Neuralab
L5 pyramidal neuron, mouse cortex, rendered in ASCIIt = 0.0 ms

Neuralab models each patient's brain circuit to decide where and how to stimulate it.

People with Parkinson's disease, tremor, epilepsy and depression already live with stimulators in or on their brains. We are building the AI that turns a model of each patient's circuit into stimulation settings a clinician can use, and then keeps those settings tuned from what the device records.

Deep brain stimulation works. It is also programmed largely by hand. A clinician picks a contact, an amplitude, a frequency and a pulse width, watches the patient, adjusts, and asks them to come back in a few weeks. Finding good settings can take many visits, and the settings rarely change again as the disease does.

The devices have changed faster than the practice. Sensing-enabled stimulators approved for Parkinson's disease now record local field potentials from the same leads they stimulate through, and adaptive systems can raise or lower stimulation in response to a signal. That creates a stream of circuit data from every implanted patient, around the clock. Very little of it is being used to decide what the device should do.

Today's tools each look at one part of the problem. Imaging shows where the lead sits. Group-average connectomes show which pathways a contact tends to reach. A single biomarker, such as beta-band power, drives a single threshold. We think these belong in one model of one patient's circuit, built around what is known about the basal ganglia and its loops, so that a proposed setting comes with a reason as well as a prediction.

Inputs, per patient
  • Lead locationimaging
  • Connectivitytractography
  • RecordingsLFP from the device
  • Symptomsscores, wearables
  • Medicationdose and timing
Model
Circuit model of the patient
basal ganglia–thalamocortical prior
Trained to fill in missing tracks, so a patient with only some data still gets a prediction, with uncertainty.
Output, on approved hardware
  • Contactwhich electrode
  • AmplitudemA or V
  • FrequencyHz
  • Pulse widthµs
  • Adaptive rulewhen to go up or down
Device recordings and symptom changes flow back into the model after every adjustment.
Every recommendation is a setting the device already supports. A clinician reviews and applies it.

The model treats each patient as a record with parallel tracks: where the lead is, what it connects to, what the circuit is doing, how the patient is doing and what medication they are on. No patient has every track filled in. The model is trained by hiding parts of these records and learning to restore them, so it can be asked a direct question: given what we know about this patient, which setting is most likely to reduce their symptoms with the fewest side effects, and how sure are we?

The output is something anyone can check. A recommended setting runs on the same approved device in any clinic, and its effect shows up in symptom scores, side effects, battery use and the number of programming visits. That makes the work reproducible in the way a designed molecule is reproducible. You can run it and measure what happened.

This is a design problem and a measurement problem at once. The model proposes, the device measures, and the next proposal is better informed. Neither half works without the other.

How we will test it

  1. Retrospective. From the data available at a patient's first programming visit, predict the settings their clinician eventually chose and how they responded. This can start on published cohorts and open datasets.
  2. Prospective pilot. Compare model-suggested starting settings with standard programming in patients who already have sensing devices. Measure visits needed and symptom scores.
  3. Adaptive. Propose adaptive rules that use more of the recorded signal than a single frequency band, and test them on devices already in use.

We have not run these studies yet. We will publish what we find, including where the model does worse than a clinician.

Where we start

First
Parkinson's disease

The largest DBS population, the best-understood circuit, and a recorded biomarker already in clinical use.

Next
Essential tremor

A well-defined circuit and a symptom that can be measured continuously.

Later
OCD, depression, epilepsy

Indications where data are scarce and circuit knowledge matters most. Noninvasive methods such as TMS and focused ultrasound follow the same logic.

Neuralab begins as decision-support software for clinicians. A system that changes settings on its own is a regulated medical device, and we will treat it as one.

This is a long project with many unknowns, and it depends on working closely with clinical teams. If you run a DBS program, build stimulation hardware, or want to spend years on this problem, get in touch.

Recent research

All work →