Blood oxygen change monitoring algorithm fused with dynamic Bayesian modeling

Abstract

The invention relates to the technical field of computer systems based on specific calculation models, and discloses a blood oxygen change monitoring algorithm fused with dynamic Bayesian modeling, which comprises the following steps: quantizing signal uncertainty by calculating entropy of pulse waveform harmonic energy distribution; and on the basis of the entropy value, driving a dynamic Bayesian network to carry out confidence coefficient evaluation and weighting processing on the blood oxygen estimation model, and finally outputting a decision pair containing a blood oxygen estimation value and the confidence coefficient thereof. According to the method, the signal uncertainty is converted into a computable entropy index, so that the system can autonomously distinguish real physiological changes and measurement noise, the problem of misjudgment caused by motion interference in traditional blood oxygen monitoring is avoided, meanwhile, non-inductive personalized calibration is achieved by utilizing the continuous learning ability of the Bayesian network, and the accuracy of the system is improved. And the reliability and the practicability of the medical wearable equipment are remarkably improved.

Description

TECHNICAL FIELD

[0001] The application relates to a blood oxygen change monitoring algorithm fusing dynamic Bayesian modeling, and belongs to the technical field of computer systems based on specific calculation models. BACKGROUND

[0002] Currently, wearable medical devices generally use photoplethysmogram (PPG) signals to analyze blood oxygen saturation, and the technical path mainly relies on time domain filtering and static threshold model calculation. For example, the mainstream scheme suppresses motion artifacts through adaptive filtering, and establishes a fixed linear regression model based on large-scale population data to output blood oxygen values. Such methods can still maintain basic accuracy in the resting state of users, but when applied to home monitoring scenarios, their inherent limitations are significantly exposed.

[0003] In the user's daily activities (such as walking or household operations), motion interference and spectral overlap of physiological signals make it difficult for traditional filtering to separate effective information. Forced noise reduction often causes waveform feature distortion, and more importantly, individual vascular characteristic differences and physiological state drift make static models unable to adapt to user-specific parameters. Slow disease patients need to be calibrated repeatedly, which deviates from the original intention of unobtrusive monitoring. To address this challenge, the industry has tried to introduce multi-sensor fusion, such as accelerometers to assist in motion recognition or complex neural networks to optimize models, but the former increases hardware costs and power consumption, and the latter is difficult to deploy on edge devices due to high computational load. Moreover, both of them have not solved the core contradiction - existing technologies always try to restore ideal data from contaminated raw signals, but lack a mechanism to quantify and utilize signal uncertainty. In a noisy environment, system performance often faces the challenges of error amplification or signal loss.

[0004] Specifically, the existing technology mainly has the following three deficiencies: 1. Time domain filtering can suppress some noise, but it cannot quantify the confidence impact of residual interference on the calculation model, resulting in a lack of reliability identification in system output; 2. Fixed threshold models ignore the specificity and changes over time of individual vascular elasticity and hemodynamics, leading to systematic bias across users and cycles; 3. Improved solutions that rely on multi-sensor or deep networks increase hardware costs and computational requirements, restricting device battery life and universal application. Therefore, how to build an uncertainty quantification mechanism inherent in the calculation model, dynamically analyze signal quality to adjust decision logic autonomously, and achieve lightweight individual adaptation, so as to achieve medical-grade reliable monitoring on resource-constrained edge devices, has become a technical problem to be solved by the present application. SUMMARY

[0005] The application provides a blood oxygen change monitoring algorithm fusing dynamic Bayesian modeling, which mainly aims to solve the problem of unreliable monitoring results caused by the lack of signal uncertainty quantization and the limitation of static modeling in the motion interference and individual difference scenarios of medical-grade wearable devices.

[0006] To achieve the above-mentioned purpose, the application provides a blood oxygen change monitoring algorithm fusing dynamic Bayesian modeling, which comprises the following steps: Step 1: obtaining an original photoelectric plethysmogram signal without complex filtering; Step 2: based on the original signal, identifying a plurality of independent pulse waveforms in a plurality of consecutive time windows, respectively; Step 3: for each independent pulse waveform, performing Fourier transform to obtain its harmonic energy distribution; Step 4: based on the harmonic energy distribution, quantitatively generating a pulse waveform morphology harmonic entropy index representing the stability of the pulse waveform morphology, wherein the pulse waveform morphology harmonic entropy index comprises a combination of harmonic entropy average value and harmonic entropy variance; Step 5: constructing a dynamic Bayesian network, wherein the internal state of the dynamic Bayesian network comprises a final blood oxygen estimation value, a current confidence level and an individual model bias for calibrating the blood oxygen estimation model; Step 6: taking the pulse waveform morphology harmonic entropy index as an input of the dynamic Bayesian network, and taking the outputs of two blood oxygen estimation models running in parallel and having different response characteristics as the other two inputs of the dynamic Bayesian network; Step 7: the dynamic Bayesian network dynamically adjusts the trust weight of the output results of the blood oxygen estimation model according to the pulse waveform morphology harmonic entropy index; and the dynamic Bayesian network performs belief updating to adjust the individual model bias according to the trust weight, and outputs a decision pair comprising the final blood oxygen estimation value and the corresponding confidence level, wherein the confidence level is determined by the pulse waveform morphology harmonic entropy index and the belief updating process.

[0007] Preferably, the dynamic Bayesian network adjusts the trust weight of the output results of the two blood oxygen estimation models running in parallel according to the numerical range of the pulse waveform morphology harmonic entropy index: when the pulse waveform morphology harmonic entropy index indicates that the pulse waveform morphology stability is high, the dynamic Bayesian network gives a higher trust weight to the stable-state blood oxygen estimation model, and iteratively adjusts the individual model bias based on a predetermined learning rate using the data of the current pulse waveform morphology stability; when the pulse waveform morphology harmonic entropy index indicates that the pulse waveform morphology stability is low, the dynamic Bayesian network gives a higher trust weight to the change rate blood oxygen estimation model, and reduces the confidence level of the final output result to a predetermined lower limit according to a predetermined algorithm.

