A precise feature point localization method based on temporal difference learning in cf-PWV

Abstract

This invention discloses a precise localization method for cf-PWV feature points based on temporal difference learning, applicable to the field of medical device signal processing. The method acquires time-aligned carotid and femoral pulse wave signals, generates candidate feature point sets within multiple consecutive cardiac cycles, and constructs multiple carotid-femoral candidate point pairs. For each candidate point pair, its propagation time difference stability information is calculated, and physiological time-corresponding information is determined based on the subsequent short-window main wave growth sequence. When a candidate point pair exhibits a stable propagation time difference but physiological time-phase mismatch, its immediate reward is reduced, and the point pair value is updated through temporal difference learning. Finally, target carotid-femoral feature point pairs with higher point pair values ​​and satisfying phase correspondence conditions are selected to determine the target propagation time difference and cf-PWV measurement value.

Description

Technical Field

[0001] This application relates to the field of medical device signal processing technology, and in particular to a method for accurate localization of cf-PWV feature points based on temporal difference learning. Background Technology

[0002] Carotid-femoral pulse wave velocity (cf-PWV) is an important physiological parameter for evaluating the elasticity of the large arteries in the human body. During cf-PWV measurement, it is typically necessary to acquire carotid and femoral pulse wave signals. The cf-PWV value is then calculated based on the time difference between the arrival times of the carotid and femoral pulse waves, as well as the propagation distance between the measurement locations in the carotid and femoral arteries.

[0003] In existing cf-PWV measurement methods, a common approach is to locate the foot point, maximum slope point, threshold intersection point, or tangent intersection point in both the carotid and femoral pulse wave signals, and then determine the propagation time difference based on the time difference between the two characteristic points. This type of method can obtain relatively stable measurement results when the waveform is regular, the noise is low, and the two signals have consistent morphology.

[0004] However, in actual medical device acquisition, there may be waveform differences between carotid and femoral pulse wave signals. Especially when a pseudo-foot point appears at the anterior shoulder of the carotid waveform's main rising edge, while the femoral waveform's foot point is clearer, if the algorithm selects the candidate feature point with the highest local confidence in a single channel from both signals, it may pair the carotid pseudo-foot point with the femoral true foot point. This candidate point pair may have a relatively stable propagation time difference over multiple cardiac cycles, but the two candidate feature points it includes are not in the same physiological phase, causing the stable propagation time difference to mask the feature point mismatch problem.

[0005] In other words, existing methods typically focus on whether the propagation time difference is stable or whether the local morphology of a single-channel feature point resembles a foot point, but they do not adequately determine whether the main wave growth process after the carotid artery candidate feature point and the main wave growth process after the femoral artery candidate feature point satisfy a preset phase correspondence. When the propagation time difference is stable but the physiological phase is mismatched, the system may output a stable but offset propagation time difference, thereby affecting the repeatability and reliability of cf-PWV measurement results.

[0006] Therefore, there is an urgent need for a feature point localization method that can identify candidate point pairs with stable propagation time differences but mismatched physiological phases during cf-PWV measurement and utilize a time-series feedback mechanism to reduce the value of erroneous candidate point pairs. Summary of the Invention

[0007] This application provides a precise localization method for cf-PWV feature points based on temporal difference learning, which reduces the propagation time difference offset problem caused by the stable propagation time difference but physiological phase mismatch between carotid artery candidate feature points and femoral artery candidate feature points during cf-PWV measurement, thereby improving the stability and repeatability of cf-PWV measurement results.

[0008] Firstly, this application provides a method for precise localization of cf-PWV feature points based on temporal difference learning. This method can be executed by a cf-PWV measurement device or by a data processing device communicatively connected to the cf-PWV measurement device. In this method, the data processing device acquires time-aligned carotid pulse wave signals and femoral pulse wave signals; for multiple consecutive cardiac cycles, it generates a set of carotid candidate feature points in the carotid pulse wave signal and a set of femoral candidate feature points in the femoral pulse wave signal; it constructs multiple carotid-femoral candidate point pairs based on the carotid and femoral candidate feature point sets; for each carotid-femoral candidate point pair, it determines propagation time difference stability information based on the candidate propagation time difference in multiple consecutive cardiac cycles, and based on the included... The carotid artery post-short window main wave growth sequence and the femoral artery post-short window main wave growth sequence after the carotid artery candidate feature point are used to determine the physiological time-corresponding information. The instantaneous reward of the carotid-femoral candidate point pair is determined based on the propagation time difference stability information and the physiological time-corresponding information. The point pair state of the carotid-femoral candidate point pair is used as the state of temporal difference learning. The point pair value of the carotid-femoral candidate point pair is updated based on the instantaneous reward and the point pair value of the next temporal state. Based on the updated point pair value, the target carotid-femoral feature point pair is determined from multiple carotid-femoral candidate point pairs.

[0009] In the above method, the data processing device does not directly select the feature points with the highest local confidence in the carotid and femoral artery channels respectively. Instead, it first constructs carotid-femoral candidate point pairs and then simultaneously analyzes the stability of the candidate propagation time difference and the physiological time correspondence of the main wave growth sequence after the two candidate feature points. When a carotid-femoral candidate point pair has a relatively stable candidate propagation time difference in multiple consecutive cardiac cycles, but its carotid post-short window main wave growth sequence and femoral post-short window main wave growth sequence do not meet the phase correspondence condition, the point pair is regarded as a point pair with a stable propagation time difference but a physiological time mismatch. In the temporal difference learning, its immediate reward or point pair value is reduced, thereby avoiding stable but erroneous candidate point pairs being mistakenly used as the basis for cf-PWV calculation.

[0010] In one possible design, acquiring time-aligned carotid and femoral artery pulse wave signals includes: acquiring the raw carotid pulse wave signal acquired by the carotid acquisition channel and the raw femoral pulse wave signal acquired by the femoral acquisition channel; and performing time compensation on the raw carotid and femoral pulse wave signals based on at least one of a synchronization trigger signal, a sampling timestamp, an ECG reference signal, a synchronization marker signal, and a pre-calibrated channel delay to obtain the time-aligned carotid and femoral pulse wave signals. This design can reduce the impact of acquisition channel delay on propagation time difference calculation.

[0011] In one possible design, for multiple consecutive cardiac cycles, a set of candidate feature points for the carotid artery pulse wave signal and a set of candidate feature points for the femoral artery pulse wave signal are generated respectively. This includes: dividing the carotid artery pulse wave signal and the femoral artery pulse wave signal into cardiac cycles; within each cardiac cycle, generating at least two candidate feature points based on at least one of the amplitude change, slope change, curvature change, local extreme value position, and threshold cross position of the corresponding pulse wave signal; wherein, the at least two candidate feature points include candidate feature points whose local morphology satisfies the condition of being at the threshold, and candidate feature points located before or near the rising edge of the main wave; the condition of being at the threshold includes at least one of the following: the average amplitude within a preset front window before the candidate feature point is lower than the peak amplitude of the current cycle; the number of sampling points with continuously positive first-order differences within a preset back window after the candidate feature point is not less than a first quantity threshold; the second-order difference near the candidate feature point undergoes a sign change; and the time position of the candidate feature point is located after the trough of the current cycle and before the peak of the main wave. With this design, the system does not prematurely determine a unique foot point within a single cardiac cycle, but instead retains multiple candidate time locations that may be disputed.

