Non-invasive continuous blood pressure estimation apparatus and method therefor

Abstract

The present invention relates to a non-invasive continuous blood pressure estimation apparatus and a method therefor. The non-invasive continuous blood pressure estimation apparatus according to an embodiment of the present invention comprises: a sensor unit comprising an electrocardiogram (ECG) sensor for detecting an ECG signal of a subject and a photoplethysmography (PPG) sensor for detecting a PPG signal of the subject; a data collection unit for collecting the ECG signal and the PPG signal from the sensor unit and storing the ECG signal and the PPG signal together with time information; a feature extraction unit for detecting an R-peak of the ECG signal and a specific point of the PPG signal and identifying a reference point of the signals; a concurrency correction unit for calculating a time mismatch between the ECG signal and the PPG signal and setting a parameter (Δ) for correcting the calculated time mismatch so as to secure concurrency between the two signals; and a blood pressure estimation unit for estimating systolic blood pressure (SBP) and diastolic blood pressure (DBP) by using the ECG signal and the PPG signal corrected by the concurrency correction unit.

Description

Non-invasive continuous blood pressure estimation device and method

[0001] The present invention relates to a biosignal-based non-invasive blood pressure measurement technology, and more specifically, to a non-invasive continuous blood pressure estimation device and method that estimates blood pressure using ECG signals and PPG signals.

[0002] Conventional blood pressure measurement methods typically involved intermittent measurement using a cuff; however, this approach cannot reflect continuous blood pressure changes in real time and causes inconvenience to users. Accordingly, research has been actively conducted to continuously estimate systolic blood pressure (SBP) and diastolic blood pressure (DBP) using biosignals such as electrocardiograms (ECG) and photoplethysmography (PPG).

[0003] In these existing studies, it has been pointed out as a problem that time delay or time discrepancy between ECG and PPG signals significantly affects the accuracy of blood pressure estimation. The R-peak of the ECG signal represents the electrical activity of cardiac contraction, and the waveform of the PPG signal reflects changes in blood flow in peripheral blood vessels, but delays may occur between the two signals depending on the measurement equipment, the attachment location of the sensor, or the physical condition.

[0004] Therefore, there is no way to guarantee that ECG and PPG signals generated by the same heartbeat were collected simultaneously, and this is the difficulty of non-invasive blood pressure estimation.

[0005] As a result, if the reference points of the two signals are not accurately aligned, the error in the blood pressure value increases.

[0006] According to an embodiment of the present invention, a non-invasive continuous blood pressure estimation device and method are provided that minimize blood pressure estimation errors caused by time discrepancies between ECG signals and PPG signals, and improve the accuracy and reliability of non-invasive continuous blood pressure estimation by ensuring simultaneity of the two signals.

[0007] A non-invasive continuous blood pressure estimation device according to an embodiment of the present invention comprises: a sensor unit including an ECG sensor that detects an electrocardiogram (ECG) signal of a subject and a PPG sensor that detects a photoplethysmography (PPG) signal of a subject; a data collection unit that collects the ECG signal and the PPG signal from the sensor unit and stores them together with time information; a feature extraction unit that detects an R-peak of the ECG signal and a specific point of the PPG signal to identify a reference point of the signal; a simultaneity correction unit that calculates a time discrepancy between the ECG signal and the PPG signal and sets a parameter (Δ; Delta) to correct the calculated time discrepancy to ensure simultaneity between the two signals; and a blood pressure estimation unit that estimates systolic blood pressure (SBP) and diastolic blood pressure (DBP) using the ECG signal and the PPG signal corrected from the simultaneity correction unit.

[0008] A non-invasive continuous blood pressure estimation method according to an embodiment of the present invention comprises the steps of detecting an electrocardiogram (ECG) signal of a subject, detecting a photoplethysmography (PPG) signal of a subject, collecting the ECG signal and the PPG signal and storing them together with time information, detecting an R-peak of the ECG signal and a specific point of the PPG signal to identify a reference point of the signal, calculating a time discrepancy between the ECG signal and the PPG signal and setting a parameter (Δ) to correct the calculated time discrepancy to ensure simultaneity between the two signals, and estimating systolic blood pressure (SBP) and diastolic blood pressure (DBP) using the corrected ECG signal and PPG signal.

[0009] According to an embodiment, the data collection unit can perform data synchronization by collecting the ECG signal and the PPG signal at the same sampling frequency and aligning the signals based on the time axis.

[0010] According to an embodiment, the feature extraction unit can extract one or more biosignal features from the ECG signal and the PPG signal, including heart rate (HR), pulse rate (PR), pulse arrival time (PAT), systolic timespan (ST), photoplethysmography intensity (PI), branch width (BW), systolic BW (SBW), diastolic BW (DBW), peak (P), valley (V), intersecting tangent (IT), and maximal derivative (MD).

[0011] According to an embodiment, the simultaneity correction unit sets the time difference from the R-peak of the ECG signal to any one of the peak (P), trough (V), cross tangent point (IT), and maximum slope point (MD) of the PPG signal as a parameter (Δ), and during the learning process, updates the parameter (Δ) through backpropagation to correct the time discrepancy between the ECG signal and the PPG signal.

[0012] According to an embodiment, the blood pressure estimation unit can estimate the systolic blood pressure and the diastolic blood pressure through a blood pressure estimation model including a Temporal Convolutional Network (TCN), using feature values ​​extracted from the corrected ECG signal and PPG signal as input.

[0013] According to an embodiment, the blood pressure estimation model includes a plurality of temporal blocks, each temporal block includes two convolutional layers, and the dilation factor may increase exponentially depending on the depth of the convolutional layers.

[0014] According to an example, the blood pressure estimation model may prevent overfitting by configuring the input sequence with 16 consecutive cardiac cycles, including three hidden layers, and applying dropout.

[0015] According to an embodiment, the blood pressure estimation model can be trained by using the Mean Squared Error (MSE) as the loss function and applying the Rectified Adam (RAdam) algorithm as the optimizer.

[0016] According to an embodiment, the blood pressure estimation unit can calculate the mean arterial pressure (MAP) using systolic blood pressure and diastolic blood pressure values ​​calculated through a learned blood pressure estimation model.

[0017] According to an embodiment, the blood pressure estimation model can be trained by dividing the training dataset into training, validation, and test sets in a 2:1:1 ratio, and can save the model at the point where validation loss is minimized.

[0018] According to an embodiment of the present invention, the simultaneity of the two signals can be ensured by quantitatively correcting the time discrepancy between the ECG signal and the PPG signal during a non-invasive continuous blood pressure estimation process. Accordingly, the estimation accuracy of systolic blood pressure (SBP) and diastolic blood pressure (DBP) is improved, and the ease of use and continuous monitoring performance are significantly enhanced compared to conventional cuff-based blood pressure measurement methods.

[0019] Furthermore, by optimizing the parameter (Δ) during the learning process, environmental variation factors such as sensor type, attachment location, and measurement delay can be automatically corrected. As a result, stable and highly reliable blood pressure estimation is possible even in diverse biosignal environments.

[0020] FIG. 1 is a block diagram showing a non-invasive continuous blood pressure estimation device according to an embodiment of the present invention.

[0021] FIG. 2 is a diagram showing features extracted from ECG signals and PPG signals in a non-invasive continuous blood pressure estimation device according to an embodiment of the present invention.

[0022] FIG. 3 is a diagram showing an example of a data window used in the feature extraction process in a non-invasive continuous blood pressure estimation device according to an embodiment of the present invention.

[0023] Figure 4 is a diagram showing the structure of a time-series convolutional neural network used in a non-invasive continuous blood pressure estimation device according to an embodiment of the present invention.

[0024] FIG. 5 is a diagram showing the overall operation flow of a non-invasive continuous blood pressure estimation device according to an embodiment of the present invention.

[0025] Figure 6 is a diagram showing the number of feature sets by age group of a non-invasive continuous blood pressure estimation device according to an embodiment of the present invention.

[0026] FIG. 7 is a histogram showing the blood pressure distribution according to age percentiles of a non-invasive continuous blood pressure estimation device according to an embodiment of the present invention.

[0027] FIG. 8 is a diagram showing the distribution of extracted feature values ​​in a non-invasive continuous blood pressure estimation device according to an embodiment of the present invention.

[0028] FIG. 9 is a scatter plot showing the actual blood pressure (BP) and the blood pressure estimated by a blood pressure estimation model in a non-invasive continuous blood pressure estimation device according to an embodiment of the present invention.

[0029] FIG. 10 shows a Bland-Altman plot indicating the degree of agreement between actual blood pressure (BP) and blood pressure estimated by a blood pressure estimation model in a non-invasive continuous blood pressure estimation device according to an embodiment of the present invention.

[0030] Figure 11 is a figure showing the results of analyzing the importance and influence of each feature used for blood pressure estimation in a blood pressure estimation model according to an embodiment of the present invention as SHAP values ​​(Shapley Additive Explanations).

[0031] FIG. 12 is a flowchart of a non-invasive continuous blood pressure estimation method according to an embodiment of the present invention.

