Radar-based non-contact cardiac condition prediction apparatus and method

Abstract

Disclosed are a radar-based non-contact cardiac condition prediction apparatus and method according to an embodiment. The radar-based non-contact cardiac condition prediction apparatus according to an embodiment comprises a processor and a memory storing a program executed by the processor, wherein the processor: irradiates a subject with a radar signal, of which frequency is modulated over time, a plurality of times for a preset period of time and then acquires a reflected signal reflected and returned from the subject; converts the reflected signal into structured data, wherein the structured data includes axes including a first axis and a second axis, the first axis represents range bins, and the second axis represents physiological variation information of the subject; and selects a target bin from among the range bins on the basis of intensity information of the reflected signal according to each distance from the subject, the intensity information being included in each of the range bins.

Description

Radar-based non-contact cardiac condition prediction device and method

[0001] The disclosed embodiments relate to a technique for predicting heart conditions in a non-contact manner using radar.

[0002] [Cross-reference to related applications]

[0003] This application claims priority to Republic of Korea Provisional Patent Application No. 10-2024-0099734, filed July 26, 2024, the entire contents of which are incorporated herein by reference.

[0004] [Description of the national research and development project that supported this invention]

[0005] This study was supported by the National Research and Development Project [Project Identification Number: 171195966, Subproject Number: 00222910, Ministry of Science and ICT, Project Management (Specialist) Agency: National Research Foundation of Korea, Project Name: Biomedical Technology Development, Project Name: Development of Medical Field Application Technology for 5 Major Diseases and Training of Physician Scientists through Customized Future Medical Research Center in the Era of 6P Medicine, Contribution Ratio: 1 / 1, Project Implementing Agency: Seoul National University Bundang Hospital, Research Period: 2023.04.01 ~ 2023.12.31].

[0006] Existing radar-based cardiac signal monitoring technologies have focused on analyzing normal cardiac signals. Therefore, these technologies rely on selecting a single range bin or simply including adjacent bins, making it easy to miss key cardiac reflection signals in cases of arrhythmias characterized by multiple reflection patterns.

[0007] In particular, since the source of the reflected signal is distributed in the case of an arrhythmic heart, the reflected signal is given complexity including nonlinearity and superposition, and therefore existing technologies designed to detect regular signals are insufficient for detecting arrhythmic heart signals.

[0008] As shown in Figure 1a, a heatmap visualizing the intensity of the reflected signal from a healthy heart by applying a fast Fourier transform (FFT) to the reflected signal from the healthy heart reveals a distinct, well-organized frequency pattern. In contrast, the reflected signal from an arrhythmic heart exhibits an unclear frequency pattern due to a mixture of irregular heartbeat signals and noise-like spectra.

[0009] Meanwhile, in Figures 1a and 1b, the fast-time axis represents the Doppler frequency change according to the heart movement within each sweep, and the slow-time axis represents the signal change according to physiological changes (e.g., respiration, heartbeat) over time. The brighter the color intensity, the stronger the reflected signal is interpreted.

[0010] Considering these differences, signal processing methods that can appropriately select the core cardiac signals and extract physiologically meaningful signal components even within nonlinear and overlapping patterns should be supplemented, as the reflection signals of arrhythmia appear across range bins.

[0011] The disclosed embodiments are for predicting heart conditions in a non-contact manner using radar.

[0012] In one embodiment, a radar-based non-contact heart condition prediction device comprises a processor; and a memory storing a program executed by the processor, wherein the processor: irradiates a radar signal, the frequency of which is modulated over time, to a subject multiple times for a preset period of time, and then obtains a reflection signal reflected back from the subject, and converts the reflection signal into structured data, wherein the structured data is composed of axes including a first axis and a second axis, the first axis representing a range bin, the second axis representing physiological variation information of the subject, and selects a target bin among the range bins based on intensity information of a reflection signal for each distance from the subject included in each of the range bins.

[0013] The above processor may be characterized by applying a Fourier transform to the reflected signal to convert the reflected signal into a channel impulse response matrix as the structured data.

[0014] The processor may be characterized in that: the channel impulse response matrix is ​​divided into windows by the preset time intervals, and the reflected signal having the largest signal intensity is identified, and the range bin containing the largest reflected signal is selected as the target bin.

[0015] The above processor may be characterized by: determining the intensity of the reflected signal for each range bin based on the size of the channel impulse response matrix, and selecting the range bin with the greatest intensity of the reflected signal as the target bin.