[0012] In one possible design, multiple carotid-femoral candidate point pairs are constructed based on the carotid artery candidate feature point set and the femoral artery candidate feature point set. This includes: determining the correspondence between the carotid artery cardiac cycle and the femoral artery cardiac cycle based on the cardiac cycle sequence number, reference trigger time, or cycle correspondence; combining the carotid artery candidate feature points and femoral artery candidate feature points in the corresponding carotid artery cardiac cycle and femoral artery cardiac cycle to obtain multiple initial carotid-femoral candidate point pairs; calculating the candidate propagation time difference based on the time position of the carotid artery candidate feature point and the time position of the femoral artery candidate feature point in each initial carotid-femoral candidate point pair; and identifying the initial carotid-femoral candidate point pairs whose candidate propagation time difference is within a preset propagation time range as carotid-femoral candidate point pairs participating in temporal difference learning. The preset propagation time range is determined based on the ratio of the preset carotid-femoral propagation distance to the preset upper limit of cf-PWV and the ratio of the preset carotid-femoral propagation distance to the preset lower limit of cf-PWV. Through this design, the system transforms the localization target from a single candidate feature point into a neck-femur candidate point pair, enabling joint evaluation of the propagation time difference and physiological time correspondence at the point pair level.

[0013] In one possible design, propagation time difference stability information is determined based on the candidate propagation time difference of each neck-femoral candidate point pair across multiple consecutive cardiac cycles. This includes: establishing a sequence of similar neck-femoral candidate point pairs across multiple consecutive cardiac cycles according to the candidate feature point's index in the corresponding candidate feature point set, the local morphological category of the candidate feature point, or the interval where the candidate propagation time difference is located; obtaining the candidate propagation time difference of the similar neck-femoral candidate point pair sequence across multiple consecutive cardiac cycles; and determining the propagation time difference stability information based on at least one of the following: fluctuation amplitude, variance, range, change between adjacent cycles, and coefficient of variation of the candidate propagation time difference. This design can identify whether a certain type of candidate point pair exhibits stable time difference within consecutive cycles, providing a basis for subsequently distinguishing between stable and correct pairs and stable but mismatched pairs.

[0014] In one possible design, physiological time-corresponding information is determined based on the carotid artery posterior short-window main wave growth sequence and the femoral artery posterior short-window main wave growth sequence. This includes: extracting a signal segment of a first preset duration starting from a candidate feature point in the carotid artery to obtain the carotid artery posterior short-window main wave growth sequence; extracting a signal segment of a second preset duration starting from a candidate feature point in the femoral artery to obtain the femoral artery posterior short-window main wave growth sequence; and determining the physiological time-corresponding information based on the sequence correlation and normalized phase deviation between the carotid artery posterior short-window main wave growth sequence and the femoral artery posterior short-window main wave growth sequence, combined with at least one of the following: consistency of growth direction and similarity of slope maintenance duration. Through this design, the system can determine whether both candidate feature points subsequently enter the corresponding main wave growth phase, rather than solely relying on the stability of the candidate propagation time difference to confirm the correctness of the feature point pair.

[0015] In one possible design, the immediate reward for the neck-thigh candidate point pair is determined based on the propagation time difference stability information and the physiological time-correspondence information. This includes: generating a first immediate reward when the propagation time difference stability information of the target neck-thigh candidate point pair meets a preset stability condition and the physiological time-correspondence information of the target neck-thigh candidate point pair meets a preset phase correspondence condition; generating a second immediate reward when the propagation time difference stability information of the target neck-thigh candidate point pair meets the preset stability condition but the physiological time-correspondence information of the target neck-thigh candidate point pair does not meet the preset phase correspondence condition, and marking the target neck-thigh candidate point pair as a neck-thigh candidate point pair with stable propagation time difference but mismatched physiological time phase; wherein, the second immediate reward is less than the first immediate reward. Through this design, neck-thigh candidate point pairs with stable propagation time difference but mismatched physiological time phase will not be continuously reinforced due to time difference stability, but will have their point pair value reduced during temporal difference learning.

[0016] In one possible design, the point-to-point state of the neck-femur candidate point pair is used as the state for temporal difference learning. The point-to-point value of the neck-femur candidate point pair is updated based on the immediate reward and the point-to-point value of the next temporal state. This includes: constructing a point-to-point state vector for each neck-femur candidate point pair, where the point-to-point state vector includes propagation time difference stabilization information and physiological time correspondence information, and includes at least one of the following: candidate propagation time difference, candidate feature point local morphology information, and post-short window main wave growth information; determining the point-to-point state of the neck-femur candidate point pair that satisfies the temporal correspondence with each neck-femur candidate point pair in the next adjacent cardiac cycle as the next temporal state; calculating the temporal difference error based on the immediate reward, the current point-to-point value, the point-to-point value of the next temporal state, and a discount factor; and updating the point-to-point value of each neck-femur candidate point pair based on the temporal difference error. This design can correlate the candidate point pair judgment result in the current cardiac cycle with the point-to-point state in subsequent cardiac cycles, using subsequent state feedback to correct the current point-to-point value.

[0017] In one possible design, the point-pair state vector also includes anterior shoulder pseudofoot risk information. This risk information is determined based on at least one of the following: slope change before and after the candidate feature point, duration of continuous rise of the main wave after the candidate feature point, amplitude of secondary fall after the candidate feature point, and position of the rising edge of the main wave where the candidate feature point is located. When the carotid artery candidate feature point or femoral artery candidate feature point in the target neck-femoral candidate point pair meets the anterior shoulder pseudofoot risk condition, the immediate reward of the target neck-femoral candidate point pair is reduced, and point-pair value inheritance blocking processing is performed on the target neck-femoral candidate point pair. This design avoids the erroneous point-pair value formed by anterior shoulder pseudofoot points in the previous cardiac cycle from being inherited into the subsequent cardiac cycle, thereby reducing the risk of pseudofoot point cross-cycle locking.

[0018] In one possible design, a target neck-femoral feature point pair is determined from multiple neck-femoral candidate point pairs based on the updated point pair value. This includes: ranking the multiple neck-femoral candidate point pairs according to the updated point pair value; selecting the neck-femoral candidate point pair with the highest point pair value and satisfying a preset phase correspondence condition from the ranked multiple neck-femoral candidate point pairs as the target neck-femoral feature point pair; determining the target propagation time difference based on the carotid artery feature point and femoral artery feature point in the target neck-femoral feature point pair; determining the cf-PWV measurement value based on the target propagation time difference and a preset neck-femoral propagation distance; determining the positioning reliability of the target neck-femoral feature point pair based on the weighted result of the point pair value, propagation time difference stability information, and physiological time correspondence information; and outputting at least one of the target neck-femoral feature point pair, target propagation time difference, cf-PWV measurement value, and positioning reliability. This design can output a target feature point pair and its reliability that can be used for cf-PWV calculation, facilitating subsequent measurement result display, quality control, or verification.

[0019] Compared with the prior art, this application has at least the following beneficial effects: This application transforms the cf-PWV feature point localization target from single-channel local optimal feature points to neck-femur candidate point pairs, avoiding the hidden biases caused by determining feature points solely based on propagation time difference stability or single-point local morphology. By introducing the post-short window main wave growth sequence, physiological time correspondence information, and propagation time difference stability information, erroneous point pairs with stable time differences but mismatched actual physiological time phases can be identified. Simultaneously, temporal difference learning is used to update the value of candidate point pairs, strengthening stable and phase-corresponding point pairs and suppressing stable but mismatched point pairs. This improves the accuracy and repeatability of neck-femur propagation time difference localization, reduces the risk of cross-cycle locking of anterior shoulder pseudofoot points, and enhances the stability and reliability of cf-PWV medical device measurement results. Attached Figure Description

[0020] Figure 1 A schematic diagram of the architecture of a cf-PWV measurement system provided in this application embodiment; Figure 2 A schematic diagram of candidate feature points for carotid artery pulse wave signal and femoral artery pulse wave signal provided for embodiments of this application; Figure 3 A flowchart illustrating a precise feature point localization method based on temporal difference learning for embodiments of this application; Figure 4 This is a schematic diagram illustrating the construction of neck-femoral candidate point pairs, provided in an embodiment of this application. Figure 5 A schematic diagram illustrating the identification of a stable propagation time difference but mismatched physiological phases, provided as an embodiment of this application; Figure 6 A schematic diagram illustrating a temporal differential learning update of point pair value provided in an embodiment of this application; Figure 7 This is a schematic diagram of a device structure provided in an embodiment of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the embodiments of this application will be further described in detail below with reference to the accompanying drawings. In the description of the embodiments of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" and "second" may explicitly or implicitly include one or more of that feature.