[0008] Preferably, the steady-state blood oxygen estimation model is a linear regression model trained based on large-scale statistical data, and the rate-of-change blood oxygen estimation model is a linear estimation model based on the rate of change of the ratio of the AC component to the DC component of the photoplethysmogram signal, and the rate-of-change blood oxygen estimation model is not sensitive to the morphology of the pulse waveform.

[0009] Preferably, the method further comprises the following steps: based on the harmonic analysis of the independent pulse waveform, extracting the energy ratio of the high-order harmonic component to the fundamental component in the harmonic energy distribution to obtain a harmonic attenuation slope index representing the energy spectrum morphology of the pulse waveform; monitoring the rate of change of the harmonic attenuation slope index in real time; and when the absolute value of the rate of change exceeds a threshold value determined based on historical data statistics within a predetermined time period, determining that a physiological transient event has occurred, and applying a dynamic compensation correction to the final blood oxygen estimation value according to the rate of change, the correction being: correction = (1 - dynamic compensation factor) * (blood oxygen estimation value - average blood oxygen value). wherein, is a dynamic compensation factor, is the rate of change of the harmonic attenuation slope index.

[0010] Preferably, the dynamic compensation factor The dynamic Bayesian network learns and updates the data after the physiological transient event based on the harmonic entropy index of the pulse waveform morphology.

[0011] Preferably, the method further comprises the following steps: obtaining the joint change trajectory of the harmonic attenuation slope index and the harmonic entropy index of the pulse waveform morphology during the physiological transient event in parallel; based on the morphological features of the joint change trajectory, classifying the physiological transient event into types, including hemodynamic transient types and cardiac electrophysiological transient types; and according to the type classification result, adaptively selecting and applying a dynamic compensation factor pre-associated with the event type to determine the dynamic compensation correction.

[0012] Preferably, the method further comprises the following steps: based on the harmonic analysis of the independent pulse waveform, extracting the phase information of the second harmonic; calculating the variance of the phase information within a predetermined time window to obtain a harmonic phase jitter index representing the phase-locked state of the photoplethysmogram signal; based on the harmonic phase jitter index, determining the source reliability of the photoplethysmogram signal to obtain a source confidence index.

[0013] Preferably, when the value of the source confidence index is higher than a predetermined threshold, the dynamic Bayesian network continues to estimate blood oxygen and output confidence; when the value of the source confidence index is not higher than the predetermined threshold, the method outputs a diagnostic information indicating that the device is abnormal, rather than a decision pair containing the blood oxygen estimation value and the confidence level.

[0014] Preferably, the method triggers a real-time alarm when the final blood oxygen estimation value is below a physiological safety threshold and its corresponding confidence level is above a credibility threshold, the alarm is sent through the user's wearable device or a remote health monitoring platform.

[0015] Preferably, the updating of the individual model bias parameters in the dynamic Bayesian network is based on the blood oxygen estimation history and the corresponding photoplethysmogram features when the pulse wave morphology harmonic entropy indicator indicates high signal stability, and the updating uses an algorithm that minimizes historical errors to ensure adaptive calibration of long-term physiological changes in the user.

[0016] Compared with the prior art, the present application has the following beneficial effects: 1. By converting the harmonic energy distribution of the pulse wave form into an entropy value measure, the system first establishes an endogenous uncertainty quantification scale in the photoplethysmogram signal processing, which enables the system to directly identify the ordered and chaotic states in the signal, rather than passively filtering out noise; when the signal is disturbed by motion, the dynamic Bayesian network automatically reduces its dependence on the precise model according to the entropy value, and instead uses a more robust change rate model for trend tracking, thereby avoiding the inevitable information loss in the signal separation process of traditional schemes.

[0017] 2. The parallel architecture of the stable state model and the change rate model is not a simple redundant design. When the harmonic entropy indicates that the signal is stable, the system preferentially calls the high-precision model to output the blood oxygen value, while continuously calibrating the individual parameters using the low-uncertainty data at that moment; when the signal chaos degree rises, the system seamlessly switches to the anti-interference model and simultaneously reduces the output confidence; this self-verification decision logic enables the device to maintain rational output in the motion scenario, allowing the user to distinguish between real physiological alarms and measurement noise, thereby addressing the trust crisis of medical-grade wearable devices from the root.

[0018] 3. The harmonic phase variance is extracted synchronously in the Fourier transform link, and the system discovers the phase-locked state information discarded by traditional schemes. When the phase jitter exceeds the threshold, the device automatically determines that the sensor contact has failed rather than ordinary motion disturbance, triggering a dedicated hardware diagnostic instruction. This mechanism reuses the intermediate products of the existing calculation process, without adding hardware or algorithm complexity, enabling the system to distinguish between signal degradation causes and providing accurate fault location basis for subsequent operations.

[0019] 4. For the scene of body position mutation, the system analyzes the coupling relationship between the dynamic trajectory and the harmonic entropy to distinguish between the two types of events, hemodynamic transients and cardiac rhythm abnormalities. This fingerprint identification based on two-dimensional phase space trajectory enables the compensation factor to adaptively adjust to the event type: slow release compensation is used when the vascular tension changes dramatically, and a fast response strategy is used when the ECG is out of sync. As a result, acute physiological changes no longer cause the model to be inaccurate, but rather, they become a new opportunity for the system to deepen individual cognition. BRIEF DESCRIPTION OF DRAWINGS

[0020] Fig. 1 The blood oxygen change monitoring time series diagram of the present application fuses dynamic Bayesian modeling; Fig. 2 The dynamic response curve of the harmonic entropy index of the present application during motion interference; Fig. 3 The blood oxygen estimation decision flowchart of the present application fuses dynamic Bayesian network.

