Deep Brain Stimulation Has Always Been Closed-Loop
Avoiding inefficient and ineffective engineering
Deep Brain Stimulation (DBS) uses implanted electrodes to deliver electrical currents to brain regions — and treats a surprisingly diverse set of neurological and psychiatric illnesses.
It’s achieved its surprising efficacy while being clinician controlled. That is, a medical doctor comes in and adjusts the stimulation parameters based on their multidimensional, subjective clinical judgement of the patient.
So, when DBS works, that’s how it’s done. If we’re trying to study, or improve, therapy, we should be very careful that we’re implementing it as-is.
Contemporary efforts in DBS often say they’re improving DBS to be “closed-loop” — implying that DBS is currently open-loop.
This is problematic for a lot of reasons — and we’re seeing this misconception even from top research labs. That’s very concerning.
Using Control Theory, we can see why and reset our understanding of the loop.
Outline
- Defining Closed-Loop
- Examples of Loops
- Why It’s Problematic
Defining Closed-Loop
When we influence or control a system, we’re generally trying to evoke in it a particular state. We do this by influencing the system some how, and then observing its state to see if it did what we thought it would.
When we observe, we’re observing some part and shadow of the system. Specifically, the system’s state: x.
When we influence, we do so through a control signal u — our instructions, designed to achieve some target state B when we start at A.
The “loop” is between us, the system, and back to us. We observe, we influence, then we observe — iterate.
This whole loop can be “open” or “closed” — and both have a specific definition in Control Theory.
Open Loop is when our u depends only on time. We never check in on the system’s state x.
Closed-Loop is when your control signal depends, in some way, on the state itself. Often, time doesn’t explicitly make its way into the control, just through the state itself.
Examples of Loops
Let’s look at three explicit examples of control loops.
Antibiotics — Open Loop
The best example of a truly open-loop control policy is antibiotics.
When your doctor thinks (or even suspects) that the underlying cause of your symptoms is a bacterial infection, she’ll prescribe an antibiotic.
When she does, she also tells you the dosing to take and for how long — sometimes the dosing even changes after a few days.
But this dosing depends only on time - not on anything even remotely related to the actual state of your body and/or infectious agent. No re-assessment of the symptoms, no attempt to estimate how much bacteria is still there.
These things only happen if you, down the road, assess your state to be unimproved. Until you assess your state, the treatment administered by the physician is open loop.
Blood Thinning — Closed-Loop
A good example of a closed-loop control policy is blood thinning with warfarin.
Warfarin thins the blood in a dose-dependent way — if the blood is too thin then we risk too much bleeding with normal activity.
The INR reflects the level of thinning, and we have a “target INR” that we’re trying to adjust the Warfarin dose to evoke. Since we come in and measure the INR alongside changes to Warfain, this is a closed-loop system.
Importantly, since we’re closing the loop with a direct physiologic measure via a blood test, this is a particular type of closed-loop.
Thermostat — Two Types of Closed
A thermostat is a common example of a closed-loop controller — the heating/cooling in your house constantly assesses the temperature of a space and then adjusts how much cool/hot air is pumped into the space to hit a target temperature.
But you can do this less precisely, to the same ultimate effect, if you just assess the comfort level of everyone in the room. If most people are too cold, then adjust the target temperature up. If most people are too hot, then adjust the target temperature down.
This is an important alternative because it’s closest to the contemporary DBS situation.
Why It’s Problematic
DBS Was Always Closed-Loop
Understanding the definition of closed-loop and seeing some examples of open loop should make it clear: DBS has always been adjusted based on state. Just not a direct measurement of state.
While the feedback used to tune DBS heretofore hasn’t been derived from physiologic state doesn’t mean the feedback isn’t reflective of the state — the clinical assessment is often done from symptoms, and symptoms are a mapping downstream of the physiologic state.
An actually open-loop DBS would be a physician setting the stimulation parameters, deciding on a time-based schedule for parameter changes, and then not assessing your symptoms ever again. In other words, your body’s underlying state never affects the decision making.
Knowing all this, we see that it is, indeed, closed-loop.
Reinventing Wheels
One obvious problem in not realizing DBS is already closed-loop is thinking that contemporary efforts are novel and without precedent. Proper credit aside, this means we deal with all the inefficiencies of reinventing the wheel.
We’ll find ourselves spending a lot of energy designing a part of a wheel that we could have easily just imported quickly. That’s wasted time and effort — at best.
Maybe we need to reinvent the wheel, but we need a targeted reason for not going with wheels that are already in motion.
That’s not what we’re seeing in the DBS space.
Starting from Scratch
Reinventing the wheel is one thing, but trying to do so while failing to realize wheels already exist moves beyond inefficiency into inefficacy.
If we’re trying to build a controller, without realizing the therapy that we want to study already has a controller inside it, then we risk something worse than inefficiency — we risk studying a strawman.
This is where thinking clinician-controlled DBS = open-loop DBS lives. It means we don’t understand the therapy’s structure right now, dooming any experimental design built downstream.
Trying to build a controller from scratch in such high dimensional spaces — think of the degrees of freedom in the brain, patient, device, stimulation waveform — is a great way to get nowhere fast.
Researchers have to build from the therapy as is — that is the nature of reverse engineering. If we don’t, we find ourselves innovating on and studying something that no one actually does.
One way to fix this problem is to see the clinician as a controller and characterizing how they make their decisions. That will give DBS the right foundation for improving our current control strategies, potentially in the direction of adaptive and automation.
Parting Thoughts
In this post, we re-affirmed the definition of “closed-loop” to that already established and useful in Control Theory. We see then that DBS has always been closed-loop — so we should be careful about ignoring the controller already a part of successful therapy.
I tend to think we should work on adaptive DBS by first characterizing clinicians as controllers; to figure out what it is they do, right now. Because whatever they are doing right now, qualms and all, achieves therapeutic response. Or at the very least, doesn’t preclude it.
I prefer the term adaptive DBS when talking about closing the DBS loop with physiologic metrics instead of behavioral ones — especially when we make it clear that we’re shifting clinical signals out for physiologic signals. More broadly, we should more explicitly rebuild DBS research on a foundation of Control Theory to avoid inefficient, ineffective, and potentially unethical efforts.