Hypoperfusion oxyhemoglobin saturation detection method based on dynamic light compensation

Abstract

The invention relates to a hypoperfusion oxyhemoglobin saturation detection method based on dynamic light compensation, which comprises the following steps of: constructing a combined characteristic relation based on multi-wavelength reflection intensity and fluctuation amplitude, and dynamically judging the optical sparseness of tissues, namely optical characterization of erythrocyte density in the tissues; reversely deducing the current perfusion degree through the morphological characteristics of the reflection waveform; establishing a perfusion optical index; analyzing the dynamic change trend of light reflection in a past period of time by adopting a time sequence modeling mode; an LED driving strategy is dynamically adjusted, and a light supplementing behavior similar to a physiological rhythm is formed in a step-by-step climbing mode, a pulse stimulation mode and a period maintaining mode; each light supplementing action is regarded as physiological stimulation; dynamic form similarity discrimination is introduced, whether tissue response is really caused by light supplement each time is quantified, and real physiological fluctuation and noise disturbance are distinguished; for low-response light supplement, the light supplement intensity is reduced in a feedback manner; on the contrary, key deformation features are extracted from high-response light supplement for signal reconstruction, and the signal recognition capability in the low-perfusion environment can be enhanced.

Description

Technical Field

[0001] The present invention relates to a method for detecting low perfusion blood oxygen saturation, in particular to a method for detecting low perfusion blood oxygen saturation based on dynamic light compensation. Background Art

[0002] Although Chinese patent CN103027690A proposes a low-perfusion blood oxygen saturation measurement method based on autocorrelation modeling, which improves the measurement accuracy in weak signal environments to a certain extent and has the advantages of signal enhancement and noise suppression, its method still has many structural deficiencies and fundamental drawbacks from multiple dimensions such as actual application scenarios, technical logic depth, physiological response credibility, system adaptability, and energy efficiency control. In particular, when faced with dynamic perfusion states, complex physiological fluctuations, and individualized responses, its inherent limitations gradually become apparent, making it difficult to meet the current development needs of smart wearables and continuous health monitoring.

[0003] First, this patented method is essentially still a static modeling + signal post-processing path, lacking a multi-dimensional signal quality control and correction framework centered on pattern recognition, pattern classification, and pattern clustering. Its key technical route is based on traditional filtering and autocorrelation enhancement of PPG signals, estimating the AC component through modeling parameters to obtain blood oxygen estimation. However, the entire process does not include dynamic adjustment and control of the input light source, nor does it lack real-time signal pattern recognition to determine waveform credibility. In other words, it cannot actively increase the light flux to improve reflection quality when the signal is weak, and can only rely on the characteristic enhancement of the signal itself. Therefore, in extremely low perfusion conditions, when the original signal is close to the noise floor, the waveform amplitude is extremely low, or the autocorrelation is already poor, the modeling parameters of this method will easily drift or be misjudged, leading to serious deviations in the estimation. Secondly, the autocorrelation method relies on the stability of the periodic structure of the signal for processing, which is effective under certain perfusion conditions. However, under low perfusion, the cardiac waveform often exhibits atypical morphological changes, such as irregular shoulder peaks, platform waveforms, and disordered micro-motion structures. At this time, without a signal pattern classification and object recognition mechanism, the autocorrelation function is prone to pseudo-peaks or weak period matching, affecting the accurate extraction of the AC component and the interpretation of its physiological significance. In addition, this method does not involve structural analysis of the waveform morphology, nor does it use object recognition or pattern clustering based on physiological priors to determine tissue reflection characteristics. It only enhances the signal from a statistical perspective and lacks a deep understanding of physiological mechanisms and response judgment. Especially in the peripheral tissue area where the capillary system dominates the reflection, the microscopic morphological characteristics of the waveform are far more diagnostically valuable than the overall energy intensity. However, this patent does not mention this at all, and is a typical signal-driven method rather than a physiological response-driven method. Third, the existing technology lacks a dynamic feedback control mechanism.

[0004] Modern blood oxygen monitoring systems, especially in wearable environments, increasingly rely on a closed-loop feedback system between light output and tissue response. This means the system should be able to distinguish valid physiological waveforms from artifactual noise in real time through signal pattern recognition and object identification, and automatically adjust light source parameters such as power, wavelength, and duty cycle based on the tissue's current physiological state to achieve precise fill light and optimize energy consumption. However, the light source parameters in this patent are fixed settings and cannot adapt to changes in conditions such as skin thickness, vascular distribution, and ambient brightness. Nor can they distinguish whether the current signal boost is truly due to increased blood flow or simply accidental noise interference. Fourthly, this method's processing is highly dependent on the stability of the modeling parameters. In actual wearable scenarios, factors such as movement, temperature, or neural tension can cause dramatic fluctuations in the reflected signal at any time. Without a quality stratification and screening mechanism based on pattern clustering, modeling all data together will lead to miscalculation of blood oxygen levels in high-noise segments. Fifth, the solution does not propose any form of signal weighting or data credibility scoring system, lacks intelligent assessment and weight allocation of multi-segment waveform quality, and ignores the core issue of significant differences in waveform quality under low perfusion conditions. In modern personalized health monitoring scenarios, especially for people with chronic diseases, their signal states are often changing. Without the support of signal pattern recognition, pattern classification, and pattern clustering, it is impossible to give priority to the use of high-quality waveforms and eliminate low-quality waveforms, which ultimately greatly limits the feasibility and practical value of this technology in clinical and consumer applications. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for detecting low perfusion blood oxygen saturation based on dynamic light compensation, thereby solving some of the drawbacks and deficiencies pointed out in the background art.