[0021] The object implementation, functional characteristics and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0022] It should be understood that the specific embodiments described herein are merely intended to explain the present application and not to limit the present application.

[0023] The embodiment of the application provides a blood oxygen change monitoring algorithm fused with dynamic Bayesian modeling, which runs on a wrist type or finger clip type wearable device integrated with photoelectric emission and receiving units, and realizes reliability self-evaluation and individualized adaptive calibration of blood oxygen estimation value by deeply analyzing the inherent uncertainty of a photoelectric plethysmogram signal, and in a specific application scenario, such as monitoring a chronic obstructive pulmonary disease patient who is at home for rehabilitation training, the algorithm can be specifically operated as follows: firstly, the photoelectric sensor of the device continuously collects a photoelectric plethysmogram original signal without complex filtering, the original signal only undergoes necessary amplification and anti-aliasing filtering of a front-end analog circuit, and the signal dynamic characteristics caused by physiological activities of a user, sensor displacement and real physiological fluctuations are completely retained, thereby providing undistorted original input for subsequent uncertainty quantification; then in a continuous rolling time window, for example, every eight seconds is a processing period, the photoelectric plethysmogram original signal collected is processed, a plurality of independent and continuous pulse waveforms are segmented out by accurately identifying wave trough or wave peak feature points in the signal sequence, and each independent pulse waveform corresponds to a complete heart beat cycle in time; then, for each independent pulse waveform segmented out, the system performs fast Fourier transform, converts the waveform signal from time domain to frequency domain, and thereby obtains a harmonic energy distribution, the energy distribution clearly reveals the energy corresponding to the heart rate as a fundamental frequency, and the energy composition of second, third and even higher harmonic components; based on the harmonic energy distribution obtained in the above step, the system calculates a pulse waveform morphology harmonic entropy index, specifically, for a single pulse waveform, first, the energy of each harmonic is normalized, so that the sum is one, thereby constructing an energy probability distribution, then, according to the calculation principle of information entropy, that is, the sum of the product of all probability items and logarithmic values is taken and then the negative value is taken, the disorder degree of the distribution is quantified, a stable and clean pulse waveform has high energy concentrated in low-order harmonics, and the distribution is ordered, and thus the harmonic entropy value is low; on the contrary, when noise such as motion artifact is mixed, the waveform is distorted, and the energy is disorderly scattered to each harmonic frequency, resulting in a significant increase in the harmonic entropy value; the algorithm further calculates the harmonic entropy of all independent pulse waveforms in a longer time window, such as thirty seconds, and finally generates a pulse waveform morphology harmonic entropy index composed of a harmonic entropy average value and a harmonic entropy variance, the average value of the index reflects the overall level of recent signal quality, and the variance reflects the fluctuation degree of signal quality, and the combination of the two provides an accurate and quantitative measurement of the current signal uncertainty for the system.

[0024] Based on this uncertainty measure, a dynamic Bayesian network is constructed and driven, which is the decision-making center of the whole algorithm, and its internal states are iteratively updated at each time step, these state variables include: the final blood oxygen estimation as the best estimation of the user's true blood oxygen saturation by the system; a quantitative indicator, which reflects the current confidence level of the system's output results; and a key calibration parameter, which is used to compensate for the systematic differences between the universal model and the user's individual physiological characteristics. At each calculation period, the network receives three key information streams: one is the aforementioned pulse waveform harmonic entropy index, which serves as an arbitrator of signal quality; the other two inputs come from the outputs of two parallel running blood oxygen estimation models with different response characteristics, the steady-state blood oxygen estimation model and the rate blood oxygen estimation model, the former is a linear regression model trained based on large-scale statistical data, which takes the ratio of the AC / DC component ratio of red and infrared light signals as the key input, and calculates the blood oxygen value through a fixed linear relationship, which has high accuracy when the signal is pure, but is very sensitive to waveform shape changes, the latter is a linear estimation model based on the rate of change of the AC / DC component ratio of the photoplethysmogram signal, which focuses on the short-term trend of the ratio rather than its absolute value, so it is relatively insensitive to pulse waveform distortion caused by motion and can provide more robust blood oxygen change direction in an interference environment; the dynamic Bayesian network performs its core reasoning and updating functions, the network first dynamically adjusts the trust weight of the outputs of the two parallel blood oxygen estimation models according to the numerical range of the input pulse waveform harmonic entropy index, when the pulse waveform harmonic entropy index indicates high stability of the pulse waveform shape, i.e. the harmonic entropy mean and variance are in a pre-set low interval, the dynamic Bayesian network gives the steady-state blood oxygen estimation model a higher trust weight, considering its output at this moment to be highly reliable, at the same time, the system uses these highly trusted data to perform belief update and iteratively adjust the individual model bias parameter based on a pre-determined learning rate, the setting of this pre-determined learning rate is the result of balancing the response speed and stability of calibration, a too high learning rate will make the model too sensitive to single measurement noise, while a too low learning rate may not be able to keep up with the user's real physiological changes in time, this updating process uses an algorithm aimed at minimizing historical errors, such as recursive least squares, to ensure adaptive calibration of long-term physiological changes in the user caused by physiological state changes or drug effects; on the contrary, when the pulse waveform harmonic entropy index indicates low stability of the pulse waveform shape, the network gives the rate blood oxygen estimation model a higher trust weight, and adopts its provided blood oxygen change trend information, at the same time, the network significantly reduces the confidence level of the final output result to a pre-determined lower limit according to a pre-determined algorithm negatively related to the harmonic entropy value, to clearly inform the user or the monitoring system that the current absolute blood oxygen reading may not be accurate, but the blood oxygen change trend is referable.

