A blood oxygenation change monitoring algorithm that integrates dynamic Bayesian modeling

Abstract

This invention relates to the field of computer system technology based on specific computational models, and discloses a blood oxygen change monitoring algorithm that integrates dynamic Bayesian modeling. The algorithm includes: quantifying signal uncertainty by calculating the entropy value of the harmonic energy distribution of the pulse waveform; driving a dynamic Bayesian network based on this entropy value to perform confidence evaluation and weighted processing on the blood oxygen estimation model; and finally outputting a decision pair containing the estimated blood oxygen value and its confidence value. This invention transforms signal uncertainty into a calculable entropy index, enabling the system to autonomously distinguish between real physiological changes and measurement noise, avoiding the misjudgment problem caused by motion interference in traditional blood oxygen monitoring. Simultaneously, it utilizes the continuous learning capability of Bayesian networks to achieve seamless personalized calibration, significantly improving the reliability and practicality of medical-grade wearable devices.

Description

Technical Field

[0001] This invention relates to a blood oxygen change monitoring algorithm that integrates dynamic Bayesian modeling, belonging to the field of computer system technology based on specific computational models. Background Technology

[0002] Currently, wearable medical devices generally use photoplethysmography (PPG) signals to analyze blood oxygen saturation. The technical approach mainly relies on time-domain filtering for noise reduction and static threshold model calculation. For example, the mainstream approach suppresses motion artifacts through adaptive filtering and outputs blood oxygen values ​​based on a fixed linear regression model established by large-scale population data. These methods can maintain basic accuracy when the user is at rest. However, when applied to home monitoring scenarios, their inherent limitations are significantly exposed.

[0003] In users' daily activities (such as walking or housework), the spectral aliasing of motion interference and physiological signals makes it difficult for traditional filtering to separate effective information. Forced noise reduction often causes waveform distortion. More importantly, individual differences in vascular characteristics and physiological state drift make static models unable to adapt to users' unique vital signs parameters. Patients with chronic diseases need repeated calibration, which deviates from the original intention of non-sensory monitoring. To address this challenge, the industry has tried to introduce multi-sensor fusion, such as accelerometer-assisted motion recognition or complex neural network optimization models. However, the former increases hardware costs and power consumption, while the latter is difficult to deploy on edge devices due to excessive computational load. Moreover, neither of them solves 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 noisy environments, system performance often faces the challenge of error amplification or signal loss.

[0004] Specifically, existing technologies suffer from the following three main shortcomings: 1. While time-domain filtering can suppress some noise, it cannot quantify the impact of residual interference on the confidence level of the computational model, resulting in a lack of reliability indicators in the system output; 2. Fixed-threshold models ignore the specificity and time-varying nature of individual vascular elasticity and hemodynamics, leading to systematic biases across users and time periods; 3. Improvement schemes relying on multi-sensor or deep networks increase hardware costs and computing power requirements, hindering device battery life and widespread application. Therefore, how to construct an uncertainty quantification mechanism endogenous to the computational model, dynamically analyze signal quality to autonomously adjust decision logic, and achieve lightweight individual self-adaptation, thereby achieving medical-grade reliable monitoring on resource-constrained edge devices, becomes the technical problem to be solved by this invention. Summary of the Invention

[0005] This invention provides a blood oxygen change monitoring algorithm that integrates dynamic Bayesian modeling. Its main purpose is to solve the problem of unreliable monitoring results caused by the lack of quantification of signal uncertainty and the limitations of static modeling in medical-grade wearable devices under scenarios of motion interference and individual differences.

[0006] To achieve the above objectives, this invention provides a blood oxygenation change monitoring algorithm that integrates dynamic Bayesian modeling, comprising the following steps:

[0007] Step 1: Obtain the raw photoplethysmography (PPG) signal without complex filtering;

[0008] Step 2: Based on the original signal, identify multiple independent pulse waveforms within multiple consecutive time windows;

[0009] Step 3: Perform a Fourier transform on each individual pulse waveform to obtain its harmonic energy distribution;

[0010] Step 4: Based on the harmonic energy distribution, a pulse waveform morphology harmonic entropy index is generated to characterize the stability of the pulse waveform morphology. The pulse waveform morphology harmonic entropy index includes a combination of the harmonic entropy mean and the harmonic entropy variance.

[0011] Step 5: Construct a dynamic Bayesian network. The internal state of the dynamic Bayesian network includes the final blood oxygen estimate, the current confidence level, and the individual model bias used to calibrate the blood oxygen estimation model.

[0012] Step 6: Use the pulse wave morphology harmonic entropy index as the input to the dynamic Bayesian network, and use the outputs of two parallel blood oxygen estimation models with different response characteristics as the other two inputs to the dynamic Bayesian network.

[0013] Step 7: The dynamic Bayesian network dynamically adjusts the trust weights of the blood oxygen estimation model output based on the pulse wave morphology harmonic entropy index; and the dynamic Bayesian network performs belief updates based on the trust weights to adjust individual model biases, and outputs a decision pair containing the final blood oxygen estimation value and its corresponding confidence level. The confidence level is jointly determined by the pulse wave morphology harmonic entropy index and the belief update process.

[0014] Preferably, the dynamic Bayesian network adjusts the confidence weights of the output results of two parallel blood oxygen estimation models based on the numerical range of the pulse waveform morphology harmonic entropy index: when the pulse waveform morphology harmonic entropy index indicates high pulse waveform morphology stability, the dynamic Bayesian network gives higher confidence weights to the steady-state blood oxygen estimation model and uses the current data with high pulse waveform morphology stability to iteratively adjust the individual model bias based on a predetermined learning rate; when the pulse waveform morphology harmonic entropy index indicates low pulse waveform morphology stability, the dynamic Bayesian network gives higher confidence weights to the rate-of-change blood oxygen estimation model and reduces the confidence level of the final output result to a predetermined lower limit according to a predetermined algorithm.

