Fagerström Test of Nicotine Dependence; MQS = Motivation of Quitting Smoking; PANAS = Positive and Negative Affect Scales; SPRSQ = The Sensitivity to Punishment and Sensitivity to Reward Questionnaire; STAI = State-Trait Anxiety Inventory.
Neurofeedback performance
During neurofeedback training, the real-time outputs of the MVPA classifier represented the calculated probabilistic score of smoking cue reactivity patterns, which indicated the extent to which the current EEG activity pattern matched the pattern for reactivity to smoking cues. A linear regression analysis revealed that there was a strong and significant negative correlation between training cycle and the mean probabilistic score across participants in the real-feedback group (r = −0.155, P = 0.001) (Fig. 3A). However, this finding was not observed in the yoked-feedback group (r = 0.015, P = 0.77) (Fig. 3B) and the correlation was significantly different from the real-feedback group (z = −2.47, P = 0.013). A two-way mixed-design ANOVA using group (real-feedback, yoked-feedback) as a between-subjects factor and cycle (first cycle of neurofeedback Visit 1, last cycle of neurofeedback Visit 2) as a within-subjects factor on the probabilistic score revealed a significant group × cycle interaction [F(1,51) = 4.04, P = 0.04, d = 0.56] (Fig. 3C). The probabilistic score for the first cycle at the first neurofeedback visit did not significantly differ between the two groups [t(51) = 1.09, P = 0.28, d = 0.30]. The interaction effect on probabilistic score was due entirely to decrease in the real-feedback group [t(27) = 2.53, P = 0.017, d = 0.69], and not the yoked-feedback group [t(24) = −0.58, P = 0.57, d = 0.18].
Neurofeedback learning. During two visits of neurofeedback training, the real-feedback group learned to deactivate the neurofeedback signal (r = −0.15, P = 1.0 × 10−3) (A), but the yoked-feedback did not (r = 0.015, P = 0.77) (B). Additionally, there was a significant group × cycle interaction effect on the probabilistic score (C). There were eight cycles in each neurofeedback visit. Each circle represented a participant. The neurofeedback signal was represented by the probabilistic score and indicated the extent to which the current brain activity pattern matched the pattern for reactivity to smoking cues. Although the neurofeedback display in the yoked-feedback group was from the real-feedback group, the probabilistic score from the yoked-feedback group in this figure was calculated by their own smoking cue activity patterns. *P < 0.05; error bar (SD). NF = neurofeedback.
Altogether, these findings indicated after two neurofeedback visits the real-feedback group successfully deactivated smoking cue reactivity patterns.
Short-term effects on cigarette craving
A two-way mixed-design ANOVA using group (real-feedback, yoked-feedback) as a between-subjects factor and time (pre-neurofeedback, post-neurofeedback) as a within-subjects factor on the cigarette craving score revealed a significant group × time interaction [F(1,51) = 4.69, P = 0.03, d = 0.61] (Fig. 4A). At pre-neurofeedback, there was no significant difference in craving score between the two groups [t(51) = 0.85, P = 0.40, d = 0.23]. After two neurofeedback visits, the real-feedback group had significantly decreased craving score from pre-neurofeedback to post-neurofeedback [t(27) = 4.28, P = 2.1 × 10−4, d = 0.69], while the yoked-feedback group did not meet statistical significance in their change in craving scores [t(24) = 1.92, P = 0.07, d = 0.27]. Furthermore, participants in the real-feedback group with lower levels of average deactivated neurofeedback performance exhibited greater decreases in craving scores (r = −0.40, P = 0.03), which was consistent with our hypothesized mechanism of action for this intervention (Fig. 4B). This correlation was not significant in the yoked-feedback group (r = −0.12, P = 0.53).
Neurofeedback effects on cigarette craving. (A) Short-term effects on cigarette craving. The craving score was assessed using the Tobacco Craving Questionnaire. The pre-neurofeedback was defined as before the first neurofeedback training session. The post-neurofeedback was after the second neurofeedback training session. (B) The correlation between the average deactivated neurofeedback performance and decreased craving. The average deactivated neurofeedback performance was defined as the average probabilistic score of smoking cue reactivity patterns across all cycles during two visits of neurofeedback. *P < 0.05; ***P < 0.005; ns = not significant; NF = neurofeedback.