[0022] It should be understood that in the embodiments of this application, "at least one" means one or more, and "more than one" means two or more. "And / or" describes the relationship between related objects, indicating that there can be three relationships, such as A and / or B, which can represent the cases where A exists alone, A and B exist simultaneously, or B exists alone, where A and B can be singular or plural. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of singular or plural items.

[0023] Figure 1 A schematic diagram of the architecture of a cf-PWV measurement system provided in this application embodiment; Reference Figure 1 The system may include cf-PWV measurement equipment and data processing equipment; The cf-PWV measurement device may include a carotid artery signal acquisition module and a femoral artery signal acquisition module; The carotid artery signal acquisition module is used to acquire carotid artery pulse wave signals, and the femoral artery signal acquisition module is used to acquire femoral artery pulse wave signals. The data processing device can be a processor installed in the cf-PWV measurement device, or a terminal device, server, or cloud processing platform that communicates with the cf-PWV measurement device.

[0024] cf-PWV is the carotid-femoral pulse wave velocity; The carotid pulse wave signal can be a waveform signal obtained by a carotid pressure sensor, tension sensor, ultrasound acquisition module or photoplethysmography pulse wave acquisition module. The femoral artery pulse wave signal can be a waveform signal obtained through a femoral artery pressure sensor, cuff pressure acquisition module, tension sensor, ultrasound acquisition module, or photoplethysmography pulse wave acquisition module. It should be noted that the specific method of acquiring pulse wave signals is not limited in the embodiments of this application, as long as the acquired signals can reflect the changes in the pulse waveforms of the carotid and femoral arteries.

[0025] In the relevant cf-PWV measurement process, such as Figure 2 As shown, the system typically locates the foot point or other feature points representing the arrival time of the pulse wave in the carotid pulse wave signal and the femoral pulse wave signal, respectively, and then determines the propagation time difference based on the time difference between the two feature points. However, when there is a pseudo-foot point in the anterior shoulder of the carotid waveform, the foot point of the femoral waveform is relatively clear, or the local morphology of the two waveforms is inconsistent, the candidate feature point with the highest local confidence in a single channel may not be in the same physiological phase as the candidate feature point in another channel. In this case, even if the propagation time difference obtained from the candidate point pair is relatively stable in multiple consecutive cardiac cycles, it may be a stable but erroneous propagation time difference.

[0026] Based on the above problems, this application provides a method for accurate localization of cf-PWV feature points based on temporal difference learning. In this method, the data processing device does not directly select the candidate points with the highest confidence in the single channel of the carotid artery channel and the femoral artery channel respectively. Instead, it first constructs multiple carotid-femoral candidate point pairs, and then judges whether there is a stable but mismatched problem in the candidate point pairs based on the stable information of propagation time difference and the corresponding information of physiological time. The value of the candidate point pairs is updated through temporal difference learning, thereby determining the target carotid-femoral feature point pair.

[0027] Figure 3 A flowchart illustrating a precise feature point localization method based on temporal difference learning for cf-PWV, provided as an embodiment of this application. (Reference) Figure 3 The method includes the following steps: S101: The data processing device acquires time-aligned carotid pulse wave signals and femoral pulse wave signals; The data processing device can acquire the raw pulse wave signal of the carotid artery acquired through the carotid artery acquisition channel and the raw pulse wave signal of the femoral artery acquired through the femoral artery acquisition channel. In order to ensure the accuracy of the propagation time difference calculation, the data processing device can perform time alignment on the raw pulse wave signals of the carotid artery and the femoral artery.

[0028] In some embodiments of this application, time alignment may include hardware synchronization alignment and software correction alignment; hardware synchronization alignment refers to the carotid artery acquisition channel and the femoral artery acquisition channel sharing the same sampling clock or synchronization trigger signal; software correction alignment refers to time compensation of the raw carotid artery pulse wave signal and the raw femoral artery pulse wave signal based on the sampling timestamp, known channel delay, ECG reference signal or synchronization marker signal. For example, when there is a fixed acquisition delay τc between the carotid artery acquisition channel and the femoral artery acquisition channel, the time position t in the original carotid artery pulse wave signal can be corrected to t-τc. Where τc represents the pre-calibrated channel delay, and t represents the sampling time position before correction. This method allows for the acquisition of carotid and femoral artery pulse wave signals at the same time reference.

[0029] In a specific example, both the carotid artery acquisition channel and the femoral artery acquisition channel are sampled at 1000Hz. The device calibration shows that the carotid artery acquisition channel has a fixed delay of 2ms relative to the femoral artery acquisition channel. When the data processing device processes the raw carotid artery pulse wave signal, it subtracts 2ms from each sampling time position to make the two signals have the same time reference before performing subsequent candidate feature point generation.

[0030] This example illustrates how time compensation is performed, without limiting the channel delay to 2ms.

[0031] S102: The data processing device generates a set of candidate feature points for the carotid artery and a set of candidate feature points for the femoral artery for multiple consecutive cardiac cycles; In this embodiment, the data processing device can first preprocess the carotid pulse wave signal and the femoral pulse wave signal; the preprocessing can include at least one of removing high-frequency noise, removing abnormal spikes, baseline correction, amplitude normalization, and sampling point relocation; after preprocessing, the data processing device can divide the carotid pulse wave signal and the femoral pulse wave signal into cardiac cycles according to the peak position, trough position, ECG reference time, or pulse wave periodicity characteristics.

[0032] Within each cardiac cycle, the data processing device can generate at least two candidate feature points based on at least one of the amplitude change, slope change, curvature change, local extremum position, and threshold cross position of the corresponding pulse wave signal. The candidate feature points are candidate time positions used to participate in the determination of the pulse wave arrival time, and they are not necessarily equivalent to the actual pulse wave arrival time.

[0033] In one example, for a pulse wave signal x(t) within any cardiac cycle, the data processing device can calculate the first-order difference d(t) = x(t) - x(t-1) and the second-order difference q(t) = d(t) - d(t-1); Where x(t) represents the pulse amplitude at sampling time position t, x(t-1) represents the pulse amplitude at the previous sampling time position, d(t) represents the first-order difference, and q(t) represents the second-order difference.

[0034] The data processing device can identify the position where d(t) enters the continuous positive value region from the low value region as the first type of candidate feature point, the position where q(t) changes sign and d(t) remains positive thereafter as the second type of candidate feature point, and the position where the amplitude reaches η times the peak amplitude of the current period as the third type of candidate feature point; where η is the amplitude proportionality coefficient, which can be 0.03 to 0.15 in one example.

[0035] In one example, foot conditions may include at least one of the following: The average amplitude within the preset front window before the candidate feature point is lower than the first proportion of the peak amplitude of this cycle. The number of sampling points with continuously positive first-order differences within the preset back window after the candidate feature point is not less than the first quantity threshold. The second-order difference near the candidate feature point changes sign. The time position of the candidate feature point is after the trough of this cycle and before the peak of the main wave. For example, the first proportion can be 3% to 15% of the peak amplitude of the current cycle, and the first quantity threshold can be determined based on the sampling frequency and the duration of the rising edge of the main wave.

[0036] The data processing device can merge the above candidate feature points and remove duplicate points with a time interval less than the preset number of sampling points to obtain a set of candidate feature points; The candidate feature point set can include candidate feature points whose local morphology satisfies the foot point condition, as well as candidate feature points located in the preceding segment of the rising edge of the main wave or near the rising edge of the main wave. In this way, the system will not directly exclude the front shoulder position or disputed foot position during the candidate generation stage, but will retain them for judgment in the subsequent point pair value update stage.

