Method for managing feedback data from wearable sensing devices after heart valve surgery

Abstract

The application discloses a wearable sensing device feedback data management method after heart valve surgery, and relates to the technical field of data processing; the management method comprises the following steps: acquiring original ECG data of a target patient continuously and dynamically after surgery, configuring an edge computing node, performing P wave feature extraction and structured processing on the original ECG data, calculating a key feature parameter set, and performing a structured packaging processing action; according to the key feature parameter set and in combination with a collected clinical feature data set; the technical key points are as follows: when the state of the Pmean trend correction mechanism and the Pmin fluctuation suppression mechanism is synchronously activated, a secondary collaborative correction strategy based on clinical credibility scoring is introduced, which can not only determine the dominant mechanism by quantitatively comparing the pathological direction of the two types of signals, but also can avoid decision conflicts through weighted fusion, solve the logical confusion problem caused by the parallel of multiple source feedback signals, and realize robust decision-making in complex scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a wearable sensing device feedback data management method after heart valve surgery. BACKGROUND

[0002] In the wearable sensing device feedback data management method after heart valve surgery, the data processing of the prior art usually refers to that the original electrocardiogram (ECG) signal collected by the wearable device is transmitted to a central server or a cloud platform through a wired or wireless manner, and signal processing and feature extraction operations such as filtering, denoising, R-wave detection and P-wave identification are performed in a remote data center; then, based on the extracted physiological parameters and clinical information, a statistical model or a machine learning algorithm is used to complete the atrial fibrillation risk assessment and early warning judgment in the centralized system; the whole process relies on stable network connection, data is uploaded in the form of original waveform or uncompressed, the processing delay is high, and there is a lack of local real-time analysis capability.

[0003] In the actual clinical application of current postoperative heart rhythm monitoring of heart valve surgery, the mainstream data processing mode still relies on the traditional centralized architecture; for example, the wearable ECG patch worn by the patient continuously uploads the original ECG signal to the hospital server through Bluetooth or Wi-Fi, and unified analysis is performed in the remote supporting data center, this mode has some practical defects: first, the original waveform data is huge, which requires a large amount of network bandwidth, once the patient is in a mobile state or the hospital network is congested, transmission delay or even packet loss is easy to occur, which leads to that the key heart rhythm event cannot be identified in time; second, the existing early warning model mostly uses fixed threshold to complete the judgment action, for example, P-wave duration > 170 ms triggers an alarm, but individual differences are not considered, for example, a young patient's maximum P-wave duration is 168 ms, which is frequently misreported due to slight fluctuations, while an old patient's average P-wave duration is continuously rising but is ignored although it does not reach the threshold; more importantly, when multiple physiological parameters change simultaneously, the system cannot distinguish which is the real pathological progression and which is the artifact interference, the traditional system often mechanically superimposes the judgment, leading to decision confusion, and lacking the comprehensive weighing ability of multi-source signals, the above problems together cause the confusion and inefficiency of current postoperative heart rhythm management. SUMMARY

[0004] To achieve the above purpose, the present application is realized by the following technical scheme:

[0005] The wearable sensing device feedback data management method after heart valve surgery comprises the following steps:

[0006] Obtaining the continuous and dynamic original ECG data of the target patient after surgery, configuring an edge computing node, performing P-wave feature extraction and structured processing on the original ECG data, calculating a set of key feature parameters, and performing structured packaging processing action;

[0007] According to the key feature parameter set and in combination with the collected clinical feature data set, a pre-constructed double-level prediction model is run, a model collaborative result is output, and a dynamic correction and adaptive optimization mechanism is triggered synchronously;

[0008] A final POAF occurrence probability value and an abnormal pattern index AI are derived;

[0009] It is judged whether the final abnormal pattern index AI exceeds a limited threshold value, and a judgment result is derived; wherein the judgment result is based on: if yes, a level III early warning is directly triggered; and if no, a hierarchical early warning mechanism is started;

[0010] The judgment result is encrypted and fed back to a target port, and data storage is completed.

[0011] Further, the original ECG data represent data obtained by implementing continuous dynamic electrocardiogram ECG monitoring of a target patient receiving a heart valve surgery for ≥7 days after the surgery; the process of performing P wave feature extraction and structured processing is as follows:

[0012] The original ECG data is subjected to two-stage filtering: adaptive notch filtering is applied, and wavelet transform is used for multi-scale denoising; based on an improved Pan-Tompkins algorithm combined with a morphological gradient method, R wave peaks are located, the boundaries of a cardiac cycle are demarcated, and a method combining dynamic template matching and first-order derivative is used to identify the P wave starting point and the P wave ending point; wherein the P wave starting point is defined as the intersection of the rising branch and the isoelectric line, and the P wave ending point is defined as the point where the descending branch returns to the isoelectric line.