[0016] The above processor may be characterized by selecting a range bin adjacent to a range bin having the greatest intensity of the reflected signal as the target bin.

[0017] The processor may be characterized in that: the first sub-model included in the pre-trained artificial intelligence model inputs the target bin to extract a first cardiac signal feature, the first sub-model includes a residual block and a graph attention block, the residual block enhances temporal features while increasing the dimension of the target bin, and the graph attention block is trained to integrate features of mutual correlation between the targets.

[0018] The processor may be characterized in that: the first cardiac signal feature is input to an encoder of a second sub-model included in the learned artificial intelligence model to compress the resolution of the first cardiac signal feature and extract a core feature, and the core feature is input to a decoder of the second sub-model to restore the resolution and extract a second cardiac signal feature, and the encoder includes a residual block and a max pooling layer, and the decoder includes an upsampling layer and a pre-transposed residual block, and the second cardiac signal feature is at a higher level than the first cardiac signal feature.

[0019] The processor may be characterized in that: the processor inputs the second cardiac signal feature into a third sub-model included in the learned artificial intelligence model to generate a cardiac pulse waveform in the form of a sequence, and the third sub-model is trained to generate the cardiac pulse waveform by modeling the temporal flow of the second cardiac signal by considering the temporal context of the second cardiac signal feature.

[0020] In one embodiment, a radar-based non-contact heart condition prediction method is provided, which is performed by a radar-based non-contact heart condition prediction device having a processor; and a memory storing a program executed by the processor, the method comprising: a step of irradiating a subject with a radar signal whose frequency is modulated over time multiple times for a preset period of time and then obtaining a reflection signal reflected back from the subject; a step of converting the reflection signal into structured data; - the structured data is composed of axes including a first axis and a second axis, the first axis representing a range bin, and the second axis representing physiological variation information of the subject - and a step of selecting a target bin among the range bins based on intensity information of a reflection signal according to a distance from the subject included in each of the range bins.

[0021] The above-described converting step may be characterized by including a step of converting the reflected signal into a channel impulse response matrix as the structured data by applying a Fourier transform to the reflected signal.

[0022] The above-described selecting step may be characterized by including: a step of identifying the largest reflection signal in each window divided by the preset time interval of the channel impulse response matrix; and a step of selecting a range bin including the largest reflection signal as the target bin.

[0023] The above selection step may be characterized by including a step of determining the intensity of the reflected signal for each range bin based on the size of the channel impulse response matrix; and a step of selecting the range bin with the greatest intensity of the reflected signal as the target bin.

[0024] The above selection step may be characterized by including a step of selecting a range bin neighboring a range bin having the greatest intensity of the reflected signal as the target bin.

[0025] The method may further include a step of extracting a first cardiac signal feature by inputting the target bin into a first sub-model included in a pre-trained artificial intelligence model, wherein the first sub-model includes a residual block and a graph attention block, wherein the residual block enhances temporal features while increasing the dimension of the target bin, and the graph attention block is trained to integrate features of mutual correlation between the targets.

[0026] The method further includes: a step of inputting the first cardiac signal feature into an encoder of a second sub-model included in a pre-learned artificial intelligence model to compress the resolution of the first cardiac signal feature and extract a core feature; and a step of inputting the core feature into a decoder of the second sub-model to restore the resolution and extract a second cardiac signal feature, wherein the encoder includes a residual block and a max pooling layer, the decoder includes an upsampling layer and a pre-transposed residual block, and the second cardiac signal feature may be characterized in that it is at a higher level than the first cardiac signal feature.

[0027] The method may further include a step of inputting the second cardiac signal feature into a third sub-model included in the learned artificial intelligence model to generate a cardiac pulse waveform in the form of a sequence, wherein the third sub-model is trained to generate the cardiac pulse waveform by modeling the temporal flow of the second cardiac signal by considering the temporal context of the second cardiac signal feature.

[0028] The disclosed embodiments can select a suitable target signal to be used for cardiac signal extraction based on the magnitude of the signal identified through the channel impulse response matrix.

[0029] The disclosed embodiments can reduce distortion of cardiac waveforms of arrhythmia patients by extracting correlations between target bins using an artificial intelligence model.

[0030] Figure 1a is a diagram showing the results of applying a fast Fourier transform to the intensity of a healthy heart signal, according to a previous study.