[0025] To cope with the sharp changes in physiological state, the algorithm also contains a set of physiological transient event detection and compensation mechanism, when performing harmonic analysis, the system synchronously extracts the energy ratio of high harmonic component and fundamental component in harmonic energy distribution, obtains a harmonic attenuation slope index representing the energy spectrum morphology of pulse waveform, the algorithm monitors the change rate of the slope index in real time, when the absolute value of the change rate exceeds a threshold value determined based on user historical data statistics within a predetermined time period, such as two to three seconds, the system determines that a physiological transient event has occurred, for example, blood pressure fluctuation caused by sudden change of body position, at this time, the system will apply a dynamic compensation correction to the final blood oxygen estimation value, the value of the correction is equal to the product of a dynamic compensation factor and the change rate of the harmonic attenuation slope index, where the dynamic compensation factor is not a fixed value, it is learned and updated by dynamic Bayesian network according to the new data of pulse waveform morphology harmonic entropy index after the physiological transient event ends, so that the compensation is more individual adaptive. Furthermore, the algorithm can classify the types of physiological transient events, by acquiring the joint change trajectory of harmonic attenuation slope index and pulse waveform morphology harmonic entropy index during the physiological transient event in parallel, and analyzing the morphological characteristics of the two-dimensional trajectory, different types of events can be distinguished, for example, the hemodynamic transient type caused by the sharp change of vascular tension, its trajectory morphology may be relatively smooth; while the cardiac electrophysiological transient type caused by arrhythmia, its trajectory may present more sharp and irregular morphology, according to the pre-established association library of trajectory morphology and event type, the system can classify the types, and adaptively select and apply a more targeted dynamic compensation factor associated with the event type to determine the dynamic compensation correction.

[0026] To ensure the reliability of the input signal source itself, the algorithm also introduces the evaluation of the source confidence. In the harmonic analysis link, the system specially extracts the phase information of the second harmonic and calculates the variance of the phase information in a predetermined time window, thereby obtaining a harmonic phase jitter index representing the phase-locked state of the photoplethysmogram (PPG) original signal rhythm. In an ideal state, the rhythm of the continuous pulse wave is stable, the phase is locked, and the index value is very low. If the sensor and the skin are not in good contact, resulting in signal sliding or falling off, it will cause signal timing disorder and phase jitter. Based on the harmonic phase jitter index, the system can judge the reliability of the photoplethysmogram (PPG) original signal source, thereby obtaining a source confidence index. When the value of the source confidence index is higher than a predetermined threshold, it indicates that the signal source is reliable, and the dynamic Bayesian network continues to estimate blood oxygen and output confidence. When the value of the source confidence index is not higher than the threshold, the system determines that the device state is abnormal, and preferentially outputs a diagnostic information indicating the abnormal state of the device, rather than outputting a decision pair containing the blood oxygen estimation value and the confidence level which may mislead people. Finally, the decision pair output by the algorithm can be used to trigger real-time and reliable alarms. When the final blood oxygen estimation value is lower than a physiological safety threshold, for example, ninety percent, and the corresponding confidence level is higher than a credibility threshold, for example, ninety-five percent, the system will trigger a real-time alarm. This double condition judgment mechanism can effectively filter out false low blood oxygen readings with low confidence caused by motion interference and the like, greatly reducing the false positive rate and improving the effectiveness of the alarm. At the same time, it should be noted that the large-scale database required for constructing the association library and training the stable-state blood oxygen estimation model is collected and used in strict accordance with the relevant laws and regulations of the state on data security and personal information protection and industry ethical norms. All the original physiological data used for offline model training and verification are collected on the premise of obtaining sufficient informed consent of the data provider. And before the data enters the algorithm model training stage, all the data are strictly anonymized or de-identified to ensure the protection of the personal privacy of the data subject. All of the above belong to the extended implementation modes known to those skilled in the art.

[0027] In this embodiment, the monitoring object is a user in the early rehabilitation stage after surgery, and the behavior pattern during the monitoring period is from long-term bed rest to the first attempt to stand up and walk a short distance with assistance. This process poses a double test of the robustness and accuracy of the monitoring algorithm in terms of technology, as it not only includes a dramatic change in signal quality from extremely high to extremely low, but also may be accompanied by real physiological transient events caused by sudden changes in body position. In the user's long-term resting state, the user's physiology is stable, the morphology of the raw photoplethysmogram signal is regular and highly repetitive. In this stage, the algorithm calculates the pulse wave morphology harmonic entropy index by performing harmonic analysis on the independent pulse waveforms. The average value and variance of the index are both low, which is interpreted in the dynamic Bayesian network as a high level of confidence in the signal source. Based on this judgment, the network assigns a high trust weight to the stable-state blood oxygen estimation model trained based on large-scale statistical data, thereby outputting a high-precision, high-confidence final blood oxygen estimation value. More importantly, the high-quality data stream in this stage is recognized as an effective calibration window. The dynamic Bayesian network uses these blood oxygen estimation historical values with a confidence level higher than the predetermined confidence threshold and their corresponding photoplethysmogram features to continuously adjust the individual model bias parameters within the network. This mechanism converts each stable rest of the user into an opportunity for unconscious personalized calibration of the model. Through an algorithm that minimizes historical errors, systematic deviations between the universal model and the user's current physiological condition are gradually avoided.