[0015] Preferably, the steady-state blood oxygen estimation model is a linear regression model trained on large-scale statistical data; the rate of change blood oxygen estimation model is a linear estimation model based on the rate of change of the ratio of AC to DC components of the photoplethysmography pulse wave signal, and the rate of change blood oxygen estimation model is not sensitive to the pulse waveform morphology.

[0016] Preferably, the method further includes the following steps: based on harmonic analysis of independent pulse waveforms, extracting the energy ratio of higher harmonic components to fundamental components in the harmonic energy distribution to obtain a harmonic attenuation slope index characterizing the energy spectrum morphology of the pulse waveform; real-time monitoring of the rate of change of the harmonic attenuation slope index; and when the absolute value of the rate of change exceeds a threshold 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 estimate based on the rate of change, wherein the correction is: correction = in, It is a dynamic compensation factor. It is the rate of change of the harmonic attenuation slope index.

[0017] Preferred dynamic compensation factor The dynamic Bayesian network learns and updates data after the physiological transient event ends, based on the pulse wave morphology harmonic entropy index after the data has recovered stability.

[0018] Preferably, the method further includes the following steps: acquiring in parallel the joint change trajectory of the harmonic attenuation slope index and the pulse wave morphological harmonic entropy index during the physiological transient event; classifying the physiological transient event based on the morphological characteristics of the joint change trajectory, including hemodynamic transient type and 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.

[0019] Preferably, the method further includes the following steps: based on harmonic analysis of independent pulse waveforms, extracting the phase information of the second harmonic in the harmonics; calculating the variance of the phase information within a predetermined time window to obtain a harmonic phase jitter index characterizing the rhythmic phase-locked state of the original photoplethysmography (PPG) wave signal; and based on the harmonic phase jitter index, determining the source reliability of the original PPG wave signal to obtain a source confidence index.

[0020] Preferably, 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 message indicating an abnormal device status, rather than a decision pair containing blood oxygen estimation value and confidence level.

[0021] Preferably, the method triggers a real-time alarm when the final blood oxygen estimate is lower than a physiological safety threshold and its corresponding confidence level is higher than a credibility threshold. The alarm is issued through the user's wearable device or a remote health monitoring platform.

[0022] Preferably, the update of the individual model bias parameters in the dynamic Bayesian network is based on the historical value of blood oxygen estimation when the pulse wave morphology harmonic entropy index indicates high signal stability and the corresponding photoplethysmography pulse wave characteristics. The update adopts an algorithm that minimizes historical errors to ensure imperceptible adaptive calibration to long-term physiological changes of individual users.

[0023] Compared with the prior art, the beneficial effects of the present invention are:

[0024] 1. By converting the harmonic energy distribution of the pulse waveform into an entropy measure, the system establishes an endogenous uncertainty quantification scale for the first time in photoplethysmography pulse wave signal processing. This mechanism 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 based on this entropy value and instead adopts a more robust rate of change model for trend tracking, thereby avoiding the information loss that is difficult to avoid in the signal separation process of traditional schemes.

[0025] 2. The parallel architecture of the steady-state model and the rate of change model is not a simple redundant design. When the harmonic entropy indicator signal is stable, the system prioritizes calling the high-precision model to output blood oxygen value, while continuously calibrating individual parameters using the low uncertainty data at this moment. When the signal chaos increases, the system seamlessly switches to the anti-interference model and simultaneously reduces the output confidence. This self-verifying decision logic enables the device to maintain rational output in motion scenarios, allowing users to distinguish between real physiological alarms and measurement noise, thus fundamentally solving the trust crisis of medical-grade wearable devices.

[0026] 3. Harmonic phase variance is extracted synchronously during the Fourier transform stage. The system discovers rhythmic phase-locked state information that has been discarded by traditional solutions. When the phase jitter exceeds the threshold, the device automatically determines that it is a sensor contact failure rather than ordinary motion interference, triggering a dedicated hardware diagnostic command. This mechanism reuses the intermediate products of the existing calculation process, enabling the system to distinguish the causes of signal degradation without adding hardware or algorithm complexity, providing accurate fault location basis for subsequent operations.

[0027] 4. For sudden changes in harmonic attenuation slope caused by scenarios such as abrupt changes in body position, the system analyzes the dynamic trajectory coupling relationship between the change and the harmonic entropy to distinguish between hemodynamic transients and cardiac rhythm abnormalities. This fingerprint recognition based on two-dimensional phase space trajectory enables the compensation factor to adaptively adjust according to the event type: slow-release compensation is used when there are sudden changes in vascular tension, and a rapid response strategy is activated when there is ECG desynchronization. As a result, acute physiological changes no longer cause model inaccuracies, but are transformed into a new opportunity for the system to deepen individual cognition. Attached Figure Description

[0028] Figure 1 This is a time series diagram of blood oxygenation change monitoring based on dynamic Bayesian modeling, as presented in this invention.

[0029] Figure 2 This is a dynamic response curve of the harmonic entropy index of the present invention during motion interference.

[0030] Figure 3 This is a flowchart of the blood oxygen estimation decision-making process that integrates dynamic Bayesian networks according to the present invention.

