Thesis lab

Rerun the study in your browser.

Follow the same workflow as Dr. Hamidi's PhD research, step by step: record a wearable session, turn motion into features, train a classifier, confirm with the ECG and clean the signal. Every run generates a new simulated volunteer.

The published research

Simulated data, generated in your browser. The features, windows and clean-up model are simplified stand-ins for the published methods, not the trained models or their settings. Scores vary from run to run.

Step 1 · Record

One chest device, five activities

As in the study protocol, a chest-worn ECG monitor with a built-in 3-axis accelerometer records a volunteer sitting, walking, climbing stairs, driving and clapping. Each stretch is labelled with the study's codes: 1 walking, 2 stairs, 4 driving, 77 clapping (the motion that corrupts the ECG) and 99 other. The accelerometer samples at 104 Hz and the ECG at 125 Hz.

One simulated session. Red band: clapping, where the ECG fills with motion noise. Sometimes the volunteer also waves their arms while sitting or driving, which looks like clapping to the accelerometer but leaves the ECG clean.

Step 2 · Describe

Motion becomes a fingerprint

The motion signal is cut into short windows. Each window is described by how much energy it holds at each time scale, a multi-scale wavelet description. The thesis uses the wavelet scattering transform, which is stable when movements shift slightly in time; this lab uses a simplified version of the same idea.

Each column is one window, each row one feature (red above average, teal below). The strip underneath shows each window's true activity; clapping windows stand out.

Step 3 · Classify and confirm

Which windows will corrupt the ECG?

A fine k-nearest-neighbour classifier (one neighbour, Euclidean distance) learns from 75% of the windows and is tested on the other 25%. Switch on the ECG confirmation to see how adding the heart signal itself removes false alarms from arm movements that are not clapping.

Accelerometer onlyFNC 2023 Accelerometer + ECGICAIMA 2023

This run

–test accuracy

Artifact class

–precision (PPV)

Artifact class

–true positive rate

Step 4 · Clean

Rebuild only what was damaged

A flagged clapping stretch goes through two steps. Step A: a biorthogonal (Haar) wavelet denoiser with a universal threshold calms the large motion swings. Step B: the thesis uses a trained denoising autoencoder to rebuild the waveform; here a simple beat-template model stands in for it.

Because this is a simulation, the true clean signal is known, so the reconstruction can be scored directly.

ECG RMS	This run	Published (ICAAI 2023)

This run

–SNR gain vs the true signal

This run

–error vs the true signal

From a lab exercise to your business

The same four steps (record, describe, classify, clean) sit behind most practical AI projects, from predictive maintenance to fraud checks. Halitech Academy teaches leaders how to run them well.

AI training for your team Live wearable demo Talk to us