[0032] FIG. 13 illustrates a non-invasive continuous blood pressure estimation device according to one embodiment of the present invention.

[0033] Hereinafter, the present invention will be described in detail with reference to the attached drawings. However, this is merely illustrative and the present invention is not limited to the specific embodiments described illustratively.

[0034] The terms used herein are for describing the embodiments and are not intended to limit the invention. In this specification, the singular form includes the plural form unless specifically stated otherwise in the text. As used herein, "comprises" and / or "made of" do not exclude the presence or addition of one or more other components, steps, actions, and / or elements to the mentioned components, steps, actions, and / or elements.

[0035] In addition, terms such as "~part," "~unit," "~block," "~part," and "~module" may refer to a unit that processes at least one function or operation. For example, the above terms may refer to at least one piece of hardware such as an FPGA (Field Programmable Gate Array) or ASIC (Application Specific Integrated Circuit), at least one piece of software stored in memory, or at least one process processed by a processor.

[0036] Unless otherwise defined, all terms used in this specification (including technical and scientific terms) may be used in a meaning commonly understood by those skilled in the art to which the present invention pertains. Additionally, terms defined in commonly used dictionaries are not to be interpreted ideally or excessively unless explicitly and specifically defined otherwise.

[0037]

[0038] FIG. 1 is a block diagram showing a non-invasive continuous blood pressure estimation device according to an embodiment of the present invention.

[0039] Referring to FIG. 1, a non-invasive continuous blood pressure estimation device according to an embodiment of the present invention comprises: a sensor unit (100) including an ECG sensor (120) for detecting an electrocardiogram (ECG) signal of a subject and a PPG sensor (140) for detecting a photoplethysmography (PPG) signal of a subject; a data collection unit (220) for collecting ECG signals and PPG signals from the sensor unit (100) and storing them together with time information; a feature extraction unit (240) for detecting an R-peak of an ECG signal and a specific point of a PPG signal to identify a reference point of the signal; a simultaneity correction unit (260) for calculating a time discrepancy between an ECG signal and a PPG signal and setting a parameter (Δ; Delta) for correcting the calculated time discrepancy to ensure simultaneity between the two signals; and a systolic blood pressure (SBP) and a PPG signal corrected from the simultaneity correction unit (260) using the ECG signal and PPG signal. It includes a blood pressure estimation unit (280) that estimates diastolic blood pressure (DBP).

[0040] A sensor unit (100) according to an embodiment of the present invention is for detecting a subject's biological signal and includes an ECG sensor (120) for detecting an ECG signal and a PPG sensor (140) for detecting a PPG signal.

[0041] The above ECG sensor (120) is attached to the surface of the subject's skin to detect changes in potential difference due to the electrical activity of the heart, and can typically be implemented in a 3-electrode or 5-electrode manner. Specifically, the ECG sensor (120) detects a waveform including an R-peak that occurs according to the heartbeat, and can identify the depolarization and repolarization activities of the heart from this waveform.

[0042] The above PPG sensor (140) is attached to or in contact with the subject's body (e.g., finger, earlobe, or wrist) to detect an optical signal corresponding to a change in blood flow in a peripheral blood vessel area. For example, it includes a light source (e.g., LED) and a photodetector, and utilizes the principle that the intensity of reflected or transmitted light changes according to fluctuations in blood flow as light emitted from the light source passes through skin tissue.

[0043] In this case, the signal detected by the PPG sensor (140) includes waveform shapes of the systole and diastole sections and consists of a continuous waveform including the rising part, peak, falling part, and intersecting tangent of the pulse wave.

[0044] The ECG sensor (120) and the PPG sensor (140) are attached to the same subject, and signal detection is performed in parallel in real time.

[0045]

[0046] The data collection unit (220) according to an embodiment of the present invention collects the ECG signal and PPG signal output from the sensor unit (100) and stores them together with time information.

[0047] The data collection unit (220) collects signals input from the ECG sensor (120) and the PPG sensor (140) at the same sampling frequency and performs data synchronization by aligning the signals based on the same time axis.

[0048] This synchronization ensures that the electrical activity of the electrocardiogram and the blood flow fluctuations of the photoplethysmography correspond within the same cardiac cycle, thereby enabling the same physiological events to be accurately compared later in the feature extraction unit (240) and the blood pressure estimation unit (280).

[0049] Specifically, the sampling frequency can be set in the range of 125 Hz to 250 Hz, and in one embodiment, it can be set to 250 Hz to acquire detailed waveforms of the two signals without distortion. In the case of the ECG signal, high frequency resolution is required to accurately distinguish heartbeat intervals, and in the case of the PPG signal, it must be possible to precisely detect minute amplitude fluctuations of changes in blood flow. In this case, even if the two signals are detected through different sensors, they are collected at the same sampling standard, so signal distortion caused by mismatch in the time axis is prevented.

[0050] According to one embodiment, the data collection unit (220) includes a function to perform time axis alignment of the signals. Time axis alignment is performed based on the start and end times of the ECG signal and the PPG signal, and is synchronized based on the time information (Time Stamp) assigned to each signal.

[0051] Specifically, the data collection unit (220) aligns specific points such as the peak, intersecting tangent (IT), and maximal derivative (MD) of the PPG signal based on the time of detection of the R-peak of the ECG signal. For example, data is arranged so that the PPG peak after a certain time corresponds to the time when the ECG R-peak is detected, and errors caused by time mismatch are minimized in this process.

[0052] According to one embodiment, the data collection unit (220) performs the function of storing the collected signal along with time information. The data stored at this time includes the voltage value, light intensity value, and time stamp at each sampling point, and subsequently, the feature extraction unit (240) uses the data to search for a reference point of the biosignal.

[0053] According to one embodiment, the data collection unit (220) may store a signal in an internal memory or an external storage device and may be configured in parallel with a buffer memory for real-time processing. Additionally, the data collection unit (220) may communicate with the sensor unit (100) via a wired interface (e.g., USB, RS-232) or a wireless interface (e.g., BLE, Wi-Fi).

[0054] In this way, the data collection unit (220) according to the embodiment of the present invention can automatically perform time-axis alignment during the collection process to increase the synchronization accuracy of ECG and PPG signals, and can improve the quality of the training data in the subsequent step using the synchronized signal.

[0055] In particular, by maintaining the same sampling frequency to prevent temporal distortion of the signal and ensuring that each physiological event corresponds consistently within the same heart rate cycle through time-axis alignment, the reliability and prediction accuracy of non-invasive continuous blood pressure estimation can be improved.

[0056]

[0057] The feature extraction unit (240) according to an embodiment of the present invention is for identifying reference points of ECG signals and PPG signals.

[0058] The feature extraction unit (240) detects the R-peak of the ECG signal and extracts specific points such as the peak, valley, intersecting tangent (IT), and maximal derivative (MD) of the PPG signal to identify each reference point. The reference points extracted in this way are used as basic data to define the temporal relationship between the two signals.

[0059] Specifically, the time interval between the R-peak point of the ECG signal and a specific point of the PPG signal can be defined by characteristics such as Pulse Arrival Time (PAT) or Systolic Timespan (ST).

[0060]

[0061] As described above, the feature extraction unit (240) according to an embodiment of the present invention can extract one or more biosignal features from the ECG signal and the PPG signal, including heart rate (HR; Heart Rate), pulse rate (PR; Pulse Rate), pulse arrival time (PAT; Pulse Arrival Time), systolic timespan (ST; Systolic Timespan), photoplethysmography intensity (PI; PPG Intensity), branch width (BW; Branch Width), systolic branch width (SBW; Systolic BW), diastolic branch width (DBW; Diastolic BW), peak (P; Peak), valley (V; Valley), intersecting tangent (IT; Intersecting Tangent), and maximal derivative (MD; Maximal Derivative).

[0062] As illustrated in FIG. 2, the feature extraction unit (240) of the present invention detects various biosignal features having physiological meaning from ECG signals and PPG signals, and constructs input data for blood pressure estimation using these features. The feature extraction unit (240) searches for reference points and singularities within the waveform of the collected signals, and derives each feature value by calculating quantitative information such as time difference, amplitude, slope, and area between the searched signal segments.

[0063] According to one embodiment, the feature extraction unit (240) analyzes both the ECG signal and the PPG signal.

[0064] The aforementioned ECG signal is a waveform reflecting the electrical activity of the heart, and analyzes the QRS complex containing the R-peak, as well as the adjacent P-wave and T-wave. The ECG signal is used as the primary signal serving as the basis for calculating the heart rate (HR). Specifically, the HR is calculated by measuring the time interval between the R-peaks of the ECG signal (RR Interval), and this value is converted into beats per minute (bpm).

[0065] The above PPG signal is an optical signal that reflects changes in the subject's blood flow and includes points such as peaks, valleys, intersecting tangents (IT), and maximal derivatives (MD) on the time axis. The PPG signal is a key signal that reflects the pulsating characteristics of blood pressure, and key parameters such as pulse arrival time (PAT) can be calculated through the time difference with the ECG signal.