[0013] Further, the process of using the improved Pan-Tompkins algorithm combined with the morphological gradient method is as follows:

[0014] The signal corresponding to the original ECG data is subjected to band-pass filtering to enhance the QRS wave, and after differential operation, it is subjected to square processing, and then envelope is obtained by moving window integration; the morphological gradient method is introduced to smooth noise by morphological closing operation and suppress baseline drift by morphological opening operation, and the difference between the upper and lower morphological gradients is calculated to highlight the steep edges of the QRS complex; in combination with the adaptive threshold detection mechanism of the original Pan-Tompkins algorithm, the peak points exceeding the dynamic threshold in the integrated signal are searched, and the R wave position is located through RR interval constraint and template matching verification;

[0015] The process of using the method combining dynamic template matching and first-order derivative is as follows:

[0016] After locating the R-wave position, search for the P-wave in the PR segment, locate the P-wave starting point using the first derivative method, when the derivative of three consecutive sampling points is greater than the set slope threshold and the signal rises, it is determined as the starting point; the end point is determined by the negative to positive of the falling branch derivative; combined with dynamic template matching, the average template is constructed with the identified P-wave, the current waveform is compared by sliding, the first derivative result is calculated, if the first derivative result and the template matching position deviation is less than 10ms, it is adopted.

[0017] Further, the calculated key feature parameter set at least includes: maximum P-wave time limit Pmax, minimum P-wave time limit Pmin, average P-wave time limit Pmean, P-wave time limit standard deviation Pstd, P-wave dispersion PWd, average P-wave peak time Pptmean and P-wave peak time standard deviation Pptstd; wherein, P-wave dispersion PWd: PWd=Pmax-Pmin;

[0018] When performing the structured packaging action, the structured JSON format is packaged, at least including: target patient ID, time stamp and 7 P-wave parameters in the key feature parameter set, and temporarily stored in the local database.

[0019] Further, the collected clinical feature data set at least includes: the age Nu of the target patient, the left atrial diameter LA, the glomerular filtration rate GFR, the urea Urea, the mechanical ventilation time Hr, and the independent prediction factor coinciding with the key feature parameter set, i.e. Pmax, Pstd and PWd; the double-layer prediction model includes a main model and an auxiliary model;

[0020] Wherein, the main model: according to the latest key feature parameter set received within a set time length, synchronously acquires the clinical feature data set of the target patient, and standardizes the 8 input variables in the clinical feature data set, substitutes the standardized input variables into the pre-trained binary Logistic regression equation, and obtains the POAF occurrence probability value RS; the auxiliary model: adopts a time series pattern recognition model based on SVM, inputs the sliding window sequence of independent prediction factors in the past predetermined time length, forms a high-dimensional feature space, and the SVM maps the sliding window sequence data to the high-dimensional space through a kernel function, finds an optimal hyperplane, and outputs an abnormal pattern index AI.

[0021] Further, the process of the triggered dynamic correction and adaptive optimization mechanism is as follows:

[0022] After receiving the key feature parameter set within a set time length, in addition to sending the independent prediction factor into the main model, the trend analysis and variation degree evaluation processing of Pmin, Pmean, Pptmean and Pptstd are performed synchronously, at least including: Pmean trend correction mechanism, Pmin fluctuation suppression mechanism and joint stability evaluation mechanism;

[0023] Continuing to analyze the results of the trend analysis and variability evaluation process, when there is a condition for simultaneous activation of the Pmean trend correction mechanism and the Pmin fluctuation suppression mechanism, a secondary collaborative correction strategy based on evidence weight fusion is started.

[0024] Further, the operation process of the Pmean trend correction mechanism is as follows:

[0025] The slope change rate k_Pmean of Pmean in the past predetermined length is calculated as a trend judgment index; if k_Pmean≤Q1, no correction is made; if Q1

[0026] The operation process of the Pmin fluctuation suppression mechanism is as follows:

[0027] The coefficient of variation CV_Pmin of Pmin is calculated: CV_Pmin=standard deviation / mean; wherein the standard deviation and the mean are calculated based on the Pmin values extracted every 5 minutes in the past 1 hour, a total of 12 data points; if CV_Pmin≤Q2, set the weight w_AI of the abnormal mode index AI corresponding to the SVM to 1; if Q2

[0028] The operation process of the joint stability evaluation mechanism is as follows:

[0029] A stability index SI is constructed: SI=Pptmean×Pptstd; when SI≤Q3, no response action is taken; when SI>Q3, a sensitivity enhancement mode is triggered, and the secondary and tertiary early warning thresholds corresponding to the main model are down-regulated; Q3 is a third standard threshold, whose value range is greater than 0.

[0030] Further, the process of the secondary collaborative correction strategy based on evidence weight fusion is as follows:

[0031] The clinical credibility scores CCS_Pmean of the Pmean trend correction mechanism and CCS_Pmin of the Pmin fluctuation suppression mechanism are calculated, and CCS_Pmean is compared with CCS_Pmin, when CCS_Pmean is greater than or equal to 2*CCS_Pmin, a first fusion sub-strategy is adopted, the POAF occurrence probability value RS after the Pmean trend correction mechanism is processed is revised twice to obtain the final determined RS_fin, when CCS_Pmin is greater than or equal to 2*CCS_Pmean, a second fusion sub-strategy is adopted, the POAF occurrence probability value RS after the Pmean trend correction mechanism is processed is revised twice to obtain the final determined RS_fin, otherwise, a balanced fusion sub-strategy is adopted, the final determined RS_fin is obtained by weighted calculation according to CCS_Pmean and CCS_Pmin.

