EMG signal processing for bruxism detection

Other Hardware Sensors
1

Hi everyone.
I've been playing with and Arduino Uno R4 WiFi and the OLIMEX-EKG-EMG shield.
I would appreciate some guidance in detecting the muscle signal properly.

Mods please feel free to relocate this post if this is the wrong category.

What I did:

  • FFT of the analog signal input using the ArduinoFFT library
  • Filtered out frequencies 0 to 60Hz
  • Sent the data over to Processing for DSP and logging

What I can't seem to accomplish:

  • Detect the jaw clenching reliably (current detection isn't continuous and it's prone to false positives, must be re-done from scratch)
  • Setting beeps and alarms to stop clenching, this is more difficult than expected because of the unreliable detection

Must mention:

  • I'm reusing the electrodes, the gel might not be fresh and that can become a problem soon.

What is the recommended approach to detect the signal in the FFT?
We're dealing with nonlinear signals, so I don't really know what I'm aiming for

Spectrogram of mostly resting jaw
Spectrogram of mostly resting jaw

Clenching and holding a couple times
Clenching and holding a couple times

Spectrogram including the frequencies 0 to 60Hz + some clenching
Spectrogram including the frequencies 0 to 60Hz + some clenching

The following snippet illustrates how the vReal float array (from the examples of ArduinoFFT) is populated before sending it to Processing to draw the spectrogram.

float vReal[samples]; // We're working with this array, of size samples/2

ArduinoFFT<float> FFT = ArduinoFFT<float>(vReal, vImag, samples, samplingFrequency);

. . . . . .


void collect_samples() {
  for (uint16_t i = 0; i < samples; i++) {
    vReal[i] = analogRead(analog_pin);
    vImag[i] = 0.0;
    delayMicroseconds(1000000 / samplingFrequency);
  }
}

. . . . . .

void loop() {
  collect_samples();

  FFT.windowing(FFTWindow::Hamming, FFTDirection::Forward);
  FFT.compute(FFTDirection::Forward);
  FFT.complexToMagnitude();

  // Erasing all bins up to 60Hz (the shield has a filter)
  uint8_t i = 255;
  while (freq_bin * (++i) <= 60)
    vReal[i] = 0;

  send_to_udp();
  
 }

The full code is a mess, partially done with GPT too. Needs to be rewritten as mentioned. I can still post it if needed.

(This post is also on reddit)

2

The FFT result is simply a different representation of the same signal, but in the frequency domain.

What are you expecting to see, and how does the FFT help?

3

Thank you for responding. The FFT is supposed to show when the muscle is activated (the signal is mostly concentrated between 50 and 150Hz).

What I am missing is a way to detect it properly, continued clenching being the hardest case because the signal seems to subside.. In the second gif I held the jaw contracted for while, you can see the signal is subtle in that phase
image

Might be relatively easy to detect, I'm new to this kind of processing.

4

I doubt that, but it will be a good learning experience. Have fun!

1 Like
5

SVM mostly did the trick. It's not perfect, but it was relatively easy indeed.

6

This project is now public on GitHub

7

SVM like in Support Vector Machine ?

8

Exactly. Have a look, this runs on the Uno R4

static const float weights[] = {
  0.00000000,0.00000000,0.00000000,0.00000000,0.00000000,-0.16666446,0.14036143,0.10275731,0.15018884,0.17533556,-0.01297485,0.29788694,-0.10588868,0.16164895,0.07309278,-0.06163993,0.72166246,-0.05186196,0.05293969,0.17246698,-0.09546843,0.25317097,-0.05021467,0.20567502,0.12985801,0.05549219,0.18976886,-0.10753195,-0.29549880,0.30059777,0.24690098,0.00517570
};
static const float bias = -0.2237529517562951;
static const int weight_length = (sizeof(weights) / sizeof(float));

static_assert(samples / 2 == weight_length, "Error: Weights are not as many as samples/2");

int classify(float input[]) {
  float sum;
  arm_dot_prod_f32(input, weights, weight_length, &sum);  // SIMD-optimized dot product
  sum += bias;

  return sum >= 55 ? 1 : 0;  // Classification threshold
}
9

That's cool :slight_smile: I used SVM for ECG/HRV classifications a decade ago, was quite a nice thing to do.

10

Can this be used to detect snoring or heart rate in some way? I don't expect to do so with only one emg sensor, but at least snoring could be detected using a microphone