[0031] Figure 1b is a diagram showing the result of applying a fast Fourier transform to the intensity of an arrhythmia cardiac signal according to a previous study.

[0032] FIG. 2 is a block diagram illustrating a radar-based non-contact heart condition prediction device according to one embodiment.

[0033] FIG. 3 is a diagram for explaining an exemplary operation flow of a radar-based non-contact heart state prediction device according to one embodiment.

[0034] FIG. 4 is a diagram illustrating a process for generating a heart pulse waveform by a radar-based non-contact heart condition prediction device according to one embodiment.

[0035] FIG. 5 is a diagram illustrating the structure of a pre-learned artificial intelligence model used by a radar-based non-contact heart condition prediction device according to one embodiment.

[0036] FIG. 6 is a flowchart illustrating a radar-based non-contact heart condition prediction method according to one embodiment.

[0037] Hereinafter, specific embodiments of one embodiment will be described with reference to the drawings. The following detailed description is provided to facilitate a comprehensive understanding of the sensor described herein. However, this is merely an example and the present invention is not limited thereto.

[0038] In describing certain embodiments, detailed descriptions of known technologies related to the present invention will be omitted if they are deemed to unnecessarily obscure the gist of the embodiments. Furthermore, numbers (e.g., "first," "second," etc.) used in the description of the embodiments are merely identifiers used to distinguish one component from another.

[0039] FIG. 2 is a block diagram illustrating a radar-based non-contact heart condition prediction device (100) according to one embodiment.

[0040] Referring to FIG. 2, a radar-based non-contact heart condition prediction device (100) according to one embodiment includes a processor (110) and a memory (120).

[0041] The processor (110) irradiates a radar signal whose frequency is modulated over time to a subject for a preset period of time and then obtains a reflected signal that is reflected back from the subject.

[0042] Here, the radar signal may include a signal in the form of a Frequency-Modulated Continuous Wave (FMCW) whose frequency linearly increases or decreases over time. In this case, the radar signal is transmitted continuously and repeatedly and may exist in a specific band (e.g., 77 GHz to 81 GHz).

[0043] The processor (110) converts the reflected signal into structured data. At this time, the structured data is composed of axes including a first axis and a second axis, the first axis representing a range bin and the second axis representing physiological fluctuation information of the subject.

[0044] Structured data can refer to data composed by converting each reflected signal into a two-dimensional matrix on the time and distance axes. Structured data can be expressed as a two-dimensional matrix with distance information along range bins as the first axis and physiological fluctuation information of the subject as the second axis.

[0045] Here, a range bin may refer to a unit cell of a distance section distinguished based on the delay time or phase difference of the reflected signal. Each range bin includes relative distance information from the transmission point of the radar signal to the location of the subject, which is the reflection point, and the intensity of the reflected signal received at the relative distance may be recorded.

[0046] For example, if the distance between the subject and the radar transmitter is 1 m and range bins are set at 5 cm intervals, a total of 20 range bins can be formed from 0 m to 1 m. In the case of a healthy subject, distinct reflection intensities appear in the range bins, whereas in the case of arrhythmia patients, the cardiac reflection signals appear scattered across the range bins, tending to form a complex and scattered reflection pattern.

[0047] That is, the first axis functions as a distance-based axis that includes reflected signal intensity or phase information at each distance interval, and can be a reference for distinguishing at what distance a signal reflected from a specific body part of the subject, for example, the chest or heart, was captured.

[0048] The physiological fluctuation information of a subject can be determined by the accumulated value of the intensity of the reflected signal over time, or by information that can be determined based on it. For example, it can include temporal changes due to respiration, heartbeat, and micro-vibrations of the body surface. The physiological fluctuation information can be expressed as the intensity or phase change of the reflected signal corresponding to each range bin.

[0049] That is, the second axis is an axis that expresses time variation information of periodic or irregular physiological signals based on the Doppler effect and micro-vibration detection, and can be used as a standard for quantitatively analyzing the intensity, period, and interval of heartbeats.

[0050] The processor (110) can apply a Fourier transform to the reflected signal to convert the reflected signal into structured data, such as a channel impulse response matrix.

[0051] The Channel Impulse Response Matrix is ​​one of the data formats structured by the processor (110), and may refer to a data structure in the form of a two-dimensional complex matrix arranged based on a fast time axis and a slow time axis.

