Wait, Automated, Personalized Remission with SAINT Might Be A Thing?
Closing the Loop on Depression, a The Frontier Psychiatrists Dispatch.
At Brain Stimulation 2025, Brandon Bentzley and colleagues presented a closed-loop neuromodulation system — and a new brain-complexity biomarker — that, together, sketch a bold architecture for keeping treatment-resistant patients well. I'm a huge fan of Brandon. I think he's both gracious and forward-thinking, a combination I aspire to. Before he was a Psychiatrist, he was a physicist, and he spent some time working on general relativity in Australia during his youth. If only all of our youthful dalliances could be so serious!
My Book Is In Print!
My 4th book, Inessential Pharmacology, is now available in paperback and hardcover. Other books by Owen Muir include an MBT Therapy manual and the poetry collections Hellthread and Why We Skeet. (affiliate links, all).
And now, today’s dispatch…
BRANDON BENTZLEY, MD, PHD — MAGNUS MEDICAL & SALMA HEALTH
Brandon opened by making the regulatory landscape explicit: SAINT — Stanford Accelerated Intelligent Neuromodulation Therapy, commercialized through Magnus Medical, which Bentzley co-founded roughly 5 years ago, has FDA breakthrough device clearance for treatment-resistant depression. In some hospitals, Medicare reimbursement is an active payment model. What he was about to describe, however, was remarkable: a personalized continuation protocol layered on top of SAINT. He noted it remains firmly experimental and carries no FDA label. He wanted that on record before a single slide advanced. [Yours truly, furiously taking notes on the future anyway!]
The problem laid out is one I recognized. Treatment-resistant depression is usually a relapsing and remitting illness. The more treatment-resistant a patient is, the higher their probability of relapse, the shorter their windows of remission. For many suffering, it’s not “true” remission at all, but a chronic state of sorta-misery that shades each day without reaching the “threshold of a full episode.”
“I think maybe as psychiatrists we’ve already resigned to the fact that this is a lifelong illness,” he said, “and we’ve got to continue to be on this journey with our patients to keep them in whatever state possible. Even if you knew the moment a patient was beginning to slip, what would you do about it? Start another antidepressant that won’t work for weeks?”
The team identified three barriers to durable remission.
The inability to intervene fast: outpatient psychiatry rarely involves more than monthly contact, and asking any clinician to see every patient daily, or asking every patient to complete a validated scale daily, is a thought experiment, not reality. Only 8% of psychiatrists complete a follow-up measure, and only 16% obtain one at baseline.
The absence of a fast-acting rescue: even with perfect knowledge that someone is beginning to slip, no currently approved treatment can reverse that course in under a week, let alone in a single day.
The heterogeneity problem: every prior attempt at “maintenance treatment” either over-treats patients who don’t need it or under-treats those who do, because people’s trajectories are deeply individual and their triggers are often completely unpredictable.
The PCT Model: Wearables, ML-Models, and Days of SAINT
The solution Bentzley’s team designed has three parts. After an initial five-day acute SAINT course, using the standard FDA-cleared protocol with individualized targeting and depth correction to maintain 90% motor threshold, patients who achieve remission (defined by MADRS score) enter a continuous monitoring program.
WHAT IS SAINT? If you are a new reader, SAINT (Stanford Accelerated Intelligent Neuromodulation Therapy) delivers accelerated, high-dose intermittent theta-burst TMS using individualized fMRI targeting. The pivotal RCT published in NEJM (Cole et al., 2022) demonstrated 78.6% remission in treatment-resistant depression after a single five-day course, vs. 13.6% for sham. It’s been replicated in an Open-Label Dose Optimization Study, on which I served as an investigator1, and in a follow-up RCT2.
That monitoring is built on medical-grade wearables from an Italian research-device company — four FDA-cleared sensors generating roughly 5.5 million data points per person per day, capturing photoplethysmography (heart rate, pulse), skin temperature, skin conductance as surrogates for sympathetic/parasympathetic balance, and multi-axis accelerometer and gyroscope data. Biometric predictions are trained against expert MADRS raters, cross-validated for inter-rater reliability, so that, each day, the system produces an estimated MADRS score for every participant. When that score crosses the threshold for early relapse (MADRS ≥ 11), the clinic receives an automated alert, and the patient is brought back for one or more additional days of SAINT to reestablish remission.
The model architecture itself is worth a moment of our time. The program maintains a generalized model, trained on the full population, which is useful for patients the system has never seen before. It also has an individualized model that updates daily for each person. The two compete: once the individualized model begins outperforming the generalized model, the system switches over.