[0028] When the user starts to stand up, the large and irregular movement of the user's limbs introduces strong motion artifacts into the photoplethysmogram raw signal. Traditional algorithms often fail at this moment because their filtering mechanism cannot balance between preserving real physiological fluctuations and filtering motion noise. In this technical solution, the pulse wave morphology harmonic entropy index will jump sharply at this moment, directly quantifying the increase in signal chaos. This jump in entropy value triggers a fundamental switch in decision logic in the dynamic Bayesian network: the network immediately deprives the stable-state blood oxygen estimation model, which is susceptible to waveform morphology, of high trust, and instead assigns trust weights to the change rate blood oxygen estimation model, which is not sensitive to pulse wave form. At the same time, the network actively lowers the confidence level in the output decision pair to the predetermined lower limit according to the magnitude of the entropy value increase. This design reflects a paradigm shift from trying to infer accurate values from unreliable data to acknowledging and managing uncertainty and giving the most reliable judgment under the current conditions, aiming to address the trust issues caused by outputting seemingly accurate but potentially inaccurate blood oxygen values in a motion state.

[0029] In this process, if the user causes a real blood flow transient due to a sudden change in body position, such as a transient postural hypotension, this event will leave a set of identifiable feature combinations in the signal level. In addition to the increase in pulse wave form harmonic entropy index due to exercise, the harmonic attenuation slope index representing the energy spectrum of the pulse wave form will also show a synchronous and different from the conventional exercise interference. The algorithm captures the joint change trajectory of the harmonic attenuation slope index and the pulse wave form harmonic entropy index during the physiological transient event, and analyzes its morphological characteristics, so as to classify the source of this event as a blood flow transient type. The event qualification ability of the system is not derived from a single index, but from the analysis of the coupling relationship between two physical meaning independent indexes. Based on this accurate classification, the system can adaptively select and apply a dynamic compensation factor associated with the event type in advance to correct the final blood oxygen estimation value, thereby capturing the real physiological event while avoiding model misalignment.

[0030] Embodiment 2: In a progressive hypoxemia event caused by mild respiratory depression, which is overlapped in time with a sudden motion artifact introduced by the user's limb position adjustment, to achieve this purpose, the test platform is built around a high-precision photoelectric pulse wave physical simulation system. The system reproduces the preset blood oxygen saturation decline curve through a programmable optical attenuation array, and synchronously superimposes a standardized motion artifact signal with a controllable amplitude to the main signal path, so as to repeatedly simulate the above-mentioned composite interference scene in the laboratory environment. First, the decline rate of the simulated blood oxygen saturation is set. The basic technical consideration of this setting is the balance between the physiological event authenticity and the algorithm response limit test. The decision rule is: the decline rate should match the time scale of the typical event in the target monitoring scene. Based on this rule, referring to the clinical data of moderate hypoxia event, a linear slope from 97% to 88% is set, with a duration of 60 seconds. This setting anchors a working condition that meets the physiological norm and is sufficient to test the dynamic tracking ability of the algorithm. Second, regarding the injection of motion artifact, the essence of the decision is to balance the damage degree of the artifact to the signal form and the recoverability of the physiological information. The decision model is: the artifact energy should be sufficient to make the pulse wave form harmonic entropy index enter the low stability interval preset by the dynamic Bayesian network, but it must ensure that the signal-to-noise ratio of the fundamental frequency is not lower than the lower limit of the effective recognition of the system. Based on this, at the moment when the blood oxygen saturation drops to 92%, a motion artifact with a duration of 10 seconds and a main frequency component concentrated at 1.5 Hz is injected. This parameter combination is verified to be able to effectively trigger the decision logic switching of the present invention.

[0031] After the test started, the algorithm began to process the composite photoplethysmogram signal output by the simulation system. In the initial stage, the blood oxygen value was stable at 97%, and the pulse waveform harmonic entropy index output by the algorithm was stable at a low level. The dynamic Bayesian network gave a stable-state blood oxygen estimation model a very high trust weight, and the final blood oxygen estimation value output was highly consistent with the simulation true value, with a confidence level of more than 98%. At the 31st second, the blood oxygen saturation began to decrease linearly, and at this time, the harmonic entropy index did not change significantly. The dynamic Bayesian network accurately tracked this physiological change based on the high-confidence stable-state model results. At the 45th second, with the injection of a pre-set motion artifact, the signal form was temporarily degraded. At this moment, the algorithm's internal key indicators showed a differentiated trend with diagnostic value: the pulse waveform harmonic entropy index jumped sharply due to waveform chaos, and the harmonic decay slope index, which represents the energy spectrum form, also had a sharp pulse in its change rate that far exceeded the historical statistical threshold. The following table records the exemplary data of the algorithm's core variables before and after this key node.

[0032] Table 1: Changes in key variables of the dynamic Bayesian monitoring algorithm and event recognition process during simulation testing.

[0033] At the 46th second, the sharp rise in harmonic entropy (from 0.85 to 2.73) triggered the dynamic Bayesian network to switch the trust weight from the stable-state model to the change rate model, and simultaneously lowered the output confidence from 97% to 82%, reflecting the system's quantification of signal uncertainty. However, the system did not classify this as ordinary motion interference, as the algorithm simultaneously detected a sharp peak in the harmonic decay slope change rate (1.88). By analyzing the joint change trajectory of these two indicators, the concurrent occurrence of a high entropy value and a high slope change rate, the system classified the root cause of the event as a hemodynamic transient type. Based on this classification, the system invoked the dynamic compensation factor associated with this event type and applied a correction to the final blood oxygen estimation value. As a result, the final blood oxygen estimation value (92.0%) closely tracked the simulation true value (92.2%), avoiding monitoring interruption or significant deviation caused by motion artifacts. After the motion artifact disappeared at the 55th second, the harmonic entropy quickly fell, and the confidence level returned to a high level, allowing the algorithm to seamlessly switch back to the high-precision monitoring mode.