[0052] The fast time axis, as an example of the first axis described above, can correspond to range bins based on distance delay time. In other words, the fast time axis of the channel impulse response matrix can represent reflection intensity or phase information for each range bin, serving as a criterion for identifying the distance interval in which the received signal originated.

[0053] The slow time axis is an example of the second axis of the above-described description, and is composed of a sequence of repeatedly measured frames or a time flow, and the slow time axis can express minute fluctuations of the body surface due to the physiological activity of the subject, such as periodic or aperiodic vibration patterns due to heartbeat or breathing.

[0054] The processor (110) selects a target bin among the range bins based on the intensity information of the reflected signal at each distance from the subject included in each range bin.

[0055] Here, the target bin may refer to a range bin corresponding to a distance interval in which physiological vibrations due to heartbeats are significantly reflected. Even in cases where the cardiac reflection signal is weakly scattered and captured at a distance, such as in patients with arrhythmia, the processor (110) may select a target bin from among the range bins included in the reflection signal to reliably extract meaningful physiological information.

[0056] The processor (110) can determine the intensity of the reflected signal for each range bin based on the size of the channel impulse response matrix, and select the range bin with the greatest intensity of the reflected signal as the target bin.

[0057] In other words, the processor (110) can calculate the size of the reflected signal component corresponding to each range bin based on the frame cumulative sum, average value, or maximum value throughout the entire section along the second axis of the channel impulse response matrix, and select the range bin with the largest intensity of the reflected signal (e.g., having the highest energy or amplitude) as the target bin.

[0058] Through this, the processor (110) can globally identify a distance section where a strong cardiac response is observed, even when the reflection signal appears irregularly at each point in time, such as in a patient with arrhythmia, thereby improving the reliability of heart position estimation and the accuracy of waveform restoration in a subsequent step.

[0059] The processor (110) can identify the largest reflection signal in each window divided into preset time intervals of the channel impulse response matrix and select a range bin containing the largest reflection signal as a target bin.

[0060] At this time, windows can be created by sliding a section corresponding to a fixed time length (e.g., tens of frames, i.e., several milliseconds) along the slow time axis of the channel impulse response matrix. At this time, each window can overlap the previous window by a certain ratio, and this overlap can increase the temporal resolution and enable continuous tracking of the cardiac response location.

[0061] The processor (110) can calculate the intensity of the reflected signal for each range bin within each window, identify the largest reflected signal for each window, identify the largest reflected signal, and select the range bin with the largest intensity of the reflected signal (e.g., having the highest energy or amplitude) as the target bin.

[0062] Through this, the processor (110) can sequentially select target bins that change over time by tracking the location of the reflected signal with the greatest signal intensity for windows divided by preset time intervals in the channel impulse response matrix.

[0063] The processor (110) can select a range bin adjacent to the range bin with the largest reflected signal intensity as a target bin.

[0064] The processor (110) can calculate the intensity of the reflected signal by calculating the sum of the squares or the absolute mean of the complex response values ​​accumulated over the time interval for each range bin. From this, the processor (110) can effectively select a distance interval where the response of the cardiac signal is consistently concentrated as a target bin, while eliminating transient peaks caused by noise.

[0065] In patients with arrhythmia, the reflection component due to the heartbeat is not limited to a single range bin, but rather, multiple scattering occurs across adjacent ranges where the response of the cardiac signal is consistently concentrated. The processor (110) can select a target bin by reflecting the multiple scattering pattern of the arrhythmia patient.

[0066] The processor (110) can generate cardiac signal features or cardiac waveforms using a pre-learned artificial intelligence model including the first to third sub-models.

[0067] First, the processor (110) can extract the first cardiac signal feature by inputting the target bin into the first sub-model included in the pre-learned artificial intelligence model.

[0068] At this time, the first sub-model may include a residual block and a graph attention block. The residual block may be pre-trained to enhance temporal features while increasing the dimensionality of target bins. The first sub-model may learn and enhance temporal features by gradually increasing the feature dimensionality of input signals (e.g., T x 1 -> T x 4 -> T x 8 -> T x 16) through the residual block (e.g., kernel size k=64, 32, 16, 16). The graph attention block may be pre-trained to integrate features of inter-correlation between targets. The attention block may receive features that have passed through the residual block as input, and integrate features by applying an attention mechanism to the inter-correlation and importance between signals from multiple target distance bins.