[0006] The present invention solves the above-mentioned technical problems by adopting the following technical solutions: a method for detecting low perfusion blood oxygen saturation based on dynamic optical compensation, comprising: constructing a combined characteristic relationship between multi-wavelength reflection intensity and fluctuation amplitude to dynamically determine tissue optical rarefaction, i.e., an optical representation of red blood cell density in tissue; inferring the current perfusion level through the morphological characteristics of the reflection waveform, including pulse edge slope, peak-to-valley ratio, and micro-fluctuation period, to establish a perfusion optical index;

[0007] Time series modeling, including gated recurrent units (GRUs) or sliding window variable-length exponential filtering, is used to analyze the dynamic trends of light reflection over time. The LED drive strategy is dynamically adjusted to create a physiological rhythm-like lighting behavior using three modes: step-by-step ramping, pulse stimulation, and periodic maintenance.

[0008] Each fill light action is regarded as a physiological stimulus. The response differences in amplitude, phase, and morphology of the capacitance gram waveforms of the three light segments are calculated in the time windows before, during, and after the onset of the fill light. Dynamic morphological similarity judgment is introduced to quantify whether each fill light action truly causes a tissue response and distinguish real physiological fluctuations from noise disturbances. For low-response fill light, the fill light intensity is reduced in feedback. Conversely, for high-response fill light, key deformation features are extracted for signal reconstruction.

[0009] Furthermore, the method for establishing the perfusion optical index includes:

[0010] S1, collecting multi-wavelength light reflection signals from the target detection part and obtaining the pulse photoplethysmography waveform;

[0011] S2. Extracting morphological features of the reflected waveform, including:

[0012] a) The slope of the rising edge of the pulse,

[0013] b) the ratio of the pulse peak to the trough amplitude,

[0014] c) Repeatability and similarity of micro-fluctuation cycles;

[0015] S3. constructing an optical perfusion state index based on the waveform morphology characteristics to determine whether the current tissue is in a low perfusion state; when the perfusion state index is lower than a preset threshold, triggering a dynamic light compensation control module to adjust the incident light power;

[0016] S4. During the light compensation process, continuously monitor the waveform morphology changes, determine the light compensation effect, and adjust the light output accordingly; calculate the blood oxygen saturation estimate based on the reflection signal optimized by light compensation.

[0017] Furthermore, the extraction of the pulse edge slope includes analyzing the time and amplitude change rate of the rising edge of the photoplethysmography waveform to determine the blood propulsion speed and tissue light response speed.

[0018] Furthermore, the amplitude ratio of the peak value to the valley value is used to represent the amplitude of tissue volume change during the cardiac cycle. The smaller the amplitude, the weaker the perfusion. The micro-fluctuation period is a secondary fluctuation outside the main pulse waveform, which is used to judge the capillary return or vascular elasticity state in the tissue.

[0019] Furthermore, the optical perfusion status index is a comprehensive score calculated based on a weighted calculation of multiple waveform features, which is used to reflect the real-time tissue perfusion quality; the dynamic light compensation control module includes a light source driver and a feedback controller, which is used to dynamically adjust the power and pulse rhythm of the LED light source according to the optical perfusion status index;

[0020] Current technologies rely heavily on the stability of tissue reflection signals. Especially in low perfusion states, the photoplethysmography (PPE) fluctuations caused by tissue blood flow are extremely weak, causing the signal to approach the noise floor, making it difficult to establish a valid ratio and easily leading to biased PEO estimates. Furthermore, traditional dynamic light compensation methods are typically static threshold-triggered, performing linear light compensation based solely on PPE intensity. This lacks physiological response relevance and cannot accurately identify whether hemodynamic feedback is truly occurring. A dynamic comprehensive scoring function is constructed based on three key timing characteristics of the PPE waveform: rising edge slope S(t), waveform amplitude ratio V(t), and micro-fluctuation period stability Δ(t) to characterize the tissue's real-time perfusion state responsiveness.

[0021] The comprehensive optical perfusion index function is defined as follows:

[0022]

[0023] in:

[0024] Π(t): represents the optical perfusion status index score at time point t; ε: historical time variable within the integration window; τ: integration review window length, representing the memory length of the perfusion response; S(ξ): waveform rising slope, reflecting the heartbeat conduction efficiency (waveform rising rate per unit time); V(ξ): waveform peak-to-valley amplitude ratio, indicating the volume blood flow change capability caused by a unit heartbeat; Δ(E): waveform microcycle variability, reflecting the stability of the local blood flow fine-tuning structure; the smaller the value, the more regular the fluctuation; Λ: normalization constant, used to limit the function result to the [0,1] interval to improve the algorithm stability; log(1+V 2 Δ): nonlinear coupling term used to emphasize the enhanced contribution of high-amplitude waveforms to the perfusion state in the presence of regular microfluctuations; e -Δ : The more stable the micro-fluctuation (the smaller the Δ), the larger this item is, indicating a higher perfusion quality;

[0025] When the scoring function Π(t) determines that the perfusion state is lower than the threshold (e.g., Π(t) < , where θ is the empirically set threshold), the system starts the light compensation control process;

[0026] The light compensation control module includes:

[0027] Light source driver: used to adjust the power and current pulse width of the LED light-emitting unit;

[0028] Feedback controller: continuously receives Π(t) and combines its time trend Determine whether the tissue responds substantially to fill light;

[0029] The controller dynamically adjusts according to the following strategies:

[0030] If Π(t) is on a downward trend, continue to increase the pulse intensity;

[0031] If there is no significant change in the Π(t) fluctuation, maintain the current rhythm;

[0032] If Π(t) increases significantly, reduce the light intensity to save energy.