[0066] Heart rate (HR) is a value calculated based on the time interval between consecutive R-peaks of the ECG signal. Pulse rate (PR) is derived from the time interval between consecutive peaks of the PPG signal and represents the frequency of blood flow pulses. Pulse time to travel (PAT) is a value calculated as the time difference from the R-peak of the ECG signal to a specific point in the PPG signal (e.g., Peak, IT, MD, etc.) and is closely related to arterial elasticity and blood flow velocity.

[0067] In the PPG signal, the peak (P) represents the highest point of the waveform and indicates the point of peak systolic blood flow, while the trough (V) represents the lowest point of the waveform and reflects the blood flow state at the end of diastole. The tangent point (IT) indicates the point where the tangent to the baseline intersects at a constant angle during the rising portion of the waveform and reflects changes in arterial wall elasticity and blood flow velocity. The maximum slope point (MD) is the point where the rate of change is maximum during the rising portion of the waveform and is used as a key factor representing the acceleration of blood flow.

[0068] Systolic interval time (ST) refers to the time of the rise interval from the trough point (V) to the peak point (P) in the PPG signal and is related to the duration of ventricular contraction. Photoplethysmography intensity (PI) is the intensity of the optical signal calculated based on the maximum amplitude of the PPG signal, quantifying changes in blood flow volume due to changes in blood light absorption. Biscuttle width (BW) refers to the width between the trough point (V) and the intersection tangent point (IT) in the PPG signal, and systolic biscuttle width (SBW) is defined as the width between the trough point (V) and the maximum slope point (MD). Diastolic biscuttle width (DBW) represents the width from the intersection tangent point (IT) to the next trough point (V) and reflects peripheral vasodilation characteristics.

[0069] The feature extraction unit (240) automatically extracts the above-mentioned biosignal features through a waveform analysis algorithm. Specifically, it includes procedures such as a peak detection algorithm, moving average-based noise removal filtering, slope detection using a differentiation technique, and threshold-based waveform segment separation. For example, in the case of an ECG signal, an R-peak can be detected using the Pan-Tompkins algorithm, and in the case of a PPG signal, the intersection tangent point (IT) and maximum slope point (MD) can be found using the first derivative curve of the waveform.

[0070] FIG. 3 is a diagram showing an example of a data window used in the feature extraction process in a non-invasive continuous blood pressure estimation device according to an embodiment of the present invention.

[0071] Figures 3A and 3B show examples of appropriate waveforms where feature extraction proceeds smoothly, and Figure 3C shows an example of an inappropriate waveform, respectively. In A, a clear valley exists between the second R-peak and the first PPG peak, so PAT V and PAT P Key features can be easily extracted. B shows a case where the corresponding bottom is absent, requiring consideration of the next PPG peak. In C, the PPG bottom corresponding to the second R-peak is located above the baseline and is considered an inappropriate signal; therefore, this signal is removed during the feature extraction process.

[0072] According to one embodiment, the feature extraction unit (240) converts data into a format suitable for machine learning while maintaining the physiological meaning of the signal. To this end, each feature value undergoes a normalization process to be scaled to a certain range and is then input into the blood pressure estimation unit (280). In this process, the extraction accuracy of each feature value depends on the simultaneity of the ECG signal and the PPG signal, and reliability is enhanced by calculating based on the signal corrected by the simultaneity correction unit (260).

[0073] In this way, the feature extraction unit (240) according to the embodiment of the present invention accurately extracts a plurality of biosignal features from the ECG signal and the PPG signal, thereby improving the quality of the input data of the blood pressure estimation model (282). Accordingly, the influence of noise or distortion within the signal is minimized, and physiological changes according to the heart rate cycle are precisely reflected, thereby improving the reliability and prediction accuracy of non-invasive continuous blood pressure estimation.

[0074]

[0075] The simultaneity correction unit (260) according to an embodiment of the present invention calculates the time discrepancy between the ECG signal and the PPG signal, and sets a parameter (Δ) to correct the calculated time difference to ensure simultaneity between the two signals.

[0076] The above simultaneity correction unit (260) calculates the time difference from the R-peak point of the ECG signal to a specific point of the PPG (peak point, trough point, intersection tangent point, maximum slope point, etc.) and defines this time difference as Δ.

[0077] Δ is not a value fixed by prior measurement, but a correction parameter that is automatically updated through backpropagation during the learning process. In other words, Δ dynamically corrects the synchronization error between signals to align the timing of occurrence of the ECG signal and the PPG signal.

[0078] In this case, as Δ is updated, the temporal discrepancy between the two signals gradually decreases, and as a result, the ECG signal and PPG signal occurring within the same cardiac cycle are aligned simultaneously.

[0079]

[0080] A simultaneity correction unit (260) according to an embodiment of the present invention sets the time difference from the R-peak of the ECG signal to any one of the peak (P), trough (V), cross tangent point (IT), and maximum slope point (MD) of the PPG signal as a parameter (Δ), and during the learning process, updates the parameter (Δ) through backpropagation to correct the time discrepancy between the ECG signal and the PPG signal.

[0081] The above-mentioned time-series correction unit calculates the time difference between two signals, defines the calculated time difference as a parameter (Δ), and can secure the temporal simultaneity of the ECG signal and the PPG signal by updating the parameter through backpropagation during the learning process.

[0082] According to one embodiment, the synchronicity correction unit (260) can improve the synchronization accuracy of the ECG signal and the PPG signal. The ECG signal represents the electrical activity of the heart, and the R-peak is used as a reference point for the heart rate cycle. The PPG signal is an optical waveform corresponding to changes in blood flow volume, and the peak (P; Peak), valley (V; Valley), intersecting tangent (IT; Intersecting Tangent), and maximal derivative (MD; Maximal Derivative) within the waveform are defined as key feature points. The synchronicity correction unit (260) sets the time difference between the R-peak of the ECG signal and any one of the peak, valley, intersecting tangent, or maximal derivative points of the PPG signal as Δ. At this time, Δ is a temporal correction variable that reflects whether the two signals actually occurred in the same cardiac cycle.

[0083] The concurrency correction unit (260) defines the calculated Δ as a learnable parameter and automatically updates Δ through backpropagation during the learning process of the blood pressure estimation model (282). That is, Δ is not a fixed value in advance, but is continuously adjusted during the learning process by reflecting the characteristics of the data and the correlation between the signals. In this process, the learning algorithm repeatedly modifies the value of Δ so that the time difference between the ECG signal and the PPG signal is minimized. For example, if the predicted blood pressure of the blood pressure estimation model (282) differs from the actual blood pressure, Δ is adjusted through backpropagation so that the synchronization precision of the signal can be improved in the next learning iteration.

[0084] Specifically, the synchronicity correction unit (260) serves to align the R-peak point of the ECG signal and the reference point (peak, trough, intersection tangent point, maximum slope point, etc.) of the PPG signal on the time axis according to the update of Δ. For example, if the time difference from the ECG R-peak point to the PPG peak is defined as Δ, the two signals are aligned to correspond to the same physiological event (e.g., heart contraction and blood flow dilation) as Δ converges to an optimal value during learning. At this time, the synchronicity correction unit (260) minimizes the time error of the signals so that the ECG signal and the PPG signal occurring within the same heart rate cycle correspond.

[0085] According to one embodiment, the simultaneity correction process is performed after signal collection and before the feature extraction step, and the value of Δ acts as a pre-correction variable to increase the consistency of input data of the blood pressure estimation model (282). That is, Δ is a correction factor to solve the non-simultaneity problem of ECG signals and PPG signals, and through this, the quality and reliability of the training data can be improved.

[0086] According to the embodiment, Δ is generally set to a small value in the order of milliseconds (ms) and can change dynamically depending on the physiological state of the subject or the delay characteristics of the measurement device. For example, the initial value of Δ may differ if ECG signals and PPG signals are collected from different channels or if the sensor attachment locations (wrist, ear, finger, etc.) are different. However, as Δ is updated through backpropagation during the learning process, it can be aligned to the same physiological timing regardless of the signal collection environment.

[0087] According to the embodiment, the implementation method of the simultaneity correction unit (260) can be hardware or software. In hardware, it may include a time difference detector or a delay corrector within the signal processing circuit, and in software, it may be configured to automatically adjust during the error backpropagation process by inserting Δ as a variable within a deep learning-based learning loop.

[0088] The simultaneity correction unit (260) according to this embodiment of the present invention is a dynamic correction through a Δ parameter. In conventional technology, the time difference between the ECG signal and the PPG signal is considered as a fixed delay value or corrected by a passive post-processing method. However, in the present invention, Δ is defined as a learnable variable and is automatically updated during the learning process, thereby allowing the time discrepancy between the two signals to be actively resolved.

