Case Study: The Clinical Translation of Transcranial Magnetic Stimulation

Key Lessons:

Though you’ve probably never heard of it, thousands of patients are successfully treated for depression every year by having magnetic fields blasted through their brains. The official name for this is Transcranial Magnetic Stimulation (TMS).

I get that treating depression with magnets sounds like crystal-healing-level BS. But in fact it’s a well-established and notoriously safe technique. It’s also notoriously noisy.

More importantly, it’s the only FDA-approved, fully noninvasive treatment for major depression.1 The saga of how it got there is worth studying for anyone interested in neurotechnology.

Super-quick Basics

TMS uses electromagnetic induction to fire neurons on the cortex noninvasively from outside the skull.

img_1.png (Source)

Different TMS coil shapes produce different electric fields in the brain, as you can see here:

coils.png (Source)

These days most clinics and researchers use a figure-eight coil, which has a focus in the center of the eight.

Single pulses of TMS can be used to fire neurons at a precise time for neuroscience experiments or diagnostics, but repetitive-TMS therapy (rTMS, though often just referred to as TMS) is FDA-approved to treat depression, OCD, smoking addiction, migraines, and depression-associated anxiety. In rTMS patients are hit with dozens of TMS pulses per second for 30+ minutes a day, 5+ days a week, for 5+ weeks.

There are other magnetic stimulation techniques which use lower-intensity magnetic fields and don’t actually fire action potentials like TMS does. But TMS is the most widely used magnetic stimulation method and the only one approved by the FDA to treat anything.

Early History

People have been putting their heads in magnetic coils for a long time, though not nearly as long as they’ve been running electrical currents through them.

Jacques-Arsène d’Arsonval is apparently the first person to have tried putting people’s heads in a big magnetic field. In 1896 he reported “phosphenes and vertigo, and in some persons, syncope” from the adventure. This effect was independently discovered in Germany and the UK soon after, because apparently reinvention is easier than learning to read French. (Source)

And then in 1914, research on TMS-avant-la-lettre completely died for 30 years, for reasons I don’t understand.

Despite the timing, it probably wasn’t WWI’s fault. The last paper on TMS in 1914 was published by researchers at the young and far-afield University of Washington in Seattle. And all the early TMS authors lived until the 1930s except Thompson, the British rediscoverer, who died in 1916 unrelated to the war. More likely it was just bad luck.

Though it made a couple reappearances after 1946, TMS might have stayed dead were it not (reportedly) reinvented again by Prof. Anthony Barker. Some technologies you just can’t keep down.

Barker’s big result (presented in a tiny a 500-word paper)2 came in 1985, when he figured out how make fingers twitch with painless magnetic stimulation of the head.

This was a big deal. People at the time could of course make fingers switch by stimulating the fingers, and they could make fingers twitch with painful electric current run across the skull, but painless finger twitches from the head were a big deal. Modern TMS was born.

rTMS Takes Off

Even though there were preliminary results of TMS raising mood in conjunction with electrical stimulation in 1987, for the next decade after Barker’s result single-pulse TMS was explored as a research tool and diagnostic tool, not as a therapeutic tool. But by 1991 Pascual-Leone et al. had shown that rTMS could interfere with visual perception and speech. And critically in 1994, they also found that the excitatory effects of rTMS in the cortex could last for minutes after the stimulation stopped. This reminded researchers of the (presumed) effects of ECT, an effective treatment for depression among other things. Concurrently, researchers had anecdotally noticed that subjects in vision or speech experiments who received TMS stimulation in their frontal cortex sometimes felt their mood change.

From 1994 to 1996 at least 6 positive clinical studies (not all controlled) were published on the effects of TMS in mood and depression. In hindsight a lot of these studies’ results were probably flukes or placebos, but that’s rarely stopped a field (and maybe shouldn’t).

In 1996 a big meeting of scientists was convened to lay out standards for rTMS use. Even though the only side effects of TMS that had been observed by then, and still now, were scalp pain and very rarely seizures, the nascent TMS field was probably very keen to avoid the cultural immolation that ECT had undergone in the 1960s. Prof. Mark George, a pioneer in the field, didn’t even patent the treatment due to fear it would damage its reputation. (Barker before him hadn’t patented his coil design either.)