[0033] Furthermore, the blood oxygen saturation calculation process introduces a response weight based on the waveform morphology changes before and after light compensation, which is used to correct the estimation deviation of the traditional ratio method.

[0034] Furthermore, the method of respectively calculating the response differences in amplitude, phase, and morphology of the capacitance waveforms of the three light segments includes:

[0035] Dynamic light compensation signals are sent to the test tissue at a preset rhythm; each light compensation behavior is divided into three time windows: before, during, and after light compensation; and multi-dimensional features of amplitude, phase, and morphology are extracted from the photoplethysmographic waveforms collected in each time window.

[0036] The characteristics of the three waveforms are compared to determine the differences in the physiological responses of tissues to the supplemental lighting behavior; the waveforms within the supplemental lighting period with physiological response characteristics are used for blood oxygen saturation estimation, and the remaining waveforms are downgraded or discarded.

[0037] Furthermore, the amplitude feature includes the amplitude change between the peak and valley values ​​of the pulse waveform, which is used to determine whether the fill light causes enhanced tissue reflection; the phase feature includes the advance or delay of the pulse wave start time relative to the reference point of the previous cycle, which is used to evaluate the reaction speed of the tissue to the fill light.

[0038] Furthermore, the morphological features include changes in the pulse waveform contour, the slope of the rising edge, and the appearance and disappearance of micro-fluctuation structures, which are used to determine whether there are changes in tissue microcirculation. If the waveform morphology in the three time windows is significantly different, it is determined to be an effective physiological response to fill light. Otherwise, it is determined to be ineffective fill light, and the fill light intensity or period is reduced.

[0039] The present invention is used to dynamically adjust the LED light compensation strategy to achieve synchronous control with the actual microcirculatory response. This method accurately identifies whether the tissue has produced a physiological response to the light compensation behavior by dividing the time period window, extracting morphological features, and constructing an integral difference function. Each time the LED dynamic light compensation behavior occurs, the system divides the photoplethysmographic waveform reflection waveform into:

[0040] Fill light front window: represents the microcirculation output in the natural state;

[0041] Fill light window: observe the instant changes of waveform under fill light stimulation;

[0042] Post-fill light window: observe the morphological recovery trend after the fill light is removed;

[0043] In each window, the system extracts the following three types of morphological features:

[0044] Pulse wave contour characteristics, including waveform symmetry, shoulder formation, and length of the steep rise period;

[0045] The rising edge slope reflects the speed of blood flowing into tissues;

[0046] The appearance, disappearance and periodic changes of micro-wave structures indicate whether the capillary system is active;

[0047] The following original integral function is constructed to quantify the differences in the three waveforms, thereby determining whether the fill light causes physiological changes in tissue microcirculation:

[0048]

[0049] in:

[0050] ΔΨ: defined waveform response morphology difference index, used to measure the significance of physiological response; t1, t3: respectively represent the start time of the window before fill light and the end time of the window after fill light; T: any time point in the integration interval; χ s (τ): The composite morphological function of the waveform in the fill light window, which includes the fusion of the contour factor, slope factor, and micro-fluctuation factor; χ b (τ), χ r (τ): composite waveform morphology value at corresponding time points before and after fill light; Ω(τ): physiological credibility weighting function, used to amplify rhythmic responses and suppress the influence of noise disturbances; |·|: represents the absolute morphology offset between the three segments, indicating the net change of the waveform caused by fill light;

[0051] This function captures the changes in the waveform structure of the tissue during the fill light period through time domain integration, and uses the average waveform before and after the fill light as a reference to evaluate whether the middle section is significantly deformed. When ΔA exceeds the system's preset response threshold ε, the fill light behavior is judged to be an effective physiological response fill light; otherwise, it is considered to be invalid fill light or optical over-fill light, and the system will reduce the LED intensity in the next fill light cycle or shorten the fill light duration.

[0052] Furthermore, the response differences of the three waveforms are used to construct a response credibility score, which is used to weight different waveforms in the calculation of blood oxygen saturation.

[0053] Beneficial effects of the present invention: Under conditions of insufficient tissue blood perfusion (such as peripheral circulatory disorders, low temperature environments, and chronic diseases), the amplitude of traditional PPG signals is extremely weak and the signal-to-noise ratio is low, which can easily lead to fluctuations or misjudgments in blood oxygen estimation. The present invention determines the tissue perfusion status in real time and adaptively controls the LED light source intensity and pulse rhythm to achieve dynamic light compensation, thereby significantly enhancing the recognition ability of weak reflection signals and ensuring that usable waveforms can still be stably acquired under low perfusion conditions. Unlike traditional passive light filling methods that use static thresholds or fixed light power, a real-time response evaluation model based on multi-dimensional features such as waveform morphology, amplitude, and phase is introduced. By performing differential analysis on the three waveforms before, during, and after the light filling, it is determined whether the tissue has produced a physiological response. The light filling behavior only continues after being identified as an "effective response", and has the ability to perform physiological feedback closed-loop control, significantly improving the intelligence and energy-saving level of the system.