[0034] Embodiment 3: This embodiment combines Figs. 1 to 3 to implement a blood oxygen change monitoring algorithm that integrates dynamic Bayesian modeling. As Fig. 1As shown, first, the photoelectric sensor is used to collect the PPG original signal, which is transmitted to the signal preprocessing module to form a transmission preprocessing signal. Then, the preprocessing signal is sent to the waveform recognition module, which analyzes the PPG signal in multiple continuous time windows and identifies independent pulse waveforms. The time domain waveform data is then sent to the Fourier transform module, in which FFT transformation is performed to obtain waveform features in the frequency domain and output harmonic energy distribution. Subsequently, the module enters the entropy calculation module, and calculates the harmonic index by performing information entropy analysis on the harmonic energy distribution. The index includes The mean and variance are calculated, and the entropy value (mean + variance) is sent to the Bayesian network module. After that, parallel processing is performed in the Bayesian network. During this processing, the waveform data is input into the steady-state model and the rate of change model respectively to generate steady-state blood oxygen estimation value and rate of change blood oxygen estimation value respectively. The Bayesian network adjusts the weight according to the entropy value of the current input to dynamically determine the degree of trust in the output of the two models, and performs belief update based on the estimation results to update the individual model deviation. Finally, the system sends the confidence-weighted result to the output module for output ( value, confidence level).

[0035] like Fig. 2 As shown in the figure, the vertical axis is the harmonic entropy value, and the horizontal axis is time (seconds). The solid line represents the harmonic entropy average value, and the dotted line represents the harmonic entropy variance, which are clearly distinguished in the legend. It can be seen from the figure that in the early resting state stage, the harmonic entropy average value is near the resting entropy value of 0.85, which is far below the high stability threshold, indicating that the signal is highly stable. At t=44s, the harmonic entropy index begins to rise and enters the motion interference stage. At t=46s, the entropy value reaches a peak of 2.73, significantly exceeding the low stability threshold, indicating that the signal stability has significantly decreased. At the same time, the entropy variance also increases synchronously, reflecting the increase in waveform chaos. During this stage, the dynamic Bayesian network will identify high uncertainty and adjust the blood oxygen estimation strategy. After t=56s, the signal gradually enters the recovery state, and the entropy average value and entropy variance gradually fall back to within the high stability threshold, indicating that the signal has become stable again.

[0036] like Fig. 3As shown, first the system acquires PPG raw signals through sensors, then pre-processes and analyzes the waveforms, identifies independent pulse waveforms, then performs Fourier transform on each waveform to obtain spectral wave distribution, based on the frequency domain distribution, the system further calculates the pulse waveform morphology spectral wave entropy (mean + variance) as an uncertainty index to measure signal stability, then the system calls two models in parallel: one is a stable-state blood oxygen model (high precision) (sensitive to morphology), suitable for high-precision estimation when the signal is stable; the second is a change rate blood oxygen model (robust) (not sensitive to morphology), used for trend judgment when the signal is disturbed, the outputs of the two models are input into the dynamic Bayesian network together with the spectral wave entropy index, which performs three key functions: dynamically adjusts the weight according to the spectral wave entropy, performs belief update, and individual model bias calibration, realizes continuous adaptation and precision optimization to individual differences of the user; finally, the system outputs the blood oxygen estimation value + confidence through the decision output module, and starts the real-time alarm triggering mechanism when the trigger condition is met, realizing high reliability and practicality of blood oxygen change monitoring and risk warning.

[0037] In this embodiment, the threshold value of the pulse waveform morphology harmonic entropy index for distinguishing whether the signal is stable or not is taken as an example. The calibration procedure first involves a data acquisition phase, which requires the tester to perform a set of pre-defined standard actions after wearing the device, including a completely static state sufficient to establish a physiological baseline, and a motion state simulating daily activities. In the data processing stage, the system calculates and records the harmonic entropy values of all independent pulse waveforms generated in these two stages respectively, forming a resting entropy value dataset and a motion entropy value dataset. The specific value of the high stability threshold is set as a higher percentile point in the resting entropy value dataset, for example, the ninetieth percentile, to ensure that most high-quality signals are correctly identified as stable. Correspondingly, the low stability threshold is set as a lower percentile point in the motion entropy value dataset, for example, the tenth percentile, to identify significant signal distortion. The determination method of the change rate of the harmonic decay slope index is to continuously collect the user's stable monitoring data in a pre-set initial calibration period, and calculate the average value and standard deviation of the change rate based on this data set. Then the threshold value is set as the value of the average value plus a predetermined multiple, for example, three times, of the standard deviation, which is used as the statistical boundary for identifying physiological transient events.

[0038] At the specific internal operation logic level of the dynamic Bayesian network, the core function is to perform weight distribution, belief update and bias calibration according to explicit rules. Regarding the dynamic adjustment of the trust weight, it is based on the aforementioned calibrated harmonic entropy threshold value. When the input pulse waveform harmonic entropy index is lower than the high stability threshold value, the trust weight of the stable state blood oxygen estimation model is assigned a preset high value, such as 0.95, and the trust weight of the change rate blood oxygen estimation model is assigned a preset low value, such as 0.05. When the harmonic entropy index is higher than the low stability threshold value, the weight distribution is reversed. In the transition interval between the two threshold values, the trust weight of the stable state model linearly decreases from the preset high value to the preset low value with the increase of the harmonic entropy index, and the weight of the change rate model increases accordingly. The sum of the two is always one. The final output of the blood oxygen estimation value is the sum of the product of the output values of the two parallel models and their trust weights at that time. The basic value of the confidence level is the trust weight of the currently dominant model. The basic value is proportionally adjusted according to the degree of harmonic entropy index exceeding the resting state baseline value to finely reflect the signal quality. Regarding the calibration mechanism of individual model bias, the bias is maintained as a specific numerical compensation value within the network. Its update and application follow the following steps: when the pulse waveform harmonic entropy index is lower than the high stability threshold value and this state has lasted for more than a preset time, for example, thirty seconds, the system determines that it has entered an effective calibration window period. In this window period, the system compares the original output value of the stable state blood oxygen estimation model with a reference value, which is determined as the moving average of the previous series of high confidence output values. The difference generated by the comparison is the estimation error at the current time. The update of the individual model bias parameter is to multiply the estimation error by a preset learning rate and add the result to the original individual model bias value. In each blood oxygen estimation, the original value of the stable state model output is first added to the real-time updated individual model bias value, and then participates in the subsequent weighted calculation, so as to realize the continuous compensation of individual physiological characteristic drift.