[0037] In a specific example, the peak amplitude of the main wave in a certain carotid artery cardiac cycle is normalized to 1.00. The data processing device uses the position where the amplitude reaches 0.05, the position where the second-order difference changes from positive to negative and the subsequent first-order difference remains positive, and the position of the local slope peak in the early part of the rising edge as candidate feature points of the carotid artery. If the above positions correspond to the 120th, 128th, and 124th sampling points, respectively, then the set of candidate feature points of the carotid artery in this cycle can include three candidate time positions: C1, C2, and C3. Among them, C3 may belong to the anterior shoulder position and is still retained for subsequent point pair construction steps.

[0038] like Figure 4 As shown, S103: The data processing device constructs multiple carotid-femoral candidate point pairs based on the carotid artery candidate feature point set and the femoral artery candidate feature point set; In this embodiment, the data processing device can determine the correspondence between the carotid artery cardiac cycle and the femoral artery cardiac cycle based on the cardiac cycle sequence number, reference trigger time, or cycle correspondence. For corresponding carotid artery cardiac cycles and femoral artery cardiac cycles, the data processing device can combine each carotid artery candidate feature point in the carotid artery cardiac cycle with each femoral artery candidate feature point in the femoral artery cardiac cycle to obtain multiple initial carotid-femoral candidate point pairs.

[0039] For each initial carotid-femoral candidate point pair, the data processing device can calculate the candidate propagation time difference based on the time position of the carotid artery candidate feature point and the time position of the femoral artery candidate feature point in the point pair. For example, if the time position of the carotid artery candidate feature point is tc and the time position of the femoral artery candidate feature point is tf, then the candidate propagation time difference Δt can be expressed as: Δt = tf - tc.

[0040] Where Δt represents the candidate propagation time difference, tf represents the time position of the femoral artery candidate feature point, and tc represents the time position of the carotid artery candidate feature point; tf and tc can be sampling timestamps or sampling point numbers. If tf and tc are sampling point numbers, then Δt can be multiplied by the sampling interval to convert it into time, and the time unit can be milliseconds or seconds.

[0041] The data processing device can identify initial neck-spine candidate point pairs whose candidate propagation time differences are within a preset propagation time range as neck-spine candidate point pairs to participate in temporal difference learning; the preset propagation time range can be determined based on the preset neck-spine propagation distance Dcf and the preset cf-PWV range. If the preset cf-PWV range is Vmin to Vmax, then the preset propagation time range can be [Dcf / Vmax, Dcf / Vmin]; Where Dcf represents the preset carotid-femoral propagation distance between the carotid artery measurement location and the femoral artery measurement location, Vmin represents the preset lower limit of cf-PWV, and Vmax represents the preset upper limit of cf-PWV.

[0042] In this way, initial neck-femur candidate point pairs that clearly do not conform to the physiological propagation relationship of cf-PWV can be excluded from the temporal difference learning process.

[0043] In a specific example, the preset carotid-femoral propagation distance (Dcf) is 0.65m, and the preset cf-PWV range is 4m / s to 20m / s. Then the preset propagation time range is [0.65 / 20, 0.65 / 4] seconds, i.e., 32.5ms to 162.5ms. If the carotid artery candidate feature point is located at 120ms and 128ms in a certain cycle, and the femoral artery candidate feature point is located at 194ms, then the candidate propagation time difference between the two initial point pairs is 74ms and 66ms, respectively, both of which are within the preset propagation time range and can be used as carotid-femoral candidate point pairs to participate in temporal difference learning.

[0044] like Figure 5 As shown, S104: The data processing device determines stable propagation time difference information for each neck-flesh candidate point pair; In this embodiment of the application, the data processing device can establish a sequence of similar neck-femoral candidate point pairs in multiple consecutive cardiac cycles according to the sequence number of the candidate feature point in the corresponding candidate feature point set, the local morphological category of the candidate feature point, or the interval where the candidate propagation time difference is located. For example, if there is a pair of points such as the carotid anterior shoulder candidate point and the femoral foot candidate point in each cardiac cycle, then such point pairs in multiple consecutive cardiac cycles can be combined into a sequence of similar carotid-femoral candidate point pairs.

[0045] For the same type of neck-femoral candidate point pair sequence, the data processing device can obtain the candidate propagation time difference of the sequence in multiple consecutive cardiac cycles, and determine the propagation time difference stability information based on at least one of the following: fluctuation amplitude, variance, range, change in adjacent cycles, and coefficient of variation of the candidate propagation time difference; the propagation time difference stability information can be a stability score or a stability marker used to indicate whether a preset stability condition is met.

[0046] In one example, if the candidate propagation time differences in M ​​consecutive cardiac cycles are Δt1, Δt2, ..., ΔtM, then the mean μΔt and the standard deviation σΔt can be calculated. Where M represents the number of cardiac cycles involved in the stability assessment, which can be taken as 3 to 10; μΔt represents the mean of M consecutive candidate propagation time differences; σΔt represents the standard deviation of M consecutive candidate propagation time differences; the propagation time difference stability score Ts can be expressed as: Ts=1 / (1+σΔt / τ0).

[0047] Where Ts represents the propagation time difference stability score, τ0 represents the preset normalized time scale, which can be the duration corresponding to a sampling interval, 5ms, 10ms, or a reference duration determined according to the device sampling frequency, and σΔt / τ0 is a dimensionless quantity. The smaller σΔt is, the larger Ts is, indicating that the candidate propagation time difference is more stable.

[0048] The above formula is only an example. In actual implementation, variance, range, changes in adjacent periods, or coefficient of variation can also be used to calculate the propagation time difference stability information.

[0049] In one example, when σΔt is less than the first stability threshold, or when σΔt / μΔt is less than the second stability threshold, it can be determined that the propagation time difference stability information meets the preset stability condition. The first stability threshold can be determined based on the sampling frequency, for example, the duration corresponding to 2 to 8 sampling intervals, and the second stability threshold can be taken as 0.03 to 0.15.

[0050] It should be noted that, in the embodiments of this application, a stable propagation time difference does not necessarily indicate that the carotid-femoral candidate point pair is correct. Although the propagation time difference of some point pairs is relatively stable in multiple consecutive cardiac cycles, their carotid artery candidate feature points and femoral artery candidate feature points may not be in the same physiological phase. Therefore, it is necessary to combine the corresponding physiological phase information for judgment.

[0051] In a specific example, the candidate propagation time differences of a certain type of neck-femoral candidate point pair in 5 consecutive cardiac cycles were 68ms, 69ms, 68ms, 70ms and 69ms, respectively, with a mean of about 68.8ms and a standard deviation of about 0.75ms. If τ0 is taken as 5ms, then Ts=1 / (1+0.75 / 5), which is about 0.87, indicating that this type of point pair has high stability of propagation time difference.

[0052] This example only indicates that the time difference is stable, and does not directly indicate that the point corresponds to the physiological phase.

[0053] S105: The data processing device determines the information corresponding to the physiological time; In this embodiment, the data processing device can extract a signal segment of a first preset duration starting from the candidate feature point of the carotid artery to obtain the main wave growth sequence of the short window after the carotid artery; and extract a signal segment of a second preset duration starting from the candidate feature point of the femoral artery to obtain the main wave growth sequence of the short window after the femoral artery; the first preset duration and the second preset duration can be the same or different, and can be determined according to the sampling frequency, cardiac cycle duration, average duration of the rising edge of the pulse wave or preset parameters of the device.

[0054] When the first preset duration and the second preset duration are different, the data processing device can resample the carotid artery posterior short window main wave growth sequence and the femoral artery posterior short window main wave growth sequence to the same length, or extract the first K common sampling points of the two and calculate the sequence correlation. Where K represents the number of uniform sampling points used to calculate sequence correlation.