George now thinks this lack of IP protection had long-term negative effects of reducing industry investment in TMS. Companies later on had to settle for patents on other parts of the system design. Despite modern startup wisdom that execution is everything, in the medical device world investment depends a great deal on the strength of IP protection. Maybe these days tech investors interested in medical devices would care less about IP, but the general lesson is be careful what you open source.

Despite tepid industry interest and the NIH not funding research into it, academic interest in rTMS grew quickly. Here you can see the numbers of published papers mentioning rTMS after the 1996 conference:


Most of these papers aren’t clinical trials, but the interest is the point. In fact this plot underestimates how much work was being done on rTMS, since within a few years of the 1996 meeting, clinics and doctors were offering off-label TMS therapy. Supposedly some countries including Australia, Israel, and Canada approved the treatment around the same time as well, though I haven’t found proof of this. Also around this time, commercial vendors started selling rTMS devices to researchers, whereas before researchers had been modifying single-pulse TMS systems or building their own.

All of which is a bit depressing.

Why? Because even with the explosion of research after 1996, it would still take another 12 years for the field to develop rTMS into a legitimate therapy for anything. In fact, as discussed below, the field was lucky the FDA approved rTMS in 2008 at all. Not to mention that the approval was very narrow: only for depression in adults who hadn’t responded to one, but no more than one, antidepressant medication. The field would have to wait another decade for TMS to be approved for OCD and smoking cessation, and I’d argue TMS is still under-researched and underused given its incredible safety profile.

I can accept the decades of work up to 1996 as the price of serendipity. But as someone who cares about the therapeutic potential of future neurotechnologies, the decades required to develop reliable TMS therapies demand analysis.

What took so long?

The major contributor to TMS’s slow path to approval was probably the field’s overwhelming focus on depression.

I’m not saying depression shouldn’t be worked on. It absolutely should be. I’m just pointing out that developing depression treatments is slow and hard. Clinical measures for depression (surveys like the MADRS and HAMD or patient self-reports) are fickle and noisy. Placebo response rates in depression trials are large, from 30% to 50% — that’s a lot for a treatment to overcome. And unless your goal is acute mood changes, it takes a long time to see whether a treatment is having an effect.

The focus on depression seems to have come from the hope that TMS could supersede the then-as-now-controversial ECT, and maybe because early TMS studies reported mood-altering effects. Some in the TMS field pointed out at the time that disorders associated with more focal neural activity like epilepsy or focal dystonia would be easier to treat than disorders with less-well-understood circuits like depression. TMS still hasn’t been approved for such focal disorders, so no idea whether that’s true.

One can debate whether the field should have gone after a presumably-easier-to-treat disorder like epilepsy first, gotten an FDA approval there, and built off that win to go after diseases like depression. I’m inclined to just accept that depression is an important and worthy target, take the focus on it for granted, and see what else we can learn.

Despite my disillusion at a 12+-year clinical translation timeline, one could argue that rTMS for depression was actually developed pretty quickly given how tedious clinical trials are. It took about 3-4 years to perform and publish a clinical trial on rTMS (not including time to secure funding), which amounts to only about 4 “cycles” of experimentation between 1996 and 2008.

Clinical trials aren’t slow for any good reasons. rTMS treatments of the kind patients were receiving in this era lasted around 4-6 weeks, with follow-up visits after maybe a few months. The rest of the 3-4 year timelines went into patient recruitment, IRB approval, writing and data analysis, and the publishing process (peer review plus time in press). And in fact publishing, which accounts for probably 6-12 months, doesn’t really count, since researchers share unpublished work with each other all the time. The first two are the big slowdowns, and I don’t think they’ve improved over time. They’ve probably worsened. This is infuriating. But no, there is no point to this paragraph other than venting. Back to the story.

Moving on from caveats about depression being hard and trials being slow: what is the parameter space for rTMS that the field had to explore? At minimum the following can be varied:

Plus things can get more complicated with current waveform and direction. This is a big parameter space to explore.