[0039] The classification of physiological transient events relies on joint change trajectory analysis and association library matching, which is specifically implemented as follows: This association library is pre-constructed during the product development stage by analyzing a large amount of clearly labeled physiological event data. For each known event type, the library stores not the original trajectory graphics, but a set of quantitative feature vectors extracted from the two time series of the harmonic entropy and harmonic attenuation slope change rate of the event; this vector contains at least dimensions such as the peak size of each peak and the time difference between the two peaks. When a physiological transient event is determined to have occurred during real-time monitoring, the system immediately calculates the actual values ​​of these features in this event to form a measured feature vector. The event type classification process is to find the pre-stored event type with the highest match to the measured feature vector in the association library by calculating a pre-defined similarity score. Once a match is successful, the system will call the dynamic compensation factor pre-associated with the event type. The learning and updating of this factor occurs after the classified transient event ends and the signal returns to stability. The algorithm determines a target estimate based on the blood oxygen readings in the stable period after the event. Then, it calculates the error between the estimate obtained using the current factor at the time of the event and this target estimate. Finally, based on the size and direction of the error, a small, iterative adjustment is applied to the factor corresponding to the event type in the library. The adjustment direction is aimed at reducing the same error the next time a similar event occurs, thereby enabling the compensation strategy to achieve personalized self-optimization.

[0040] Example 5: In another specific embodiment of the present invention, in order to ensure the universality and high accuracy of the algorithm across different hardware platforms and individual users, a systematic offline calibration and initialization process is required before deployment, and a precise internal operation logic is followed during operation. The process first involves the construction of a physiological transient event association library. By collecting a database covering annotated physiological event types, for each segment of event data, the harmonic entropy time series is synchronously extracted. Time series of the rate of change of harmonic attenuation slope indicator , calculate a standardized feature vector from these two time series, which contains Peak 、 Peak ,as well as Peak relative to Peak time difference Then, the feature vectors of the same type of events are clustered to form a standard fingerprint of each type of event, which is composed of the mean and covariance matrix of the feature vectors, and is stored in the correlation library of the device. When a physiological transient event is judged to occur in real-time monitoring, the system calculates the measured feature vector of the event, and obtains the similarity score by calculating the Mahalanobis distance between the feature vector and each standard fingerprint in the library. The matching item with the highest score is the classification result of the event.

[0041] In the initial calibration phase of the individual deployment for a specific user, the user needs to complete a standardized test containing rest, specific limb activity and posture transformation according to the guidance. The core task of the algorithm in this phase is to calibrate the key parameters for subsequent operation. First, based on the data collected in this phase, the statistical method described in Embodiment 4 is used to determine the high stability threshold and the low stability threshold. Second, the predetermined learning rate for individual model bias update is calibrated During the rest period of the test process, the system regards the continuous output of the stable-state blood oxygen estimation model as a time series and calculates the autocorrelation coefficient thereof. The value of the learning rate is set to be inversely proportional to the autocorrelation coefficient, so as to ensure the stability of the calibration process when the user's physiological state is stable. For the dynamic compensation factor , its initial value is calibrated in the posture transformation phase. The system optimizes the smooth transition of blood oxygen readings before and after the user's posture transformation as the target, and solves a that can minimize the reading jump through the gradient descent optimization method.

[0042] After the above calibration and entering the daily monitoring, each operation cycle of the dynamic Bayesian network follows the following steps. The network first receives the pulse wave form harmonic entropy index , the original output value of the stable-state blood oxygen estimation model and the output value of the change rate blood oxygen estimation model . Then, according to and the calibrated high stability threshold and low stability threshold, the trust weight of the stable-state model is calculated, which linearly decreases with the increase of in the transition interval between the two thresholds. Next, the system applies the individual model bias for calibration, and the calibrated output value of the stable-state model is , wherein is the currently stored individual model bias value. The final output blood oxygen estimation value is synthesized by the following weighting formula: At the same time, the system calculates the final confidence , which is is the base value, and is scaled down proportionally according to the degree of deviation from the resting state baseline, to finely reflect the reliability of the result, if the current period is determined to be a valid calibration window, i.e. If the duration is continuously below the high stability threshold, the network performs an update of the individual model bias. Its update logic will take the current , compare it with a moving average based on recent high-confidence output values, and derive an estimated error , and update the bias value accordingly: If the system matches a once event to a specific type of physiological transient based on the aforementioned correlation library, and detects signal recovery to stability after the event, it will initiate an iterative optimization of the dynamic compensation factor associated with the event type, the algorithm will compare the compensated blood oxygen readings with the readings during the post-event stable period, and according to the difference, apply a small adjustment to the corresponding value stored in the library, to make it more adaptive to the specific physiological response pattern of this user, thus achieving personalized self-optimization of the compensation strategy.

[0043] It is apparent for those skilled in the art that the present application is not limited to the details of the foregoing exemplary embodiments, and the present application can be implemented in other concrete forms without departing from the spirit or essential characteristics of the present application.

