Health AI research
Movement corrupts the heart signal a wearable records. Dr. Hamidi's PhD research built a complete answer in three steps: find the corrupted moments, confirm them, and rebuild the clean signal.
Illustration only, not patient data.
The problem
A chest patch records clean ECG while the wearer is still. Walking, climbing stairs or moving the arms shifts the electrodes against the skin, and the recording fills with noise that can hide real heart events or trigger false alarms. Filtering everything all the time is no answer either: it throws away good signal along with the bad.
How it works
Step 1 · Detect · FNC 2023
94.7%test accuracy · 93% precision on artifact moments
Using only the device's 3-axis accelerometer, wavelet scattering features and a fine k-nearest-neighbour classifier recognise the activity and flag the moments likely to corrupt the ECG. No extra sensors are needed.
Step 2 · Confirm · ICAIMA 2023
98.8%test accuracy · AUC 0.995 for artifact moments
Feeding the ECG together with the accelerometer into the same approach makes detection sharper, so a monitor can trust its own flags.
Step 3 · Clean · ICAAI 2023
+12.64%SNR improvement · RMSE 0.0058
Flagged stretches pass through a biorthogonal wavelet filter, then a denoising autoencoder (a deep-learning model) rebuilds the ECG while keeping the shape of each heartbeat.
The evidence
Signal strength (RMS) of the recorded ECG. Movement pushed it from 306 to 442 microvolts; after the two-step clean-up it returns to 313, within a few percent of the clean recording.
| Method | RMSE | SNR improvement |
|---|---|---|
| Prior study A | 0.037 | 6.49% |
| Prior study B | 0.0189 | 12.57% |
| Prior study C | 0.063 | 9.55% |
| Hamidi et al., two-step method | 0.0058 | 12.64% |
Data: a chest-worn ECG monitor with a built-in accelerometer, recorded from one healthy adult volunteer while walking, climbing stairs, driving and clapping (about 309,000 motion samples). Results are from that controlled dataset; testing on larger and clinical groups is the next step.
Published work
With co-authors Bill Robertson and Jacek Ilow, Dalhousie University.
FNC 2023 · Halifax · Open access
18th International Conference on Future Networks and Communications (FNC). Procedia Computer Science, vol. 224, pp. 60–67 · doi:10.1016/j.procs.2023.09.011
ICAIMA 2023 · July 14, 2023 · Scopus-indexed
International Conference on Artificial Intelligence in Medical Applications (ICAIMA)
ICAAI 2023 · October 10, 2023
International Conference on Advance Artificial Intelligence (ICAAI 2023). International Journal of Engineering Research in Computer Science and Engineering (IJERCSE), ISSN 2394-2320
PhD thesis · Dalhousie University · 2024
Research stage. Halitech's ECG methods are not a medical device and have not been approved or cleared by Health Canada, the FDA or any other regulator. Nothing on this site is medical advice.
From the lab to your company
Most AI projects fail for the reasons this research had to solve: unclear goals, messy data and models that look good in a demo but not in daily use. The habits behind these papers are the ones Halitech brings to companies adopting AI.
Not "clean the ECG" but "find the seconds when movement corrupts it". For you: pick one decision or task where AI can measurably help, before buying any tool.
The method works with the accelerometer already inside the device. For you: start with the data your systems already collect; most companies have more than they think.
Every activity was labelled and timed. For you: a short data audit shows whether your records can support AI, and what to fix first.
A fine-KNN classifier beat heavier approaches on this data, and deep learning was used only where it earned its place. For you: right-sized tools, lower cost, fewer surprises.
Accuracy, precision and false alarms are reported separately, with the dataset's limits stated. For you: pilots with clear success measures before scaling.
The method targets wearables with limited power and no extra sensors. For you: plans that fit your people, budget and systems, not a lab.