Effects on event-related potential correlates of cigarette craving
While neurofeedback training reduced ratings of cigarette craving, we were interested as to whether training also altered the neural correlates of cigarette craving. Therefore, we compared the P300 component evoked by smoking-related cues at electrode site Pz (Fig. 5A and B), which is the most frequently reported electrode site for analysis in previous studies (Littel et al., 2012), between groups from pre-neurofeedback to post-neurofeedback. A two-way mixed-design ANOVA using group (real-feedback, yoked-feedback) as a between-subjects factor and time (pre-neurofeedback, post-neurofeedback) as a within-subjects factor on average amplitude at selected group peak time window (350 − 450 ms) revealed a significant group × time interaction effect [F(1,51) = 5.13, P = 0.028, d = 0.64]. Permutation tests showed that the real-feedback group exhibited a significantly decreased P300 amplitude (P < 0.005) with decreased P300 amplitudes being evident in frontal-parietal regions on a scalp topographical analysis (Fig. 5A); in contrast, the yoked-feedback group did not exhibit an effect (P > 0.1; Fig. 5B). The significantly decreased P300 amplitude was not observed for neutral-related cues, regardless of group assignment (Supplementary Fig. 5). Additionally, participants in the real-feedback group with a greater decrease in the mean P300 amplitude exhibited a greater decrease in craving score (r = 0.43, P = 0.02; Fig. 5C). No significant correlation was observed in the changed P300 latency and decreased cigarette craving in either group (all P > 0.30). Bootstrapped mediation analyses showed that the P300 amplitude change partially mediated the relationship between average neurofeedback performance and decreased craving (Fig. 5D).
![Neurofeedback effects on P300 correlates of cigarette craving. The neurofeedback effects on P300 correlates of cigarette craving appeared in the real-feedback group and was evident in frontal-parietal regions (A), but not in the yoked-feedback group (B). (C) The correlation between the mean decreased P300 amplitudes and decreased craving score. The mean P300 activity was calculated across the electrodes, which were significant in scalp topography of the real-feedback group (frontal-parietal regions). (D) The association between the average neurofeedback performance, P300 change, and craving change. Bootstrapped mediation analyses showed that there was a significant indirect path from average neurofeedback performance to decreased cigarette craving, via decreased P300 amplitudes [B = −17.90; 95% confidence interval, CI: (−48.49, −0.39)]. Furthermore, after controlling for the indirect path, the correlation between the average neurofeedback performance and decreased craving remained significant [B = −37.54; 95% CI: (−99.39, −0.05)], revealing the partial mediation of P300 amplitude. Numbers above the arrows indicate beta coefficients, and numbers in parentheses indicate standard error (SE). *P < 0.05; ***P < 0.005; NF = neurofeedback; ns = not significant.](https://oup.silverchair-cdn.com/oup/backfile/Content_public/Journal/brain/142/6/10.1093_brain_awz114/1/m_awz114f5.jpeg?Expires=1747854652&Signature=uNdODKcPyU19W9e8CmU-Y3IzclJfYHtVCSUpXNEnHGZw~mLoE6vt8yfX0jjpisI2WaiFenIeRvvwv98krutZ81mAeMmsZhLkvqytxNjv6Rx~EFG~W6bCmtCu6T11xmQqfE7fxfBo9SNg0Tk7Omlxpa-A4X5dC5NS4EEycemvyhHN235eNt-XzHs4ynDXlTg2YqlfO38uz3LC4x90IEaNZKK~Ubc~IkIEHMBH-K0IJWtdtaW2kS4IucSU3dJmVvKMI6RohR2ctN~XSNKoWqEu2G01jr5PYjkkNOs-PUOzc0WE60HfRruzsyomE6I4t8G7L2m3vqGccEi7rh5UpzlE4w__&Key-Pair-Id=APKAIE5G5CRDK6RD3PGA)
Neurofeedback effects on P300 correlates of cigarette craving. The neurofeedback effects on P300 correlates of cigarette craving appeared in the real-feedback group and was evident in frontal-parietal regions (A), but not in the yoked-feedback group (B). (C) The correlation between the mean decreased P300 amplitudes and decreased craving score. The mean P300 activity was calculated across the electrodes, which were significant in scalp topography of the real-feedback group (frontal-parietal regions). (D) The association between the average neurofeedback performance, P300 change, and craving change. Bootstrapped mediation analyses showed that there was a significant indirect path from average neurofeedback performance to decreased cigarette craving, via decreased P300 amplitudes [B = −17.90; 95% confidence interval, CI: (−48.49, −0.39)]. Furthermore, after controlling for the indirect path, the correlation between the average neurofeedback performance and decreased craving remained significant [B = −37.54; 95% CI: (−99.39, −0.05)], revealing the partial mediation of P300 amplitude. Numbers above the arrows indicate beta coefficients, and numbers in parentheses indicate standard error (SE). *P < 0.05; ***P < 0.005; NF = neurofeedback; ns = not significant.
