US-12622625-B2 - Machine learning techniques for electrocardiogram (ECG) analysis

US12622625B2US 12622625 B2US12622625 B2US 12622625B2US-12622625-B2

Abstract

Described herein are techniques for analyzing at least one electrocardiogram (ECG) signal. In some embodiments, the techniques include: receiving at least one ECG signal; encoding the at least one ECG signal using the encoder to obtain a numeric encoding of the at least one ECG signal; and processing the numeric encoding of the at least one ECG signal using at least one trained machine learning model to obtain: (i) at least one denoised ECG signal corresponding to the at least one ECG signal, and/or (ii) characteristics of the at least one ECG signal, the characteristics comprising: (i) rhythm types including a respective rhythm type for each of at least some segments of the at least one ECG signal, and/or (ii) sample-level ECG labels including a respective sample-level ECG label for each of at least some of the plurality of samples.

Inventors

John Paul Duffy
Michael Feist
Esmatullah Naikyar

Assignees

NEURALCLOUD SOLUTIONS INC.

Dates

Publication Date: 20260512
Application Date: 20250815

Claims (10)

1 . A Holter monitor, comprising: electrocardiogram (ECG) leads comprising at least three ECG leads; at least one processor; and at least one non-transitory computer-readable storage medium storing: at least one trained machine learning model; a residual vector quantizer; and processor-executable instructions that, when executed by the at least one processor, cause the at least one processor to perform a method comprising: measuring at least one ECG signal using the ECG leads at a sampling rate of between 128 Hz and 256 Hz and for a duration of between one and fourteen days, the at least one ECG signal comprising a plurality of segments and a plurality of samples, wherein a sample is an ECG measurement at a single time point, and wherein a segment comprises at least some samples of the plurality of samples; encoding the at least one ECG signal to obtain a numeric encoding of the at least one ECG signal; quantizing the numeric encoding of the at least one ECG signal using the residual vector quantizer to obtain a quantized numeric encoding of the at least one ECG signal; processing the quantized numeric encoding of the at least one ECG signal using the at least one trained machine learning model to obtain one or more characteristics of the at least one ECG signal, obtaining the one or more characteristics comprising: (i) determining, for each of at least some of the plurality of segments, a respective rhythm type selected from a group of rhythm types including: normal sinus, atrial fibrillation, atrial flutter, premature atrial contraction (PAC), and premature ventricular contraction (PVC), and/or (ii) assigning, for each of at least some of the plurality of samples, a respective sample-level ECG label indicative of a respective portion of the at least one ECG signal to which a respective sample corresponds; and outputting the one or more characteristics of the at least one ECG signal.
2 . The Holter monitor of claim 1 , further comprising a user interface, wherein the method further comprises generating an indication of the one or more characteristics of the at least one ECG signal, wherein outputting the one or more characteristics of the at least one ECG signal comprises outputting the indication via the user interface.
3 . The Holter monitor of claim 1 , wherein; the at least one trained machine learning model comprises a trained rhythm classifier; obtaining the one or more characteristics comprises determining the respective rhythm types, and processing the quantized numeric encoding of the at least one ECG signal using the at least one trained machine learning model to obtain the one or more characteristics of the at least one ECG signal comprises: positionally encoding the quantized numeric encoding of the at least one ECG signal to obtain a positionally encoded quantized numeric encoding of the at least one ECG signal; and processing the positionally encoded quantized numeric encoding of the at least one ECG signal using the trained rhythm classifier to obtain, for each of the at least some of the plurality of segments, a respective output indicative of the respective rhythm type.
4 . The Holter monitor of claim 3 , wherein the trained rhythm classifier is a classification head comprising one or more convolutional layers.
5 . The Holter monitor of claim 1 , wherein: the at least one trained machine learning model comprises a trained sample-level ECG classifier; obtaining the one or more characteristics comprises assigning the respective sample-level ECG labels, and processing the quantized numeric encoding of the at least one ECG signal using the at least one trained machine learning model to obtain the one or more characteristics of the at least one ECG signal comprises: processing the quantized numeric encoding of the at least one ECG signal using the trained sample-level ECG classifier to obtain the respective sample-level ECG label for each of the at least some of the plurality of samples.
6 . The Holter monitor of claim 5 , wherein the trained sample-level ECG classifier is a classification head comprising one or more convolutional layers, one or more encoder layers, and one or more decoder layers.
7 . The Holter monitor of claim 1 , wherein the residual vector quantizer is a grouped residual vector quantizer.
8 . The Holter monitor of claim 1 , wherein: the at least one non-transitory computer-readable storage medium is further storing: a decoder; and the method further comprises: decoding the quantized numeric encoding of the at least one ECG signal using the decoder to obtain at least one denoised ECG signal.