[0031] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0032] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0033] This application provides a blood oxygen saturation monitoring algorithm that integrates dynamic Bayesian modeling. This algorithm runs on a wrist-worn or finger-clip wearable device that integrates photoelectric transmitting and receiving units. By deeply analyzing the inherent uncertainties of the photoplethysmography (PPG) signal, it achieves self-assessment of the reliability of the estimated blood oxygen value and individualized adaptive calibration. In a specific application scenario, such as monitoring a chronic obstructive pulmonary disease (COPD) patient undergoing home rehabilitation training, the algorithm operates as follows: First, the device's photoelectric sensor continuously acquires the raw PPG signal without complex filtering. This raw signal has only undergone necessary amplification by the front-end analog circuitry. Anti-aliasing filtering completely preserves the dynamic characteristics of the signal caused by the user's physiological activities, sensor displacement, and real physiological fluctuations, providing undistorted raw input for subsequent uncertainty quantification. Then, within a continuously rolling time window, for example, every eight seconds as a processing cycle, the algorithm processes the acquired photoplethysmography (PPG) pulse wave raw signal. By accurately identifying trough or peak feature points in the signal sequence, it segments multiple independent, continuous pulse waveforms, each corresponding to a complete heartbeat cycle in time. Subsequently, for each segmented independent pulse waveform, the system performs a Fast Fourier Transform (FFT) to transform the waveform signal from the time domain... The frequency domain is transformed to obtain its harmonic energy distribution, which clearly reveals the energy corresponding to the fundamental frequency of the heart rate, as well as the energy composition of its second, third, and even higher harmonic components. Based on the harmonic energy distribution obtained in the previous step, the system calculates a pulse wave morphological harmonic entropy index. Specifically, for a single pulse waveform, the energy of each harmonic is first normalized so that its sum is one, thereby constructing an energy probability distribution. Then, according to the principle of information entropy calculation, that is, summing the products of all probability terms and their logarithms and taking the negative value, the degree of disorder of the distribution is quantified. A pulse waveform with a stable and clean morphology has high energy. The intensity is concentrated in low-order harmonics and is distributed in an orderly manner, resulting in low harmonic entropy values. Conversely, when noise such as motion artifacts is mixed in, the waveform is distorted, and energy is randomly distributed across various harmonic frequencies, leading to a significant increase in harmonic entropy values. The algorithm further calculates the harmonic entropy of all independent pulse waveforms within a relatively long time window, such as 30 seconds, and finally generates a pulse waveform morphology harmonic entropy index composed of the average harmonic entropy and the variance of harmonic entropy. The average value of this index reflects the overall level of recent signal quality, while the variance characterizes the degree of fluctuation in signal quality. The combination of the two provides the system with a precise and quantitative measure of the uncertainty of the current signal.

[0034] Based on this uncertainty measure, a dynamic Bayesian network is constructed and driven, which serves as the decision-making center of the entire algorithm. Its internal state is iteratively updated at each time step. These state variables include: the final blood oxygen estimate, representing the system's best estimate of the user's true blood oxygen saturation; a quantitative index reflecting the system's current confidence level in its output; and a crucial calibration parameter, used to compensate for the individual model bias, which compensates for systematic differences between the universal model and the user's individual physiological characteristics. In each computation cycle, the network receives three key information streams: the first is the aforementioned pulse wave morphology harmonic entropy index, which acts as an arbiter of signal quality; the other two inputs come from two parallel blood oxygen estimates with different response characteristics. The model outputs two models: a steady-state blood oxygen estimation model and a rate-of-change blood oxygen estimation model. The former is a linear regression model trained on large-scale statistical data. It uses the ratio of the AC to DC components of red and infrared light signals as the key input and calculates blood oxygen values ​​through a fixed linear relationship. It has extremely high accuracy when the signal is pure, but is very sensitive to changes in waveform morphology. The latter is a linear estimation model based on the rate of change of the AC to DC component ratio of the photoplethysmography pulse wave signal. It 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 a more robust direction of blood oxygen change in interference environments. The dynamic Bayesian network performs its core inference and update functions. The function involves the network dynamically adjusting the trust weights of the outputs of two parallel blood oxygen estimation models based on the numerical range of the input pulse wave morphology harmonic entropy index. When the pulse wave morphology harmonic entropy index indicates high stability of the pulse waveform morphology, i.e., the mean and variance of the harmonic entropy are both in a preset low range, the dynamic Bayesian network assigns a higher trust weight to the steady-state blood oxygen estimation model, considering its output at this moment to be highly reliable. Simultaneously, the system uses this highly trusted data to perform belief updates, iteratively adjusting the individual model bias parameters based on a predetermined learning rate. This predetermined learning rate is the result of a trade-off between calibration response speed and stability; an excessively high learning rate will cause the model to become overly sensitive to noise from a single measurement. The network is sensitive to changes in blood oxygen saturation, but if the value is too low, it may not be able to keep up with the user's actual physiological changes in a timely manner. This update process uses an algorithm designed to minimize historical errors, such as recursive least squares, to ensure imperceptible adaptive calibration for long-term physiological changes caused by changes in the user's physiological state or the effects of medication. Conversely, when the pulse waveform morphology harmonic entropy index indicates low stability of the pulse waveform morphology, the network assigns a higher confidence weight to the rate of change blood oxygen estimation model, adopting the blood oxygen change trend information it provides. At the same time, the network significantly reduces the confidence level of the final output result to a predetermined lower limit according to a predetermined algorithm that is negatively correlated with the harmonic entropy value, thereby clearly informing the user or monitoring system that the current absolute blood oxygen reading may be inaccurate, but the trend of blood oxygen change is referential.