[0089] In this way, the simultaneity correction unit (260) according to the embodiment of the present invention defines the time discrepancy between the ECG signal and the PPG signal as a parameter (Δ), and automatically updates Δ through error backpropagation during the learning process, thereby ensuring temporal simultaneity between the two signals. Accordingly, the ECG signal and the PPG signal are aligned as physiological events occurring in the same heart rate cycle, and as a result, the consistency of the input data and the accuracy of signal correspondence are improved, thereby improving the reliability and prediction accuracy of non-invasive continuous blood pressure estimation.

[0090]

[0091] The blood pressure estimation unit (280) according to an embodiment of the present invention can estimate systolic blood pressure (SBP) and diastolic blood pressure (DBP) using the ECG signal and PPG signal corrected from the simultaneity correction unit (260).

[0092] The above blood pressure estimation unit includes a computation unit equipped with a machine learning or deep learning-based blood pressure estimation model, and specifically, may include a structure such as a time series convolutional neural network (TCN).

[0093] The blood pressure estimation unit (280) takes feature values ​​(e.g., PAT, ST, PI, BW, etc.) extracted from two signals as input and calculates blood pressure through a blood pressure estimation model (282) including a Temporal Convolutional Network (TCN).

[0094]

[0095] Figure 4 is a diagram showing the structure of a Temporal Convolutional Network (TCN) used in a non-invasive continuous blood pressure estimation device according to an embodiment of the present invention.

[0096] As described above, the blood pressure estimation unit (280) according to an embodiment of the present invention can estimate the systolic blood pressure and the diastolic blood pressure through a blood pressure estimation model (282) including a Temporal Convolutional Network (TCN) using feature values ​​extracted from the corrected ECG signal and PPG signal as input.

[0097] According to one embodiment, the blood pressure estimation unit (280) can predict blood pressure from the feature values ​​of the input biosignals. The blood pressure estimation unit (280) receives the feature values ​​of the ECG signal and PPG signal transmitted from the feature extraction unit (240) as input and applies a deep learning model for analyzing time series data.

[0098] According to one embodiment, the blood pressure estimation model (282) included in the blood pressure estimation unit (280) includes a Temporal Convolutional Network (TCN). The Temporal Convolutional Network has a structure capable of maintaining time dependency while processing time-series input data in parallel. The TCN applied in this embodiment processes the input sequence according to time order and is designed so that the signal at each time point influences the prediction at subsequent times points.

[0099] According to one embodiment, the blood pressure estimation unit (280) is composed of an input layer, a hidden layer, and an output layer. The input layer receives feature values ​​extracted from corrected ECG signals and PPG signals, and these feature values ​​include heart rate (HR), pulse rate (PR), pulse arrival time (PAT), systolic timespan (ST), PPG intensity (PI), branch width (BW), systolic BW (SBW), diastolic BW (DBW), peak (P), valley (V), intersecting tangent (IT), and maximal derivative (MD). The hidden layer consists of multiple convolutional layers, activation functions, and normalization layers, and converts input time series data into learnable feature maps. The output layer calculates systolic and diastolic blood pressure based on the learned feature maps.

[0100] A blood pressure estimation unit (280) according to one embodiment recognizes the temporal pattern of input data by applying a filter in each convolutional layer. At this time, the convolutional kernel is extended in the direction of the time axis and can extract multiple levels of temporal information depending on the length of the signal. For example, the initial layer learns short-term signal patterns (e.g., changes within an individual heart rate cycle), and the deep layer learns long-term patterns (e.g., fluctuations between consecutive heart rate cycles). Through this hierarchical learning structure, the blood pressure estimation unit (280) can reflect not only instantaneous fluctuations of individual heart rates but also long-term blood pressure trends.

[0101] According to one embodiment, the blood pressure estimation unit (280) processes a fixed number of cardiac cycles as a learning unit by fixing the length of the input sequence. Specifically, the input sequence consists of 16 consecutive cardiac cycles, which is set to a range sufficient to capture the temporal context of blood pressure changes. Additionally, the blood pressure estimation unit (280) can prevent overfitting by applying a dropout technique.

[0102] The blood pressure estimation unit (280) according to an embodiment of the present invention may have a structure that gradually increases the dilation factor for input feature values ​​in each temporal block. This structure enables efficient learning of the long-term dependency of the signal and can more precisely reflect the correlation between consecutive heart rate cycles.

[0103] As such, the blood pressure estimation unit (280) according to the embodiment of the present invention can reflect both the temporal variation of the ECG signal and the PPG signal and the morphological information of the waveform by estimating blood pressure through a time-series convolutional neural network using feature values ​​extracted from corrected ECG and PPG signals as input. Accordingly, changes in systolic and diastolic blood pressure can be predicted continuously and non-invasively, and precise blood pressure estimation based on temporal continuity is possible.

[0104]

[0105] A blood pressure estimation model (282) according to an embodiment of the present invention includes a plurality of temporal blocks, each temporal block includes two convolutional layers, and the dilation factor may increase exponentially depending on the depth of the convolutional layers.

[0106] As illustrated in FIG. 4, the blood pressure estimation model (282) according to an embodiment of the present invention includes a plurality of temporal blocks, and each temporal block may be structured to include two convolutional layers. In addition, the model is configured so that the dilation factor increases exponentially with the depth of each convolutional layer, thereby enabling efficient learning of the long-term dependency of the input temporal signal.

[0107] A blood pressure estimation model (282) according to one embodiment includes a structure for recognizing temporal patterns and maintaining long-term signal correlations. Specifically, a time series block is a minimum unit for processing feature values ​​extracted from input ECG signals and PPG signals, and can analyze the time series characteristics of the signal in a multi-layered manner. Each time series block consists of two consecutive convolutional layers, and a temporal feature map is extracted from input data through convolutional operations and passed through an activation function to perform a non-linear transformation. The first convolutional layer learns short-term time series dependencies by applying a relatively small dilation factor, and the second convolutional layer reflects long-term dependencies by applying a larger dilation factor.

[0108] According to one embodiment, the dilation factor applied in the time series block increases exponentially with the depth of the convolutional layer. For example, the dilation factor may increase to 1 in the first time series block, 2 in the second block, 4 in the third block, and 8 in the fourth block. This dilation factor structure enables the blood pressure estimation model (282) to capture long-term dependencies of the input signal and to simultaneously learn blood pressure change patterns over multiple cardiac cycles as well as short-term fluctuations.

[0109] According to one embodiment, each time series block includes a normalization layer and an activation function layer (ReLU, Rectified Linear Unit, etc.) in addition to the convolutional layer to increase learning stability and prevent overfitting.

[0110] According to one embodiment, the time series block within the blood pressure estimation model (282) is designed to perform parallel operations, resulting in higher computational efficiency compared to a traditional recurrent neural network (RNN). That is, since calculations for all time points of the input sequence can be performed simultaneously, long-term time series data of ECG signals and PPG signals can be processed quickly and efficiently.

[0111] According to one embodiment, the blood pressure estimation model (282) may be composed of a total of 4 to 8 time series blocks, and the filter size of each block may be set to a range of 3 to 5. This setting is a structure optimized to reflect both short-period characteristics (e.g., interval between waveform peaks, amplitude change, etc.) and long-period characteristics (e.g., average pulse wave transit time, heart rate variability, etc.) of the ECG signal and PPG signal.

[0112] According to one embodiment, the blood pressure estimation model (282) may apply a padding technique to maintain a constant output length regardless of the length of the input signal. As a result, the output of the time series block is maintained to be the same as the length of the input sequence, and the corresponding blood pressure value for each heart rate cycle can be stably predicted. In addition, the blood pressure estimation model (282) is suitable for real-time blood pressure estimation environments because it continuously predicts systolic blood pressure (SBP) and diastolic blood pressure (DBP) by arranging the feature values ​​of the input ECG signal and PPG signal in chronological order.

[0113] According to one embodiment, the blood pressure estimation model (282) can simultaneously learn short-term or long-term signal dependencies through an exponential increase in the latency rate within a time series block. That is, as one moves toward deeper layers of the blood pressure estimation model (282), signal patterns covering a wider time range are learned, which accurately reflect the complex interaction between the ECG signal and the PPG signal. Due to these structural features, the blood pressure estimation model (282) can achieve the high precision and stability required for non-invasive continuous blood pressure estimation.

[0114]

[0115] A blood pressure estimation model (282) according to an embodiment of the present invention can be trained by using the Mean Squared Error (MSE) as the loss function and applying a normalized Adam (RAdam; Rectified Adam) algorithm as the optimizer.

[0116] The blood pressure estimation model (282) can predict systolic blood pressure (SBP) and diastolic blood pressure (DBP) using feature values ​​of input ECG and PPG signals. At this time, the blood pressure estimation model (282) can quantify the prediction error by using the mean squared error (MSE) as a loss function. The mean squared error is the average of the squared differences between the predicted value and the actual value, and is the most common regression loss function for evaluating the magnitude of the overall error that the blood pressure estimation model (282) generates in the prediction.