9 . The Holter monitor of claim 8 , wherein the method further comprises: transmitting, to at least one remote device, an indication of the one or more characteristics of the at least one ECG signal and/or the at least one denoised ECG signal.
10 . The Holter monitor of claim 1 , wherein: the at least one non-transitory computer-readable storage medium is further storing: a convolutional neural network encoder; and encoding the at least one ECG signal to obtain the numeric encoding of the at least one ECG signal comprises encoding the at least one ECG signal using the convolutional neural network encoder.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS The present application claims the benefit of and priority under 35 U.S.C. § 119(e) of U.S. Provisional Patent Application No. 63/684,432 filed Aug. 18, 2024, entitled “METHODS AND SYSTEMS FOR ANALYZING ECG SIGNALS USING NEURAL NETWORKS,” and U.S. Provisional Patent Application No. 63/819,451 filed Jun. 6, 2025, entitled “MACHINE LEARNING TECHNIQUES FOR ELECTROCARDIOGRAM (ECG) ANALYSIS,” each of which is incorporated by reference herein in its entirety. This Application is a Continuation-in-part of U.S. Non-Provisional patent application Ser. No. 18/896,784, filed Sep. 25, 2024, entitled “METHODS AND SYSTEMS FOR ANALYZING ECG SIGNALS USING NEURAL NETWORKS,” which claims the benefit of and priority under 35 U.S.C. § 11(e) of U.S. Provisional Patent Application No. 63/684,432, filed Aug. 18, 2024, entitled “METHODS AND SYSTEMS FOR ANALYZING ECG SIGNALS USING NEURAL NETWORKS.” BACKGROUND Heartbeats are triggered by electrical impulses that travel through conduction pathways in the heart, causing coordinated contractions of the atria and ventricles. An electrocardiogram (ECG) is a tool that measures signals (“ECG signals”) representing the electrical activity of the heart over time. An ECG signal typically includes at least one waveform representing at least one measured heartbeat. In a healthy subject, an ECG waveform is expected to have several different portions, each of which represents a particular function that contributes to the heartbeat. ECG waveform portions may include the P-wave, the QRS complex, and the T-wave. The P-wave represents the depolarization of the atria, which precedes atrial contraction. The QRS complex typically includes three waves (e.g., the Q-wave, the R-wave, and the S-wave) and represents the rapid depolarization of the ventricles, leading to ventricular contraction. The T-wave represents the repolarization of the ventricles, which allows the ventricles to reset for the next cycle of depolarization. ECG waveforms can be analyzed and interpreted to diagnose and treat cardiac issues. An ECG system measures ECG signals using one or more leads. A “lead” refers to the electrical viewpoint of the heart's activity. A lead measures the voltage difference between two or more points on the body. For example, the voltage difference may be measured between two or more ECG sensors (e.g., electrodes) placed in direct or indirect contact with the body. Different ECG systems may use different numbers of leads. For example, some ECG systems may use 1, 3, 6, or 12 leads. SUMMARY Some embodiments provide for a method for denoising at least one electrocardiogram (ECG) signal, the method comprising: using at least one processor to perform: obtaining the at least one ECG signal, the at least one ECG signal having been previously measured using ECG sensors; processing the at least one ECG signal using a trained denoising model to obtain at least one denoised ECG signal, the trained denoising model comprising an encoder, a residual vector quantizer, and a decoder, the processing comprising: encoding the at least one ECG signal using the encoder to obtain a numeric encoding of the at least one ECG signal; quantizing the numeric encoding of the at least one ECG signal using the residual vector quantizer to obtain a quantized numeric encoding of the at least one ECG signal; and decoding the quantized numeric encoding of the at least one ECG signal using the decoder to obtain the at least one denoised ECG signal; and outputting the at least one denoised ECG signal. The trained denoising model may be a denoising model that has been trained using a method of training a denoising model to denoise at least one electrocardiogram (ECG) signal as described herein. Methods according to the present embodiments may further comprise obtaining the trained denoising model using a method of training a denoising model to denoise at least one electrocardiogram (ECG) signal as described herein. Some embodiments provide for at least one non-transitory computer-readable storage medium storing processor-executable instructions that, when executed by at least one processor, cause the at least one processor to perform a method for denoising at least one electrocardiogram (ECG) signal, the method comprising: obtaining the at least one ECG signal, the at least one ECG signal having been previously measured using ECG sensors; processing the at least one ECG signal using a trained denoising model to obtain at least one denoised ECG signal, the trained denoising model comprising an encoder, a residual vector quantizer, and a decoder, the processing comprising: encoding the at least one ECG signal using the encoder to obtain a numeric encoding of the at least one ECG signal; quantizing the numeric encoding of the at least one ECG signal using the residual vector quantizer to obtain a quantized numeric encoding of the at least one ECG signal; and decoding the quantized numeric encoding of the at l