[0032] Further, the process of starting the hierarchical early warning mechanism is as follows: when the final determined RS_fin is in the range of 40% to 60% of the standard risk threshold, a first-level warning is triggered, and is marked as a low-to-medium risk; when the final determined RS_fin is in the range of 60% to 80% of the standard risk threshold, a second-level warning is triggered, and is marked as a medium-to-high risk; when the final determined RS_fin exceeds 80% of the standard risk threshold, a third-level warning is triggered, and is marked as an extremely high risk.

[0033] The application provides a wearable sensing device feedback data management method after heart valve surgery, which has the following advantages:

[0034] (1) The scheme realizes local feature extraction and structured packaging, which not only significantly reduces data transmission load and privacy leakage risk to a certain extent, but also improves system real-time response capability, thereby solving the problems of high network delay and difficult data security guarantee in the traditional device cloud architecture, and realizing efficient and safe postoperative ECG monitoring;

[0035] (2) After the intelligent analysis in stages and in a progressive manner, the scheme realizes efficient, accurate and low-power local identification of new-onset atrial fibrillation POAF risk after surgery, and the double-level model architecture is composed of two logical levels: the first layer is clinical+ECG, which fuses a Logistic regression risk score model and is responsible for basic risk assessment; the second layer is a support vector machine SVM dynamic pattern recognition model, which is used to capture nonlinear ECG evolution trends, and the two layers of models work cooperatively, which not only guarantees interpretability and computing efficiency, but also improves the sensitivity to complex electrophysiological precursors;

[0036] (3) This scheme uses four P-wave parameters not included in the main model for dynamic correction, on the one hand, through Pmean trend amplification risk score to improve sensitivity, on the other hand, through Pmin fluctuation inhibition abnormal index to enhance specificity, so as to solve the false alarm or missing alarm problem of traditional early warning scheme with fixed threshold when facing individual differences, realize the dynamic balance of early warning accuracy and individual adaptability;

[0037] (4) When this scheme deals with the state of synchronous activation of Pmean trend correction mechanism and Pmin fluctuation inhibition mechanism, the secondary synergistic correction strategy based on clinical credibility score is introduced, which can not only determine the dominant mechanism by quantitatively comparing the pathological direction of the two types of signals, but also can avoid decision conflict through weighted fusion, solve the logical confusion problem caused by parallel of multi-source feedback signals, realize the robust decision in complex scene. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 The figure is a schematic diagram of the brief process of the heart valve postoperative wearable sensor device feedback data management method. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0040] Please refer to Figure 1The embodiment provides a wearable sensing device feedback data management method after heart valve surgery. The wearable electrocardiogram device applied in the management method is a patch type single-lead wearable electrocardiogram device. The device adopts a double-electrode patch design, is attached to the chest at an angle of 45° in actual use, can continuously record electrocardiogram signals for up to 14 days, and significantly improves the detection rate of postoperative new-onset atrial fibrillation POAF. For example, the detection rate given in the present case is 31%, while the detection rate of traditional monitoring is 19%. The wearable sensing device can identify short, asymptomatic or nighttime POAF events, and make up for the defects of missed diagnosis caused by short monitoring time of regular ICU telemetry or intermittent 12-lead electrocardiogram ECG. In the present scheme, the wearable sensing device is optimized, a low-noise amplifier and a 16-bit high-resolution ADC module can be used to improve the collection accuracy of weak P-wave signals, and a three-axis acceleration sensor can be additionally arranged in the wearable sensing device to identify the body position and motion state of the patient, and assist in distinguishing physiological artifacts from real arrhythmia. The device supports Bluetooth and Wi-Fi dual-mode communication to ensure stable uploading of data to a local edge computing terminal. The original ECG data is collected at a high sampling rate. The high sampling rate in the present case can be 1000 Hz, which retains the details of the complete waveform and provides high-quality input for subsequent P-wave form analysis. At the same time, the device is equipped with a real-time clock RTC module to stamp each piece of data with an accurate UTC timestamp, ensuring accurate alignment with clinical events such as surgery time and medication records, facilitating multi-modal data fusion analysis, and all data are encrypted and stored on the device to protect patient privacy and safety.

[0041] The specific steps of the management method are as follows:

[0042] S1, constructing a cooperative data collection architecture of a wearable sensing device after heart valve surgery and an edge computing node:

[0043] Deploying the wearable sensing device at the target patient to obtain postoperative continuous dynamic original ECG data, configuring the edge computing node, performing P-wave feature extraction and structured processing on the original ECG data, and calculating a set of key feature parameters to perform structured packaging processing action;

[0044] The specific steps of S1 are as follows:

[0045] In the original ECG data obtained by performing postoperative continuous dynamic electrocardiogram ECG monitoring on the target patient receiving heart valve surgery for ≥7 days, the original ECG data is used to capture the subtle changes in atrial electrophysiology during the postoperative period; the edge computing node is deployed or configured at the patient's local terminal, such as a smart phone or a dedicated gateway; the process of performing P-wave feature extraction and structured processing is as follows:

[0046] S1.1, secondary filtering of original ECG data: adaptive notch filtering is applied, and multi-scale denoising is performed by wavelet transform, which is applied to improve the signal-to-noise ratio of P wave; wherein the adaptive notch filtering is used to eliminate power frequency interference such as 50 / 60Hz, and the wavelet transform uses db4 wavelet;