[0054] The reliability of each waveform segment is dynamically assessed and its weight in the blood oxygen estimation process is adjusted accordingly, effectively preventing interference from invalid waveforms. This waveform quality-based weighting strategy can significantly improve the stability and accuracy of blood oxygen estimation in situations with severe perfusion fluctuations and significant differences in waveform quality. This method not only evaluates the effectiveness of a single illumination session but also integrates the temporal trend of the tissue's response to continuous illumination, creating a physiological illumination rhythm with gradual increases, periodic maintenance, or decreases. This ensures that the LED output more closely matches the tissue's true state, preventing interference or discomfort caused by excessive illumination. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a functional diagram of dynamic light compensation for low perfusion blood oxygen detection according to the present invention.

[0056] Figure 2 This is a flow chart of the dynamic light compensation control based on time series analysis of the present invention.

[0057] Figure 3 This is a diagram showing the relationship between the intelligent fill light function based on three-segment waveform analysis of the present invention.

[0058] Figure 4 This is a schematic diagram of the operation of the home blood oxygen monitoring and intelligent light compensation system for Mr. Wang in a low perfusion state in Example 1.

[0059] Figure 5 This is a structural diagram of Mr. Wang's layered dynamic light-filled blood oxygen monitoring system in Example 2. DETAILED DESCRIPTION

[0060] The following is a detailed description of the specific embodiments of the present invention with reference to the accompanying drawings.

[0061] Combined with attachment Figure 1This method, based on dynamic optical compensation, uses the physical and morphological characteristics of the tissue-reflected light waveform to infer the local tissue blood perfusion state, guiding dynamic optical compensation. This method accurately estimates blood oxygen saturation even under extremely low perfusion conditions. This method does not rely on external indicators such as pressure or skin temperature, but instead builds on multi-wavelength optical reflection characteristics and PPG waveform details to estimate tissue optical sparsity through a fusion data model, indirectly characterizing the dynamic changes in red blood cell density distribution, i.e., tissue perfusion level. The tissue's optical response is converted into a perfusion state perception indicator, which in turn drives optical compensation behavior, forming a physiologically meaningful closed-loop control system. Data acquisition is based on a three-wavelength optical detection system, typically consisting of three LED channels: red (660nm), near-infrared (880-940nm), and a reference wavelength (e.g., 530nm or 590nm). A photodiode (PD) simultaneously collects the reflection intensity of each wavelength on the target tissue (e.g., a fingertip). The device continuously collects the reflected light intensity of each wavelength at a high sampling rate (e.g., 100 Hz to 250 Hz) and records the PPG waveform with time stamp alignment.

[0062] The data preprocessing stage begins with noise suppression and baseline drift correction, primarily using wavelet transforms and adaptive filtering to remove motion artifacts and ambient light interference. Subsequently, three core morphological features are extracted within each sampling cycle: the pulse edge slope, or the slope of the rising segment of the PPG waveform, represents the dynamic characteristics of blood entering peripheral tissues; the peak-to-valley ratio, or AC / DC amplitude ratio, reflects the magnitude of local volume changes within each cardiac cycle; and the microfluctuation periodicity, which reflects the microrhythm of capillary blood flow in tissues. Its variations indicate whether the local perfusion system maintains stable microcirculatory activity. All features are normalized and a sliding window is constructed along the time dimension to form a structured input vector. The model utilizes a lightweight neural network architecture to support efficient operation on embedded platforms such as MCUs. The model input is a structured feature sequence within a time series window (e.g., the past 5 seconds or 10 cardiac cycles), including the aforementioned multi-wavelength reflectance intensity and morphological features. The output is a scalar, the perfusion optical score index, ranging from 0 to 1, representing the optical sparsity of the tissue at that point. Lower values ​​indicate a sparser blood signal in the reflected light and weaker perfusion. The model architecture utilizes a three-layer time-series-aware network (e.g., GRU or simplified LSTM) stacked with a multi-layer perceptron (MLP) output layer. Model training utilizes a dataset with real perfusion labels. These labels are obtained from standard clinical perfusion instruments or PI sensors and are paired with the acquired reflectance waveforms to construct supervised learning samples. The dataset includes a variety of perfusion scenarios under different temperatures, pressures, and pathological conditions to enhance the model's generalization capabilities. The training loss function is a mean squared error plus a regularization term. A weighting mechanism is introduced to enhance sensitivity for detecting low perfusion states. Iterative training is performed using the Adam optimizer. The final model can output a perfusion score per second in practical applications, which is used to determine in real time whether the light compensation state should be entered, and to control parameters such as LED drive current and pulse width to dynamically adjust the luminous intensity.