[0055] In one example, the main wave growth sequence Xc of the short window of the carotid artery can be represented as a normalized amplitude increment sequence of K consecutive sampling points starting from the candidate feature point of the carotid artery, and the main wave growth sequence Xf of the short window of the femoral artery can be represented as a normalized amplitude increment sequence of K consecutive sampling points starting from the candidate feature point of the femoral artery. Where Xc represents the main wave growth sequence of the short window in the carotid artery, Xf represents the main wave growth sequence of the short window in the femoral artery, and K represents the number of sampling points in the short window, which can be determined based on the sampling frequency and the duration of the rising edge of the main wave. For example, when the sampling frequency is 1000Hz, K can be between 20 and 80.

[0056] The data processing device can determine the physiological time-corresponding information based on the sequence correlation and normalized phase deviation between the main wave growth sequence Xc of the carotid artery posterior short window and the main wave growth sequence Xf of the femoral artery posterior short window, and by combining at least one of the consistency of growth direction and similarity of slope maintenance duration. The physiological time-corresponding information is not used to determine whether the amplitudes of the two waveforms are exactly the same, but rather to determine whether the main wave growth process after the candidate feature point of the carotid artery and the main wave growth process after the candidate feature point of the femoral artery both enter a continuous rising phase, and whether the two satisfy a preset correspondence in terms of growth direction, slope maintenance duration, and sequence correlation.

[0057] In one example, the physiological time-corresponding score Ps can be determined by the following formula: Ps=w1·Corr(Xc,Xf)+w2·Gc+w3·Sc-w4·Pdn; Wherein, Ps represents the score corresponding to the physiological time; Corr(Xc,Xf) represents the sequence correlation between the carotid artery posterior short window main wave growth sequence Xc and the femoral artery posterior short window main wave growth sequence Xf; Gc represents the growth direction consistency score, used to characterize whether the growth directions of the two sequences are consistent at the corresponding sampling positions; Sc represents the slope duration similarity score, used to characterize whether the duration of the slope remaining positive in the two sequences is close; Pdn represents the normalized phase deviation, Pdn=Pd / P0; Pd represents the number of offset sampling points or offset duration of the two posterior short window main wave growth sequences at the position of maximum correlation; P0 represents the preset phase normalization scale, which can be K sampling points, the posterior short window length, or the preset maximum allowable phase deviation; w1, w2, w3, and w4 are non-negative weight coefficients; Corr(Xc,Xf), Gc, Sc, and Pdn can all be normalized to the range of 0 to 1.

[0058] The larger the value of Ps, the more closely the main wave growth processes following the two candidate feature points correspond, and the more likely they are to belong to the same physiological phase.

[0059] In some embodiments of this application, w1, w2, w3 and w4 can be normalized so that w1+w2+w3+w4=1; the calculated Ps can be restricted to the range of 0 to 1 by a truncation function, that is, when Ps is less than 0, it is taken as 0, and when Ps is greater than 1, it is taken as 1, and the truncated score is used as the corresponding score of the final physiological time.

[0060] In one example, when Corr(Xc,Xf) is greater than the first phase threshold and Pdn is less than the second phase threshold, it can be determined that the physiological time-related information meets the preset phase correspondence condition. For example, when Corr(Xc,Xf) ≥ 0.75 and Pdn ≤ 0.2, it can be determined that the physiological time-related information meets the preset phase correspondence condition; when Corr(Xc,Xf) < 0.75 or Pdn > 0.2, it can be determined that the physiological time-related information does not meet the preset phase correspondence condition.

[0061] It should be noted that the above threshold is only an example and can also be determined based on equipment calibration samples, historical measurement data, or verification samples.

[0062] In some embodiments of this application, w1, w2, w3, and w4 can be determined by calibration samples. Calibration samples may include carotid pulse wave signals and femoral pulse wave signals of target neck-femoral feature point pairs that have been manually confirmed. The data processing device can use manually confirmed target point pairs as positive samples and point pairs with stable propagation time differences but mismatched physiological time phases as negative samples, and adjust w1, w2, w3, and w4 so that the physiological time phase corresponding scores of positive samples are higher than those of negative samples.

[0063] In a specific example, for candidate point pair A, the sequence correlation of the main wave growth sequence of the carotid artery posterior short window and the main wave growth sequence of the femoral artery posterior short window is 0.82, the normalized phase deviation is 0.12, the consistency score of the growth direction is 0.90, and the similarity score of the slope maintenance duration is 0.84. Therefore, this point pair can meet the preset phase correspondence condition. For candidate point pair B, although its propagation time difference is stable, the sequence correlation is 0.58, the normalized phase deviation is 0.31, and there is a plateau and fallback in the carotid artery posterior short window. Therefore, this point pair does not meet the preset phase correspondence condition.

[0064] S106: Data processing equipment identifies neck-femoral candidate point pairs with stable propagation time difference but mismatched physiological phases; In this embodiment of the application, when the propagation time difference stability information of the target neck-thigh candidate point pair meets the preset stability condition, and the physiological time-phase correspondence information of the target neck-thigh candidate point pair does not meet the preset phase correspondence condition, the data processing device can identify the target neck-thigh candidate point pair as a neck-thigh candidate point pair with stable propagation time difference but mismatched physiological time phase. Among them, the preset stability condition is used to limit the fluctuation of the candidate propagation time difference in multiple consecutive cardiac cycles, and the preset phase correspondence condition is used to limit the phase correspondence or sequence growth correspondence between the carotid artery posterior short window main wave growth sequence and the femoral artery posterior short window main wave growth sequence. For example, in five consecutive cardiac cycles, the candidate propagation time differences of a certain type of neck-femoral candidate point pair are 68ms, 69ms, 68ms, 70ms, and 69ms, respectively, indicating that the propagation time difference of the candidate point pair fluctuates little and meets the preset stability condition. However, if the main wave growth sequence after the carotid artery candidate feature point in the candidate point pair shows a brief plateau and a secondary decline, while the main wave growth sequence after the femoral artery candidate feature point continues to rise, and the sequence correlation of the two main wave growth sequences is lower than the first phase threshold, or the normalized phase deviation is greater than the second phase threshold, then the data processing device identifies the candidate point pair as a neck-femoral candidate point pair with a stable propagation time difference but a physiological phase mismatch.

[0065] In this way, the system can avoid directly equating the stability of propagation time difference with the physiological correctness of feature points, thereby reducing the risk of stable but erroneous point pairs entering the final cf-PWV calculation process.

[0066] S107: Instant reward for data processing equipment determining neck-femur candidate point pairs; In this embodiment of the application, the data processing device can determine the instantaneous reward of the neck-femur candidate point pair based on the propagation time difference stability information and the physiological time correspondence information: When the propagation time difference stability information of the target neck-thigh candidate point pair meets the preset stability condition, and the physiological time-corresponding information of the target neck-thigh candidate point pair meets the preset phase correspondence condition, the data processing device can generate the first instant reward. When the propagation time difference stability information of the target neck-thigh candidate point pair meets the preset stability condition, and the physiological time-corresponding information of the target neck-thigh candidate point pair does not meet the preset phase correspondence condition, the data processing device can generate a second instant reward and mark the target neck-thigh candidate point pair as a neck-thigh candidate point pair with stable propagation time difference but mismatched physiological time phase; the second instant reward is less than the first instant reward. When the propagation time difference stability information of the target neck-thigh candidate point pair does not meet the preset stability condition, and the physiological time-corresponding information of the target neck-thigh candidate point pair meets the preset phase correspondence condition, the data processing device can generate a third instant reward. When the propagation time difference stability information of the target neck-thigh candidate point pair does not meet the preset stability condition, and the physiological time-corresponding information of the target neck-thigh candidate point pair does not meet the preset phase correspondence condition, the data processing device can generate a fourth instant reward. Among them, the third instant reward is less than the first instant reward, and the fourth instant reward is less than or equal to the second instant reward.

[0067] In this way, each neck-femur candidate point pair participating in temporal difference learning receives a corresponding immediate reward.