[0035] To address rapid changes in physiological state, this algorithm also includes a mechanism for detecting and compensating for physiological transient events. During harmonic analysis, the system simultaneously extracts the energy ratio of higher harmonic components to the fundamental component in the harmonic energy distribution, obtaining a harmonic attenuation slope index that characterizes the energy spectrum morphology of the pulse waveform. The algorithm monitors the rate of change of this slope index in real time. When the absolute value of this rate of change exceeds a threshold determined based on user historical data statistics within a predetermined time period (e.g., two to three seconds), the system determines that a physiological transient event has occurred, such as blood pressure fluctuations caused by a sudden change in body position. In this case, the system applies a dynamic compensation correction to the final blood oxygen estimate. The value of this correction is equal to the product of a dynamic compensation factor and the aforementioned rate of change of the harmonic attenuation slope index. The dynamic compensation factor is not a fixed value; it is learned and updated by a dynamic Bayesian network after the physiological transient event ends, based on new data after the pulse waveform morphology harmonic entropy index has recovered stability, thus making the compensation more individualized. Even more remarkably, the algorithm can classify physiological transient events. By acquiring the joint change trajectory of the harmonic attenuation slope index and the pulse wave morphology harmonic entropy index during physiological transient events in parallel, and analyzing the morphological characteristics of this two-dimensional trajectory, it can distinguish events of different natures. For example, the trajectory of hemodynamic transients caused by drastic changes in vascular tension may be relatively smooth, while the trajectory of cardiac electrophysiological transients caused by arrhythmias may exhibit a sharper and more irregular shape. Based on a pre-established association library between trajectory shape and event type, the system can classify the type and adaptively select and apply a more targeted dynamic compensation factor pre-associated with the event type to determine the dynamic compensation correction amount.

[0036] To ensure the reliability of the input signal source itself, the algorithm also introduces a source confidence assessment. In the harmonic analysis stage, the system specifically extracts the phase information of the second harmonic and calculates the variance of this phase information within a predetermined time window. This yields a harmonic phase jitter index characterizing the phase-locked state of the original photoplethysmography (PPG) signal. Ideally, the continuous pulse wave rhythm is stable and the phase is locked, resulting in a very low index value. If poor contact between the sensor and the skin causes signal slippage or detachment, it will lead to signal timing disorder and severe phase jitter. Based on this harmonic phase jitter index, the system can determine the source reliability of the original PPG signal, thus obtaining a source confidence index. When the value of this source confidence index is higher than a predetermined threshold, it indicates that the signal source is reliable, and the dynamic Bayesian network continues to perform blood oxygen estimation and confidence output. However, when its value is not higher than the threshold, the system determines that the device status is abnormal and prioritizes outputting a diagnostic message indicating the device status abnormality, rather than outputting a decision pair containing blood oxygen estimation and confidence level that may be misleading. Ultimately, the decision pairs output by the algorithm can be used to trigger real-time, reliable alarms. When the final estimated blood oxygen saturation value is below a physiological safety threshold, such as 90%, and its corresponding confidence level is above a credibility threshold, such as 95%, the system will trigger a real-time alarm. This dual-condition judgment mechanism can effectively filter out false low blood oxygen readings with low confidence caused by motion interference, greatly reducing the false alarm rate and improving the effectiveness of the alarm. It should also be noted that the large-scale database required for constructing the association library and training the steady-state blood oxygen estimation model strictly adheres to national laws and regulations and industry ethical standards regarding data security and personal information protection. All raw physiological data used for offline model training and verification is collected with the full informed consent of the data providers. Furthermore, before the data enters the algorithm model training stage, all data undergoes strict anonymization or de-identification processing to ensure the protection of the data subject's personal privacy. These are all extension implementation methods known to those skilled in the art.

[0037] Example 1: In this example, the monitoring subject is a user in the early recovery stage after surgery. The behavior pattern during the monitoring period is from long periods of bed rest to the first attempt to stand up and walk a short distance with assistance. This process constitutes a dual test of the robustness and accuracy of the monitoring algorithm from a technical point of view, because it not only includes a drastic change in signal quality from extremely high to extremely low, but may also be accompanied by real physiological transient events caused by sudden changes in body position. When a user is at rest for an extended period, their physiological state is stable, and the original photoplethysmography (PPG) signal exhibits regular morphology and high repeatability. During this phase, the algorithm performs harmonic analysis on independent pulse waveforms, calculating a pulse waveform morphological harmonic entropy index that remains consistently low, with both its mean and variance being low. In the dynamic Bayesian network, this is interpreted as indicating a signal source with extremely high reliability. Based on this judgment, the network assigns a very high trust weight to the steady-state blood oxygen estimation model trained on large-scale statistical data, thereby outputting a high-precision, high-confidence final blood oxygen estimation value. More importantly, the high-quality data stream at this stage is identified as an effective calibration window. The dynamic Bayesian network utilizes these historical blood oxygen estimation values ​​with confidence levels exceeding a predetermined confidence threshold, along with their corresponding PPG characteristics, to continuously iteratively adjust the individual model bias parameters within the network. This mechanism transforms each period of stable rest for the user into an opportunity for seamless, personalized model calibration. By minimizing historical errors, the algorithm gradually avoids systematic deviations between the universal model and the user's current physiological state.