Long-term effects on smoking behaviour
A two-way mixed-design ANOVA using group (real-feedback, yoked-feedback) as a between-subjects factor and time (pre-neurofeedback, 1-week follow-up, 1-month follow-up, 4-month follow-up visit) as a within-subjects factor on daily cigarette consumption revealed a significant group × time interaction [F(3,126) = 3.68, P = 0.01, d = 0.59] (Fig. 6A). Groups did not differ in the number of cigarettes smoked per day at pre-neurofeedback [t(51) = −0.81, P = 0.42, d = 0.22]. After two neurofeedback training visits, the real-feedback group showed significantly decreased cigarette consumption per day compared to the yoked-feedback group at the 1-week follow-up [t(48) =−2.53, P = 0.01, d = 0.72], 1-month follow-up [t(47) = −2.98, P < 0.005, d = 0.86], and 4-month follow-up [t(42) = −2.21, P = 0.03, d = 0.67]. A maximum 38.2% decrease in daily cigarette consumption from pre-neurofeedback to follow-up was observed in the real-feedback group, which was significantly different from the consumption reported at baseline [t(26) = 6.49, P < 0.001, d = 1.24]. In addition, the real-feedback group showed a significant correlation between the average deactivated neurofeedback performance and the current cigarette amount at the 4-month follow-up (r = 0.58, P = 0.004) (Fig. 6B). The temporal trajectory of changes in smoking behaviour induced by neurofeedback displayed a significant u-shape pattern, with maximal effects seen at 1-month post-neurofeedback (Supplementary Fig. 6). To exclude potentially confounding effects from the participants lost to follow-up, we excluded all participants lost to follow-up and still found that neurofeedback produced significant long-term effects on daily cigarette consumption (Supplementary Fig. 10). In addition, long-term effects on EEG patterns of smoking cue reactivity were observed at the 4-month follow-up (Supplementary Fig. 7).
Neurofeedback effects on smoking behaviour. (A) Long-term effects on smoking behaviour. Smoking behaviour was assessed by the average number of cigarettes smoked per day from the last visit/telephone interview to the current one. Error bar (standard deviation, SD). (B) The correlation between the average deactivated neurofeedback performance and the number of cigarettes smoked per day at the 4-month follow-up. *P < 0.05; **P < 0.01; ***P < 0.005. NF = neurofeedback.
Individual-level prediction of short- and long-term effects
Figure 7A shows that the classification accuracy of the pre-neurofeedback classifier significantly predicted decreased craving scores in the real-feedback group (r = 0.40, P = 0.03), while the same prediction was not significant in the yoked-feedback group (r = 0.13, P = 0.54). Moreover, the correlation analysis revealed that the degree of deactivation during the first cycle of the first neurofeedback successfully predicted the number of cigarettes smoked per day at the 4-month follow-up (r = 0.45, P = 0.03; Fig. 7B) in the real-feedback group, but not in the yoked-feedback group (r = 0.16, P = 0.45). After controlling for the all psychological scales in Table 1, the partial correlation analysis revealed that the prediction of short- and long-term effects using classification accuracy (r = 0.53, P = 0.03) and initial neurofeedback performance (r = 0.63, P = 0.04), respectively were still significant (Supplementary Fig. 8).
