Programmable stimulation waveformsmy own work

Most brain stimulators produce the waveform their output stage naturally produces. In TMS it is the damped ringdown of a capacitor into a coil; in implanted stimulators, a rectangular current pulse with a charge-recovery phase added afterwards. In neither case was the shape chosen because neurons respond to it best. It is what the circuit does.

This is my own work, and the reason mine exists as a field on this site.

Why the shape was never a choice

A conventional TMS device is, electrically, a capacitor discharging through an inductor. That fixes the pulse to a damped sinusoid whose width follows from $\sqrt{LC}$, and fixes the energy per pulse at

$$E = \tfrac{1}{2} C V^2$$

almost all of which ends up as heat in the coil. Coil temperature, not physiology, is what limits how long a session can run — so pulse shape, repetition rate, and session length are not independent knobs but a single thermal compromise.

The consequence for neuroscience is that a large part of the stimulation parameter space has gone unexplored for reasons that are purely about power electronics. Questions like which pulse shape recruits which neural population have been hard to ask, not because they are uninteresting, but because no device could vary the shape.

Synthesising the pulse instead

A modular multilevel converter changes the situation: rather than one capacitor discharging, many modules switch in sequence to approximate a commanded waveform. Ours runs at 2.8 kV and 6 kA, which is enough to reach stimulation thresholds while still following an arbitrary command — closer to a digital-to-analog converter operating at a few megawatts than to a capacitor discharge. Pulse shape, phase timing, inter-pulse interval, and burst structure all become parameters rather than constants.

Two problems had to be solved to make that usable.

Getting monophasic pulses cheaply enough to repeat. Monophasic pulses are what neuroscientists generally want, and also the ones that heat the coil fastest, which is why repetitive protocols have largely had to use biphasic pulses instead. Giving a sinusoidal current a long, low pre-phase in the direction opposite to its short main phase produces an equivalent electric field in the brain while shifting the current waveform down by nearly half. Because loss goes as the square of RMS current, that removes roughly three quarters of the coil loss, which is what makes rapid-rate monophasic stimulation practical at all.

Quantizing the command well without switching constantly. A modular converter must repeatedly choose which discrete output level to switch to next. Nearest-level modulation switches rarely but quantizes coarsely; carrier-based PWM quantizes finely but switches hard. Feeding the nearest-level residue through a low-pass filter before a one-bit PWM stage keeps most of the accuracy at most of the low switching rate — the filter matters because one-bit PWM only helps when its input bandwidth is constrained, and without it the scheme collapses back into level-shifted-carrier PWM.

What it is for

The point is not the converter. It is that once shape and timing are free parameters, closed-loop stimulation becomes a design problem rather than an aspiration: you can adjust the stimulus in response to what the brain just did, which is the natural thing to want when the system you are controlling adapts back.

That work is ongoing, and this entry will grow as it does.