[0063] Combined with attachment Figure 2As shown in the figure, by introducing a time series modeling mechanism, this approach addresses the issue of delayed and inaccurate light compensation in traditional blood oxygen monitoring devices during hypoperfusion, caused by weak or unstable light reflection signal fluctuations. Using time series analysis methods such as gated recurrent units (GRUs) or sliding window variable-length exponential filtering, the dynamic trends of the light reflection signal over a period of time are modeled. This allows for real-time prediction of the tissue's potential response to light compensation. Based on this prediction, the LED light source's driving strategy is dynamically adjusted to achieve light compensation behavior that mimics a physiological rhythm. This ensures that physiologically meaningful PPG waveforms and accurate blood oxygen estimation can be continuously acquired even in low perfusion conditions. Data acquisition relies on a multi-channel optical detection system consisting of LEDs for red, near-infrared, and a reference wavelength channel, and a photodetector (PD). The light source and detector are installed in the wearable device and continuously record the reflected light intensity of each channel at a high sampling frequency (e.g., 200Hz). Simultaneously, multi-dimensional signals such as the dynamic LED light compensation control signal (actual luminous power), ambient light intensity, and cardiac rhythm timestamps are collected to establish a dynamic correspondence with the reflected light waveform. During the data preprocessing phase, signal denoising is first performed, using adaptive bandpass filtering and wavelet packet decomposition to remove motion artifacts and baseline drift. A time series window is then constructed for each sampling point (e.g., approximately 1,000 sampling points within the past 5 seconds). Structured data extraction is performed on the rate of change of light reflection intensity, local extreme value density, and short-term average amplitude change within the window. Previously output LED light intensity and control instructions are then added as covariates to form a time series input tensor. A sliding window approach is used to construct a time-sliding input stream, which is fed into the time series model to learn light reflection trends. One implementation utilizes a gated recurrent unit (GRU) neural network, which has superior memory capacity and lower computational complexity than traditional RNNs, making it suitable for embedded platforms. The GRU input is a T×F-dimensional tensor, where T is the time window step size and F is the feature dimension of each time point. The output is a high-order state estimate of the light reflection trend at the current moment. Another implementation involves an adaptive trend-tracking model based on variable-length exponential filtering. This model dynamically adjusts the exponential decay factor based on the historical rate of change at each time point, achieving a weighted response to local fluctuations. This approach is suitable for modeling tissue responses sensitive to frequency variations. Regardless of the model's output, an intermediate quantity describing the trend of tissue light reflectance changes, called the dynamic reflectance trend response score, is ultimately used to guide subsequent lighting adjustments.In terms of light compensation control, this response score is used to drive the LED output strategy to dynamically switch between three states: When the score indicates a negative change in tissue reflectance (i.e., the light reflectance signal continues to decrease or fluctuations disappear), the system enters a step-by-step ramp-up mode, increasing the LED drive current in a gentle, linear increment. When the score exhibits short, intense fluctuations without a periodic pattern, the system determines that the tissue is in a delayed or compressed phase and switches to a pulse stimulation mode, using short, intense light stimulation with a low duty cycle and high amplitude to attempt to elicit a tissue response. When the score stabilizes and the light reflectance morphology exhibits a rhythmic output, the system enters a cycle maintenance mode, outputting light compensation at a constant frequency and power to maintain a balance between energy efficiency and tissue adaptability. This three-stage light compensation mechanism can be viewed as a photophysiological response that mimics the rhythm of real microcirculation, enabling light compensation and tissue blood flow to form a feedback-like synergistic mechanism.

[0064] Combined with attachment Figure 3As shown, each LED lighting event is treated as a controllable low-light physiological stimulus to observe whether the tissue produces a substantial hemodynamic response under the light stimulation. This allows for the determination of physiological significance and potential use in signal enhancement or blood oxygen estimation. Starting with signal acquisition, a three-stage waveform analysis mechanism is constructed: Each lighting event is separated into a baseline window before the lighting event begins, a stimulation window during the lighting event, and a recovery window after the lighting event ends. Within each window, three key features of the tissue reflectance waveform are collected and calculated: amplitude change, phase change, and morphological structure change. This is used to quantify whether the lighting event induces true tissue blood flow changes, rather than merely physical light enhancement or system noise perturbations. Data acquisition is based on a PPG (photoplethysmography) system, which simultaneously collects reflected light signals in red, near-infrared, and a reference wavelength channel at a sampling frequency of 125-250Hz. Each lighting event is time-stamped on the wearable device, marking the data ranges before, during, and after the lighting event. When extracting features in each window, the waveform is first pre-processed for denoising, including baseline correction based on sliding minimum filtering, wavelet packet decomposition to remove artifacts, and the use of bandpass filters to extract the effective components of the main pulse frequency band (about 0.5-5Hz). Subsequently, three types of waveform indicators are extracted in the three windows: (1) Amplitude features are the mean and standard deviation of the pulse amplitude in the current window, which measures the strength of blood flow fluctuations; (2) Phase features are the time offset of the current cycle waveform relative to the reference point of the previous pulse, reflecting the delay or advance of local blood flow; (3) Morphological features include secondary morphological factors such as rising edge slope, shoulder ratio, waveform width, and bimodal structure, which are used to determine whether blood flow changes actually occur at the microstructural level. In order to systematically quantify the differences between the three waveforms and make validity judgments, this method introduces a model module called dynamic morphological similarity discriminator, which determines whether the fill light has caused meaningful physiological changes by calculating the distance between the morphological features of the stimulus window and the previous and next windows. Its core is to construct a similarity scoring function, and denote the feature vectors of the three windows as F b , F s ,F r , by defining the morphological difference vector Δ s =F s -(F b +F r ) / 2, and then uses a nonlinear mapping function to map the difference into a normalized score, Ψ. A higher score indicates a true response in the waveform morphology. This function can be implemented as a simple feedforward neural network or heuristic rule logic. Its training data comes from multiple real-world fill light events. The supervised learning samples are then manually or semi-annotated to determine whether the fill light was effective, i.e., whether a physiological response occurred.