[0047] S1.2, based on the improved Pan-Tompkins algorithm combined with the morphological gradient method, the R wave peak is located, the cardiac cycle boundary is determined, and the method of dynamic template matching combined with the first derivative is used to identify the P wave starting point and the P wave ending point; wherein the P wave starting point is defined as: the intersection of the rising branch and the isoelectric line, and the P wave ending point is defined as: the isoelectric line where the descending branch returns; it should be noted that this process needs to be processed with reference to the Ecglab software standard; the process of using the improved Pan-Tompkins algorithm combined with the morphological gradient method is: band-pass filtering is performed on the signal corresponding to the original ECG data to enhance the QRS wave, differential operation is performed to highlight the waveform slope, square processing is performed to strengthen the high-frequency component, and then moving window integration is performed to obtain the envelope; On this basis, the morphological gradient method is introduced to smooth the noise by morphological closing operation and suppress the baseline drift by morphological opening operation, and the difference between the upper and lower morphological gradients is calculated to highlight the steep edge of the QRS complex; combined with the adaptive threshold detection mechanism of the original algorithm, the peak points exceeding the dynamic threshold are searched in the integral signal, and the R wave position is accurately located through RR interval constraint and template matching verification, and the robustness under low signal-to-noise ratio is improved; the process of using the dynamic template matching combined with the first derivative method is: after the R wave is located, the P wave is searched in the PR segment, the first derivative method is used to locate the P wave starting point, when the derivative of three consecutive sampling points is greater than the set slope threshold and the signal rises, it is determined as the starting point; The end point is determined by the negative to positive or the isoelectric line of the descending branch; combined with dynamic template matching, an average template is constructed from the identified P wave, the current waveform is compared by sliding, the first derivative result is calculated, and if the first derivative result and the template matching position deviation is <10ms, it is adopted; otherwise, the template is fine-tuned as the center to realize high-precision P wave boundary identification; wherein the action of fine-tuning can be determined according to actual needs, the purpose is to make the first derivative result and the template matching position deviation <10ms after fine-tuning.

[0048] The key feature parameter set is calculated as follows:

[0049] a maximum P-wave duration Pmax, a minimum P-wave duration Pmin, an average P-wave duration Pmean, a P-wave duration standard deviation Pstd, a P-wave dispersion PWd, an average P-wave peak time Pptmean, and a P-wave peak time standard deviation Pptstd; wherein the P-wave dispersion PWd is PWd = Pmax - Pmin; each parameter in the set of key feature parameters is closely related to the occurrence of POAF, wherein the Pmax, Pstd, and PWd three indicators are potential independent predictors; therefore, the three features are optimized in the scheme, a sliding window can be used, for example: updated every 5 minutes, to dynamically calculate the value, and combined with the Z-score standardization method to identify abnormal fluctuation trend, to provide basis for subsequent risk early warning;

[0050] When performing the structured packaging action, the seven P-wave parameters mentioned in the set of key feature parameters are packaged in a structured JSON format, including: target patient ID, timestamp, and the like, and are temporarily stored in the local database, only the structured features are uploaded to the cloud instead of the original waveform, which greatly reduces the transmission load and privacy risk.

[0051] In the embodiment, the local feature extraction and structured packaging can be realized by setting an edge computing node, which not only significantly reduces the data transmission load and privacy leakage risk to a certain extent, but also improves the real-time response capability of the system, thereby solving the problems of high network delay and difficult data security guarantee in the traditional device cloud architecture, realizing efficient and safe postoperative ECG monitoring. The effect is the basis of the whole scheme. The original ECG data is converted into seven P-wave parameters and encrypted for uploading, avoiding long-distance transmission of massive waveform data, and providing a low-delay data input channel for subsequent real-time early warning.

[0052] S2, constructing a double-level prediction model:

[0053] According to the set of key feature parameters and the collected clinical feature data set, the pre-constructed double-level prediction model is run to output the model collaborative result, and a dynamic correction and self-adaptive optimization mechanism is triggered synchronously to correct the model collaborative result;

[0054] The collected clinical feature data set includes:

[0055] The age Nu of the target patient, the left atrial diameter LA, the glomerular filtration rate GFR, the urea Urea, the mechanical ventilation time Hr, the Pmax, the Pstd, and the PWd; the above pre-constructed double-level prediction model is usually deployed on an edge computing node and belongs to a double-level model architecture;

[0056] Specifically, the running process of the double-level prediction model is as follows:

[0057] The first level (main model): According to the set time length, for example, every hour, the latest set of key feature parameters are received, the clinical feature data set of the target patient is synchronously obtained from the local cache or the HIS system matched with the hospital where the target patient is located, and the 8 input variables in the clinical feature data set are standardized to eliminate the dimensional difference. The standardized input variables are substituted into the pre-trained binary Logistic regression equation, which is based on:

[0058] ;

[0059] Wherein, RS represents the POAF occurrence probability value, e is the base number of natural logarithm, β0 is the intercept term, which represents the logarithmic ratio of the event occurrence when all input variables X i are 0, β i represents the regression coefficient of the i-th input variable, the value range is -1~1, which can be determined according to actual requirements, X i represents the i-th input variable, that is, the clinical feature data set used by the model, i=1, 2,..., N, N is the number of data types in the clinical feature data set, N=8 in the embodiment, and i represents the data number of the corresponding clinical feature data set; The role of the above regression equation is to convert the linear prediction value into a POAF occurrence probability value RS between 0 and 1 through the Logistic function, which is used to quantify the individual POAF risk;