One thing that’s clear in reviews and meta-analyses from the time is that the easier a parameter was to vary, the more it was explored. For instance: almost every study used commercially available figure-8 coils. This isn’t surprising since changing the coil shape is really hard: you have to re-engineer the high-voltage electronics every time you do since the coil resistance changes. At the other extreme, pulse intensity and pattern are the easiest to change and were the most varied between studies. But there’s also no necessary correlation between importance of a parameter and how easy it is to vary.

Even among the easy-to-vary parameters, though, the field simplified its exploration early on in two key ways.

One was the division of pulse patterns into “high” (>1 Hz) and “low” (<=1 Hz) frequency. Early observations from 1996 suggested that high-frequency rTMS was excitatory and low-frequency rTMS was inhibitory, and that excitation and inhibition could last for minutes after stimulation. At least this is what was observed in the motor cortex, where excitation and inhibition could be measured via the impact on finger twitching. The field stuck to this dichotomy pretty strictly, and also frequently suggested that the excitatory and inhibitory effects were associated with long-term potentiation and depression. This conclusion is almost certainly oversimplified, but it seems to have been good enough.

The other early assumption that was made was that the left dorsolateral prefrontal cortex (DLPFC) should be excited and the right DLPFC should be inhibited to treat depression. This idea goes all the way back to 1995 and was inspired by lesion and functional imaging work. The left DLPFC has remained the main target of TMS for depression — it’s what the FDA approved in 2008.

Whether it was scientific skill or luck that caused the field to make these two assumptions, in hindsight they seem to have been spot on. The earliest TMS trials for depression from 1995 and 1996 used nearly exactly the same stimulation setup as would eventually be approved by the FDA in 2008. The main innovation in those 12 years was turning up the pulse intensity by ~40% and doing 4-6 weeks of treatment rather than 1. There were also big innovations in the quality of sham treatments during this time. (“Sham” is the medical device world term for placebo.)

It’s hard to know what to make of this, other than maybe wishing trials with sufficient power had been run sooner. Though Prof. George thinks running trials sooner would have been premature, pointing to what happened with DBS rushing from open-label trials to big trials that flopped.

In hindsight, two big things the field does seem to have underrated at the time were the importance of coil positioning and what the patient thinks about during the treatment.

As simple as it sounds, getting a TMS coil over the right part of the brain, orienting it correctly, and keeping it in place during a session is hard.3 The way they used to find a patient’s DLPFC was just by measuring 5 cm in front of wherever TMS made the patient’s thumb twitch. But this isn’t where the PFC is in 1/3 of patients, not to mention the DLPFC. They then started using the 10–20 system, but we know from newer results that even that is pretty inaccurate. Even with modern computers and neuronavigation systems, targeting is tricky, and for large trials one has to train technicians/clinicians to target correctly and carefully.

We also know now that what a patient does during stimulation can have a big effect on a TMS treatment. The approved TMS treatment for OCD involves a “Personalized Symptom Provocation” — in other words, the doctors make you think about something gross before they zap your brain. Between 1996 and 2008, as far as I can tell, patients could do whatever they wanted during TMS treatments: ruminate, read, sleep, etc.

What’s the opposite of a slam dunk?

In 2007, a company called Neuronetics took the big leap. They applied to the FDA for approval of their NeuroStar TMS system to treat treatment-resistant depression.

The triple-blind, multi-site clinical trial their application was based on was the largest that had been run to that point by far. It took almost 2 years, starting in January 2004, to recruit all 325 patients. The IRBs from all 23 (!) sites involved in the study had to approve the trial. (I cannot even imagine how much work…)

But the massive investment was worth it, because when the trial was finally finished it…uh…

…still failed its primary endpoint and arguably all its secondary endpoints too.

And then the FDA approved NeuroStar anyway.

If you want all the juicy details you can read the transcript of FDA deliberations on NeuroStar, Jan 2007. But here’s an artistic sketch of what happened:

Folks running the trial: “Hey, FDA, if we do rTMS on some people, and they feel better than placebo according to this MADRS survey with a p-value of < 0.05, you’ll approve our device, right?”

