ML for Brain Physiology

Pavel Rudych28 Oct 2023

Problem

A small team in Novosibirsk works with doctors on EEG, fMRI and balance data. In a small remote clinic only a nurse may watch a patient, so fatigue goes unnoticed and exercises go wrong. After the two-week insurance-funded course patients regress at home. Diagnosis from functional MRI and EEG lacks the mature models that anatomical MRI has, and doctors must learn to trust model hints. Recordings are too few to train a network on raw signals.

Example from the article: stroke patients finish two weeks of biofeedback training, go home with a 1000-ruble stabilometer made from scales and an Arduino, connect it to a website, and the doctor follows progress there and in WhatsApp.

Solution in brief

The talk applies the standard ML project cycle to physiological signals. Clean the data conservatively, reduce it to features, train a simple model, and validate with a placebo group or an independent metric, so the model does not "recognize a wolf in the forest" instead of physiology.

Questions this case answers
  • How to train a model on very small recordings of physiological signals without fitting artifacts, and verify that it reflects the physiology
  • How to deliver a low-cost home monitoring device for rehabilitation with remote supervision
  • How to automate analysis of EEG and fMRI signals for rehabilitation monitoring and diagnostic support
  • How to clean noisy EEG and fMRI recordings before training a model on a small dataset

Tools and methods

  • Independent component analysis (ICA) — removes muscle, heart, blink, mains and scanner artifacts from EEG
  • Spectral power by rhythm — main EEG features (theta, alpha, beta)
  • Stabilometry metrics — balance features, e.g. energy spent on keeping balance
  • SVM, CatBoost, TabNet — models on tabular features
  • ResNet, OpenPose — posture from video
  • Wav2Vec2 — transformer for audio and EEG streams
  • Placebo-controlled biofeedback — checks that an EEG pattern matters clinically
Continue reading — it's free

Google or a one-time email code. No password.