[0065] During the model training process, the input is the waveform feature vector extracted within three time windows, and the label is whether the fill light causes a valid waveform change (1 or 0). The network structure adopts a two-layer fully connected neural network, and the activation function adopts Swish or Tanh to enhance the perception of small-scale nonlinear changes. The loss function uses weighted cross entropy to enhance the accuracy of the model in identifying low-response events, and uses the Adam optimizer for iteration. The training dataset contains fill light response behaviors under various environments, fittings, and perfusion states, and evaluates the accuracy, recall rate, and AUC curve in the test set to ensure that the model is still real-time and accurate after deployment on the edge device. During runtime, the DMSE model calculates the response score Ψ immediately after each fill light. If the score is lower than the set threshold θ l , it is judged as low response fill light, the system will feedback reduce the LED light intensity, shorten the duration, and enter energy saving mode; if the score is higher than the set threshold θ h , it is determined to be a high-response fill light. The system will record the key waveform changes caused by this fill light, such as the steep increase in slope and microstructure enhancement, as the target structural features of the subsequent signal enhancement module, promoting the physiological reconstruction of the PPG signal at the structural level, and restoring the waveform input with physiological rhythm even when the perfusion is insufficient.

[0066] Example 1:

[0067] Combined with attachment Figure 4 As shown, the user is Mr. Wang, a 65-year-old diabetic patient. Due to long-term impaired peripheral microcirculatory function, he experiences typical hypoperfusion symptoms in winter, such as cold fingertips and hypoperfusion. He uses the technology based on the present invention at home. The system executes step S1, which collects multi-wavelength reflection signals from the target detection site (the tip of the left index finger). The LED in the device emits pulsed light of 660nm (red light), 880nm (near-infrared light), and 530nm (reference green light), each wavelength alternating at a frequency of 50Hz. Simultaneously, a coaxial photodetector collects the reflection intensity signal at a sampling frequency of 250Hz. The signal is recorded continuously for 20 seconds to form the initial data segment, and a PPG waveform is extracted. Preliminary data shows that the reflection intensity in the near-infrared band fluctuates within ±0.15V, the red light fluctuates within ±0.10V, and the green light exhibits a nearly flat waveform with fluctuations not exceeding ±0.03V, indicating the possibility of initial hypoperfusion.

[0068] In step S2, the system enters the waveform morphology feature extraction stage. First, the edge slope of the rising edge of the pulse waveform is analyzed, and the average rising slope of the 3rd, 4th, and 5th cardiac cycles in the near-infrared band is measured to be 0.012V / ms, which is significantly slower than the normal value (such as >0.025V / ms); secondly, the peak-to-valley amplitude ratio of each cardiac cycle is analyzed, and it is found that the average amplitude is 0.09V, while the baseline DC value is 0.95V. The AC / DC ratio obtained is about 0.095, which is lower than the conventional low perfusion warning threshold of 0.12; finally, the similarity of the micro-fluctuation cycle is analyzed. The system determines that the fine fluctuation cycle of the PPG waveform has been structurally unstable in the past 5 cardiac cycles through wavelet packet decomposition and zero crossing density analysis, and the average cycle variance is 0.14s 2 , indicating an irregular, fluctuating state. After aggregating the three indicators, the system constructed a feature vector [0.012, 0.095, 0.14] and fed it into a trained perfusion state scoring model (a two-layer feedforward neural network trained on 20,000 data samples). The model outputted a perfusion optical index value of 0.36, far below the normal perfusion threshold of 0.70, indicating severe hypoperfusion.

[0069] In step S3, the system triggers the dynamic light compensation module. After reading the perfusion index value, the light source controller linearly increases the current drive of the LED red and near-infrared channels from the original 8mA to 16mA. It uses a dual-strategy output strategy of step-by-step ramping and pulse intervention, increasing the ramp by 1mA per second for 3 seconds. Subsequently, a short pulse with an intensity of 22mA, a duty cycle of 30%, and a duration of 600ms is introduced to induce a tissue light response. During this process, the system continues to execute S4, monitoring the changes in the waveform morphology during the fill light period in real time and determining the fill light effect.

[0070] According to the real-time collected data, the system found that in the 6th to 8th cardiac cycles, the rising slope of the near-infrared waveform increased to 0.026V / ms, returning to normal levels, the AC / DC ratio increased to 0.13, the similarity of the micro-fluctuation cycle was significantly improved, and the cycle variance decreased to 0.06s. 2 These changes in morphological features indicate that the tissue has produced a clear physiological response to the supplemental lighting behavior. The system inputs the before-and-after morphological differences into the morphological similarity discriminator DMSE module, which calculates a dynamic response index of 0.82, far exceeding the preset threshold of 0.60, confirming high-responsive supplemental lighting. Based on this, the system decides to maintain the current LED output level and use the current waveform for blood oxygen calculation. The blood oxygen estimation model uses the ratio method plus a correction coefficient to calculate the R value. The AC / DC ratio of the three cycles after the supplemental lighting is taken for R calculation, and the corrected R value is 0.41. After substituting it into the calibration curve regression model, the output SpO2 is 93%, which is consistent with the patient's actual historical blood oxygen data. At the same time, the system indicates that perfusion has recovered and has entered stable supplemental lighting mode.