[0038] When a user attempts to stand up, the large, irregular movements of their limbs introduce strong motion artifacts into the original photoplethysmography (PPG) signal. Traditional algorithms often fail at this point because their filtering mechanisms struggle to balance preserving true physiological fluctuations with filtering out motion noise. In this technical solution, the pulse wave morphology harmonic entropy index jumps sharply at this moment, directly quantifying the increase in signal chaos. This entropy jump triggers a fundamental shift in decision logic within the dynamic Bayesian network: the network immediately abandons its high trust in the steady-state blood oxygen estimation model, which is susceptible to wave morphology, and instead assigns trust weights to the rate-of-change blood oxygen estimation model, which is insensitive to pulse waveform morphology. Simultaneously, based on the magnitude of the entropy increase, the network proactively lowers the confidence level of the output decision pairs to a predetermined lower limit. This design reflects a paradigm shift in the algorithm, moving from attempting to infer precise values ​​from unreliable data to acknowledging and managing uncertainty and providing the most reliable judgment under the current conditions. It aims to address the trust problem caused by seemingly accurate but potentially inaccurate blood oxygen values ​​during movement.

[0039] During this process, if a user experiences a real hemodynamic transient due to a sudden change in body position, such as brief orthostatic hypotension, this event will leave a set of identifiable features at the signal level. In addition to the pulse wave morphology harmonic entropy index increasing due to movement, the harmonic attenuation slope index, which characterizes the energy spectrum morphology of the pulse waveform, will also show synchronous and dramatic changes that differ from conventional movement interference. By capturing the joint change trajectory of the harmonic attenuation slope index and the pulse wave morphology harmonic entropy index during the physiological transient event and analyzing their morphological characteristics, the algorithm can classify the root cause of this event as a hemodynamic transient type. The system's ability to qualitatively classify events does not stem from a single index, but from its analysis of the coupling relationship between two physically independent indices. Based on this accurate classification, the system can adaptively select and apply a dynamic compensation factor pre-associated with the event type to correct the final blood oxygen estimation value, thereby capturing real physiological events while avoiding model inaccuracies.

[0040] Example 2: In a progressive hypoxemia event caused by mild respiratory depression, there is a temporal overlap with a sudden motion artifact introduced by the user's limb position adjustment. To achieve this, the experimental platform is built around a high-precision photoplethysmography (PPG) physical simulation system. This system reproduces the preset blood oxygen saturation decline curve through a programmable optical attenuation array, and simultaneously superimposes the standardized motion artifact signal onto the main signal path with a controllable amplitude, thereby repeatably simulating the aforementioned composite interference scenario in a laboratory environment. First, the fundamental technical consideration in setting the simulated rate of decrease in blood oxygen saturation lies in balancing the realism of physiological events with the limit test of algorithm response. The decision rule is that the rate of decrease should match the time scale of typical events in the target monitoring scenario. Based on this rule, referring to clinical data of moderate hypoxia events, a linear ramp from 97% to 88% is set, lasting 60 seconds. This setting anchors a working condition that is both physiologically consistent and sufficient to test the algorithm's dynamic tracking ability. Second, regarding the injection of motion artifacts, the essence of the decision lies in balancing the degree of damage to the signal morphology caused by artifacts with the recoverability of physiological information. The decision model is that the artifact energy should be sufficient to bring the pulse wave morphology harmonic entropy index into the low stability range preset by the dynamic Bayesian network, but it must ensure that the signal-to-noise ratio of the fundamental frequency of the signal is not lower than the lower limit of effective recognition by the system. Accordingly, when the blood oxygen saturation drops to 92%, a motion artifact lasting 10 seconds with the main frequency component concentrated at 1.5Hz is injected. This parameter combination has been verified to effectively trigger the decision logic switching of this invention.

[0041] After the experiment started, the algorithm began processing the composite photoplethysmography (PPG) signal output by the simulation system. In the initial stage, the blood oxygen value was stable at 97%, and the pulse wave morphology harmonic entropy index output by the algorithm remained stable at a low level. The dynamic Bayesian network assigned a very high confidence weight to the steady-state blood oxygen estimation model, and the final blood oxygen estimation value output was highly consistent with the simulation true value, with a confidence level maintained above 98%. At the 31st second, the blood oxygen saturation began to decrease linearly. At this time, the harmonic entropy index did not show a significant change. Based on the high-confidence steady-state model results, the dynamic Bayesian network accurately tracked this physiological change. At the 45th second, with the injection of preset motion artifacts, the signal morphology deteriorated instantaneously. At this moment, the key indicators inside the algorithm showed a differentiation trend with diagnostic value: the pulse wave morphology harmonic entropy index jumped sharply due to waveform chaos, while the harmonic attenuation slope index, which characterizes the energy spectrum morphology, also showed a sharp pulse in its rate of change that far exceeded the historical statistical threshold. The following table records exemplary data of the algorithm's core variables before and after this key node.

[0042] Table 1: Changes in key variables and event recognition process of dynamic Bayesian monitoring algorithm in simulation test.

[0043]

[0044] At 46 seconds, the sharp increase in harmonic entropy (from 0.85 to 2.73) triggered the dynamic Bayesian network to switch the trust weights from a steady-state model to a rate-of-change model, and simultaneously reduced 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, because the algorithm simultaneously detected a sharp peak (1.88) in the rate of change of the harmonic attenuation slope. By analyzing the joint change trajectory of these two indicators, i.e., the concurrence of a high entropy value and a high rate of change of slope, the root cause of the event was classified as a hemodynamic transient. Based on this classification, the system called the dynamic compensation factor associated with this event type to correct the final blood oxygen estimate. As a result, the final blood oxygen estimate (92.0%) still closely tracked the simulated true value (92.2%), avoiding monitoring interruption or significant deviation of values ​​due to motion artifacts. After the motion artifacts disappeared after 55 seconds, the harmonic entropy quickly dropped back, and the confidence returned to a high level. The algorithm seamlessly switched back to high-precision monitoring mode.

