CN-121987170-A - System and method for non-contact arterial blood pressure waveform monitoring based on machine learning

CN121987170ACN 121987170 ACN121987170 ACN 121987170ACN-121987170-A

Abstract

The invention provides a system and a method for non-contact arterial blood pressure waveform monitoring based on machine learning. The method includes transmitting a millimeter wave signal waveform with a millimeter wave radar and receiving a reflected signal from a chest of a subject, extracting complex-valued millimeter wave IQ data from the received reflected signal, steering the complex-valued millimeter wave IQ data with a beamforming-based data enhancement module to form a target signal beam, and estimating an arterial blood pressure waveform from the target signal beam with a millimeter wave-arterial blood pressure waveform transformer. During a training phase, the beamforming-based data enhancement module is configured to generate millimeter wave signal beam data from different angles for training the millimeter wave-arterial blood pressure waveform transformer, and to generate a teacher model with a cross-modal knowledge migration module to collectively oversee training of the millimeter wave-arterial blood pressure waveform transformer.

Inventors

HU QINGYONG
ZHANG QIAN
ZHOU YUXUAN

Assignees

香港科技大学

Dates

Publication Date: 20260508
Application Date: 20251030
Priority Date: 20241101

Claims (20)

1. A system for non-contact arterial blood pressure waveform monitoring based on machine learning, comprising: Millimeter wave radar; one or more processors in communication with the millimeter wave radar; A machine learning tool in communication with the one or more processors; a memory device storing the machine learning tool and cooperatively coupled with the one or more processors; Wherein: the machine learning tool comprises a millimeter wave-arterial blood pressure waveform converter and a data enhancement module based on wave beam forming; the millimeter wave radar is configured to transmit a millimeter wave signal waveform and receive a reflected signal from a subject's chest; the memory device further stores instructions that, when executed, cause the one or more processors to: Extracting complex-valued millimeter wave IQ data from the received reflected signals; steering the complex-valued millimeter wave IQ data with the beamforming-based data enhancement module to form a target signal beam, and An arterial blood pressure waveform is estimated from the target signal beam using the millimeter wave-arterial blood pressure waveform transformer.
2. The machine learning based system for non-contact arterial blood pressure waveform monitoring of claim 1, wherein: The transmitted millimeter wave signal waveform is a frequency modulated continuous waveform, and The millimeter wave radar is further configured to waveform mix the received reflected signal with the transmitted millimeter wave signal to obtain an intermediate frequency signal.
3. The machine learning based system for contactless arterial blood pressure waveform monitoring of claim 2, wherein the one or more processors are configured to: performing distance-fast fourier transform on the intermediate frequency signal to obtain the frequency and phase of the intermediate frequency signal; detecting a signal region of the chest of the subject from the frequency and phase of the intermediate frequency signal using a constant false alarm rate algorithm; searching a distance bin having the greatest reflected energy from the detected signal region, and And extracting the complex-valued millimeter wave IQ data from the searched distance bin.
4. The machine learning based system for non-contact arterial blood pressure waveform monitoring of claim 1, wherein the millimeter wave-arterial blood pressure waveform transformer comprises an arterial blood pressure waveform feature extractor comprising: an encoder configured to encode the target signal beam and generate a compressed feature tensor; A self-attention transformer configured to perform global context modeling and spatial attention optimization on the compressed feature tensor to obtain an optimized feature tensor; a decoder configured to decode the optimized feature tensor and generate an arterial blood pressure waveform feature tensor, and A plurality of spatially aware attention modules bridging intermediate layers of the encoder and the decoder to facilitate the arterial blood pressure waveform feature extractor stitching feature tensors of different spatial dimensions by selectively re-weighting feature activations according to the importance of the target signal beams from different reflection regions.
5. The machine learning based system for contactless arterial blood pressure waveform monitoring of claim 4, wherein the millimeter wave-arterial blood pressure waveform transformer further comprises a personalized waveform regressor comprising: A plurality of expert regressor models respectively trained by using a plurality of training data sets corresponding to respective blood pressure categories, and Explicit gating configured to select an appropriate expert regressor model for arterial blood pressure waveform estimation based on the rough blood pressure class of the subject.
6. The machine learning based system for non-contact arterial blood pressure waveform monitoring of claim 5, wherein the appropriate expert regressor model is further trained using arterial blood pressure waveform data of the subject.
7. The machine learning based system for contactless arterial blood pressure waveform monitoring of claim 1, wherein the beamforming based data enhancement module comprises: a first neural network configured to estimate a heart rate of the subject from the complex-valued millimeter wave IQ data, and A second neural network configured to identify spectral features of the complex-valued millimeter wave IQ data and to identify the target signal beam based on the estimated heart rate and the identified spectral features.
