System and method for continuous atrial fibrillation detection via PPG to ECG signal translation
A system and a method are disclosed for AFib detection using ECG signals generated from monitored PPG signals. A method includes receiving PPG signals of a user measured by a PPG sensor; translating the measured PPG signals into ECG signals using a dynamic model; analyzing the translated ECG signals using an AFib detection model, which is trained on measured ECG signals for AFib detection; and providing the analyzed AFib detection results to the user.
1 . A method for atrial fibrillation (AFib) detection, the method comprising:
receiving photoplethysmogram (PPG) signals of a user measured by a PPG sensor;
translating the measured PPG signals into electrocardiogram (ECG) signals using a dynamic model;
analyzing the translated ECG signals using an AFib detection model, which is trained on measured ECG signals for AFib detection, wherein analyzing the translated ECG signals comprises replacing a translated ECG signal corresponding to a time t with a measured ECG signal; and
providing the analyzed AFib detection results to the user.
2 . The method of claim 1 , wherein analyzing the translated ECG signals using the AFib detection model further comprises:
determining that the measured ECG signal is available for the time t;
replacing the translated ECG signal corresponding to the time t with the measured ECG signal; and
analyzing the measured ECG signal and the translated ECG signals, except for the translated ECG signal replaced by the measured ECG signal, using the AFib detection model.
3 . The method of claim 1 , wherein analyzing the translated ECG signals using the AFib detection model further comprises:
fusing the translated ECG signal at the time t with the measured ECG available for the time t, wherein the fused ECG signal at the time t is the same as the translated ECG signal, if a measured ECG is not available at the time t;
retraining the AFib detection model on the fused ECG signals; and
analyzing the fused ECG signals using the retrained AFib detection model.
4 . The method of claim 1 , wherein generating the dynamic model comprises combining a sequential deep generative model with a state-space model augmented by an attention mechanism.
5 . The method of claim 4 , wherein the state-space model is generated using an ECG generative decoding process and a latent state inference posterior encoding process.
6 . The method of claim 5 , wherein the ECG generative decoding process includes determining ECG segments based on PPG segments and latent states.
7 . The method of claim 6 , wherein an ECG segment is defined by a time elapsed between two successive R peaks on an ECG signal.
8 . The method of claim 6 , wherein a PPG segment is defined by a time elapsed between two successive systolic peaks on a PPG signal.
9 . The method of claim 5 , wherein the latent state inference posterior encoding process includes determining latent states corresponding to ECG segments based on the ECG segments and recurrent states.
10 . The method of claim 4 , wherein the attention mechanism includes attention weights defining a quantity of each source segment to be considered for each output interval.
11 . A system for atrial fibrillation (AFib) detection, the system comprising:
a photoplethysmogram (PPG) sensor; and
a processor configured to:
receive PPG signals of a user measured by the PPG sensor,
translate the measured PPG signals into electrocardiogram (ECG) signals using a dynamic model,
analyze the translated ECG signals using an AFib detection model, which is trained on measured ECG signals for AFib detection, wherein analyzing the translated ECG signals comprises replacing a translated ECG signal corresponding to a time t with a measured ECG signal, and
provide the analyzed AFib detection results to the user.
12 . The system of claim 11 , wherein the processor is further configured to analyze the translated ECG signals using the AFib detection model by:
determining that the measured ECG signal is available for the time t,
replacing the translated ECG signal corresponding to the time t with the measured ECG signal, and
analyzing the measured ECG signal and the translated ECG signals, except for the translated ECG signal replaced by the measured ECG signal, using the AFib detection model.
13 . The system of claim 11 , wherein the processor is further configured to analyze the translated ECG signals using the AFib detection model by:
fusing the translated ECG signal at the time t with the measured ECG available for the time t, wherein the fused ECG signal at the time t is the same as the translated ECG signal, if a measured ECG is not available at the time t;
retraining the AFib detection model on the fused ECG signals, and
analyzing the fused ECG signals using the retrained AFib detection model.
14 . The system of claim 11 , wherein the processor is further configured to generate the dynamic model by combining a sequential deep generative model with a state-space model augmented by an attention mechanism.
15 . The system of claim 14 , wherein the processor is further configured to generate the state-space model using an ECG generative decoding process and a latent state inference posterior encoding process.
16 . The system of claim 15 , wherein the ECG generative decoding process includes determining ECG segments based on PPG segments and latent states.
17 . The system of claim 16 , wherein an ECG segment is defined by a time elapsed between two successive R peaks on an ECG signal.
18 . The system of claim 16 , wherein a PPG segment is defined by a time elapsed between two successive systolic peaks on a PPG signal.
19 . The system of claim 15 , wherein the latent state inference posterior encoding process includes determining latent states corresponding to ECG segments based on the ECG segments and recurrent states.
20 . The system of claim 14 , wherein the attention mechanism includes attention weights defining a quantity of each source segment to be considered for each output interval.