FDA: “No promises, but I bet we’d accept that.”

Folks: *Run study*

Folks: *Get p-value of 0.057*

Folks: “Err, hey, check out all these other surveys like the HAMD we gave the patients. According to some of them, the rTMS patients felt way better!”

FDA: “Uh, you can’t just do that. If you run enough surveys you’ll eventually find whatever you want just by random chance. And when you correct for all that double-dipping, rTMS still doesn’t beat placebo.”

Folks: …

FDA: “Actually never mind. Approved.”

Obviously I’m being flippant here. I definitely think approving TMS for depression was the right thing to do from a cost-benefit perspective. rTMS had a fantastic safety profile, and even though the totality of the evidence at the time was noisy, it seemed like rTMS was doing something positive for at least some patients. Post-market studies have borne this out: TMS works as much as most depression treatments “work”. It’s just hard to wrap one’s head around the FDA sometimes.

One thing I still don’t get about the trial is why the researchers didn’t survey patients about what treatment they thought they got in order to test the blind. I have no idea why a trial wouldn’t do this, and doing so might have let them avoid critiques like this:

“36% of the active group compared to only 4% of the sham group complained of [head] pain… When [you adjust for patients feeling pain], p-values for the treatment effect became less significant. In particular, the p-value for the primary MADRS endpoint went from borderline significant to [really bad].” — FDA

To be fair, the FDA never sounds excited about anything. Even in their deliberations on the Pfizer COVID vaccine EUA in December 2020 — the most slam-dunk clinical trial in recent memory — their average tone never exceeded “fairly into it”, and only 17 out of the 22 committee members voted to approve.

But the bottom line is that the NeuroStar approval was pretty far from a slam dunk. Plenty of folks let the FDA know it after the fact, too. And another multi-center trial using very similar stimulation parameters showing TMS had no antidepressant effect was published right after the NeuroStar trial. (Though it was for patients who were also taking antidepressant drugs, whereas the NeuroStar trial was for patients not taking drugs.)

Nevertheless, by 2010, the American Psychiatric Association (APA) had included TMS in its Practice Guidelines. Insurance seems to have taken a little while to come around, but it was mostly covered by 2015, though even today it can apparently be hard to get insurance to cover it.

Recent developments

This infographic tells the story after 2008 better than I can:


TMS has gone on to be approved by the FDA for migraines, OCD, smoking cessation, and depression-associated anxiety. And TMS is regularly used off-label for a lot of things.

There’s no way in the near future for TMS to become miniaturized or a consumer device. The high-voltage, quick-discharging power systems it requires are cabinet-sized and not conceivably shrinkable. And Maxwell’s equations mean TMS faces a fundamental tradeoff between depth and focality. The outermost parts of the cortex will always be stimulated more strongly than deeper brain structures, and the deeper one tries to stimulate, the more diffuse the stimulation is.

But TMS still has lots of room for growth. The more we can learn from neuroimaging about the cortical circuits involved in neuropsychiatric diseases, the more TMS protocols can be developed to target them. And personalized targeting is showing impressive results that might hugely reduce the treatment time and increase efficacy. If the treatment time can be brought down to just a few days, it’d be hard to argue that TMS shouldn’t become a first-line therapy for depression. It’s far safer than any antidepressant drug — the cost and time are what make it infeasible now.

Let’s just hope we can start measuring this growth in years, not decades.

Have feedback? Find a mistake? Please let me know!

  1. Pharmaceuticals and (anesthetized) ECT don’t count as fully noninvasive since you’re putting chemicals into your body for both. ↩︎

  2. Actually 500 words and a picture of the device, which the Lancet told Barker he should include “if it wasn’t too dull." ↩︎

  3. Though at least TMS has undeniable neurological effects (e.g. making fingers twitch) that can be used to calibrate it! (Usually the strength of TMS stimulation used on a patient is proportional to how much current is required to make their thumb twitch.) Compare that to most tES techniques, where you don’t have any idea where the current is going. ↩︎