“In the long term,” Bentzley noted, “the individualized model always works much better.”
Dimensionality reduction was handled carefully—overfitting3 is a risk when there are far more features than participants. Careful feature selection was essential to prevent overfitting, with the system dynamically prioritizing the biometric features most predictive for each individual. For some patients, heart rate variability dominated. For others, activity data carried the signal.
The Results: A Year in Remission
The accuracy of the biometric prediction system improved steadily over the year-long trial, ultimately exceeding 95% — with one notable dip mid-trial when a code error propagated, causing a temporary drop that reversed rapidly once corrected. Regarding clinical outcomes, participants entered with an average MADRS of approximately 30 and maintained an average MADRS in the remitted range throughout the year. On self-report and CGI-S measures, results were similar. Participants spent an estimated 85.9% of the year in remission.
“We don’t give people with spinal cord injuries a treatment that repairs the cord and then send them home without physical therapy. But we often expect exactly that of people with disabling mental illness.”
One of the more striking findings was that the number of additional SAINT days required per continuation course did not increase over time. In fact, self-report MADRS scores at the end of each continuation course trended lower across successive courses, a result that reached statistical significance. Bentzley offered two interpretations: a biological sensitization effect from repeatedly targeting the same circuit, and a psychological one.
“When you take people who are used to relapsing and clamp them in remission,” he said, “they have a lot more time to learn how not to be depressed.” Patients started traveling again. Getting jobs. Returning to school. Sometimes sending photos from places they hadn’t been able to go in years.
Consistent with a large body of prior literature, including decades of ECT maintenance studies, baseline treatment resistance was the strongest predictor of the amount of continuation treatment a patient would need. Bentzley closed with the spinal cord metaphor: if we had a treatment that could immediately repair a spinal cord injury, we wouldn’t expect someone who hadn’t walked in thirty years to run a marathon the next day. We’d give them rehabilitation.
“That’s one of the things I really want to think deeply about,” he said. “How do we go from biological induction of remission to actual rehabilitation?”
From the Floor: Selected Q&A
After the talk, a question arose about whether separate models were trained per individual or across the full cohort. Bentzley explained the two-model competing architecture: the generalized model handles new patients on day one, while the individualized model updates continuously and takes over once it demonstrates superior predictive performance.
A second question from the floor asked what specific biometric features drove predictions across the population. Bentzley’s answer was carefully hedged: feature importance varies substantially by individual. Heart rate variability dominates for some; for others, it contributes nothing. The system itself determines which features to prioritize for each person.
Your author, Owen Muir, closed the Q&A with a question cut to the chase:
When the two models compete, is the underlying math essentially a Tinder-style pairwise ranking?
Brandon confirmed my hunch. The models compete on held-out performance metrics in a winner-takes-all structure, not unlike ensemble methods in which models are scored against each other, with the winner deployed for live predictions until the balance of evidence shifts again.
It seems depression care, online dating, and StarCraft matches have more in common than we might have guessed.
We offer SAINT at Radial, and Brandon’s Salma Health does too! It’s also available from friends and co-authors of the PCT trial at Acacia Clinics.
Don’t sleep on study investigator David Carreon, MD’s substack!
You can see Salma, Radial, Definium, NRx, Beckley Clinical, Ampa, SimplePractice, GrayMatters Health, MDhub, Psyrin, and more teams speak live and in person on the first night of the APA annual meeting at RAMHT 2026 SF—May Edition!
Get your tickets now!
Solvason, H. B., Pottanat, R., Muir, O. S., MacMillan, C., Cook, I. A., Carreon, D. M., ... & Bentzley, B. S. (2025). Assessing the Effectiveness of the SAINT® Neuromodulation System to Treat Major Depressive Disorder in a Real-World Setting: A Multisite, Open-Label Optimization Study. Transcranial Magnetic Stimulation, 3.
Kratter, I. H., Austelle, C. W., Lissemore, J. I., Wada, M., Geoly, A., Chaiken, A., ... & Williams, N. R. (2026). Stanford neuromodulation therapy for treatment‐resistant depression: a randomized controlled trial confirming efficacy, and an EEG study providing insight into mechanism of action and a potentially predictive biomarker of efficacy. World Psychiatry, 25(1), 105-116.
This is a well-established statistical principle. When you have more features (predictors) than observations (participants), a model can fit the noise in the training data rather than the true underlying signal — essentially memorizing the data rather than learning generalizable patterns. This is sometimes called the "curse of dimensionality" or the p >> n problem, and it's a foundational concern in machine learning and statistics. It's why techniques like regularization, cross-validation, and dimensionality reduction exist.