[0068] In one example, the first instant reward can be set to a positive value R1, and the second instant reward can be set to a value R2 that is less than R1; R2 can be 0 or a negative value. When the candidate point pair satisfies the propagation time difference stability condition and the preset phase correspondence condition, R1 is used as the immediate reward. When a candidate point pair satisfies the propagation time difference stability condition but does not satisfy the preset phase correspondence condition, R2 is used as the immediate reward.

[0069] In this way, candidate point pairs with stable propagation time differences but mismatched physiological phases will not be continuously reinforced in temporal difference learning due to the stability of the time difference.

[0070] In a specific example, the first immediate reward R1 can be 1.0, the second immediate reward R2 can be -0.3, the third immediate reward can be 0.2, and the fourth immediate reward can be -0.5. If a candidate point pair has a stable propagation time difference but does not correspond to the physiological time phase, then -0.3 is used as the immediate reward, causing that point pair to be downweighted in subsequent TD updates. The above values ​​are only used to illustrate the execution method of the reward branch and do not limit the actual reward to the above values.

[0071] like Figure 6 As shown, S108: The data processing device updates the value of point pairs based on temporal difference learning; In this embodiment, the data processing device can construct a point-pair state vector for each neck-femur candidate point pair; the point-pair state vector includes propagation time difference stabilization information and physiological time corresponding information, and includes at least one of candidate propagation time difference, candidate feature point local morphology information and post-short window main wave growth information; the candidate feature point local morphology information may include at least one of slope, curvature, amplitude change, local extreme value position and threshold cross position near the candidate point; the post-short window main wave growth information may include at least one of the main wave continuous rise duration after the candidate point, cumulative amplitude growth, slope maintenance duration and secondary fall amplitude.

[0072] The data processing device can determine the state of the neck-femur candidate point pairs that satisfy the temporal correspondence with the current neck-femur candidate point pair in the next adjacent cardiac cycle as the next temporal state. Satisfying the temporal correspondence can include at least one of the following: the candidate propagation time difference falls within the same interval, the candidate feature points have the same local morphological category, and the candidate feature points have the same index in the corresponding candidate feature point set.

[0073] In one example, the timing difference error δ can be expressed as: δ=r+γ·V(s')-V(s); Where δ represents the temporal difference error; r represents the immediate reward; γ represents the discount factor, which can be 0≤γ≤1; V(s) represents the point-pair value corresponding to the current point-pair state s; V(s') represents the point-pair value corresponding to the next temporal state s'; s represents the point-pair state of the current neck-stock candidate point pair; and s' represents the point-pair state that satisfies the temporal correspondence in the next adjacent cardiac cycle.

[0074] The data processing device can update the point-pair value of the current point-pair state according to the following formula: Vnew(s) = V(s) + α·δ; Where Vnew(s) represents the updated point-pair value; α represents the learning rate, which can be greater than 0 and less than or equal to 1.

[0075] In some embodiments of this application, the initial point pair value V0(s) can be uniformly initialized to 0, or it can be weighted and initialized according to whether the candidate propagation time difference is within a preset propagation time range, the local morphological score of the candidate feature point, and the corresponding score of the physiological time. For multiple consecutive cardiac cycles, the data processing device can perform point-to-point value updates in chronological order; When there is no candidate point pair for the next cardiac cycle that satisfies the temporal correspondence, the state of the point pair can be taken as the termination state, and the point pair value of the next temporal state can be set to 0.

[0076] In a specific example, the current pair value V(s) of a pair with a stable propagation time difference but a physiological time mismatch is 0.45, the pair value V(s') of the next time state is 0.30, the immediate reward r is -0.3, the discount factor γ is 0.8, and the learning rate α is 0.4. Then δ = -0.3 + 0.8 × 0.30 - 0.45 = -0.51, and the updated pair value Vnew(s) = 0.45 + 0.4 × (-0.51) = 0.246.

[0077] This shows that although stable but mismatched point pairs have stable time differences, their point pair value is reduced due to the mismatch in physiological phases.

[0078] In another specific example, given a stable propagation time difference and a point-to-point value V(s) corresponding to a physiological time of 0.45, a point-to-point value V(s') for the next time-series state of 0.55, an immediate reward r of 1.0, a discount factor γ of 0.8, and a learning rate α of 0.4, then δ = 1.0 + 0.8 × 0.55 - 0.45 = 0.99, and the updated point-to-point value Vnew(s) = 0.45 + 0.4 × 0.99 = 0.846.

[0079] This demonstrates that stable and phase-corresponding point pairs are reinforced in temporal difference learning.

[0080] S109: The data processing device performs point-to-point value inheritance blocking processing based on the risk information of the pseudofoot point on the anterior shoulder; In some embodiments of this application, the point-to-state vector may also include risk information of pseudo-foot points in the front shoulder; a pseudo-foot point in the front shoulder refers to a candidate feature point located before the starting point of the continuous rise of the main wave, which meets the foot point condition in the local slope or curvature, but does not form a continuous rise of the main wave within a preset short window thereafter.

[0081] Risk information for the pseudofoot point in the anterior shoulder can be determined based on at least one of the following: the slope change before and after the candidate feature point, the duration of the continuous rise of the main wave after the candidate feature point, the magnitude of the secondary fall after the candidate feature point, and the position of the rising edge of the main wave where the candidate feature point is located.

[0082] In one example, if the candidate feature point is located after the trough of the current cycle and before the peak of the current cycle, and the number of consecutively rising sampling points within L sampling points after the candidate feature point is less than L / 2, and there is a case where the amplitude drop after the candidate feature point exceeds 3% to 10% of the peak amplitude of the current cycle, then the candidate feature point can be determined to meet the risk conditions of the pseudofoot point of the anterior shoulder. Where L represents the number of subsequent sampling points used to determine the risk of pseudofoot points in the anterior shoulder, which can be determined based on the sampling frequency and the duration of the rising edge of the main wave.

[0083] When the carotid artery candidate feature point or femoral artery candidate feature point in the target neck-femoral candidate point pair meets the risk condition of the anterior shoulder pseudofoot point, the data processing device can reduce the immediate reward of the target neck-femoral candidate point pair and perform point pair value inheritance blocking processing on the target neck-femoral candidate point pair.

[0084] The point-to-point value inheritance blocking process is used to prohibit or attenuate the impact of the updated point-to-point value of the target neck-femur candidate point-to-point pair on the initial point-to-point value of the time-series corresponding point-to-point pair in the next cardiac cycle.

[0085] In one example, the point-to-point value inheritance blocking process may include: when initializing the point-to-point value in the next cardiac cycle, prohibiting the use of the updated point-to-point value of the target neck-thigh candidate point-to-point pair as the initial point-to-point value of the neck-thigh candidate point-to-point pair corresponding to its time sequence; In another example, the point-pair value inheritance blocking process may include: reducing the impact of the updated point-pair value of the target neck-femur candidate point pair on the next cardiac cycle according to a preset decay coefficient, wherein the preset decay coefficient is less than 1; For example, when a candidate feature point of the carotid artery is located in the early part of the rising edge of the main wave, and then a small rise appears in the short window followed by a plateau or fall back, while the candidate feature point of the femoral artery enters the continuous rising phase of the main wave, the candidate feature point of the carotid artery may be a pseudofoot point of the anterior shoulder. If the system does not block the inheritance of its point-pair value, the erroneous point-pair may be continuously reinforced in subsequent cycles.

[0086] Therefore, by blocking point-to-point value inheritance, the risk of pseudo-foot points being locked across cycles can be reduced.

[0087] In a specific example, among the 40 sampling points following a certain carotid artery candidate feature point, only 14 sampling points showed consecutive upward movement, less than 40 / 2, and there was a subsequent amplitude drop of 6% of the peak amplitude of the current cycle. The data processing device marked this candidate feature point as a risk point of anterior shoulder pseudofoot and blocked the value inheritance of the carotid-femoral candidate point pair containing this candidate feature point, so that the updated point pair value of this erroneous point pair would no longer be used as the initial point pair value of the same type of point pair in the next cardiac cycle.