[0069] The processor (110) can input the first cardiac signal feature into the encoder of the second sub-model included in the pre-trained artificial intelligence model, thereby compressing the resolution of the first cardiac signal feature and extracting key features. For example, the processor (110) can expand the feature dimension to T x 32 through a residual block (e.g., kernel size k=8) included in the encoder. In addition, the encoder can repeatedly use max pooling to reduce the temporal resolution by half.

[0070] Meanwhile, here, the first cardiac signal may mean a low-dimensional feature vector extracted by learning a pattern on the time axis based on a time series reflection signal input from a target bin.

[0071] For example, primary cardiac signal features may include RR interval variations due to heartbeats and irregularities in the abnormal rhythm.

[0072] The processor (110) can input the core features into the decoder of the second sub-model to restore the resolution and extract the second cardiac signal features.

[0073] The decoder can gradually restore temporal resolution using upsampling layers. The decoder can then process the upsampled features using a preprocessed residual block to increase their resolution.

[0074] The encoder and decoder are connected via a skip connection, and detailed information that may be lost during the encoding process can be directly transmitted to the decoder through the skip connection. The processor (110) can input the second cardiac signal features into the third sub-model included in the pre-trained artificial intelligence model to generate a cardiac pulse waveform in the form of a sequence.

[0075] At this time, the third sub-model can be trained to generate a cardiac pulse waveform by modeling the temporal flow of the second cardiac signal by considering the temporal context of the second cardiac signal features.

[0076] The processor (110) generates a time-series aligned heart pulse waveform from the third sub-model, and can further quantitatively calculate various heart condition indicators, such as heart rate, RR interval, and rhythm type, through the heart waveform.

[0077] The memory (120) stores instructions executed by the processor (110).

[0078] The memory (120) can store various data used by the processor (110). For example, the memory (120) can include software (e.g., input data or output data for a program executed by the processor (110) and / or instructions related to the program).

[0079] FIG. 3 is a diagram for explaining the operation flow of a radar-based non-contact heart state prediction device according to one embodiment.

[0080] The processor (110) can irradiate a frequency modulated continuous wave (FMCW) radar signal to the chest of the subject for a preset period of time and obtain a reflected signal that is reflected back.

[0081] At this time, the processor (110) can acquire the subject's electrocardiogram signal in parallel using a separate contact electrocardiogram device. Here, the electrocardiogram signal can be utilized as the correct answer value of the artificial intelligence model of the heartbeat.

[0082] Thereafter, the processor (110) can apply a Fourier transform to the reflected signal to convert the reflected signal into a channel impulse response matrix as structured data.

[0083] The processor (110) can select target bins sensitive to cardiac activity from among the range bins within the channel impulse response matrix generated as structured data. Even in the case of arrhythmia patients, where the reflected signal is not concentrated in a single location but is scattered across locations, the process of selecting target bins allows signal components with high correlation to be utilized for analysis.

[0084] The processor (110) can extract a phase component from a complex signal included in a target bin. The processor (110) can extract the phase component using the real and imaginary parts of the complex component included in each target bin, and based on the physiological basis that minute movements due to heartbeats are expressed as phase changes, amplitude information can be excluded.

[0085] The phase signal extracted by the processor (110) can be filtered by applying a band-pass filter that passes a preset frequency band (e.g., 1 to 2 Hz). From this, the processor (110) can remove frequency components unrelated to heartbeat, such as those due to breathing or unnecessary biological movements.

[0086] The processor (110) may perform downsampling on the phase signal to save computational resources. The processor (110) may perform downsampling to a frequency that can minimize information loss of the phase signal, for example, 20 Hz.

[0087] The processor (110) can input a target bin into a pre-trained artificial intelligence model (hereinafter referred to as HPR-net) to generate a heart pulse waveform corresponding to a reflection signal. Specifically, the processor (110) can input a phase signal of the target bin into HPR-net to generate a heart pulse waveform corresponding to a reflection signal.

[0088] At this time, HPR-net can generate a heart pulse waveform corresponding to the reflected signal through a series of processes using sub-models and blocks. A more detailed description is provided later in Figure 4.

[0089] The processor (110) can extract biomarkers based on the generated cardiac pulse waveform to evaluate and classify the heart condition. For example, the processor (110) can detect peaks based on the cardiac pulse waveform to calculate heartbeat timing. The processor (110) can calculate heart rate (HR) and heart rate variability (HRV) based on RR intervals based on the heartbeat timing. The processor (110) can input combinations of the calculated results into a pre-trained classifier to determine whether the heart is in a normal or abnormal state (e.g., arrhythmia).

