Real-Time PPG-to-ECG Reconstruction Model With On-Device Recalibration Facility

Reconstruction of physiological signals can be useful for biomedical applications from two perspectives, i.e., fewer sensor attachments for wearable health applications and lost data recovery in case of signal corruption in one channel. One important aspect to consider is periodic recalibration for continuous usage, rarely addressed in reported works. This article describes photoplethysmogram (PPG) to electrocardiogram (ECG) reconstruction model (P2ERM) using a data-driven model developed using deep learning (DL) tools. It employs a feature translation from PPG to ECG using a novel memory-based autoencoder, proposed for the first time, that tunes a multilayer perceptron neural net (MLPNN) for PPG-to-ECG feature translation. The final stage uses a long short-term memory (LSTM) for the feature to raw sample prediction, with heart rate (HR) adjustment. The model was evaluated with 142 data records (42-h duration) for the MIMIC-II/III waveform dataset. The average MSE and MAE for ECG sample prediction were found as
$1.28\times 10^{-3}$
and
$17.81\times 10^{-3}$
mV, respectively. The average deviation of ten clinical features in the predicted ECG was around 4.97%. A notable feature of the P2ERM is on-device recalibration (in situ), which can provide improved results in case the PPG morphology varies with time. An end-to-end hardware implementation using an ARM v6 controller resulted in energy savings even with a hardware ECG sensor. The latency for a single ECG beat prediction was found as 216.05 ms. P2ERM also provides superior performance over reported works on PPG-to-ECG signal reconstruction and can be useful for rehabilitation monitoring applications.