[0045] Example 3: This example combines Figures 1 to 3 This paper describes the implementation of a blood oxygenation change monitoring algorithm that integrates dynamic Bayesian modeling. Figure 1As shown, firstly, a photoelectric sensor is used to acquire the raw PPG signal. This signal is transmitted to a signal preprocessing module to form a preprocessed signal. Next, the preprocessed signal is sent to a waveform recognition module, which analyzes the PPG signal across multiple consecutive time windows and identifies independent pulse waveforms. Subsequently, the time-domain waveform data is sent to a Fourier transform module, where an FFT transform is performed to acquire waveform features in the frequency domain and output the harmonic energy distribution. Then, the data enters an entropy calculation module, which performs information entropy analysis on the harmonic energy distribution to calculate harmonic indices. These indices include... The system calculates the mean and variance, and sends the entropy value (mean + variance) to the Bayesian network module. Thereafter, parallel processing occurs within the Bayesian network. During this process, waveform data is input into both the steady-state model and the rate-of-change model, generating steady-state and rate-of-change blood oxygen estimates respectively. The Bayesian network adjusts the weights based on the current input entropy value to dynamically determine the level of confidence in the outputs of the two models. It also performs belief updates based on the estimation results to correct individual model biases. Finally, the system sends the confidence-weighted result to the output module for output. Value, confidence level.

[0046] like Figure 2 As shown in the figure, the vertical axis represents harmonic entropy values, and the horizontal axis represents time (seconds). The solid line represents the average harmonic entropy, and the dashed line represents the harmonic entropy variance, clearly distinguished in the legend. The figure shows that in the initial resting state, the average harmonic entropy is around 0.85, far below the high stability threshold, indicating a highly stable signal. At t=44s, the harmonic entropy begins to rise, entering the motion interference stage, reaching a peak entropy of 2.73 at t=46s, significantly exceeding the low stability threshold, indicating a significant decrease in signal stability. Simultaneously, the entropy variance also increases, reflecting increased waveform chaos. During this stage, the dynamic Bayesian network identifies high uncertainty and adjusts the blood oxygen estimation strategy. After t=56s, the signal gradually enters a recovery state, and the average entropy and entropy variance gradually fall back below the high stability threshold, indicating that the signal has stabilized again.

[0047] like Figure 3As shown, the system first acquires the raw PPG signal through sensors, then performs preprocessing and waveform analysis to identify independent pulse waveforms. Next, a Fourier transform is performed on each waveform to obtain the spectral distribution. Based on this frequency domain distribution, the system further calculates the pulse wave morphological spectral entropy (mean + variance) as an uncertainty indicator to measure signal stability. Then, the system calls two models in parallel: one is a steady-state blood oxygenation model (high accuracy) (sensitive to morphology), suitable for high-precision estimation when the signal is stable; the other is a rate-of-change blood oxygenation model (robust) (insensitive to morphology), used for trend judgment when the signal is disturbed. The outputs of these two models, along with the spectral entropy indicator, are input into a dynamic Bayesian network. This network performs three key functions: dynamically adjusting weights based on spectral entropy, updating beliefs, and calibrating individual model biases, achieving continuous adaptation and accuracy optimization to individual user differences. Finally, the system outputs the estimated blood oxygen value plus confidence level through decision-making, and activates a real-time alarm trigger mechanism when trigger conditions are met, achieving highly reliable and practical blood oxygenation change monitoring and risk warning.

[0048] Example 4: This example uses the threshold of the pulse wave morphological harmonic entropy index, used to distinguish whether a signal is stable or not, as an example. Its calibration procedure first involves a data acquisition phase, requiring the test subject to perform a set of predefined standard actions after wearing the device. This includes a period of complete rest sufficient to establish a physiological baseline, and a period of motion simulating daily activities. In the data processing phase, the system calculates and records the harmonic entropy values ​​of all independent pulse waveforms generated in these two phases, forming a resting entropy dataset and a motion entropy dataset. The specific value of the high stability threshold is set as a relatively high percentile in the resting entropy dataset. For example, the 90th percentile is used to ensure that the vast majority of high-quality signals are correctly identified as stable. Correspondingly, the low stability threshold is set at a lower percentile in the motion entropy value dataset, such as the 10th percentile, to identify significant signal distortion. The threshold for judging the rate of change of the harmonic attenuation slope index is determined by continuously collecting stable monitoring data from users within a preset initial calibration period, and calculating the mean and standard deviation of the rate of change based on this dataset. The threshold is then set as the mean plus a predetermined multiple, such as three times, the standard deviation, as the statistical boundary for judging physiological transient events.

[0049] At the level of the specific internal operational logic of the dynamic Bayesian network, its core function is to perform weight allocation, belief update, and bias calibration according to explicit rules. Regarding the dynamic adjustment of trust weights, it is based on the previously calibrated harmonic entropy threshold. When the harmonic entropy index of the input pulse wave morphology is lower than the high stability threshold, the trust weight of the steady-state blood oxygen estimation model is assigned a preset high value, such as 0.95, while the trust weight of the rate-of-change 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, this weight allocation is reversed. Within the transition range between the two thresholds, the trust weight of the steady-state model decreases linearly from the preset high value to the preset low value as the harmonic entropy index increases, while the weight of the rate of change model increases accordingly. The sum of the two is always one. The final output blood oxygen estimate is the sum of the products of the output values ​​of the two parallel models and their current trust weights. The final output confidence level is based on the trust weight of the currently dominant model. This base value is then proportionally adjusted according to the degree to which the harmonic entropy index exceeds the resting state baseline value to accurately reflect the signal quality. Regarding the calibration mechanism for individual model bias, this bias is maintained as a specific numerical compensation within the network. Its update and application follow these steps: When the pulse wave morphology harmonic entropy index is lower than the high stability threshold, and this state has lasted for more than a preset duration, such as 30 seconds, the system recognizes that it has entered an effective calibration window. Within this window, the system compares the original output value of the steady-state blood oxygen estimation model with a benchmark value, which is determined to be the moving average of a series of previous high-confidence output values. The difference generated by the comparison is the estimation error at the current moment. 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. Each time blood oxygen is estimated, the original value output by the steady-state model is first added to this real-time updated individual model bias value, and then participates in the subsequent weighted calculation, thereby achieving continuous compensation for the drift of individual physiological characteristics.