[0060] The second level (auxiliary model): a time sequence pattern recognition model based on support vector machine SVM is used, and a sliding window sequence of independent prediction factors within a predetermined time length, for example, 6 hours, is input, for example: sampling every 10 minutes, a total of 36 time points, forming a high-dimensional feature space, and the support vector machine SVM maps the sliding window sequence data to the high-dimensional space through the kernel function, finds the optimal hyperplane, and is used to distinguish between stable and unstable atrial electrical activity states, and outputs an abnormal pattern index AI;

[0061] Wherein, the abnormal pattern index AI is used to represent the degree of deviation of the current ECG rhythm from the normal baseline;

[0062] After the phased, progressive intelligent analysis, the efficient, accurate and low-power local identification of postoperative new-onset atrial fibrillation POAF risk is realized, and the double-level model architecture is composed of two logical levels: the first level is clinical+ECG, which integrates a Logistic regression risk score model, responsible for basic risk assessment; the second level is a support vector machine SVM dynamic pattern recognition model, used to capture nonlinear electrocardiogram evolution trends, and the two-layer model works together to ensure interpretability and computational efficiency, and to improve the sensitivity to complex electrophysiological precursors; then, the POAF occurrence probability value RS and the abnormal pattern index AI provided by the double-level prediction model are the output model collaborative results.

[0063] In the above-mentioned double-level prediction model:

[0064] The core input variables of the first level Logistic regression model of clinical+ECG and the second level time series pattern recognition model based on support vector machine SVM are focused on the independent prediction factors selected by single factor analysis, i.e. Pmax, Pstd, PWd three P-wave related parameters, because they are significantly related to POAF occurrence in statistics and can effectively improve the model AUC; however, among the original 7 P-wave parameters, Pmin, Pmean, Pptmean, Pptstd four parameters are not directly included in the main model, if these data are simply discarded, it will cause information waste, this scheme proposes an innovative improvement: do not use these four parameters as direct inputs of the main model, but convert them into feedback adjustment signals of model performance, which are used to dynamically modify the decision threshold and weight configuration of the double-level model, so as to realize the dynamic balance of early warning accuracy and individual adaptability, not only improving the system intelligence level, but also enhancing the adaptability to different patients' electrocardiogram baseline differences.

[0065] The content of the triggered dynamic modification and adaptive optimization mechanism is as follows:

[0066] First, after the edge computing node receives the key feature parameter set for a certain period of time, in addition to sending Pmax, Pstd, PWd into the main model, it also synchronously performs trend analysis and variability assessment processing on Pmin, Pmean, Pptmean, Pptstd four parameters, the content of the trend analysis and variability assessment processing includes:

[0067] Pmean trend modification mechanism:

[0068] A slope change rate k_Pmean of Pmean in a predetermined time length in the past is calculated as a trend judgment index; if k_Pmean≤Q1, it is considered that the atrial conduction is stable, no correction is made, and the original POAF occurrence probability value RS is maintained; if Q1

[0069] Pmin fluctuation suppression mechanism:

[0070] A coefficient of variation CV_Pmin of Pmin is calculated: CV_Pmin=standard deviation / mean; wherein the standard deviation and the mean are calculated based on Pmin values extracted every 5 minutes in the past 1 hour, and there are 12 data points; it should be noted that the time window is sufficient to reflect the short-term fluctuation trend and avoid long delay, the coefficient of variation CV_Pmin of Pmin is used instead of the absolute standard deviation, which can standardize individual differences and more accurately reflect the relative fluctuation degree, is suitable for patients with different baseline Pmin, and ensures the universality and rationality of the above operation mechanism; at the same time, Pmin represents the shortest P-wave time limit, and its sharp fluctuation may be caused by respiratory movement, body position change or local muscle bundle ectopic excitation, which is easy to cause false positive AI increase, so the coefficient of variation CV_Pmin of Pmin is introduced to evaluate its stability; if CV_Pmin≤Q2, it is considered that Pmin is stable, the weight w_AI of the SVM corresponding to the abnormal mode index AI is set to 1, that is, no adjustment is made; if Q2 Based on this, the dynamically corrected abnormal mode index AI=original AI×w_AI; it should be noted that high CV_Pmin indicates that the P-wave lower limit is unstable, but if Pmax is not extended synchronously, it is more likely to be an artifact rather than a real electrical reconstruction; therefore, by reducing the weight corresponding to AI, it is avoided to trigger excessive early warning due to temporary fluctuation; Q2 represents a second standard threshold, which has a value range of 0~0.1, and in the embodiment, the value is usually 0.08;

[0071] Combined stability evaluation mechanism:

[0072] Constructing stability index SI: SI=Pptmean×Pptstd; wherein, the higher SI is, the more disordered atrial activation is; when SI≤Q3, it indicates relative stability, maintaining the default warning threshold, i.e. no response action; when SI>Q3, it indicates entering an unstable state, triggering a sensitivity enhancement mode, which down-regulates the threshold corresponding to the secondary and tertiary warning of the main model; wherein, the down-regulation is: reducing the threshold corresponding to the original secondary and tertiary warning according to the set ladder index; it should be noted that, in the sub-group of SI>Q3 and upward trend, when the threshold is down-regulated by 0.08, the sensitivity reaches 88% and the specificity remains 76%, which is better than the performance under the fixed threshold, and the 0.08 at this point is the specific value of the set ladder index; Q3 represents the third standard threshold, which has a value greater than 0, and in this embodiment, it is usually 1.8; the combined stability evaluation mechanism at this point is to correct the corresponding threshold in the subsequent hierarchical warning mechanism.