Individual-level prediction of short-term (A) and long-term (B) effects. (A) Correlation between classification accuracy of pre-neurofeedback classifier and decreased craving. (B) Correlation between the mean probabilistic score at the first cycle of the first neurofeedback visit and the number of cigarettes smoked per day at the 4-month follow-up. NF = neurofeedback.
Discussion
In this study, we developed and evaluated a new closed-loop, smoking cue reactivity pattern EEG-based neurofeedback protocol. To our knowledge, this is the first randomized clinical trial of neurofeedback therapy for nicotine addiction. Smokers completing two training sessions (∼1 h/session) with our novel neurofeedback paradigm reported significant behavioural and neurophysiological changes during both short- and long-term periods, up to 4 months post-intervention. Furthermore, these effects were predicted by classification accuracy at pre-neurofeedback and deactivated performance at the beginning of the neurofeedback training.
Neurofeedback learning with a specific process classifier
In the current study, brain activity patterns for smoking cue reactivity identified via EEG recording were successfully deactivated after two neurofeedback training sessions. Although previous studies (e.g. alpha training) suggest that EEG neurofeedback is a promising technique for treating mental disorders, these studies at best reveal influences of single and fixed EEG frequency on human behaviour, which is a limitation for an individually adapted pattern approach. In contrast, the current neurofeedback method manipulates smoking cue reactivity patterns, allowing us to influence specific functions. Similarly, previous functional MRI-based neurofeedback studies showed that brain activity patterns in vision and attention were successfully manipulated after only several training sessions (deBettencourt et al., 2015; Amano et al., 2016). These findings support the hypothesis that complex multivariate brain activity patterns can be self-regulated in addition to a signal from a single component or brain region. This extends results from previous studies focused on self-regulation of activity patterns in healthy participants aimed at improving cognitive abilities (Sitaram et al., 2017), and suggests that clinical populations may also benefit from multivariate pattern neurofeedback training.
Although some traditional EEG-neurofeedback studies required dozens of training sessions, the skill learning (Hinterberger et al., 2005; Koralek et al., 2012) explanation for neurofeedback suggests that participants can indeed learn skills that are maintained for long periods after only a small number of training sessions (Miyachi et al., 2002; Philippens and Vanwersch, 2010; Koralek et al., 2012; Birbaumer et al., 2013; Witte et al., 2013). Additionally, the quality of the feedback provided is critical to effective training (Oblak et al., 2017). The current novel neurofeedback protocol, which combined a specific process of addiction with MVPA methods and a closed-loop feedback concept, is different from previous EEG-neurofeedback protocols. Similarly, numerous recent functional MRI-based neurofeedback studies showed significant training effects after only a few training sessions (Beatty et al., 1974; Shibata et al., 2011; deBettencourt et al., 2015; Amano et al., 2016; Keynan et al., 2016; Yanagisawa et al., 2016; Koush et al., 2017). In particular, some functional MRI-based neurofeedback studies showed that two sessions seem to be sufficient (Subramanian et al., 2011; Scheinost et al., 2013; Young et al., 2017a, b).
One may consider that participants in the yoked-control group realized the feedback signals were not theirs, raising the possibility that the successful neurofeedback learning and effects were driven by the belief of controlling one’s own neural activity. To test this possibility, we examined the perceived control ratings of regulating the feedback signal between the two groups and found that there was no difference in perceived control rating after either training session (Supplementary material). The yoked-control design has the advantage of ensuring the same feedback stimuli were used, and controlled for variations in task difficulty and feedback reward levels. We can conclude that the results were not due to differences in perceived self-efficacy, though future studies using different control conditions (such as strategy only or feedback of a different pattern) may provide useful comparisons.
Although there was no difference in perceived control rating after either training session between the groups, both groups perceived some degree of brain control (Supplementary material). Given that the neurofeedback signal in the yoked-feedback group was from another participant’s brain activity pattern, rather than random noise signals, the participants in the yoked-feedback group may get some illusive feeling that is reacting to and following the intensity of the stimuli presented during neurofeedback. This possible explanation was supported by our finding that the control rating for the yoked-feedback group was significantly correlated with the false neurofeedback signal that the yoked-feedback group received, but not with their own brain activity patterns (Supplementary material).