8. The machine learning based system for contactless arterial blood pressure waveform monitoring of claim 1, wherein the beamforming based data enhancement module is configured to generate millimeter wave signal beam data from different angles for training the millimeter wave-arterial blood pressure waveform converter.
9. The machine learning based system for contactless arterial blood pressure waveform monitoring of claim 1, wherein the machine learning tool further comprises a cross-modal knowledge migration module configured to generate a teacher model to collectively oversee training of the millimeter wave-arterial blood pressure waveform transformer.
10. The machine learning based system for contactless arterial blood pressure waveform monitoring of claim 1, wherein the teacher model is a transducer model trained based on a common ECG/PPG dataset to predict ECG/PPG signals from millimeter wave data.
11. A machine learning-based method for non-contact arterial blood pressure waveform monitoring, comprising: transmitting a millimeter wave signal waveform using a millimeter wave radar and receiving a reflected signal from the chest of the subject; Extracting complex-valued millimeter wave IQ data from the received reflected signals; Steering the complex-valued millimeter wave IQ data with a beamforming-based data enhancement module to form a target signal beam, and Estimating an arterial blood pressure waveform from the target signal beam using a millimeter wave-arterial blood pressure waveform transducer; wherein the beamforming-based data enhancement module and the millimeter wave-arterial blood pressure waveform transformer are machine learning tools in communication with one or more processors and stored in a memory device cooperatively coupled with the one or more processors.
12. The machine learning based method of claim 11 wherein: The transmitted millimeter wave signal waveform is a frequency modulated continuous waveform, and The machine learning based method further includes waveform mixing the received reflected signal with the transmitted millimeter wave signal to obtain an intermediate frequency signal.
13. The machine learning based method of claim 12, wherein the complex-valued millimeter wave IQ data is extracted by the one or more processors by: performing distance-fast fourier transform on the intermediate frequency signal to obtain the frequency and phase of the intermediate frequency signal; detecting a signal region of the chest of the subject from the frequency and phase of the intermediate frequency signal using a constant false alarm rate algorithm; searching a distance bin having the greatest reflected energy from the detected signal region, and And extracting the complex-valued millimeter wave IQ data from the searched distance bin.
14. The machine learning based method of claim 11, wherein, The millimeter wave-arterial blood pressure waveform transformer comprises an arterial blood pressure waveform feature extractor, wherein the arterial blood pressure waveform feature extractor comprises an encoder, a self-attention transformer, a decoder and a plurality of space perception attention modules bridging intermediate layers of the encoder and the decoder; The estimating of the arterial blood pressure waveform from the target signal beam comprises: Encoding the target signal beam with the encoder and generating a compressed feature tensor; Performing global context modeling and spatial attention optimization on the compressed feature tensor with the self-attention transformer to obtain an optimized feature tensor; Decoding the optimized feature tensor and generating an arterial blood pressure waveform feature tensor using the decoder, and The arterial blood pressure waveform feature extractor is facilitated by the plurality of spatially aware attention modules to stitch feature tensors of different spatial dimensions by selectively re-weighting feature activations according to the importance of the target signal beams from different reflection regions.
15. The machine learning based method of claim 14, wherein, The millimeter wave-arterial blood pressure waveform converter further comprises a personalized waveform regressor, wherein the personalized waveform regressor comprises a plurality of expert regressor models; the machine learning based method further comprises Training the plurality of expert regressor models with a plurality of training data sets corresponding to respective blood pressure categories, respectively, and Based on the rough blood pressure category of the subject, an appropriate expert regressor model is selected from a plurality of trained expert regressor models for arterial blood pressure waveform estimation.
16. The machine learning based method of claim 15, further comprising training the appropriate expert regressor model using arterial blood pressure waveform data of the subject.
17. The machine learning based method of claim 11, wherein, The beamforming-based data enhancement module includes a first neural network and a second neural network, and The target signal beam is formed by: Estimating a heart rate of the subject from the complex-valued millimeter wave IQ data using the first neural network, and Utilizing the second neural network, identifying spectral features of the complex-valued millimeter wave IQ data, and identifying the target signal beam based on the estimated heart rate and the identified spectral features.
18. The machine learning based method of claim 11, further comprising generating millimeter wave signal beam data from different angles with the beamforming-based data enhancement module for training the millimeter wave-arterial blood pressure waveform transformer.
19. The machine learning based method of claim 11, further comprising generating a teacher model with a cross-modal knowledge migration module to collectively oversee training of the millimeter wave-arterial blood pressure waveform transformer.
20. The machine learning based method of claim 11, wherein the teacher model is a transducer model trained based on a common ECG/PPG dataset to predict ECG/PPG signals from millimeter wave data.