[0073] Specifically, the four P-wave parameters not included in the main model are used for dynamic correction, on the one hand, the risk score is amplified by Pmean trend to improve sensitivity, and on the other hand, the abnormal index is suppressed by Pmin fluctuation to enhance specificity, so as to solve the problem of false positives or false negatives of the traditional warning scheme with fixed threshold when facing individual differences, and realize the dynamic balance of warning accuracy and individual adaptability.

[0074] Secondly, continue to analyze the results of trend analysis and variability evaluation process, when there is a condition that Pmean trend correction mechanism and Pmin fluctuation suppression mechanism are activated simultaneously, then start the secondary collaborative correction strategy based on evidence weight fusion;

[0075] The process of the above-mentioned secondary collaborative correction strategy based on evidence weight fusion is as follows:

[0076] First step: calculate the clinical credibility score of Pmean trend correction mechanism and Pmin fluctuation suppression mechanism:

[0077] The clinical credibility score CCS Pmean of the Pmean trend correction mechanism is based on the linear regression slope k Pmean and the goodness of fit R2, and the basis is CCS Pmean=k Pmean×R2; wherein R2 reflects the stability of the trend, and the closer it is to 1, the more reliable it is; the clinical credibility score CCS Pmin of the Pmin fluctuation suppression mechanism is calculated based on CV Pmin and the accompanying P wave form distortion rate DR, and the basis is CCS Pmin=CV Pmin×DR; wherein DR is defined as the proportion of positioning failure of the start / end point of the P wave in the past 1 hour;

[0078] Second step: determine the dominant correction mechanism and set the fusion weight:

[0079] By comparing CCS Pmean and CCS Pmin, when CCS Pmean is greater than or equal to 2×CCS Pmin, it is determined that the Pmean trend is the dominant signal, and the first fusion sub-strategy is adopted, which is a risk amplification-based fusion strategy; when CCS Pmin is greater than or equal to 2×CCS Pmean, it is determined that the Pmin fluctuation is the dominant signal, and the second fusion sub-strategy is adopted, which is a risk suppression-based fusion strategy; otherwise, the balanced fusion sub-strategy is adopted.

[0080] Third step: execute the corresponding fusion sub-strategy:

[0081] When the first fusion sub-strategy is adopted, the POAF occurrence probability value RS obtained after the Pmean trend correction mechanism is processed is modified twice, and the basis is RS_fin=RS×Fs_t1; wherein RS_fin represents the final determined RS, RS represents the POAF occurrence probability value RS obtained after the Pmean trend correction mechanism is processed, and Fs_t1 represents the first amplification coefficient, and the value range is usually between 1~2; when the second fusion sub-strategy is adopted, the POAF occurrence probability value RS obtained after the Pmean trend correction mechanism is processed is modified twice, and the basis is RS_fin=RS×Fs_t2; Fs_t2 represents the second amplification coefficient, and the value range is usually between 0~1, that is, not amplified but slightly reduced; when the balanced fusion sub-strategy is adopted, the weighted calculation is performed according to CCS Pmean and CCS Pmin, and the basis is RS_fin=a1×CCS Pmean+a2×CCS Pmin; wherein a1 and a2 are weight coefficients, and the value range is 0~1, which can be set according to actual needs to obtain the required final determined RS, that is, RS_fin.

[0082] RS_fin=a1×CCS Pmean+a2×CCS Pmin; wherein a1 and a2 are weight coefficients, and the value range is 0~1, which can be set according to actual needs to obtain the required final determined RS, that is, RS_fin.

[0083] By adopting the technical scheme, when the Pmean trend correction mechanism and the Pmin fluctuation suppression mechanism are activated synchronously, a secondary synergistic correction strategy based on clinical credibility score is introduced, which can not only determine the dominant mechanism by quantitatively comparing the pathologic directivity of the two types of signals, but also avoid decision conflicts by weighted fusion, solve the logical confusion problem caused by parallel multi-source feedback signals, and realize robust decision-making in complex scenarios.