[0117] According to one embodiment, the mean squared error loss function used in the learning process of the blood pressure estimation model (282) can adjust the output of the blood pressure estimation model (282) by simultaneously considering the prediction accuracy of systolic blood pressure and diastolic blood pressure. That is, in the learning process, the difference between the predicted blood pressure value and the actual blood pressure value is calculated for each learning iteration (Epoch), and the average of the squared values ​​is calculated as the loss value. A smaller loss value means that the blood pressure estimation model (282) predicts the actual blood pressure more accurately, and the learning can be repeated until the loss value is minimized.

[0118] According to one embodiment, the blood pressure estimation model (282) may apply a Rectified Adam (RAdam) algorithm as an optimizer to improve the efficiency and stability of learning. Adaptive Moment Estimation (Adam) optimizes using an adaptive learning rate. However, the existing Adam algorithm may have a slow convergence speed or oscillations due to an unstable adaptive rate in the early stages of learning. To improve this problem, an embodiment of the present invention applies a Rectified Adam (RAdam) algorithm to automatically adjust the learning rate in the early stages of learning and ensure stable convergence.

[0119] This RAdam algorithm can solve the problem of excessive fluctuation in the learning rate by correcting the bias of the moment estimate during the early stages of learning. Through this, the blood pressure estimation model (282) can learn stably while responding to the diversity of the data and the variability of the biosignal. In addition, RAdam has improved learning speed and convergence stability compared to the existing Adam, so the learning efficiency of the blood pressure estimation model (282) from ECG and PPG signals can be improved.

[0120] According to the embodiment, the training of the blood pressure estimation model (282) is repeated approximately 200 times (epochs), and the learning rate can be set within the range of 0.0005. At this time, the parameters of the blood pressure estimation model (282) are updated through the backpropagation process by the mean squared error loss function, and the optimizer finely adjusts the parameters at each training step to minimize the loss. In addition, a validation dataset is used in parallel during the training process to prevent overfitting and maintain the generalization performance of the blood pressure estimation model (282).

[0121] According to the embodiment, after the completion of training, the blood pressure estimation model (282) predicts blood pressure based on actual ECG and PPG signals during the verification phase, and the performance of the blood pressure estimation model (282) is evaluated through the mean squared error between the predicted value and the actual value. As in the embodiment, a structure that applies the mean squared error as a loss function and uses a normalized Adam algorithm as an optimizer enables stable training while reflecting both the characteristics of non-linear biological signals and time-series patterns.

[0122] As such, by using the mean squared error (MSE) as a loss function and applying a normalized Adam (RAdam) algorithm as an optimizer, the learning stability of the blood pressure estimation model (282) can be improved and the prediction error reduced. Accordingly, the accuracy of systolic and diastolic blood pressure estimated non-invasively from ECG and PPG signals is increased, and the reliability and reproducibility of continuous blood pressure estimation can be improved.

[0123]

[0124] The blood pressure estimation unit (280) according to an embodiment of the present invention can calculate the mean arterial pressure (MAP) using systolic blood pressure and diastolic blood pressure values ​​calculated through a learned blood pressure estimation model (282).

[0125] The blood pressure estimation unit (280) according to the embodiment may include a computation module for calculating a physiologically valid mean arterial pressure based on a blood pressure value estimated non-invasively. Mean arterial pressure refers to the average pressure over a certain period of time while blood circulates through the artery, and is used as a key indicator for evaluating the cardiovascular condition. Generally, mean arterial pressure is calculated using systolic blood pressure and diastolic blood pressure, and in the embodiment of the present invention, it is calculated by applying the two blood pressure values ​​estimated through a learned blood pressure estimation model (282) to a computation formula.

[0126] According to one embodiment, the mean arterial pressure is a representative physiological indicator that reflects changes in blood pressure during a heart rate cycle, and the blood pressure estimation unit (280) includes a structure for precisely calculating it. Specifically, the mean arterial pressure can be calculated according to the following relationship.

[0127] MAP=DBP+1 / 3(SBP-DBP)

[0128] In the above equation, SBP represents systolic blood pressure and DBP represents diastolic blood pressure, and the equation is a value calculated by weighting the intervals in which blood pressure changes during the heart rate cycle. The blood pressure estimation unit (280) can perform this calculation process to calculate the mean arterial pressure in real time based on the SBP and DBP estimated from the ECG signal and PPG signal.

[0129] According to the embodiment, the blood pressure estimation unit (280) can update the mean arterial pressure at each point in time based on continuous time-series data input, thereby reflecting the trend of blood pressure fluctuation over time in real time. In addition, the calculated mean arterial pressure can be provided as a reference signal for evaluating the user's condition, feedback control, or linking with medical equipment. This allows the non-invasive blood pressure estimation device of the present invention to go beyond simple blood pressure value prediction and produce a blood pressure indicator with physiological reliability.

[0130] According to an embodiment, the blood pressure estimation unit (280) includes a dedicated computation module for calculating the mean arterial pressure, and the computation module mathematically processes the blood pressure value output from the learned blood pressure estimation model (282). Additionally, a moving average or filtering technique may be additionally applied to remove signal outliers during the calculation of the mean arterial pressure. In this case, averaging is performed on the data of consecutive heart rate intervals so that distortion caused by outliers in a single heart rate cycle is minimized.

[0131] A blood pressure estimation unit (280) according to one embodiment may store the calculated mean arterial pressure together with other physiological signal indicators (e.g., heart rate, pulse wave transit time, etc.) or transmit it to an external display and control system. This allows the user to monitor the blood pressure status in real time and analyze the trend of blood pressure changes over the long term. Additionally, since mean arterial pressure is a clinically important indicator, the device of the present invention is also applicable to medical diagnosis and preventive health management.

[0132] In this way, the blood pressure estimation unit (280) according to the embodiment of the present invention calculates the average arterial pressure using the systolic blood pressure and diastolic blood pressure values ​​calculated through the learned blood pressure estimation model (282), thereby providing accurate and reliable average blood pressure information in a non-invasive manner.

[0133]

[0134] A blood pressure estimation model (282) according to an embodiment of the present invention can be trained by dividing the training dataset into training, validation, and test sets in a 2:1:1 ratio, and can save the model at the point where the validation loss is minimized.

[0135] According to one embodiment, the training dataset of the blood pressure estimation model (282) includes various biosignal feature values ​​extracted from ECG signals and PPG signals, and these feature values ​​are used as input data to learn the relationship with systolic blood pressure (SBP) and diastolic blood pressure (DBP). The dataset is organized into units of data windows of a certain length, taking into account temporal continuity, and each window consists of multiple biosignal samples. In the step of constructing the training dataset, about 50% of the total data is allocated as the training set, 25% as the validation set, and the remaining 25% as the test set.

[0136] The above training set is used to train the parameters of the blood pressure estimation model (282). The blood pressure estimation model (282) repeatedly learns the relationship between the input signals included in the training set and the actual blood pressure values, and updates the weights of the blood pressure estimation model (282) in a direction that minimizes the loss function.

[0137] The above validation set is used to evaluate the intermediate results of the model's training and serves as a criterion for determining whether overfitting has occurred. That is, to prevent the blood pressure estimation model (282) from fitting well to the training data but having a large error on new data, the validation loss is monitored at each training iteration (Epoch).

[0138] The above test set is used for final performance evaluation after the model training is completed. The data included in the test set consists of independent data that was not used during the training and validation process, and is used as a standard for evaluating the generalization ability of the blood pressure estimation model (282). In one embodiment of the present invention, by configuring the dataset in a ratio of 2:1:1 as described above, sufficient training data can be secured while maintaining the statistical reliability of the validation and test data.

[0139] According to one embodiment, the blood pressure estimation model (282) is designed to store the optimal blood pressure estimation model (282) based on the point in time when the validation loss is minimized. Specifically, during the training process, the loss value of the validation set is calculated for each Epoch, and when a point is detected where the loss value no longer decreases compared to the previous Epoch, the blood pressure estimation model (282) at that point in time is stored. This prevents overfitting, where the blood pressure estimation model (282) is excessively fitted to the training set, and maximizes blood pressure prediction performance in a real environment.

[0140] As such, according to an embodiment of the present invention, the blood pressure estimation model (282) learns by dividing the learning dataset into training, validation, and test sets in a ratio of 2:1:1, and saves the blood pressure estimation model (282) at the point where validation loss is minimized, thereby improving the learning efficiency and prediction accuracy of the blood pressure estimation model (282). Accordingly, the reliability of systolic and diastolic blood pressure estimated in a non-invasive manner is increased, and stable performance can be maintained even in a long-term continuous blood pressure monitoring environment.

[0141]

[0142] Figure 5 is a diagram showing the overall procedure flow of a non-invasive continuous blood pressure estimation device according to an embodiment of the present invention.

[0143] As described above, the data used includes ECG, PPG, and ABP (Arterial Blood Pressure) signals collected from the public biosignal database VitalDB and the pediatric intensive care database (VitalDB PICUs Dataset). After selecting subject data containing signals necessary for blood pressure estimation from the said database, multiple feature sets are extracted for each subject based on ECG, PPG, and ABP signals.

