Closed-Loop Neurofeedback Targeting the Right Dorsolateral Prefrontal Cortex for Cardiac Autonomic Modulation in Coronary Artery Disease With Anxiety

This study is a randomised, sham-controlled, parallel-group neurofeedback trial designed to test whether real-time fNIRS-BCI targeting the right dorsolateral prefrontal cortex (right DLPFC), using active volitional control during a slow-wave auditory acoustic paradigm, can suppress cardiac sympathetic activity and improve autonomic regulation in right-handed patients with stable coronary heart disease and comorbid DSM-5 anxiety. Eligibility criteria, outcome measures, arms and interventions, and statistical analysis are specified in their dedicated fields of this record to avoid duplication.

Participants complete screening and a three-day adaptation phase, followed by two formal experimental sessions on consecutive days. No verbal guidance is provided on test days. During each formal session, participants undergo 15 block cycles of 60 seconds each, comprising 20 seconds of rest and 40 seconds of auditory stimulation. The auditory stimulus is a 1 Hz amplitude-modulated pure tone at approximately 60 dB. Within the 40-second window, a 10-second white-noise burst is randomly embedded while avoiding the first and last 2 seconds; start times are selected without replacement from a discrete set to minimise predictability and sequence effects and to probe interference resistance.

Neurofeedback targets the right DLPFC. Real-time statistical values of cortical activation are mapped to a visual energy bar that serves as the feedback signal. Participants use active volitional strategies to push the energy bar below an unlabeled threshold line during stimulation. In the real-feedback group the statistical threshold is set to T = -3.3, producing a sensitive, contingent mapping between right DLPFC activity and the energy bar. In the sham-feedback group the threshold is set to T = -1, weakening the effective coupling between neural state and on-screen feedback while keeping the interface identical. Only the session operator is aware of group assignment; outcome assessors and data analysts are blinded.

Signals are acquired synchronously. fNIRS uses a NirSmart system with 7 sources and 7 detectors forming 19 channels, dual wavelengths at 730 nm and 850 nm, source-detector separation of 3 cm, and an 11 Hz sampling rate with avalanche photodiode detectors. Optodes are positioned over bilateral prefrontal cortex using the 10-20 system with FPz as the central reference. Channels are co-registered to MNI space, with right DLPFC (Brodmann area 46) designated a priori as the region of interest; channels 11 and 13 represent right DLPFC for feedback, and channels 17 and 18 represent left DLPFC for reference. ECG is recorded using a 3-lead configuration at 1,000 Hz.

Online processing and feedback are implemented with the NeuroMind-NIRS pipeline. A MATLAB timer provides a fixed update cadence. The streaming architecture is layered into data, visualisation, control, and persistence channels. A sliding-window general linear model is applied to fNIRS signals with low-pass filtering, linear detrending, and within-window normalisation. A boxcar task regressor convolved with a canonical haemodynamic response function is used; the statistical readout at the end of each window updates the energy bar and threshold decision. Statistical inference is prioritised ahead of rendering within a fixed real-time budget to maintain a stable feedback cadence. Per-frame compute time, render time, total pipeline time, latency, and dropped-frame rate are logged. The acoustic stimulus is not triggered or gated by the statistical readout.

Offline fNIRS preprocessing is performed in MATLAB using the Homer2 toolbox. Steps include conversion to optical density, motion artefact detection by channel, spline interpolation correction, band-pass filtering at 0.01 to 0.08 Hz, and calculation of oxyhaemoglobin concentration via the modified Beer-Lambert law. Analyses focus on HbO due to its higher signal-to-noise ratio. For activation analysis, normalised HbO time series from DLPFC channels enter a general linear model with a task regressor based on the convolved canonical haemodynamic response function and its derivatives; preprocessed ECG-derived covariates are included to account for low-frequency physiological oscillations. Subject-level task β estimates are taken forward to group-level linear mixed-effects models.

ECG processing includes R-peak detection by the Pan-Tompkins algorithm, correction of spurious detections, construction of RR interval series, ectopic beat removal using robust filters, and resampling as needed for alignment with fNIRS. Heart rate and heart rate variability are computed within block-aligned windows with baseline correction using the pre-stimulation interval. For spectral analysis, time-frequency methods are applied to quantify power around 0.0167 Hz, corresponding to the 60-second block rhythm within the very low frequency band.

Safety is monitored continuously from the start of adaptation through the end of Session 2. Anticipated adverse events related to sympathetic attenuation include mild dizziness or drowsiness, transient bradycardia or hypotension-related discomfort, headache, and nausea. Standard operating procedures specify pausing or terminating sessions and initiating clinical assessment as needed; serious adverse events trigger immediate discontinuation and follow-up. Quality control includes unscheduled supervisory checks, standardised operator training, checkpointed critical steps, real-time records, de-identified data handling, tiered access control, and auditable transfer logs. De-identified individual data and analysis scripts will be shared on a recognised repository at publication in accordance with data use and ethics policies.