[0050] The classification of physiological transient events relies on joint trajectory analysis and association database matching, which is implemented as follows: This association database is pre-built during the product development phase by analyzing a large amount of clearly labeled physiological event data. For each known event type, the database does not store the original trajectory graph, but a set of quantified feature vectors extracted from the event's harmonic entropy and harmonic attenuation slope change rate. This vector contains at least the peak size and the time difference between the two peaks. When a physiological transient event is detected in 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 involves finding the pre-stored event type with the highest matching degree to the measured feature vector in the association database by calculating a predefined similarity score. Once a match is successful, the system calls 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 stabilizes. Based on the blood oxygen reading during the stable period after the event, the algorithm determines a target estimate. Subsequently, 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 magnitude and direction of this error, a small, iterative adjustment is applied to the factor of the corresponding event type in the database. The adjustment aims to reduce the same error when a similar event occurs in the next time, thereby enabling the compensation strategy to achieve personalized self-optimization.

[0051] Example 5: In another specific embodiment of the present invention, to ensure the universality and high accuracy of the algorithm across different hardware platforms and individual users, a systematic offline calibration and initialization process needs to be executed before deployment, and a precise internal calculation logic needs to be followed during runtime. This process first involves the construction of a physiological transient event association library. By collecting a database covering labeled physiological event types, harmonic entropy time series are extracted synchronously for each event data segment. Time series of the rate of change of harmonic attenuation slope index Calculate a standardized feature vector from these two time series, which contains peak , peak ,as well as peak relative to Peak time difference Subsequently, cluster analysis is performed on the feature vectors of similar events to form a standard fingerprint for each event type. This fingerprint is composed of the mean and covariance matrix of the feature vectors and is stored in the device's association database. When a physiological transient event is detected in real-time monitoring, the system calculates the measured feature vector of this event and obtains a similarity score by calculating its Mahalanobis distance with each standard fingerprint in the database. The matching item with the highest score is the classification result of this event.

[0052] In the initial calibration phase of individualized deployment for specific users, users need to complete a standardized test including resting, specific limb movements, and posture changes as instructed. The core task of the algorithm in this phase is to calibrate key parameters for subsequent calculations. First, based on the data collected in this phase, the high stability threshold and low stability threshold are determined using the statistical methods described in Example 4. Second, the predetermined learning rate used for updating individual model biases is calibrated. During the resting period of the test process, the system treats the continuous output of the steady-state blood oxygen estimation model as a time series and calculates its autocorrelation coefficient and learning rate. The value is set to be inversely proportional to the autocorrelation coefficient to ensure the stability of the calibration process when the user's physiological state is stable, for dynamic compensation factors. Its initial value Calibration is performed during the attitude change phase. The system aims to optimize the smooth transition of blood oxygen readings before and after user attitude changes. It uses a gradient descent optimization method to find a solution that minimizes reading jumps. .

[0053] After completing the above calibration and entering daily monitoring, each operation cycle of the dynamic Bayesian network follows the following steps: the network first receives the pulse wave morphological harmonic entropy index. The raw output values ​​of the steady-state blood oxygen estimation model and the output value of the rate of change blood oxygen estimation model Subsequently, according to Calculate the trust weights of the steady-state model using the calibrated high stability threshold and low stability threshold. The weight varies within the transition range between the two thresholds. The value decreases linearly with the increase of the value. Next, the system applies individual model bias for calibration, and the calibrated output value of the steady-state model is... ,in This is the currently stored individual model bias value. The final output is the estimated blood oxygen level. Synthesized using the following weighted formula: At the same time, the system calculates the final confidence level. The confidence level is based on This is used as the base value, and then proportionally adjusted downwards based on the degree to which the signal quality deviates from the resting baseline to finely reflect the reliability of the results. If the current period is determined to be a valid calibration window, then... If the value remains below the high stability threshold, the network will update the individual model biases. The update logic will then adjust the current... The estimation error is obtained by comparing it with a moving average based on recent high-confidence output values. And update the deviation value accordingly: If the system successfully classifies an event into a specific type of physiological transient based on the aforementioned association database, and detects that the signal has stabilized after the event ends, then the dynamic compensation factor associated with that event type will be activated. Through iterative optimization, the algorithm compares the compensated blood oxygen readings with the readings during the post-event stabilization period, and adjusts the corresponding values ​​stored in the database based on the differences between them. The value is slightly adjusted to better suit the user's specific physiological response pattern, thereby achieving personalized self-optimization of the compensation strategy.