[0144] In this embodiment, a total of 157,678,505 feature sets were generated from approximately 3,766 subjects, and a balanced data composition was secured by limiting the number of sets to a maximum of 200,000 per age interval (in 5-year increments). As a result, approximately 2,312,381 feature sets were retained, and among these, 71,559 outliers falling outside the age percentiles (1st to 99th percentile) were removed, resulting in a final total of 2,240,822 feature sets being included in the analysis. Subsequently, the entire feature set is divided into 1,115,681 sets for training, 574,967 sets for validation, and 550,174 sets for testing, which are then used for training a Temporal Convolutional Network (TCN).

[0145]

[0146] Figure 6 is a diagram showing the number of feature sets by age group of a non-invasive continuous blood pressure estimation device according to an embodiment of the present invention.

[0147] As illustrated, the results show that the entire data collected from the biosignal databases (VitalDB and VitalDB PICUs Dataset) is organized into 5-year intervals based on age ranges (0 years and older, but less than 95 years). The feature sets extracted for each age range are classified into the Open Dataset, the PICUs Dataset, and outliers, with each color in the bar graph representing one of these three categories. In the example, up to 200,000 feature sets were randomly extracted for each age range, and outliers are defined as cases where systolic blood pressure (SBP) or diastolic blood pressure (DBP) falls outside the age-specific percentile range (1st to 99th percentile). As illustrated, the PICUs Dataset has a high proportion in the age range of 0 years and older, while the Open Dataset has a relatively high proportion in the age range of 30 years and older. Furthermore, the total number of data points tends to decrease as age increases, which reflects the actual clinical data collection environment. Therefore, this indicates that the blood pressure estimation device according to the embodiment of the present invention is designed to ensure blood pressure estimation accuracy that reflects physiological differences according to age by configuring balanced training data considering age-specific characteristics.

[0148]

[0149] FIG. 7 is a histogram showing the blood pressure distribution according to age percentiles of a non-invasive continuous blood pressure estimation device according to an embodiment of the present invention.

[0150] As described, Figure A shows the distribution of systolic blood pressure (SBP), and Figure B shows the distribution of diastolic blood pressure (DBP). In one embodiment, the blood pressure data of each subject was classified according to the 1st to 99th percentile intervals based on age, and if it falls within the above range, it is indicated as 'Within (normal range)', and if it falls outside the above range, it is indicated as 'Outside (outlier)'.

[0151] As can be seen in the illustrated histogram, most systolic and diastolic blood pressures are distributed within the normal range (Within), while outliers (Outside) exist in a relatively small proportion. These data distribution characteristics are utilized as a preprocessing step to ensure the stability and reliability of the training data in the blood pressure estimation device of the present invention. Therefore, this drawing visually demonstrates that the blood pressure estimation model (282) of the present invention can improve blood pressure prediction accuracy by systematically excluding data outliers based on the physiological blood pressure range by age.

[0152]

[0153] FIG. 8 is a diagram showing the distribution of extracted feature values ​​in a non-invasive continuous blood pressure estimation device according to an embodiment of the present invention, visually representing the statistical distribution of major biosignal features derived from ECG signals and PPG signals.

[0154] The above features include heart rate (HR), pulse arrival time (PAT), branch width (BW), systolic branch width (SBW), diastolic branch width (DBW), photoplethysmogram intensity (PI), acceleration plethysmogram intensity (AI), and systolic timespan (ST).

[0155] Also, Fig. 8 shows ST defined as a new feature. VIT and ST VMD Values ​​are plotted together, representing changes in the systole interval time based on the Intersecting Tangent (IT) and Maximal Derivative (MD) points within the PPG waveform, respectively. Each feature value is expressed in milliseconds (ms), beats per minute, or dimensionless values ​​depending on the unit; for example, the PAT, BW, SBW, DBW, and ST series are expressed in time units (ms), HR and PR in frequency units (beats / min), and the remaining PI, AI, PIR, BWR, etc. are calculated as dimensionless ratios.

[0156] The distribution shown in FIG. 8 reflects physiological variability for each subject, and through the standardization process of the above feature values, the blood pressure estimation model (282) of the present invention is designed to enable stable and precise blood pressure estimation in response to various physical conditions and age groups.

[0157]

[0158] FIG. 9 is a scatter plot showing the actual blood pressure (BP) and the blood pressure estimated by the blood pressure estimation model (282) in a non-invasive continuous blood pressure estimation device according to an embodiment of the present invention.

[0159] As shown, Fig. 9A shows a scatter plot for systolic blood pressure (SBP), and Fig. 9B shows a scatter plot for diastolic blood pressure (DBP).

[0160] In Fig. 9, each point corresponds to an individual feature set, and the point density is represented by the color gradient shown on the right. A higher point density indicates a higher degree of agreement between the actual blood pressure and the estimated blood pressure, and in the case of the systolic blood pressure in Fig. 9A, the coefficient of determination (R²) 2 ) is found to be approximately 0.931, confirming that the blood pressure estimation model (282) of the present invention has a very high correlation with the actual systolic blood pressure. In addition, in the case of the diastolic blood pressure in FIG. 9 B, R 2 With a value of approximately 0.910, it can be seen that the prediction accuracy with respect to actual diastolic blood pressure is also excellent.

[0161] Accordingly, FIG. 9 demonstrates that a blood pressure estimation model (282) according to an embodiment of the present invention can predict actual blood pressure with high precision based on feature values ​​extracted from ECG signals and PPG signals, and visually illustrates the reliability and prediction performance of a non-invasive continuous blood pressure estimation device.

[0162] FIG. 10 shows a Bland-Altman plot indicating the degree of agreement between actual blood pressure (BP) and blood pressure estimated by a blood pressure estimation model (282) in a non-invasive continuous blood pressure estimation device according to an embodiment of the present invention.

[0163] As illustrated, Figure 10 A shows the relationship between actual systolic blood pressure (SBP) and estimated systolic blood pressure, and Figure 10 B shows the relationship between actual diastolic blood pressure (DBP) and estimated diastolic blood pressure.

[0164] In each plot, the horizontal axis represents the mean value of actual blood pressure and estimated blood pressure, and the vertical axis represents the difference between the two blood pressure values; the color of the points indicates the density level of the data through the color gradient shown on the right. The points in the plot correspond to feature sets for each subject, and the horizontal dashed line represents the mean difference (bias) and the limit of agreement (±1.96 standard deviations from the mean difference). This indicates that it falls within the range of the 95% confidence interval.

[0165] As illustrated, it can be seen that most points are distributed within the agreement range for both systolic and diastolic blood pressure, which indicates that the blood pressure estimation model (282) according to the embodiment of the present invention has a statistically significant level of consistency and accuracy with actual blood pressure. Thus, FIG. 10 indicates that a non-invasive continuous blood pressure estimation device can provide stable blood pressure estimation performance within a clinically acceptable error range, as well as a high correlation between the model prediction value and the actual measurement value.

[0166]

[0167] FIG. 11 is a diagram showing the results of analyzing the importance and influence of each feature used for blood pressure estimation in a blood pressure estimation model (282) according to an embodiment of the present invention as SHAP values ​​(Shapley Additive Explanations).

[0168] As described, the graph on the left quantitatively expresses the overall Feature Importance by representing the average absolute SHAP value of each feature, and the graph on the right visually indicates the direction and magnitude of the influence of individual features on the model output (blood pressure estimate) by plotting the distribution of SHAP values ​​of each feature. The color gradient in the figure is distinguished according to the magnitude of each feature value, with blue representing low values ​​and red representing high values. Key features include Pulse Arrival Time (PAT), PI Ratio (PIR), Acceleration Plethysmogram Intensity (AI), Branch Width (BW), Systolic Branch Width (SBW), Diastolic Branch Width (DBW), Photoplethysmogram K Value (PPGK), Intersecting Tangent (IT), and Systolic Timespan (ST).

[0169] According to the above analysis results, PAT P , PAT MD and PIR P Time-based and intensity-based features such as those mentioned above have the greatest influence on blood pressure estimation, which means that the blood pressure estimation model (282) according to the embodiment of the present invention accurately predicts blood pressure by utilizing the time difference and waveform shape change of the ECG signal and PPG signal as key indicators. Accordingly, FIG. 11 shows that the blood pressure estimation model (282) of the present invention secures both reliability and interpretability of non-invasive blood pressure estimation by systematically reflecting the relative importance of each biosignal feature.

[0170]

[0171] As such, the non-invasive continuous blood pressure estimation device according to the embodiment of the present invention can ensure the simultaneity of the two signals by automatically correcting the time discrepancy between the ECG signal and the PPG signal through the Δ parameter.

[0172] Accordingly, by aligning biosignals generated within the same heart rate cycle, the accuracy of input data can be improved, thereby enhancing the reliability and prediction accuracy of non-invasive continuous blood pressure estimation.

[0173]

[0174] FIG. 12 is a flowchart of a non-invasive continuous blood pressure estimation method according to an embodiment of the present invention. Since it overlaps with the above-described embodiment, a detailed description of this embodiment will be omitted.