Short-term effects on cigarette craving
The real-feedback group significantly deactivated smoking cue reactivity patterns, which resulted in greater decreases in cigarette craving relative to the yoked-feedback group, which did not significantly deactivate smoking cue reactivity patterns. The significant difference indicates neurofeedback of smoking cue reactivity patterns played a crucial role in craving change. The importance of smoking cue reactivity neurofeedback to nicotine addiction is further highlighted by the model of cue reactivity, which indicates that smoking cues are processed at both bottom-up and top-down levels to amplify the conditional value, and has treatment implications for addictive behaviours (Chiamulera, 2005). Moreover, the results of recent studies that found that self-regulating haemodynamic activity related to smoking cue reactivity in the anterior cingulate cortex or prefrontal cortex decreased cigarette craving (Hartwell et al., 2016). Furthermore, the significant correlation between average deactivation of activity patterns and decreased craving was observed in the real-feedback group, but not in the yoked-feedback group, supporting the hypothesis that deactivation of smoking cue reactivity patterns led to decreases in cigarette craving. Additionally, these results support the hypothesis that learning control over specific neural substrates could change specific health-related aspects of mental function resulting in psychological or psychiatric benefit (Sitaram et al., 2017).
We also found significantly decreased P300 amplitudes immediately after neurofeedback training. This result, in conjunction with cigarette craving change, demonstrates that neurofeedback training produces effects on both subjective psychological feelings as well as objective physiological performance. The significant correlation between mean P300 amplitude change and cigarette craving change is in-line with previous studies showing the P300 during smoking cue reactivity task is robustly associated with cigarette craving (Knott et al., 2008; Field et al., 2009). Altogether, these reports suggest the potential of neurofeedback to affect both neural and psychological measures of cigarette craving in nicotine addicts.
The mediation analysis indicates that P300 partially mediates the relationship between average deactivated neurofeedback performance and craving change. Although P300 was one of the EEG components used to construct the smoking cue reactivity patterns and significantly overlapped for distinguishing smoking and neutral EEG activity (Supplementary Fig. 9), alpha, beta, and gamma power features were also critical components in the smoking cue reactivity pattern (Supplementary Fig. 2). Altogether, these findings suggest that P300 plays an important role in craving change, but is not the only factor. Cigarette craving involves many complex cognitive components, including memory, attention, and emotion, etc. Smoking cue-elicited P300 has been previously related to attention. We also found the scalp distribution of P300 change occurred primarily in frontal-parietal regions, the site where attention networks are localized (Dixon et al., 2018). Although the current study focused on the clinical benefits of this novel neurofeedback, to fully understand the brain network mechanisms of action for this neurofeedback intervention, concurrent EEG-functional MRI experiments could be designed in future studies.
One may wonder whether the mental strategy alone during neurofeedback explained the significant neurofeedback effects. However, we found that the reported deactivation strategies during neurofeedback in the yoked-feedback group were similar to the real-feedback group (Supplementary Table 1), but short- and long-term effects on nicotine addiction improved to a greater extent in the real-feedback group. This finding suggests that strategy alone could not lead to the successful deactivation of activity patterns and significant neurofeedback effects. This novel neurofeedback itself was crucial.
Long-term effects on smoking behaviour
After neurofeedback training, the average decrease in the number of cigarettes smoked per day during the 4-month follow-up was nearly 40% in the real-feedback group. A recent study that used multivariate pattern neurofeedback training across several days found that altered associative learning behaviour can persist 3–5 months post-training (Amano et al., 2016), supporting long-term effects following acute neurofeedback training may occur. Additionally, the classification accuracy at 4-month follow-up was lower in the real-feedback group than that in the yoked-feedback group (Supplementary Fig. 7), which indicated that the evoked smoking and neutral cue reactivity patterns at pre-neurofeedback did not classify the pattern of measured EEG signals at the 4-month follow-up into one of two types of cue. Thus, this novel neurofeedback training may affect brain activity patterns for long-lasting periods. Furthermore, we found the average neurofeedback performance significantly correlated with the number of cigarettes smoked per day at the 4-month follow-up visit. These results support the notion that a small number of neurofeedback sessions have the potential to produce long-term effects in both brain activity and behaviour (Sitaram et al., 2017).