[0088] S110: The data processing equipment determines the target neck-thigh feature point pair and outputs the corresponding results; In this embodiment of the application, the data processing device can sort multiple neck-femoral candidate point pairs according to the updated point pair value, and select the neck-femoral candidate point pair with the highest point pair value and satisfying the preset phase correspondence condition from the sorted multiple neck-femoral candidate point pairs as the target neck-femoral feature point pair; the target neck-femoral feature point pair includes the target carotid artery feature point and the target femoral artery feature point.

[0089] The data processing device can determine the target propagation time difference based on the time position of the target carotid artery feature point and the time position of the target femoral artery feature point; for example, if the time position of the target carotid artery feature point is Tc and the time position of the target femoral artery feature point is Tf, then the target propagation time difference ΔT can be expressed as: ΔT = Tf - Tc; Where ΔT represents the target propagation time difference, Tf represents the time position of the target femoral artery feature point, and Tc represents the time position of the target carotid artery feature point; Tf and Tc can be sampling timestamps or sampling point numbers; if Tf and Tc are sampling point numbers, then ΔT can be multiplied by the sampling interval to convert it into time.

[0090] The data processing equipment can also determine the cf-PWV measurement value based on the target propagation time difference and the preset neck-femur propagation distance; for example, the cf-PWV measurement value can be determined according to the following formula: cf-PWV=Dcf / ΔT; Wherein, Dcf represents the preset neck-femur propagation distance, which can be in meters; ΔT represents the target propagation time difference, which can be in seconds; cf-PWV represents the neck-femur pulse wave propagation velocity, which can be in meters per second; Dcf can be obtained by manual input, body surface distance measurement, device preset, or user parameter estimation.

[0091] The data processing equipment can also determine the reliability of the positioning of the target neck-femur feature point pair based on the weighted result of the point pair value, propagation time difference stability information, and physiological time-corresponding information of the target neck-femur feature point pair.

[0092] In one example, the location confidence C can be determined according to the following formula: C=a·Vn+b·Ts+c·Ps; Where C represents the location reliability; Vn represents the normalized point-pair value; Ts represents the propagation time difference stability score; Ps represents the physiological time phase corresponding score; a, b, and c are non-negative weight coefficients, and a+b+c=1; Vn, Ts, and Ps can all be normalized to the range of 0 to 1, so that the value of C is also within the range of 0 to 1; when the target neck-thigh feature point pair is identified as having a stable propagation time difference but a physiological time phase mismatch, C can be multiplied by the penalty coefficient β, where β is less than 1.

[0093] In one example, the normalized point-pair value Vn can be obtained by max-min normalization of the point-pair values ​​of multiple neck-femoral candidate point pairs within the same cardiac cycle or the same processing window, i.e., Vn=(V(s)-Vmin) / (Vmax-Vmin), where Vmax represents the maximum point-pair value and Vmin represents the minimum point-pair value; when Vmax equals Vmin, Vn can be set to a preset constant.

[0094] In a specific example, the data processing device ultimately selects the target carotid artery feature point time position as 128ms and the target femoral artery feature point time position as 194ms. Therefore, the target propagation time difference ΔT is 66ms, or 0.066s. If the preset carotid-femoral propagation distance Dcf is 0.65m, then the cf-PWV measurement value is 0.65 / 0.066, approximately 9.85m / s. This cf-PWV measurement value is only used as a medical device measurement parameter output and is not directly used as a disease diagnosis conclusion.

[0095] The data processing device can output at least one of the following: target neck-femur feature point pair, target propagation time difference, cf-PWV measurement value, and positioning reliability; the output results can be used for medical device display interface display, quality control, measurement record storage, or verification prompts.

[0096] Figure 7 A schematic diagram of a device structure provided in an embodiment of this application; Reference Figure 7The device may include a processor, memory, and a communication interface; The processor, memory, and communication interface can be connected via a bus; the memory can store computer program instructions, and when the processor executes the computer program instructions, it implements the cf-PWV feature point accurate localization method based on temporal difference learning as described in any of the above embodiments.

[0097] In some implementations, the communication interface can be used to receive carotid pulse wave signals and femoral pulse wave signals, and the processor can be used to perform processing steps such as time alignment, candidate feature point set generation, carotid-femoral candidate point pair construction, propagation time difference stability information determination, physiological time corresponding information determination, instant reward determination, point pair value update, and target carotid-femoral feature point pair output.

[0098] The processor can be a central processing unit, a digital signal processor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices. The memory can be read-only memory (ROM), random access memory (RAM), flash memory, or other suitable types of memory.

[0099] Based on the above embodiments, this application can also provide a computer-readable storage medium storing a computer program, which, when executed by a computer, causes the computer to perform the cf-PWV feature point accurate localization method based on temporal difference learning as described in any of the above embodiments.

[0100] The above description is merely a specific embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A precise feature point localization method based on temporal difference learning for cf-PWV, characterized in that, The method includes: Acquire time-aligned carotid pulse wave signals and femoral pulse wave signals; For multiple consecutive cardiac cycles, a set of candidate feature points for the carotid artery is generated in the carotid pulse wave signal, and a set of candidate feature points for the femoral artery is generated in the femoral pulse wave signal. Both the set of candidate feature points for the carotid artery and the set of candidate feature points for the femoral artery include multiple candidate time positions used to participate in the determination of the arrival time of the pulse wave. Based on the carotid artery candidate feature point set and the femoral artery candidate feature point set, multiple carotid-femoral candidate point pairs are constructed, each carotid-femoral candidate point pair includes one carotid artery candidate feature point and one femoral artery candidate feature point; For each carotid-femoral candidate point pair, propagation time difference stability information is determined based on the candidate propagation time difference in multiple consecutive cardiac cycles. Physiological time correspondence information is determined based on the main wave growth sequence of the short window after the carotid artery candidate feature point and the main wave growth sequence of the short window after the femoral artery candidate feature point. Physiological time correspondence information is used to characterize whether the main wave growth process after the carotid artery candidate feature point and the main wave growth process after the femoral artery candidate feature point satisfy the preset phase correspondence relationship. Based on the stable propagation time difference information and the corresponding physiological time information, the instantaneous reward of the neck-thigh candidate point pair is determined. Among them, the instantaneous reward of the neck-thigh candidate point pair with stable propagation time difference but mismatched physiological time is lower than the instantaneous reward of the neck-thigh candidate point pair with stable propagation time difference and corresponding physiological time. The point-to-point state of the neck-stock candidate point pair is used as the state of the temporal difference learning. The point-to-point value of the neck-stock candidate point pair is updated according to the immediate reward and the point-to-point value of the next temporal state. Based on the updated point-pair value, the target neck-stock feature point pair is determined from multiple neck-stock candidate point pairs.

2. The method according to claim 1, characterized in that, Acquire time-aligned carotid and femoral artery pulse wave signals, including: Acquire raw carotid pulse wave signals from the carotid artery acquisition channel and raw femoral pulse wave signals from the femoral artery acquisition channel; Based on at least one of the synchronization trigger signal, sampling timestamp, ECG reference signal, synchronization marker signal and pre-calibrated channel delay, time compensation is performed on the original carotid pulse wave signal and the original femoral pulse wave signal to obtain time-aligned carotid pulse wave signal and femoral pulse wave signal.