[0054] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A blood oxygenation change monitoring algorithm integrating dynamic Bayesian modeling, characterized in that, Includes the following steps: Step 1: Acquire the raw photoplethysmography (PPG) signal; Step 2: Based on the original signal, identify multiple independent pulse waveforms within multiple consecutive time windows; Step 3: Perform a Fourier transform on each individual pulse waveform to obtain its harmonic energy distribution; Step 4: Based on the harmonic energy distribution, a pulse waveform morphology harmonic entropy index is generated to characterize the stability of the pulse waveform morphology. The pulse waveform morphology harmonic entropy index includes a combination of the harmonic entropy mean and the harmonic entropy variance. Step 5: Construct a dynamic Bayesian network. The internal state of the dynamic Bayesian network includes the final blood oxygen estimate, the current confidence level, and the individual model bias used to calibrate the blood oxygen estimation model. Step 6: Use the pulse wave morphology harmonic entropy index as the input to the dynamic Bayesian network, and use the outputs of two parallel blood oxygen estimation models with different response characteristics as the other two inputs to the dynamic Bayesian network. Step 7: The dynamic Bayesian network dynamically adjusts the trust weights of the blood oxygen estimation model output based on the pulse wave morphology harmonic entropy index. The dynamic Bayesian network performs belief updates based on trust weights to adjust individual model biases and outputs a decision pair containing the final blood oxygen estimate and its corresponding confidence level. The confidence level is jointly determined by the pulse wave morphology harmonic entropy index and the belief update process.

2. The blood oxygenation change monitoring algorithm based on dynamic Bayesian modeling as described in claim 1, characterized in that, The dynamic Bayesian network adjusts the confidence weights of the outputs of two parallel blood oxygen estimation models based on the numerical range of the pulse waveform morphology harmonic entropy index: when the pulse waveform morphology harmonic entropy index indicates high pulse waveform morphology stability, the dynamic Bayesian network assigns higher confidence weights to the steady-state blood oxygen estimation model and uses the current data with high pulse waveform morphology stability to iteratively adjust the individual model bias based on a predetermined learning rate; when the pulse waveform morphology harmonic entropy index indicates low pulse waveform morphology stability, the dynamic Bayesian network assigns higher confidence weights to the rate-of-change blood oxygen estimation model and reduces the confidence level of the final output to a predetermined lower limit according to a predetermined algorithm.

3. The blood oxygenation change monitoring algorithm based on dynamic Bayesian modeling as described in claim 2, characterized in that, The steady-state blood oxygen estimation model is a linear regression model trained on large-scale statistical data; the rate of change blood oxygen estimation model is a linear estimation model based on the rate of change of the ratio of AC to DC components of the photoplethysmography pulse wave signal. The rate of change blood oxygen estimation model is not sensitive to the morphology of the pulse waveform.

4. The blood oxygenation change monitoring algorithm based on dynamic Bayesian modeling as described in claim 1, characterized in that, The method also includes the following steps: based on harmonic analysis of independent pulse waveforms, extracting the energy ratio of higher harmonic components to the fundamental component in the harmonic energy distribution to obtain a harmonic attenuation slope index characterizing the energy spectrum morphology of the pulse waveform; real-time monitoring of the rate of change of the harmonic attenuation slope index; and when the absolute value of the rate of change exceeds a threshold 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 estimate based on the rate of change, the correction being: correction = in, It is a dynamic compensation factor. It is the rate of change of the harmonic attenuation slope index.

5. The blood oxygenation change monitoring algorithm based on dynamic Bayesian modeling as described in claim 4, characterized in that, Dynamic compensation factor The dynamic Bayesian network learns and updates data after the physiological transient event ends, based on the pulse wave morphology harmonic entropy index after the data has recovered stability.

6. The blood oxygenation change monitoring algorithm based on dynamic Bayesian modeling as described in claim 4, characterized in that, The method also includes the following steps: acquiring the joint change trajectory of the harmonic attenuation slope index and the pulse wave morphological harmonic entropy index during physiological transient events in parallel; classifying physiological transient events based on the morphological characteristics of the joint change trajectory, including hemodynamic transient type and 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 based on the type classification result.

7. The blood oxygenation change monitoring algorithm based on dynamic Bayesian modeling as described in claim 1, characterized in that, The method also includes the following steps: based on harmonic analysis of independent pulse waveforms, extracting the phase information of the second harmonic in the harmonics; calculating the variance of the phase information within a predetermined time window to obtain a harmonic phase jitter index characterizing the rhythmic phase-locked state of the original photoplethysmography (PPG) wave signal; and based on the harmonic phase jitter index, determining the source reliability of the original PPG wave signal to obtain a source confidence index.

8. The blood oxygenation change monitoring algorithm based on dynamic Bayesian modeling as described in claim 7, characterized in that, When the source confidence index value is higher than a predetermined threshold, the dynamic Bayesian network continues to perform blood oxygen estimation and confidence output; when the source confidence index value is not higher than the predetermined threshold, the method outputs a diagnostic message indicating abnormal device status, rather than a decision pair containing blood oxygen estimation value and confidence level.

9. The blood oxygenation change monitoring algorithm based on dynamic Bayesian modeling as described in claim 1, characterized in that, The method triggers a real-time alarm when the final blood oxygen estimate is below a physiological safety threshold and its corresponding confidence level is above a credibility threshold. The alarm is issued through the user's wearable device or a remote health monitoring platform.

10. The blood oxygenation change monitoring algorithm based on dynamic Bayesian modeling as described in claim 1, characterized in that, The update of the individual model bias parameters in the dynamic Bayesian network is based on the historical value of blood oxygen estimation when the pulse wave morphology harmonic entropy index indicates high signal stability and the corresponding photoplethysmography pulse wave characteristics. The update adopts an algorithm that minimizes the historical error to ensure imperceptible adaptive calibration to long-term physiological changes of individual users.

Images