[0071] The pulse edge slope directly reflects the speed and intensity of tissue response to light after blood is pushed to the detection site by the heartbeat within the capillaries during each cardiac cycle. Therefore, it is used by the present invention as an important signal for determining microcirculatory efficiency. In Mr. Wang's device, the slope extraction process begins with the acquisition of the original PPG waveform. This waveform is illuminated by light emitted by a multi-wavelength LED light source (660nm, 880nm, 530nm), and the reflected signal is collected by a photodiode to form a continuous waveform. The system records the reflection intensity at a sampling rate of 250Hz, forming an independent time series on each optical channel. The original waveform is first processed by noise suppression filtering, including a bandpass filter (0.5-8Hz) to remove motion and baseline drift interference, and sharp mutations are smoothed by wavelet transform or local weighted regression filtering to make the extracted rising edge more physiologically meaningful. The system then executes the pulse recognition logic, locking the starting point and peak position of each cardiac cycle through zero crossing point and derivative extreme value method. For each detected pulse cycle, the system extracts the rising segment from the trough to the peak as the analysis interval. This waveform segment usually exhibits a nonlinear rising shape. At this time, the system does not use the traditional linear fitting method, but instead divides the rising segment into equally spaced time windows (such as once every 4ms) and calculates the instantaneous slope value within each window. That is, the reflection intensity change ΔI is divided by the time increment Δt to obtain the growth rate of the waveform per unit time. To adapt to the situation where the waveform amplitude is small and the change is slow under low perfusion, the present invention introduces a weighted time-amplitude response rate algorithm (WTVR): within each micro-window, the current slope value is compared with the average value of its previous window to form an acceleration factor. If the acceleration value is continuously positive, it means that the blood propulsion speed is still increasing and the tissue is gradually responding to the supplemental light behavior; on the contrary, if the acceleration is negative or close to zero, it can be preliminarily judged that the blood propulsion is limited or the response is weak. The system detected that in the fifth cardiac cycle, the duration of the rising segment was 96ms, the maximum instantaneous slope was 0.013V / ms, and the average slope was 0.009V / ms. Compared with its personal benchmark model (the average value is 0.021V / ms under normal conditions), the system judged that the edge slope was slowed, indicating that the blood flow to local tissues was slow. This slope information was input into the optical perfusion model as part of the three-dimensional feature vector, and was compared with the amplitude ratio (such as 0.095) and micro-fluctuation period similarity (such as the period variance of 0.14s) extracted at the same time. 2) together constitute the perfusion status assessment vector. The perfusion recognition model structure is a type of shallow neural network, consisting of a three-layer fully connected network: the input layer is a 3D waveform feature vector, the first hidden layer contains 16 nodes, the activation function is ReLU, the second layer is 8 nodes and the Dropout mechanism is introduced to enhance generalization ability, the output layer is a single-node Sigmoid mapping, and the output is a normalized perfusion score. The training data is collected from real clinical settings and contains more than 20,000 PPG signal segments under normal and low perfusion states. Each segment of data is labeled with its true perfusion status (based on the perfusion index PI meter or manual expert interpretation). The model is trained by supervised learning, and the loss function uses weighted binary cross entropy. The Adam optimizer is used for 100 rounds of iterative training to converge to the optimal parameters. The final deployed model can run on the MCU platform with a delay of less than 30ms. Therefore, during Mr. Wang's wearing process, when the system judged that the slope continued to decline based on edge slope extraction, and the vector output perfusion score composed of the other two features was lower than the set threshold (for example, <0.5), the system immediately activated the dynamic light compensation mechanism and continuously tracked the degree of slope recovery after fill light. If the slope rose back to above 0.021V / ms, the system judged that the tissue response was obvious and the fill light was effective, and the subsequent waveform could be used as the basis for blood oxygen estimation.

[0072] The system further relies on two key waveform features—the amplitude ratio of the pulse peak to the trough value, and the micro-fluctuation periodicity feature—to comprehensively assess the microcirculatory status of the fingertip and drive subsequent light compensation and blood oxygen estimation strategies. During the raw data acquisition phase, the multi-wavelength LED in the device operates sequentially with red light (660nm), near-infrared (880nm), and reference green light (530nm). The photodetector collects the reflected signal of each wavelength at a sampling rate of 250Hz to form the PPG raw waveform. In the 20-second data segment collected by Mr. Wang, the reflected signal of the near-infrared channel exhibited basic periodic fluctuations, but the overall amplitude was low. The system preprocessed this waveform segment in three steps: first, using a bandpass filter (0.5-5Hz) to remove low-frequency drift and high-frequency noise; second, using the empirical mode decomposition (EMD) method to separate the main wave component from the secondary micro-fluctuations; and third, segment-by-segment normalization for each cardiac cycle to ensure consistent feature extraction at different amplitude scales.