[0175] As described above, a non-invasive continuous blood pressure measurement device according to an embodiment of the present invention comprises the steps of: detecting an electrocardiogram (ECG) signal of a subject (S910); detecting a photoplethysmography (PPG) signal of a subject (S920); collecting the ECG signal and the PPG signal and storing them together with time information (S930); detecting an R-peak of the ECG signal and a specific point of the PPG signal to identify a reference point of the signal (S940); calculating a time discrepancy between the ECG signal and the PPG signal and setting a parameter (Δ) to correct the calculated time discrepancy to ensure simultaneity of the two signals (S950); and estimating systolic blood pressure (SBP) and diastolic blood pressure (DBP) using the corrected ECG signal and PPG signal (S960).

[0176] According to an embodiment of the present invention, the step (S910) of detecting an electrocardiogram (ECG) signal of a subject is a process of detecting the electrical activity of the heart using an ECG sensor (120). The ECG signal reflects the potential change occurring according to the depolarization and repolarization processes of the myocardium and is used as basic data for analyzing heart rate cycles and rhythm changes. The ECG signal is digitized at a constant sampling frequency and transmitted to a data collection unit of the next step.

[0177] According to an embodiment of the present invention, the step (S920) of detecting a photoplethysmography (PPG) signal of a subject is a process of detecting a change in blood flow within a blood vessel using a PPG sensor (140). The PPG signal is obtained by detecting a change in the amount of light reflected or transmitted from the surface of the skin and provides waveform information according to the change in blood flow within the blood vessel. Similar to an ECG signal, the PPG signal can also be collected at a constant sampling frequency and aligned according to the same time axis.

[0178] According to an embodiment of the present invention, in the step (S930) of collecting ECG signals and PPG signals and storing them together with time information, a data collection unit acquires the ECG signals and PPG signals transmitted from the sensor unit at the same sampling frequency and records the time information at the time the signals were collected together to form a data set. The collected data is subsequently utilized to ensure simultaneity between signals and to calculate time discrepancies, and is stored as raw data for future feature extraction and signal correction steps.

[0179] According to an embodiment of the present invention, the step (S940) of identifying a reference point of a signal by detecting an R-peak of an ECG signal and a specific point of a PPG signal is performed by a feature extraction unit. In the ECG signal, the R-peak represents the moment of ventricular depolarization, and the specific point of the PPG signal may be one of the rising or falling points of the pulse wave. By accurately detecting the reference points of the two signals in this step, subsequent calculation of time discrepancies and synchronous correction become possible.

[0180] According to an embodiment of the present invention, in the step (S950) of calculating the time discrepancy between an ECG signal and a PPG signal and setting a parameter (Δ) to correct the calculated time discrepancy to ensure simultaneity of the two signals, the simultaneity correction unit is the performing entity. Since collection delays may occur in ECG and PPG signals depending on the type of sensor and attachment location, the time difference (discrepancy) between the two signals is defined as the parameter Δ. During the learning process, this Δ(Δ) value is repeatedly updated using a backpropagation method, and finally, the two signals are corrected so that they are synchronized.

[0181] According to an embodiment of the present invention, the step (S960) of estimating systolic blood pressure (SBP) and diastolic blood pressure (DBP) using corrected ECG signals and PPG signals is performed by a blood pressure estimation unit. The blood pressure estimation unit receives the corrected signals as input and predicts blood pressure values ​​through a blood pressure estimation model based on a Temporal Convolutional Network (TCN). The SBP and DBP calculated as the prediction results can subsequently be utilized for the calculation of Mean Arterial Pressure (MAP), etc.

[0182] According to the embodiments, the non-invasive continuous blood pressure estimation method is applicable in both medical institutions and home environments, and ECG and PPG sensors can be attached to various body parts, such as the wrist, finger, or ear, respectively. In addition, the data collection and processing can be performed in real time, enabling continuous blood pressure monitoring.

[0183]

[0184] Table 1 below shows the results of evaluating the blood pressure (BP) estimation performance of the blood pressure estimation model according to an embodiment of the present invention in accordance with international standards.

[0185]

[0186]

[0187] Table 1 shows the key indicators, such as the mean error (ME), standard deviation error (SDE), mean absolute deviation (MAD), and cumulative error (CE), for the systolic blood pressure (SBP) estimation model and the diastolic blood pressure (DBP) estimation model, respectively.

[0188] According to the results in Table 1, the blood pressure estimation model according to the embodiment of the present invention satisfied a 'PASS' or 'A' grade in all indicators according to international blood pressure measurement standards (e.g., AAMI, BHS standards). In particular, high precision was secured with a mean absolute deviation (MAD) of 3 mmHg or less for both systolic and diastolic blood pressure, and recorded 91.5% and 97.2%, respectively, within the cumulative error range of 10 mmHg, demonstrating excellent agreement in terms of clinical reliability.

[0189] Therefore, Table 1 above is evidence showing that the blood pressure estimation model of the present invention maintains statistically high correlation and accuracy with actual blood pressure measurements, despite being a non-invasive method.

[0190]

[0191] Table 2 below shows the results of analyzing the blood pressure estimation performance of the blood pressure estimation model according to an embodiment of the present invention, based on age group and gender.

[0192]

[0193]

[0194] As shown in Table 2 above, the performance evaluation is based on the coefficient of determination (R²) for systolic blood pressure (SBP) and diastolic blood pressure (DBP). 2 It includes the mean error (ME), standard deviation error (SDE), mean absolute deviation (MAD), and cumulative error (CE). CE represents the percentage of estimates that fall within ±5 mmHg, ±10 mmHg, and ±15 mmHg of the actual blood pressure value.

[0195] According to the results in Table 2, the highest coefficient of determination (R²) for both systolic and diastolic blood pressure was found in the age group of 15 to under 20 years. 2 The model estimating blood pressure by combining ECG and PPG signals showed the highest accuracy in the corresponding age group, exhibiting a low mean error (ME) and a low mean error (ME). On the other hand, in the age group of 45 years and older, a tendency for the mean error and mean absolute deviation (MAD) to increase slightly was observed due to the influence of physiological changes and a decrease in vascular elasticity.

[0196] As a result of the gender analysis, the coefficient of determination (R²) for systolic blood pressure in the female group 2The value was 0.934, which was higher than the 0.920 of the male group, and the mean absolute deviation (MAD) also showed a slight difference, with 2.718 mmHg for women and 2.759 mmHg for men. In the case of diastolic blood pressure as well, the mean error (ME) and standard deviation error (SDE) between the two groups remained within an error range of 0.1 mmHg or less, confirming that there is no significant difference in performance based on gender.

[0197] In addition, regarding diastolic blood pressure estimation, most age groups showed a cumulative accuracy of over 90% in the CE10 and CE15 results, and in particular, over 95% was satisfied across all age groups according to the CE15 criteria. This is a result of the biosignal characteristics based on ECG and PPG signals increasing sensitivity to changes in blood pressure and improving the consistency of estimation results.

[0198] Therefore, the ECG-PPG-based non-invasive continuous blood pressure estimation device according to the present invention can be confirmed to be suitable for real-time continuous blood pressure monitoring by maintaining high correlation and low error across various age and gender ranges.

[0199]

[0200] FIG. 13 illustrates a non-invasive continuous blood pressure estimation device according to an embodiment of the present invention. Referring to FIG. 13, the non-invasive continuous blood pressure estimation device (10) may include a sensor unit (100), a processing unit (200), a storage unit (300), and a communication unit (400).

[0201] The processing unit (200) can perform signal preprocessing, feature extraction, synchronicity correction, blood pressure estimation, and mean arterial pressure calculation operations based on the electrocardiogram (ECG) signal and photoplethysmography (PPG) signal collected from the ECG sensor (120) and the PPG sensor.

[0202] The processing unit (200) first aligns the time axes of the ECG signal and PPG signal input from the data collection unit, and detects signal features such as R-peak, peak (P), valley (V), intersecting tangent (IT), and maximal derivative (MD) through the feature extraction unit. Subsequently, the time discrepancy between the ECG signal and the PPG signal is set as a parameter (Δ) through the simultaneity correction unit, and the simultaneity of the two signals is secured by updating the corresponding parameter through backpropagation.

[0203] The processing unit (200) transmits these corrected signals to the blood pressure estimation unit to execute a blood pressure estimation model configured based on a Temporal Convolutional Network (TCN). The processing unit (200) continuously estimates systolic blood pressure (SBP) and diastolic blood pressure (DBP) through the learned blood pressure estimation model, and calculates the mean arterial pressure (MAP) from the calculated blood pressure values.

[0204] The storage unit (300) is configured to store data necessary for the operation of the non-invasive continuous blood pressure estimation device (10), and may include, for example, ECG and PPG raw signal data, feature extraction results, learned model parameters, simultaneity correction parameters (Δ), blood pressure estimation results, mean arterial pressure calculation results, time information, user identification information, etc. Additionally, the storage unit (300) may store the configurations of the training, validation, and test datasets used for model learning in a 2:1:1 ratio, and preserve the weight values ​​of the model with minimized validation loss so that they can be utilized for the retraining and updating of the device.