[0084] S3, a hierarchical early warning mechanism is configured, and a differentiated clinical response process is triggered according to the risk level:

[0085] It is judged whether the finally derived abnormal pattern index AI exceeds the limited threshold value, and a judgment result is obtained; the judgment result is based on: if yes, a level three early warning is triggered directly; if no, a hierarchical early warning mechanism is started; the judgment result is fed back to the target port in an encrypted manner, and data storage is completed;

[0086] The process of the hierarchical early warning mechanism is as follows:

[0087] When the finally derived POAF occurrence probability value is between 40% and 60% of the standard risk threshold value (the endpoints 40% and 60% are included), a level one early warning is triggered, marked as medium-low risk, and a report is automatically recorded and generated, pushed to the supervisor doctor APP, and a prompt is given to pay attention to the P-wave parameter change trend; when the finally derived POAF occurrence probability value is between 60% and 80% of the standard risk threshold value (the endpoint 80% is included, and the endpoint 60% is not included), a level two early warning is triggered, marked as medium-high risk, and the perioperative management strategy is suggested to be adjusted, such as optimizing electrolyte balance, starting beta blocker or amiodarone prevention, and automatically calling the clinical data of the target patient LA, GFR and other data to assist decision-making; when the finally derived POAF occurrence probability value is more than 80% of the standard risk threshold value, a level three early warning is triggered, marked as extremely high risk, an emergency notification is immediately triggered, the on-duty doctor and the patient's family are contacted through the APP pop-up window, short message and telephone multiple channels, and the recent 1 hour electrocardio characteristics and abnormal waveform segments are automatically uploaded to the cloud for remote consultation by the cardiology expert; for the case that the finally derived POAF occurrence probability value is less than 40% of the standard risk threshold value, no response is made; for the differentiated clinical response process after triggering the corresponding level early warning, only the target process is called from the existing database, and no detailed description is made here; it should be noted that the finally derived POAF occurrence probability value of 60% to 80% of the standard risk threshold value is the level two early warning threshold corresponding to the main model proposed in S2; the finally derived POAF occurrence probability value more than 80% of the standard risk threshold value is the level three early warning threshold corresponding to the main model proposed in S2.

[0088] When the judgment result is fed back to the target port, the target port can refer to an edge node or a medical station terminal in the embodiment, and can be set according to actual requirements, and the data is stored in a secure manner, and then the doctor can check at any time.

[0089] To sum up, according to the above-mentioned scheme, first, an efficient and secure data base is constructed, second, a precise prediction core engine is established, third, individual calibration intelligent adjustment is realized, and finally, decision fusion in complex scenarios is completed. The above-mentioned scheme works cooperatively to form an indivisible technical whole, realizes a full-process closed-loop data management scheme from data acquisition to intelligent early warning, and can improve the precision and clinical practicability of heart valve postoperative arrhythmia management to a certain extent.

[0090] The above embodiments can be realized wholly or partially by software, hardware, firmware or any combination thereof. When realized by software, the above embodiments can be realized in the form of a computer program product wholly or partially. Those skilled in the art can realize that the units and algorithm steps of the examples described in connection with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solutions.

[0091] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, which can be located in one place or distributed on multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0092] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for managing feedback data from wearable sensing devices after heart valve surgery, characterized in that, The steps are as follows: Obtain the original ECG data of the target patient after surgery, configure the edge computing node, perform P wave feature extraction and structured processing on the original ECG data, calculate the key feature parameter set, and perform structured packaging processing action; According to the key feature parameter set and combined with the collected clinical feature data set, run the pre-constructed double-level prediction model, output the model collaborative result, and trigger the dynamic correction and adaptive optimization mechanism synchronously; The collected clinical feature data set at least includes: the age Nu of the target patient, the left atrial diameter LA, the glomerular filtration rate GFR, the urea Urea, the mechanical ventilation time Hr, and the independent prediction factor coinciding with the key feature parameter set, that is, Pmax, Pstd and PWd; The double-level prediction model includes a main model and an auxiliary model; Wherein, the main model: according to the latest key feature parameter set received within a set time, synchronously obtains the clinical feature data set of the target patient, and standardizes 8 input variables in the clinical feature data set, and substitutes the standardized input variables into the pre-trained binary Logistic regression equation to obtain the POAF occurrence probability value RS; The auxiliary model: a time series pattern recognition model based on SVM is used, the sliding window sequence of independent prediction factors in the past predetermined time is input, a high-dimensional feature space is formed, the SVM maps the sliding window sequence data to the high-dimensional space through the kernel function, finds the optimal hyperplane, and outputs the abnormal pattern index AI; The process of the triggered dynamic correction and adaptive optimization mechanism is as follows: After receiving the key feature parameter set within a set time, in addition to sending the independent prediction factor to the main model, the trend analysis and variability evaluation processing of Pmin, Pmean, Pptmean and Pptstd are also performed, at least including: Pmean trend correction mechanism, Pmin fluctuation suppression mechanism and joint stability evaluation mechanism; Continue to analyze the results of trend analysis and variability evaluation processing, when there are conditions for the simultaneous activation of Pmean trend correction mechanism and Pmin fluctuation suppression mechanism, then start the secondary collaborative correction strategy based on evidence weight fusion; The running process of Pmean trend correction mechanism is as follows: Calculate the slope change rate k_Pmean of Pmean within the past predetermined time as the trend judgment index; If k_Pmean≤Q1, no correction is performed; If Q1 The running process of Pmin fluctuation suppression mechanism is as follows: Calculate the coefficient of variation CV_Pmin of Pmin: CV_Pmin=standard deviation / mean; wherein the standard deviation and the mean are calculated based on the Pmin values extracted every 5 minutes in the past 1 hour, a total of 12 data points; if CV_Pmin≤Q2, set the weight w_AI of the abnormal mode index AI corresponding to the SVM=1; if Q2 The running process of the joint stability evaluation mechanism is as follows: Construct the stability index SI: SI=Pptmean×Pptstd; when SI≤Q3, no response action is taken; when SI>Q3, the sensitivity enhancement mode is triggered, and the secondary and tertiary early warning thresholds corresponding to the main model are down-regulated; Q3 is the third standard threshold, and its value range is greater than 0; The process of the secondary collaborative correction strategy based on evidence weight fusion is as follows: Calculate the clinical credibility score CCS_Pmean of the Pmean trend correction mechanism and the clinical credibility score CCS_Pmin of the Pmin fluctuation suppression mechanism; compare CCS_Pmean and CCS_Pmin; when CCS_Pmean≥2×CCS_Pmin, the first fusion sub-strategy is adopted: the POAF occurrence probability value RS after the Pmean trend correction mechanism is processed is corrected again, and the basis is: RS_fin=RS×Fs_t1; wherein RS_fin represents the final determined RS, RS represents the POAF occurrence probability value RS after the Pmean trend correction mechanism is processed, and Fs_t1 represents the first amplification coefficient of the reference, which is in the value range of 1-2, and the final determined RS_fin is obtained; when CCS_Pmin≥2×CCS_Pmean, the second fusion sub-strategy is adopted: the POAF occurrence probability value RS after the Pmean trend correction mechanism is processed is corrected again, and the basis is: RS_fin=RS×Fs_t2; Fs_t2 represents the second amplification coefficient of the reference, which is in the value range of 0-1, and the final determined RS_fin is obtained; otherwise, the balanced fusion sub-strategy is adopted: weighted calculation is performed according to CCS_Pmean and CCS_Pmin, and the basis is: RS_fin=a1×CCS_Pmean+a2×CCS_Pmin; wherein a1 and a2 are weight coefficients, and the value range is 0-1, and the final determined RS_fin is obtained; determining whether the final abnormal pattern index AI exceeds a defined threshold, to obtain a determination result; wherein the determination result is based on: if yes, directly triggering a third-level warning; if no, starting a hierarchical warning mechanism; the process of starting the hierarchical warning mechanism is: when the finally determined RS_fin is within 40% to 60% of the standard risk threshold, a first-level warning is triggered, marked as low-medium risk; when the finally determined RS_fin is within 60% to 80% of the standard risk threshold, a second-level warning is triggered, marked as medium-high risk; when the finally determined RS_fin exceeds 80% of the standard risk threshold, a third-level warning is triggered, marked as extremely high risk; encrypting the determination result and feeding back to the target port, and completing data storage.

