A Moscow biofeedback protocol printed down to the threshold that counts as success
- The study and its design. An open-label randomised parallel-group trial at a single centre, the psychotherapy department of the A.Ya. Kozhevnikov Clinic of Nervous Diseases at the Sechenov University Clinical Centre in Moscow, published in Brain Sciences in July 2026 by Kotelnikova and colleagues. Data were collected between July 2022 and February 2025. 188 inpatients with a first-time diagnosis of anxiety disorder, ICD-10 F40 or F41, were randomised to biofeedback alone (76), escitalopram 10 mg per day alone (46), or the two together (66); 180 finished. Registration on ClinicalTrials.gov, NCT07628153, was retrospective and no protocol was posted in advance.
- The procedure, written out. Ten individual 60-minute sessions on ten working days across two weeks, Monday to Friday with the weekend off, run by a medical psychologist certified in the method. Each session held an introduction and a review of the previous day's homework, a 20-minute biofeedback run, and the next assignment. The equipment was the Reacor system from Medicom MTD, Russian registration certificate FSR 2009/05647. Background was recorded for 45–60 seconds before each controlled stage, and the procedure itself took 20–30 minutes.
- The parameter and the threshold. The controlled parameter was alpha power at 8–13 Hz, taken from occipital leads O1 and O2 with central leads C3 and C4 added. Target thresholds were set at 110–120% of each patient's own mean background, and success was counted as a rise of at least 15% in mean power over that background. Auditory feedback grew louder as alpha rose. For heart rate no fixed percentage threshold was used; the feedback ran inversely to the current rate.
- The clinical outcome. HAM-A total, rated by a psychiatrist blinded to allocation. From baseline to day 10 the biofeedback group moved from 26.55 (95% CI 25.14–27.96) to 14.87 (13.60–16.13) and the medication group from 26.85 (24.88–28.82) to 21.17 (18.83–23.51). The day-10 difference between medication and biofeedback was 6.30 points (95% CI 3.89–8.72), Cohen's d 0.94 (0.57–1.31), p = 0.00046. At one month that difference was 4.46 points (2.20–6.72), p = 0.074.
- The self-report outcome, also declared primary. The FFMQ total, filled in by patients who knew their allocation, ended 30.28 points further ahead in the biofeedback group than in the medication group (95% CI 23.91–36.65, p < 0.000001), the medication group having fallen. Heart rate was the one recorded parameter whose course did not differ between the groups, F(4, 352) = 0.949, p = 0.436.
The device is the Reacor, built by Medicom MTD and carrying Russian registration certificate FSR 2009/05647, and what it hands back to the patient is the power of their own alpha rhythm between 8 and 13 Hz, filtered in real time and turned into a sound that swells as the rhythm swells. Electrodes sit at O1 and O2, with C3 and C4 added. Heart rate is fed back in the same run by a signal that moves the other way, thinning as the pulse climbs.
Everything downstream of that is written to a number. Before each controlled stage the system takes 45–60 seconds of background and sets the target at 110–120% of that patient's own average. Success is not left to the clinician's reading of the session: it is a rise of at least 15% in mean alpha power over that background. Heart rate is handled differently and the authors say so – no fixed percentage there, the threshold shifted by algorithm with the patient's own course.
Around that sits a course of ten 60-minute sessions on ten working days, Monday to Friday. Twenty of the sixty minutes are the machine. The other forty are an introduction, a review of yesterday's homework and psychological work, then the next assignment: recorded meditations, videos of proprioceptive exercises, reading tied to the session topic, and a list of daily tasks with space to write. The person at the console is a medical psychologist certified in the method. The patients were inpatients at one Moscow clinic with a first-time F40 or F41 diagnosis; photosensitive epilepsy ruled a patient out, as did severe impairment of attention or memory.
Of the two outcomes declared primary, one was rated blind. The Hamilton scale, scored by a psychiatrist who did not know the allocation, separated biofeedback from medication by 6.30 points at day 10 (95% CI 3.89–8.72). The mindfulness questionnaire separated them by 30.28 points (23.91–36.65), and patients filled it in knowing their arm, so the larger movement is the one measured without that safeguard. Heart rate, fed back throughout, moved alike in all three groups, F(4, 352) = 0.949, p = 0.436. At one month the Hamilton gap reached p = 0.074; with all 188 carried in, medication against the two active arms returned 0.077 and 0.093, in the paper's order. Its interval, 2.20 to 6.72, still sits above zero. What that supports is that the advantage was not shown to hold, not that it went away.
Delivery is reported apart from effect. Recruiting the 188 ran from July 2022 to February 2025 at this single clinic, and the registry entry, NCT07628153, was filed after the trial had finished. The schedule itself held. All 76 patients in the biofeedback arm attended all ten sessions, the combined arm's completers averaged 9.2, range 8 to 10, and in the medication arm 44 of 46 took at least 80% of the prescribed dose by pill count.
Success in the session was defined before the session: a rise of at least 15% in mean alpha power over the 45–60 seconds of background just recorded.
The trial was open-label, so patients and therapists knew the allocation; only the psychiatrist rating the Hamilton scale was blinded, and the second primary outcome was a self-report filled in by unblinded patients. There was no sham biofeedback arm and no arm receiving the same sixty minutes of attention without a device, so the 20 minutes at the machine cannot be separated from the 40 minutes of psychological work and the homework. Randomisation was simple and unstratified, which left the arms unequal in size. Registration on ClinicalTrials.gov came after the fact and no protocol was public in advance. All eight dropouts fell in the combined arm; the authors compared them with completers on age, sex, HAM-A and FFMQ and found no differences at p > 0.10, and repeated the analysis on all 188 with baseline values carried forward. In that intention-to-treat version the one-month advantage over medication reached p = 0.077 and p = 0.093 for the two active arms; the paper lists the pair in that order without naming either arm at that point. No global correction was applied across the secondary outcomes. Anxiety subtypes were not recorded, so nothing here separates panic disorder from generalised anxiety. Follow-up stopped at one month, and the paper reports no functional, occupational, relapse or readmission outcome. This was one centre, inpatients only, first-time diagnoses only, and the authors limit their conclusion to that population. The data sit with the corresponding author on request rather than in a repository.