[0205] The communication unit (400) can transmit and receive data between the non-invasive continuous blood pressure estimation device (10) and a user terminal, an external analysis server, or a cloud-based data storage. The communication unit (400) can transmit measured ECG signals, PPG signals, blood pressure estimation results, and mean arterial pressure (MAP) data to the outside via wireless communication such as Wi-Fi, Bluetooth, LTE, 5G, or a wired network (Ethernet, LAN), or update measurement conditions and reference data set externally to the device.

[0206] The processing unit (200) can process ECG signals and PPG signals in real time by referring to reference data and control conditions stored in the storage unit (300), perform simultaneity correction, calculate continuous blood pressure values ​​through a blood pressure estimation model, and then provide the results to a user interface (UI) through the communication unit (400). Additionally, measurement results such as ECG signal and PPG signal waveforms, systolic blood pressure, diastolic blood pressure, mean arterial pressure, heart rate cycle, and pulse wave transit time (PAT) can be visually displayed on the user interface (UI).

[0207] The non-invasive continuous blood pressure estimation device (10) according to an embodiment of the present invention can be applied to application environments such as hospital patient monitoring systems, home healthcare devices, smartwatches, and wearable devices, and can be used for biosignal-based health monitoring and data analysis by continuously and non-invasively measuring the blood pressure of a subject.

[0208]

[0209] As such, the non-invasive continuous blood pressure estimation method according to an embodiment of the present invention can improve the accuracy and reliability of non-invasive blood pressure estimation by quantitatively correcting the time discrepancy between the ECG signal and the PPG signal to ensure the simultaneity of the two signals. In addition, continuous blood pressure measurement in real time becomes possible, allowing for stable monitoring of the user's blood pressure changes over a long period.

[0210]

[0211] The description above is merely an example of applying the principles of the present invention, and other configurations may be further included without departing from the scope of the present invention. For example, at least some of the various embodiments of the present invention described above may be combined.

Claims

1. A sensor unit comprising an ECG sensor that detects the subject's electrocardiogram (ECG) signal and a PPG sensor that detects the subject's photoplethysmography (PPG) signal; A data collection unit that collects ECG signals and PPG signals from the sensor unit and stores them together with time information; A feature extraction unit that identifies a reference point of the signal by detecting the R-peak of the ECG signal and a specific point of the PPG signal; A simultaneity correction unit that calculates the time discrepancy between the ECG signal and the PPG signal, and sets a parameter (Δ; Delta) for correcting the calculated time discrepancy to ensure simultaneity between the two signals; and A blood pressure estimation unit that estimates systolic blood pressure (SBP) and diastolic blood pressure (DBP) using ECG signals and PPG signals corrected from the above-mentioned simultaneity correction unit. A non-invasive continuous blood pressure estimation device comprising 2. In Paragraph 1, The above data collection unit is, Data synchronization is performed by collecting the ECG signal and the PPG signal at the same sampling frequency and aligning the signals based on the time axis. Non-invasive continuous blood pressure estimation device.

3. In Paragraph 1, The above feature extraction unit is, Extracting one or more biosignal features from the above ECG signal and the above PPG signal, including heart rate (HR), pulse rate (PR), pulse arrival time (PAT), systolic timespan (ST), PPG intensity (PI), branch width (BW), systolic branch width (SBW), diastolic branch width (DBW), peak (P), valley (V), intersecting tangent (IT), and maximal derivative (MD). Non-invasive continuous blood pressure estimation device.

4. In Paragraph 1, The above simultaneity correction unit is, A parameter (Δ) is set as the time difference from the R-peak of the ECG signal to any one of the peak (P), trough (V), cross tangent point (IT), and maximum slope point (MD) of the PPG signal, and the parameter (Δ) is updated through backpropagation during the learning process to correct the time discrepancy between the ECG signal and the PPG signal. Non-invasive continuous blood pressure estimation device.

5. In Paragraph 1, The above blood pressure estimation unit is, Estimating the systolic blood pressure and the diastolic blood pressure through a blood pressure estimation model including a Temporal Convolutional Network (TCN) using feature values ​​extracted from the corrected ECG signal and PPG signal as input, Non-invasive continuous blood pressure estimation device.

6. In Paragraph 5, The above blood pressure estimation model is, It includes multiple temporal blocks, each temporal block includes two convolutional layers, and the dilation factor increases exponentially with the depth of the convolutional layers, Non-invasive continuous blood pressure estimation device.

7. In Paragraph 5, The above blood pressure estimation model is, Consisting of an input sequence of 16 consecutive cardiac cycles, including 3 hidden layers, and applying dropout to prevent overfitting, Non-invasive continuous blood pressure estimation device.

8. In Paragraph 5, The above blood pressure estimation model is, Learned by using Mean Squared Error (MSE) as the loss function and applying the Rectified Adam (RAdam) algorithm as the optimizer, Non-invasive continuous blood pressure estimation device.

9. In Paragraph 5, The above blood pressure estimation unit is, Calculating Mean Arterial Pressure (MAP) using systolic and diastolic blood pressure values ​​derived through a learned blood pressure estimation model, Non-invasive continuous blood pressure estimation device.

10. In Paragraph 9, The above blood pressure estimation model is, Training by splitting the training dataset into training, validation, and test sets in a 2:1:1 ratio, and saving the model at the point where the validation loss is minimized, Non-invasive continuous blood pressure estimation device.

11. A step of detecting the subject's electrocardiogram (ECG) signal; A step of detecting the photoplethysmography (PPG) signal of a subject; A step of collecting the ECG signal and the PPG signal and storing them together with time information; A step of identifying a reference point of the signal by detecting the R-peak of the ECG signal and a specific point of the PPG signal; A step of calculating the time discrepancy between the ECG signal and the PPG signal, and setting a parameter (Δ; Delta) to correct the calculated time discrepancy to ensure the simultaneity of the two signals; and A method comprising the step of estimating systolic blood pressure (SBP) and diastolic blood pressure (DBP) using the above-mentioned corrected ECG signal and PPG signal, Non-invasive continuous blood pressure estimation method.

12. In Paragraph 11, The step of collecting the ECG signal and the PPG signal is, A method comprising the step of collecting the above signals at the same sampling frequency and aligning them based on the time axis to perform data synchronization. Non-invasive continuous blood pressure estimation method.

13. In Paragraph 11, The step of identifying the reference point of the above signal is, A method comprising the step of extracting one or more biosignal features from the ECG signal and the PPG signal, including heart rate (HR), pulse rate (PR), pulse arrival time (PAT), systolic timespan (ST), PPG intensity (PI), branch width (BW), systolic branch width (SBW), diastolic branch width (DBW), peak (P), valley (V), intersecting tangent (IT), and maximal derivative (MD). Non-invasive continuous blood pressure estimation method.

14. In Paragraph 11, The step of correcting the above time discrepancy is, A step comprising setting a parameter (Δ) for the time difference between the R-peak of the ECG signal and any one of the peak (P), trough (V), cross tangent point (IT), and maximum slope point (MD) of the PPG signal, and updating the parameter (Δ) through backpropagation during the learning process to correct the time discrepancy between the ECG signal and the PPG signal. Non-invasive continuous blood pressure estimation method.

15. In Paragraph 11, The step of estimating the above blood pressure is, A step comprising estimating the systolic blood pressure and the diastolic blood pressure using a blood pressure estimation model including a Temporal Convolutional Network (TCN) with feature values ​​extracted from the corrected ECG signal and PPG signal as input, Non-invasive continuous blood pressure estimation method.

16. In Paragraph 15, The step of using the above blood pressure estimation model is, A step comprising applying a model having a structure in which a dilation factor increases exponentially with the depth of the convolutional layers, wherein the model includes a plurality of temporal blocks, each temporal block includes two convolutional layers. Non-invasive continuous blood pressure estimation method.

17. In Paragraph 15, The step of using the above blood pressure estimation model is, A step comprising configuring the input sequence into 16 consecutive cardiac cycles, including 3 hidden layers, and applying dropout to prevent overfitting, Non-invasive continuous blood pressure estimation method.

18. In Paragraph 15, The step of training the above blood pressure estimation model is, A step including a learning step that uses Mean Squared Error (MSE) as the loss function and applies the Rectified Adam (RAdam) algorithm as the optimizer, Non-invasive continuous blood pressure estimation method.

19. In Paragraph 15, The step of estimating the above blood pressure is, A method comprising the step of calculating Mean Arterial Pressure (MAP) using systolic and diastolic blood pressure values ​​calculated through a learned blood pressure estimation model, Non-invasive continuous blood pressure estimation method.

20. In Paragraph 19, The step of training the above blood pressure estimation model is, A method comprising the step of splitting a training dataset into training, validation, and test sets in a 2:1:1 ratio for training, and saving the model at the point where the validation loss is minimized. Non-invasive continuous blood pressure estimation method.

Images