Description

System and method for non-contact arterial blood pressure waveform monitoring based on machine learning Cross reference to related applications The present application claims priority from U.S. provisional patent application No. 63/714,886 filed on month 11 of 2024, the disclosure of which is incorporated herein by reference in its entirety. Technical Field The present invention relates generally to machine learning techniques and, more particularly, to machine learning based systems and methods for contactless arterial blood pressure waveform (arterial blood pressure waveform, ABPW) monitoring. Background Blood Pressure (BP) monitoring is critical for assessing the health status of the heart and cerebrovascular systems. Traditionally, blood pressure measurements have focused on discrete Systolic (SBP) and Diastolic (DBP) values, but these indicators are not sufficient to fully reflect the heart's functional status. More complex cardiac parameters, such as stroke volume, cardiac Output (CO) and vascular resistance, can more accurately reflect circulatory system dynamics and are closely related to common cardiovascular diseases such as heart failure, cardiogenic shock, etc. Currently, the global population affected by this disease has exceeded six thousand four million. . In addition to discrete blood pressure values, arterial blood pressure waveforms contain fine-grained information that delineates the complete cardiac cycle. As shown in fig. 1, the arterial blood pressure waveform reflects the rising of blood pressure caused by ejection of blood during systole, the falling of pressure after closure of the aortic valve, and the trough state formed when blood flows out of the aorta. Unlike discrete blood pressure, which only provides a single measurement, arterial blood pressure waveforms can reveal abnormal conditions inside the cardiovascular system by dynamic changes between successive heart beats, where some of the abnormalities are not even directly observable from an Electrocardiogram (ECG) signal. Thus, arterial blood pressure waveforms provide a more comprehensive and diagnostically valuable perspective for cardiovascular health assessment. As shown in fig. 2, even if two groups of individuals have similar discrete blood pressure values, their arterial blood pressure waveforms may still differ significantly, revealing a potential heart dysfunction. Thus, continuous arterial blood pressure waveform monitoring is of great importance in assessing overall cardiovascular status and aiding in the diagnosis of related heart diseases. Existing arterial blood pressure waveform monitoring methods typically rely on invasive procedures or continuous skin contact, which are inconvenient and uncomfortable. The arterial catheter-based monitoring method, which realizes real-time blood pressure waveform measurement by inserting a catheter into a blood vessel, is regarded as clinical gold standard, but is limited to specific scenes such as an Intensive Care Unit (ICU) due to its invasiveness. For non-invasive methods, existing studies (including cuff-based auscultation and oscillography, wearable methods, and non-contact solutions, etc.) generally only use discrete Systolic (SBP) and Diastolic (DBP) estimates as a rough approximation of arterial blood pressure waveforms, failing to obtain important information reflecting the trend of blood pressure dynamics. In particular, current non-contact methods cannot achieve truly continuous arterial blood pressure waveform monitoring because they rely on Pulse Transit Time (PTT) methods, which require that a single discrete result be calculated after a steady signal is acquired over several cardiac cycles. In recent years, arterial blood pressure waveform monitoring technology has advanced, and non-invasive schemes such as volume-clamp (v-clamp) and wearable monitoring have been developed to obtain arterial blood pressure waveforms without puncturing blood vessels. However, these methods often rely on specialized and costly hardware equipment (about $40,000), or require intimate contact with the skin, resulting in discomfort and operational inconvenience in everyday use scenarios. Arterial blood pressure waveforms, on the other hand, are found to be closely related to changes in blood volume in blood vessels and mechanical activity of the heart. Figure 3 illustrates a typical arterial blood pressure waveform over a single cardiac cycle. In the systole stage, the left ventricle injects blood into the main artery, the kinetic energy of the blood causes elastic expansion of the aortic wall, resulting in an increase in blood pressure, and when the aortic valve closes, part of the blood flows back to the heart, resulting in a decrease in blood pressure. The highest, lowest and average values of the waveform correspond to systolic pressure (SBP), diastolic pressure (DBP) and Mean Arterial Pressure (MAP), respectively. The arterial blood pressure waveform contains a plurality of phys