[0090] FIG. 4 is a diagram illustrating a process for generating a heart pulse waveform by a radar-based non-contact heart condition prediction device according to one embodiment.

[0091] As illustrated in FIG. 4, a radar-based non-contact heart condition prediction device according to one embodiment can generate a heart pulse waveform corresponding to a reflected signal using a pre-learned model (HPR-net).

[0092] The data depicted in the left region of Figure 4 is input data used in HPR-net and may represent selected target bins. Each cell in a target bin can visually represent the intensity of data sorted based on time and distance. For example, a brightly colored area may indicate a relatively high intensity. The red highlighted portion of the target bin may represent an area that can be used as input feature information for generating a heart pulse waveform.

[0093] The block depicted in the central region of Figure 4 is a HPR-Net (a neural network-based artificial intelligence model) that generates a corresponding heart pulse waveform based on the input target bin. The HPR-Net can be structured hierarchically. During the HPR-Net processing process, input data features can be extracted and transformed in stages, and HPR-Net can refine the heartbeat information contained in the input data.

[0094] The upper right corner of Figure 4 shows an electrocardiogram waveform collected based on actual cardiac activity, which can be utilized in the training and / or evaluation of HPR-Net. The electrocardiogram waveform can be compared with the cardiac pulse waveform generated by HPR-Net to evaluate the training and / or accuracy of HPR-Net.

[0095] The curved waveform shown in the lower right corner of Figure 4 is the heart pulse waveform output by HPR-net. The heart pulse waveform can be displayed sequentially, that is, by visualizing changes in heart rate over time. At this time, the heart pulse waveform can include peak points indicating the time of heart contraction. The points displayed on the heart pulse waveform represent the results of identifying the peak points and can be used as basic information for calculating heart rate and heart rate variability.

[0096] FIG. 5 is a diagram illustrating the structure of a pre-learned artificial intelligence model used by a radar-based non-contact heart condition prediction device according to one embodiment.

[0097] The processor (110) can input a target bin into a pre-learned artificial intelligence model to generate a heart pulse waveform corresponding to a reflection signal.

[0098] At this time, the pre-learned artificial intelligence model can be composed of a first sub-model, a second sub-model, and a third sub-model, as shown in Fig. 5.

[0099] The first sub-model (Heart Signal Extractor) takes as input a target bin of size TХ1 or a phase signal extracted from a target bin, and sequentially passes the input data through residual blocks with kernel sizes K=64, 32, and 16, thereby sequentially expanding the dimension of the input data to TХ4, TХ8, and TХ16, and strengthening temporal features. Afterwards, the first sub-model can learn the interrelationship between target bins and integrate features through a graph attention block. At this time, the result can be refined and output as a first heart signal feature of size TХ8, for example, through an average operation.

[0100] The second sub-model may be composed of an encoder and a decoder. The encoder may repeatedly apply a residual block and a max pooling layer to reduce the resolution of the input features, and the decoder may restore the resolution by including an upsampling layer and a transposed residual block. Additionally, the encoder and decoder may include skip connections between corresponding levels, and the skip connections may preserve the structural information of the original signal. The final output of the second sub-model may be, for example, a second cardiac signal feature of the size of TX32. The second cardiac signal feature may be composed of high-dimensional features with a higher level of abstraction than the first cardiac signal feature.

[0101] The third sub-model can generate a time-series cardiac pulse waveform based on the second cardiac signal features output from the second sub-model. The third sub-model can include a bidirectional recurrent neural network (BiLSTM) to bidirectionally learn temporal context and gradually reduce the output dimension to TX1 through fully connected layers. The intermediate layers can be applied with a LeakyReLU activation function (α=0.2) and a dropout layer (p=0.2), and the final output can be applied with a sigmoid function to output a normalized cardiac pulse waveform between 0 and 1.

[0102] FIG. 6 is a flowchart illustrating a radar-based non-contact heart condition prediction method according to one embodiment.

[0103] Referring to FIG. 6, the method of FIG. 3 can be performed by a radar-based non-contact heart condition prediction device (100) according to an embodiment of FIG. 2.