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not to limit the present application, and although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. A blood oxygenation change monitoring algorithm fusing dynamic Bayesian modeling, characterized in that, The method comprises the following steps: Step 1, obtaining a raw photoplethysmogram signal; Step 2, based on the raw signal, identifying a plurality of independent pulse waveforms in a plurality of consecutive time windows; Step 3, for each independent pulse waveform, performing Fourier transform to obtain its harmonic energy distribution; Step 4, based on the harmonic energy distribution, quantitatively generating a pulse waveform morphology harmonic entropy index representing the stability of the pulse waveform morphology, the pulse waveform morphology harmonic entropy index comprising a combination of harmonic entropy mean value and harmonic entropy variance; Step 5, constructing a dynamic Bayesian network, the internal state of the dynamic Bayesian network comprising a final blood oxygen estimation value, a current confidence level, and an individual model bias for calibrating the blood oxygen estimation model; Step 6, taking the pulse waveform morphology harmonic entropy index as an input of the dynamic Bayesian network, and taking the outputs of two parallel running blood oxygen estimation models with different response characteristics as the other two inputs of the dynamic Bayesian network; Step 7, the dynamic Bayesian network dynamically adjusts the trust weight of the output results of the blood oxygen estimation models according to the pulse waveform morphology harmonic entropy index; and the dynamic Bayesian network performs belief updating to adjust the individual model bias according to the trust weight, and outputs a decision pair comprising the final blood oxygen estimation value and the corresponding confidence level, the confidence level being determined by the pulse waveform morphology harmonic entropy index and the belief updating process.

2. The fusion dynamic Bayesian modeling of blood oxygenation change monitoring algorithm of claim 1, wherein, The dynamic Bayesian network adjusts the trust weight of the output results of the two parallel running blood oxygen estimation models according to the numerical range of the pulse waveform morphology harmonic entropy index: when the pulse waveform morphology harmonic entropy index indicates that the pulse waveform morphology stability is high, the dynamic Bayesian network gives a higher trust weight to the stable-state blood oxygen estimation model, and iteratively adjusts the individual model bias based on a predetermined learning rate using the data of the current pulse waveform morphology stability; when the pulse waveform morphology harmonic entropy index indicates that the pulse waveform morphology stability is low, the dynamic Bayesian network gives a higher trust weight to the change rate blood oxygen estimation model, and reduces the confidence level of the final output result to a predetermined lower limit according to a predetermined algorithm.

3. The fusion dynamic Bayesian modeling of blood oxygenation change monitoring algorithm of claim 2, wherein, The stable-state blood oxygen estimation model is a linear regression model trained based on large-scale statistical data; the change rate blood oxygen estimation model is a linear estimation model based on the change rate of the AC / DC component ratio of the photoplethysmogram signal, and the change rate blood oxygen estimation model is not sensitive to the pulse waveform morphology.

4. The fusion dynamic Bayesian modeling of blood oxygenation change monitoring algorithm of claim 1, wherein, The method further comprises the steps of: based on the harmonic analysis of the independent pulse waveform, extracting the energy ratio of the high harmonic component to the fundamental component in the harmonic energy distribution to obtain a harmonic attenuation slope index representing the energy spectrum shape of the pulse waveform; monitoring the change rate of the harmonic attenuation slope index in real time; and when the absolute value of the change rate exceeds a threshold determined based on historical data statistics within a predetermined time period, determining that a physiological transient event occurs, and applying a dynamic compensation correction amount to the final blood oxygen estimation value according to the change rate, the correction amount being: correction amount = (1 - dynamic compensation factor) * (1 - change rate of harmonic attenuation slope index) wherein, the dynamic compensation factor is, the change rate of the harmonic attenuation slope index.

5. The fusion dynamic Bayesian modeling of blood oxygenation change monitoring algorithm of claim 4, wherein, Dynamic compensation factor The dynamic Bayesian network learns and updates from the data after the physiological transient event ends according to the pulse wave morphology harmonic entropy index.

6. The fusion dynamic Bayesian modeling of blood oxygenation variation monitoring algorithm of claim 4, wherein, The method further comprises the following steps: simultaneously obtaining the joint change trajectory of the harmonic attenuation slope index and the pulse waveform morphology harmonic entropy index during a physiological transient event; based on the morphological features of the joint change trajectory, classifying the physiological transient event into a type, the type classification comprising a hemodynamic transient type and a cardiac electrophysiological transient type; and adaptively selecting and applying a dynamic compensation factor pre-associated with the event type to determine the dynamic compensation correction amount according to the type classification result.

7. The fusion dynamic Bayesian modeling of blood oxygenation variation monitoring algorithm of claim 1, wherein, The method further comprises the steps of: extracting phase information of the second harmonic in the harmonics based on harmonic analysis of the independent pulse waveform; calculating variance of the phase information within a predetermined time window to obtain a harmonic phase jitter index representing the phase-locked state of the PPG original signal; and determining the source reliability of the PPG original signal based on the harmonic phase jitter index to obtain a source confidence index.

8. The fusion dynamic Bayesian modeling of blood oxygenation variation monitoring algorithm of claim 7, wherein, When the value of the source confidence index is higher than a predetermined threshold, the dynamic Bayesian network continues to perform blood oxygen estimation and confidence output; when the value of the source confidence index is not higher than the predetermined threshold, the method outputs a diagnostic information indicating that the device is abnormal, instead of a decision pair containing the blood oxygen estimation value and the confidence level.

9. The fusion dynamic Bayesian modeling of blood oxygenation variation monitoring algorithm of claim 1, wherein, When the final blood oxygen estimation value is lower than a physiological safety threshold and the corresponding confidence level is higher than a credibility threshold, the method triggers a real-time alarm, which is sent through a user wearable device or a remote health monitoring platform.

10. The fusion dynamic Bayesian modeling of blood oxygenation variation monitoring algorithm as claimed in claim 1, wherein, The update of the individual model bias parameter in the dynamic Bayesian network is based on the blood oxygen estimation history value and the corresponding PPG feature when the pulse waveform morphology harmonic entropy index indicates that the signal stability is high, and the update uses an algorithm that minimizes historical errors to ensure adaptive calibration of long-term physiological changes of the user.

Images