The number of cigarettes smoked during follow-up displayed a significant u-shape change (Supplementary Fig. 6). This is consistent with a previous report that the time-dependent trajectory of behavioural effects from neurofeedback training may not be a simple decrease or increase following neurofeedback (Rance et al., 2018), highlighting the importance of monitoring the temporal patterns of effects induced by neurofeedback. This monitoring would benefit the treatment of nicotine addiction and may provide information as to the best time for patients to possibly select other treatment approaches or receive booster sessions.
Individual-level prediction of outcomes
Short- and long-term effects were correlated with classification accuracy at pre-neurofeedback and neurofeedback performance during the first training cycle, respectively. The deactivated neurofeedback signal was decoded by the constructed classifier at pre-neurofeedback. Thus, high classification accuracy of the classifier would accurately decode the real-time brain state with high probability, which may result in better neurofeedback self-regulation. This finding that the classification accuracy at pre-neurofeedback significantly correlated with short-term craving change suggests that improving the classification accuracy may enhance neurofeedback effects.
A fundamental goal of this neurofeedback method was to deactivate brain activity patterns of smoking cue reactivity in nicotine addicts; thus, neurofeedback was perhaps particularly successful in patients with superior activity pattern deactivation capacities, as reflected by deactivation performance at the beginning of training. The finding that neurofeedback performance during the first training cycle significantly correlated with the number of cigarettes smoked per day at the 4-month follow-up visit suggests that a capacity-related biomarker could predict trait-related performance in the long-term. In addition, the result that no significant correlation was observed between the classification accuracy at pre-neurofeedback and initial neurofeedback performance during the first training cycle, in conjunction with the finding that after controlling for all psychological scales these two predictors still significantly predict the short- and long-term effects, demonstrated that these two predictors specifically predicted the novel neurofeedback effects. Collectively, these individual-level prediction findings may provide help in personalized treatment.
The present study has several limitations. First, there were no neurofeedback transfer runs to see the generalized effect without feedback at the 4-month follow-up. One may wonder if the long-term effects could be accounted for by other factors. However, we did find a significant correlation between the average neurofeedback performance and the cigarette consumption at 4-month follow-up, supporting the view that long-term improvements were driven by the hypothesized mechanism of action. Second, the outcome measurements in this study did not follow Russell Standard (West et al., 2005) exactly: self-report of smoking abstinence and biochemical verification data were not collected and participants were followed-up only three times: at 1 week, 1 month, and 4 months. The main reason for this methodological difference is that only 2 h of neurofeedback training was provided in the present study, which may not produce strong effects on smoking abstinence. Further optimizing the training sessions and extending the follow-up times is important in developing this intervention as a clinical treatment. Third, we only included young male smokers. And there may be limited generalizability of the novel neurofeedback effects on older smokers and female smokers. Additionally, it was unclear if the participants were interested in smoking cessation in the near future. Further studies need to consider this information. Fourth, it is interesting to explore whether participants used the learned neurofeedback strategy to handle craving in the real-world during the follow-up period. Although we asked participants to practice the strategies offline that worked best for them during neurofeedback training in the short-term, we did not provide instructions to participants during the follow-up period. However, this interesting question deserves investigation in future studies.
In conclusion, we developed and tested a novel neurofeedback protocol to deactivate EEG activity patterns of smoking cue reactivity, which produced short- and long-term effects on cigarette craving and smoking behaviour. In particular, the rate of smoking amount decreased as much as 38.2% during the 4-month follow-up period after only 2 h of neurofeedback training. These results suggest that this novel neurofeedback intervention is a promising treatment for addiction, with potential to be a low-cost and high-portability brain-based treatment for addiction. This approach therefore merits further testing.
Acknowledgements
We acknowledge Lizhuang Yang, Guanbao Cui, Nan Wu, Siyu Wang and Hongge Xu for their assistance with statistical approaches used and subject assessment.
Funding
This work was supported by grants from The National Key Basic Research Program (2016YFA0400900 and 2018YFC0831101), The National Natural Science Foundation of China (31771221, 31471071, 61773360, and 71874170), The Fundamental Research Funds for the Central Universities of China. A portion of the numerical calculations in this study were performed with the supercomputing system at the Supercomputing Centre of USTC.
Competing interests
All authors report no biomedical financial interests or potential conflicts of interest.