[0104] First, a radar-based non-contact heart condition prediction device (100) according to one embodiment irradiates a radar signal whose frequency is modulated over time to a subject for a preset period of time and then obtains a reflected signal reflected back from the subject (610).

[0105] Thereafter, a radar-based non-contact heart condition prediction device (100) according to one embodiment converts the reflected signal into structured data (620). At this time, the structured data is composed of axes including a first axis and a second axis, the first axis representing a range bin, and the second axis representing physiological fluctuation information of the subject.

[0106] Thereafter, a radar-based non-contact heart condition prediction device (100) according to one embodiment selects (630) a target bin among the range bins based on the intensity information of the reflected signal according to the distance from the subject included in each range bin.

[0107] Although the method is described as being divided into multiple steps in the illustrated Figure 6, at least some of the steps may be performed in a different order, combined with other steps and performed together, omitted, divided into detailed steps and performed, or additional steps not illustrated may be added and performed.

[0108] Embodiments of the present invention may include a program for performing the methods described herein on a computer, and a computer-readable recording medium including the program. The computer-readable recording medium may include program instructions, local data files, local data structures, etc., alone or in combination. The medium may be specially designed and configured for the present invention, or may be one commonly used in the field of computer software. Examples of the computer-readable recording medium include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical recording media such as CD-ROMs and DVDs, and hardware devices specially configured to store and execute program instructions such as ROMs, RAMs, and flash memories. Examples of programs may include not only machine language codes such as those generated by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like.

[0109] While representative embodiments of the present invention have been described in detail above, those skilled in the art will appreciate that various modifications to the above-described embodiments are possible without departing from the scope of the present invention. Therefore, the scope of the present invention should not be limited to the described embodiments, but should be determined not only by the claims set forth below but also by equivalents thereof.

[0110] The terms described below are terms defined in consideration of their functions in the present invention, and may vary depending on the operator's intention or custom. Therefore, their definitions should be made based on the contents throughout this specification. The terminology used in the detailed description is only for the purpose of describing one embodiment and should never be limited. Unless clearly used otherwise, the singular form includes the plural form. In this description, expressions such as "comprises" or "having" are intended to indicate certain components, numbers, steps, operations, elements, parts or combinations thereof, and should not be construed to exclude the presence or possibility of one or more other components, numbers, steps, operations, elements, parts or combinations thereof other than those described.

[0111] Additionally, the embodiments described herein may have aspects that are entirely hardware, partially hardware and partially software, or entirely software. As used herein, the term "unit" or the like refers to a computer-related entity, such as hardware, a combination of hardware and software, or software.

[0112] A radar-based non-contact cardiac condition prediction device and method according to one embodiment reconstructs cardiac signals of a patient with arrhythmia through a mathematical model and / or an artificial intelligence model, and is applicable to the digital medical industry.

Claims

1. Processor; and A radar-based non-contact heart condition prediction device having a memory for storing a program executed by the above processor, The above processor: After irradiating a radar signal whose frequency is modulated over time to a subject multiple times for a preset period of time, a reflected signal reflected back from the subject is obtained. Converting the above reflected signal into structured data, - The above structured data is composed of axes including a first axis and a second axis, wherein the first axis represents a range bin, and the second axis represents physiological fluctuation information of the subject, A radar-based non-contact heart condition prediction device characterized in that a target bin is selected from among the range bins based on the intensity information of the reflected signal at each distance from the subject included in each of the range bins.

2. In paragraph 1, The above processor: A radar-based non-contact heart state prediction device characterized in that the reflected signal is converted into a channel impulse response matrix as the structured data by applying a Fourier transform to the reflected signal.

3. In paragraph 2, The above processor: Identify the reflected signal having the largest signal intensity in each window divided by the above-described time interval of the above-described channel impulse response matrix, A radar-based non-contact heart condition prediction device characterized in that the range bin containing the largest reflection signal is selected as the target bin.

4. In paragraph 2, The above processor: The intensity of the reflected signal for each range bin is determined based on the size of the channel impulse response matrix, A radar-based non-contact heart condition prediction device characterized in that the range bin with the greatest intensity of the reflected signal is selected as the target bin.

5. In paragraph 3, The above processor: A radar-based non-contact heart condition prediction device characterized in that the range bin adjacent to the range bin with the greatest intensity of the reflected signal is selected as the target bin.