2. The method of claim 1, wherein the wearable sensor feedback data management method for post-cardiac valve surgery is characterized by: The original ECG data represents data obtained by implementing continuous dynamic electrocardiogram (ECG) monitoring on a target patient receiving a heart valve surgery for ≥7 days after the surgery; the process of performing P-wave feature extraction and structured processing is as follows: Secondary filtering is performed on the original ECG data: adaptive notch filtering is applied, and wavelet transform is used for multi-scale denoising; based on an improved Pan-Tompkins algorithm combined with a morphological gradient method, R-wave peaks are located, the boundaries of the cardiac cycle are determined, and a method combining dynamic template matching and first-order derivative is used to identify the P-wave start point and the P-wave end point; wherein the P-wave start point is defined as the intersection of the rising branch and the isoelectric line, and the P-wave end point is defined as the point where the descending branch returns to the isoelectric line.

3. The method of claim 2, wherein: The process of using the improved Pan-Tompkins algorithm combined with the morphological gradient method is as follows: The signal corresponding to the original ECG data is band-pass filtered to enhance the QRS wave, squared after differential operation, and then integrated through a moving window to obtain an envelope; the morphological gradient method is introduced to smooth noise through morphological closing operation and suppress baseline drift through morphological opening operation, and the difference between the upper and lower morphological gradients is calculated to highlight the steep edges of the QRS complex; combined with the adaptive threshold detection mechanism of the original Pan-Tompkins algorithm, the peak points exceeding the dynamic threshold in the integrated signal are searched, and the R-wave position is located through RR interval constraint and template matching verification; The process of using the method combining dynamic template matching and first-order derivative is as follows: After locating the R-wave position, the P-wave is searched within the PR segment, the first-order derivative method is used to locate the P-wave start point, and when the derivative of three consecutive sampling points is greater than the set slope threshold and the signal is rising, it is determined as the start point; the end point is determined by the descending branch derivative turning from negative to positive; combined with dynamic template matching, an average template is constructed using the identified P-wave, the current waveform is compared by sliding, the first-order derivative result is calculated, and if the first-order derivative result and the template matching position deviation are <10 ms, it is adopted.

4. The method of claim 1, wherein: The calculated key feature parameter set at least includes: maximum P-wave time limit Pmax, minimum P-wave time limit Pmin, average P-wave time limit Pmean, P-wave time limit standard deviation Pstd, P-wave dispersion PWd, average P-wave peak time Pptmean, and P-wave peak time standard deviation Pptstd; wherein the P-wave dispersion PWd is: PWd=Pmax-Pmin; When performing the structured packaging processing action, at least 7 P-wave parameters in the key feature parameter set are packaged in a structured JSON format, including a target patient ID, a timestamp, and the 7 P-wave parameters, and are temporarily stored in a local database.

Images