New ECG signal derivation reconstruction from ppg using deep learning network AI approach

Introduction

Non-invasive ECG recordings are still required in the area of atrial fibrillation detection . PPG signal via smart watch technologies have been developed therefore many automated diagnoses are deemed inconclusive, despite yielding a readable single lead electrocardiogram tracing.

Objective

Based on the existing relationship between electrical ECG signal and PPG signal measured using light sensors, we sought to set up a new signal transformation method based on artificial intelligence (AI) algorithms in order to elaborate a reproducible DII ECG derivation.

Method

Our work was two fold: (1) First, based on data (PPG and ECG collected simultaneously) from thousands of patients with and without cardiac abnormalities, which have been pooled in databases (MIMIC and Vital DB). Pre-processing steps ranging from filtering to slicing of these millions of coupled PPG-ECG data segments preceded the training processes of the Deep Learning Network AI (DLN AI) models for signal transformation; (2) Second, we compared the lead II ECG reference with the transformed signal from PPG-AI via the sensitivity, specificity, the positive and negative predicted values and intra-observer variability based on 11 cardiologists expertise. For each unique signal be it reference or transformed signal, we collected all classifications, confidence scores, and quality ratings. The reference signals was based on a consensus of electrophysiologists expert. We tested the average confidence score and average quality rating from the confidence scores and quality ratings of all the experts who labelled the signal. Results (1). We tested the new DLN AI model in a large series of ECG datasets via MIMIC III subset, VitalIDB, PulseDB, and MIMIC III subset vs. VitalIDB together for a sequence train and test. The performance was tested on RR plus interpolation, RR plus padding, 2 seconds and 4 seconds respectively for MIMIC III subset (p=0,92; rRMSE=0.08), (p=0.94; rRMSE=0.16), (p=0.84; rRMSE=0.10), (p=0.86; rRMSE=0.09); for VitalDB (p=0,93; rRMSE=0.09), (p=0.91; rRMSE=0.10), (p=0.88; rRMSE=0.11), (p=0.89; rRMSE=0.10); for pulsed DB (p=0,91; rRMSE=0.10), (p=0.92; rRMSE=0.09), (p=0.84; rRMSE=0.13), (p=0.88; rRMSE=0.11). Mixed performance was also tested via Mimic III for train and vitalDB subset for testing respectively: (p=0,93; rRMSE=0.09), (p=0.82; rRMSE=0.7), (p=0.76; rRMSE=0.20), p=0.75; rRMSE=0.18) and via vitalDB for train and Mimic III for test respectively: (p=0,83; rRMSE=0.17), (p=0.69; rRMSE=0.19), (p=0.69; rRMSE=0.19), (p=0.68; rRMSE=0.19).[EV1] [LO2] [LO3] The correlation values of PPG and ECG derivation reconstruction varied from 0.69 to 0.94 with a strong positive correlation and rRMSE values closer to 0.

Results (2). Figures 1 and 2.

Conclusion

This preliminary study found that reconstructing the PPG signals into an ECG derivation using algorithms based on an AI model were well correlated to the true ECG derivation. The overall accuracy is close to 90%.