6. In paragraph 1, The above processor: By inputting the target bin into the first sub-model included in the learned artificial intelligence model, the first cardiac signal feature is extracted, A radar-based non-contact heart condition prediction device, characterized in that the first sub-model includes a residual block and a graph attention block, wherein the residual block is trained to enhance temporal features while increasing the dimension of the target bin, and the graph attention block is trained to integrate features of mutual correlation between the targets.

7. In paragraph 6, The above processor: By inputting the first cardiac signal feature into the encoder of the second sub-model included in the above-mentioned learned artificial intelligence model, the resolution of the first cardiac signal feature is compressed to extract the core feature, By inputting the above core features into the decoder of the second sub-model, the resolution is restored and the second cardiac signal features are extracted, The above encoder includes a residual block and a max pooling layer, The above decoder includes an upsampling layer and a pre-transposed residual block, A radar-based non-contact heart condition prediction device, characterized in that the second heart signal feature is at a higher level than the first heart signal feature.

8. In paragraph 7, The above processor: By inputting the second cardiac signal feature into the third sub-model included in the above-mentioned learned artificial intelligence model, a cardiac pulse waveform in the form of a sequence is generated, A radar-based non-contact heart condition prediction device, characterized in that the third sub-model is trained to generate the heart pulse waveform by modeling the temporal flow of the second heart signal by considering the temporal context of the second heart signal features.

9. Processor; and A method performed by a radar-based non-contact heart state prediction device having a memory for storing a program executed by the processor, The above method is: A step of irradiating a subject with a radar signal whose frequency is modulated over time multiple times for a preset period of time and then obtaining a reflected signal reflected back from the subject; A step of converting the above reflection signal into structured data; - The structured data is composed of axes including a first axis and a second axis, wherein the first axis represents a range bin, and the second axis represents physiological fluctuation information of the subject; and A radar-based non-contact heart condition prediction method, characterized in that it comprises a step of selecting a target bin among the range bins based on the intensity information of the reflected signal according to the distance from the subject included in each of the range bins.

10. In paragraph 9, The above conversion step is, A radar-based non-contact heart state prediction method, characterized in that it includes a step of converting the reflected signal into a channel impulse response matrix as the structured data by applying a Fourier transform to the reflected signal.

11. In paragraph 10, The above selection step is, A step of identifying the largest reflection signal in each window divided by the above-described time interval of the above-described channel impulse response matrix; and A radar-based non-contact heart condition prediction method, characterized in that it comprises a step of selecting a range bin containing the largest reflection signal as the target bin.

12. In paragraph 10, The above selection step is, A step of determining the intensity of the reflected signal for each range bin based on the size of the channel impulse response matrix; and A radar-based non-contact heart state prediction method, characterized in that it comprises a step of selecting a range bin with the greatest intensity of the reflected signal as the target bin.

13. In paragraph 11, The above selection step is, A radar-based non-contact heart state prediction method, characterized in that it comprises a step of selecting a range bin neighboring a range bin having the greatest intensity of the reflected signal as the target bin.

14. In paragraph 9, The above method is: Further comprising a step of extracting a first cardiac signal feature by inputting the target bin into a first sub-model included in the learned artificial intelligence model, A radar-based non-contact heart state prediction method, characterized in that the first sub-model includes a residual block and a graph attention block, wherein the residual block is trained to enhance temporal features while increasing the dimension of the target bin, and the graph attention block is trained to integrate features of mutual correlation between the targets.

15. In paragraph 14, The above method is: A step of inputting the first cardiac signal feature into the encoder of the second sub-model included in the above-mentioned learned artificial intelligence model to compress the resolution of the first cardiac signal feature and extract the core feature; and Further comprising a step of inputting the above core features into the decoder of the second sub-model to restore the resolution and extract the second cardiac signal features, The above encoder includes a residual block and a max pooling layer, The above decoder includes an upsampling layer and a pre-transposed residual block, A radar-based non-contact heart condition prediction method, characterized in that the second heart signal feature is at a higher level than the first heart signal feature.

16. In paragraph 15, The above method is: Further comprising a step of generating a heart pulse waveform in the form of a sequence by inputting the second heart signal feature into a third sub-model included in the above-mentioned learned artificial intelligence model, A radar-based non-contact heart condition prediction method, characterized in that the third sub-model is trained to generate the heart pulse waveform by modeling the temporal flow of the second heart signal by considering the temporal context of the second heart signal features.

Images