3. The method according to claim 1, characterized in that, For multiple consecutive cardiac cycles, candidate feature point sets for the carotid artery are generated in the carotid pulse wave signal, and candidate feature point sets for the femoral artery are generated in the femoral pulse wave signal, including: Cardiac cycle segmentation was performed on carotid artery pulse wave signals and femoral artery pulse wave signals; Within each cardiac cycle, at least two candidate feature points are generated based on at least one of the following: amplitude change, slope change, curvature change, local extremum location, and threshold crossover location of the corresponding pulse wave signal. Among them, at least two candidate feature points include candidate feature points whose local morphology satisfies the foot point condition, and candidate feature points located in the front part of the rising edge of the main wave or near the rising edge of the main wave. The conditions for a candidate feature point include at least one of the following: the average amplitude within a preset front window before the candidate feature point is lower than a first proportion of the peak amplitude of the current cycle; the number of sampling points with continuously positive first-order differences within a preset back window after the candidate feature point is not less than a first quantity threshold; the second-order differences near the candidate feature point undergo a sign change; and the time position of the candidate feature point is after the trough of the current cycle and before the peak of the main wave.

4. The method according to claim 1, characterized in that, Based on the carotid artery candidate feature point set and the femoral artery candidate feature point set, multiple carotid-femoral candidate point pairs are constructed, including: The correspondence between the carotid artery cardiac cycle and the femoral artery cardiac cycle is determined based on the cardiac cycle sequence number, reference trigger time, or cycle correspondence. The carotid candidate feature points and femoral candidate feature points in the corresponding carotid and femoral cardiac cycles are combined to obtain multiple initial carotid-femoral candidate point pairs. The candidate propagation time difference is calculated based on the time positions of the carotid artery candidate feature points and the femoral artery candidate feature points in each initial carotid-femoral candidate point pair. Initial neck-thigh candidate point pairs whose candidate propagation time difference is within the preset propagation time range are identified as neck-thigh candidate point pairs participating in temporal difference learning. The preset propagation time range is determined based on the ratio of the preset neck-thigh propagation distance to the preset upper limit of cf-PWV, and the ratio of the preset neck-thigh propagation distance to the preset lower limit of cf-PWV.

5. The method according to claim 1, characterized in that, For each neck-femoral candidate point pair, propagation time difference stability information is determined based on its candidate propagation time difference across multiple consecutive cardiac cycles, including: Based on the sequence number of the candidate feature point in the corresponding candidate feature point set, the local morphological category of the candidate feature point, or the interval of the candidate propagation time difference, establish the same neck-femoral candidate point pair sequence in multiple consecutive cardiac cycles; Obtain the candidate propagation time difference of similar neck-femoral candidate point pair sequences in multiple consecutive cardiac cycles; Based on at least one of the following: fluctuation range, variance, range, change in adjacent periods, and coefficient of variation, the propagation time difference stability information is determined. The propagation time difference stability information includes a propagation time difference stability score or a stability marker used to indicate whether a preset stability condition is met.

6. The method according to claim 1, characterized in that, Based on the carotid artery candidate feature points followed by the main wave growth sequence of the short window after the carotid artery and the femoral artery candidate feature points followed by the main wave growth sequence of the short window after the femoral artery, the corresponding physiological information is determined, including: Starting from the candidate feature points of the carotid artery, a signal segment of the first preset duration is extracted to obtain the main wave growth sequence of the short window of the carotid artery. Starting from the candidate feature points of the femoral artery, a signal segment of the second preset duration is extracted to obtain the femoral artery posterior short window main wave growth sequence; Based on the sequence correlation and normalized phase deviation between the main wave growth sequence of the posterior short window of the carotid artery and the main wave growth sequence of the posterior short window of the femoral artery, and combined with at least one of the consistency of growth direction and similarity of slope maintenance duration, the physiological time-corresponding information is determined; The normalized phase deviation is determined based on the ratio of the phase deviation to the preset phase normalization scale. The phase deviation is used to characterize the number of offset sampling points or the offset duration of the main wave growth sequence in the posterior short window of the carotid artery and the main wave growth sequence in the posterior short window of the femoral artery at the position of maximum correlation.

7. The method according to claim 1, characterized in that, Based on stable information about the time difference of propagation and corresponding physiological time information, the immediate reward for the neck-femur candidate point pair is determined, including: When the propagation time difference stability information of the target neck-thigh candidate point pair meets the preset stability condition, and the physiological time-corresponding information of the target neck-thigh candidate point pair meets the preset phase correspondence condition, the first instant reward is generated. When the propagation time difference stability information of the target neck-thigh candidate point pair meets the preset stability condition, and the physiological time-corresponding information of the target neck-thigh candidate point pair does not meet the preset phase correspondence condition, a second instant reward is generated, and the target neck-thigh candidate point pair is marked as a neck-thigh candidate point pair with stable propagation time difference but mismatched physiological time phase. When the propagation time difference stability information of the target neck-thigh candidate point pair does not meet the preset stability condition, and the physiological time-corresponding information of the target neck-thigh candidate point pair meets the preset phase correspondence condition, a third instant reward is generated. When the propagation time difference stability information of the target neck-thigh candidate point pair does not meet the preset stability condition, and the physiological time-corresponding information of the target neck-thigh candidate point pair does not meet the preset phase correspondence condition, a fourth instant reward is generated. Among them, the second instant reward is less than the first instant reward, the third instant reward is less than the first instant reward, and the fourth instant reward is less than or equal to the second instant reward.

8. The method according to claim 1, characterized in that, Using the point-to-point state of the neck-stock candidate point pair as the state for temporal difference learning, the point-to-point value of the neck-stock candidate point pair is updated based on the immediate reward and the point-to-point value of the next temporal state, including: For each neck-femur candidate point pair, a point pair state vector is constructed. The point pair state vector includes propagation time difference stability information and physiological time corresponding information, and includes at least one of the following: candidate propagation time difference, candidate feature point local morphology information, and post-short window main wave growth information. The point pair state of the neck-femur candidate point pair that satisfies the temporal correspondence with each neck-femur candidate point pair in the next adjacent cardiac cycle is determined as the next temporal state. The temporal correspondence includes at least one of the following: the candidate propagation time difference falls into the same interval, the candidate feature points have the same local morphological category, and the candidate feature points have the same index in the corresponding candidate feature point set. The time series difference error is calculated based on the immediate reward, the current point-to-point value, the point-to-point value of the next time series state, and the discount factor. The point-pair value of each neck-stock candidate point pair is updated based on the time-series difference error.

9. The method according to claim 8, characterized in that, The point-to-state vector also includes the risk information of the pseudofoot point in the front shoulder. The risk information of the pseudofoot point in the front shoulder is determined based on at least one of the following: the slope change before and after the candidate feature point, the duration of the continuous rise of the main wave after the candidate feature point, the magnitude of the second fall after the candidate feature point, and the position of the rising edge of the main wave where the candidate feature point is located. The method further includes: When the carotid artery candidate feature point or femoral artery candidate feature point in the target neck-femoral candidate point pair meets the risk condition of the anterior shoulder pseudofoot point, the immediate reward of the target neck-femoral candidate point pair is reduced, and the point pair value inheritance blocking process is performed on the target neck-femoral candidate point pair. Among them, the point-to-point value inheritance blocking process is used to prohibit or attenuate the influence of the updated point-to-point value of the target neck-femur candidate point-to-point pair on the initial point-to-point value of the time-series corresponding point-to-point pair in the next cardiac cycle.

10. The method according to any one of claims 1-9, characterized in that, Based on the updated point-pair values, target neck-stock feature pairs are determined from multiple neck-stock candidate point pairs, including: Multiple neckline-stock candidate point pairs are sorted based on the updated point-pair value; From the sorted multiple neck-stock candidate point pairs, select the neck-stock candidate point pair with the highest point pair value and that meets the preset phase correspondence condition as the target neck-stock feature point pair; Determine the target propagation time difference based on the carotid artery feature points and femoral artery feature points in the target neck-femoral feature point pair; The cf-PWV measurement value is determined based on the target propagation time difference and the preset neck-femur propagation distance; The reliability of the target neck-thigh feature point pair location is determined based on the weighted results of the point pair value, propagation time difference stability information, and physiological time-corresponding information. Output at least one of the following: target neck-thigh feature point pair, target propagation time difference, cf-PWV measurement, and positioning confidence level.

Images