[0073] In the extraction of peak-to-valley amplitude ratio features, the system first detects the starting point and peak point of each cardiac cycle, quickly locates them through the zero-crossing method combined with the first-order derivative extreme point, and then measures the amplitude difference ΔA between the peak value (P) of each cycle and its previous valley value (V), and normalizes it with the average DC level DC of the cycle to calculate the AC / DC ratio. For Mr. Wang's actual data, in the near-infrared waveforms of the 5th to 8th cycles, the average ΔA was 0.085V, the average DC was 0.96V, and the AC / DC ratio was 0.088. This ratio is much lower than the typical value of healthy people in the perfusion state (usually 0.12-0.25). Based on the empirical threshold, the system judges that the volume fluctuation amplitude is weak, reflecting a decrease in the local perfusion level. In the dynamic perfusion model, this ratio is normalized to a characteristic value f1, which is used as one of the inputs of the perfusion state indicator. The system then analyzes the periodic characteristics of micro-fluctuations in the PPG waveform, that is, after removing the main heartbeat frequency component, the residual waveform contains secondary low-amplitude, high-frequency periodic fluctuations. These fluctuations are usually caused by non-main heartbeat mechanisms such as capillary return, venous filling, and changes in vascular elasticity. The system uses wavelet packet decomposition (WPD) to decompose the PPG signal to the fourth layer, selects the wavelet component with a frequency band of 1.5-3.5Hz as the main component of micro-fluctuations, and then uses autocorrelation analysis to detect periodic repetitive structures. Mr. Wang's data showed irregular micro-fluctuations in this analysis, with large variations between cycles. The system measured a period variance of 0.15s 2, the autocorrelation peak dropped to 0.42, far below the typical range of 0.75-0.9 under high perfusion conditions. This indicates that the microcirculatory activity of the fingertip capillaries is weakened and the elastic response of the vascular wall is not obvious. The system combines the above two indicators f1 (amplitude ratio) and f2 (micro-fluctuation period stability) with the previously extracted edge slope f3 to form a three-dimensional feature vector [f1, f2, f3] = [0.088, 0.15, 0.009], which is fed into the optical perfusion state model as input. The model structure is a three-layer feedforward neural network with 3 nodes in the input layer, 16 neurons in the first hidden layer (activation function is ReLU), 8 neurons in the second hidden layer, and a Sigmoid normalized node output perfusion score value in the output layer. The lower the score, the worse the perfusion. The model was trained using approximately 30,000 segments of labeled data collected by the hospital. The training data included real PPG waveform segments with different perfusion states. Professional physicians manually annotated the presence of effective tissue blood perfusion (using the perfusion index PI as the standard). The model was trained using a weighted cross-entropy loss function, the Adam optimizer, and 150 training rounds, ultimately achieving a test accuracy of 92.4%. After edge optimization and compression, it can be deployed in low-power MCU devices for real-time operation. In Mr. Wang's test, the model output a perfusion score of 0.38, which was lower than the fill light start threshold of 0.5. The system immediately activated the LED dynamic light compensation mechanism, increased the near-infrared LED drive current from the original 8mA to 18mA, and adopted a periodic strong pulse stimulation mode. During the fill light process, the system continued to monitor new PPG data and found that the peak-to-valley ratio from the 9th to the 11th cycle increased to 0.12, and the micro-fluctuation period variance decreased to 0.07s. 2 , the autocorrelation rose to 0.66, indicating that tissue blood flow fluctuations and microcirculation were significantly enhanced, and the perfusion score also rebounded to 0.64. The system confirmed that the supplemental light was effective and used the updated PPG waveform for subsequent blood oxygen estimation, and obtained an SpO2 of 94%, an increase of 3% compared to before the supplemental light.

[0074] During Mr. Wang's continuous monitoring process, in order to overcome the problem of unstable blood oxygen estimation caused by insufficient blood perfusion in his fingertip tissue, the three temporal physiological characteristics of the PPG waveform - the rising edge slope S(t), the peak-to-valley amplitude ratio V(t), and the micro-fluctuation period stability Δ(t) - were sampled in real time and continuously integrated to form an optical perfusion state scoring function Π(t). Based on this score, the fill light behavior of the LED light source was driven and regulated, so that the light output was highly matched with the physiological response of the tissue, ensuring high-quality PPG signal acquisition and accurate SpO2 estimation even when perfusion was insufficient.

[0075] During the acquisition process, the system refreshes the waveform state every 100ms, and sets the integration window T = 6s, that is, the Π(t) value at each time t is determined by the data within the previous 6 seconds. The device worn by Mr. Wang at this time records the following waveform feature data fragment (for ease of explanation, a simplified time series is used): In the integration interval [4s, 10s] at time t0 = 10s, the following mean value sequence is collected: S(ξ) = 0.012V / ms, V(ξ) = 0.093, Δ(ε) = 0.16s 2 . Substitute into the scoring function model:

[0076]

[0077] Numerically, the integral can be performed using the equal-interval numerical integration method (set to integrate in units of 0.5s). The system accumulates the integral results within the function point by point, obtaining 6 integral terms, each with a value of approximately 0.0071, 0.0070, 0.0072, 0.0073, 0.0071, and 0.0072, with a total integral result of approximately 0.0429. Setting the normalization constant Λ = 0.2, we finally obtain:

[0078]

[0079] The perfusion threshold θ∈[0.4,0.6] is set at this time, and the system defaults to the median value θ=0.5. Since Π(t0)<, it is judged that Mr. Wang’s current perfusion state is poor, and the system immediately triggers the dynamic light compensation control module.

[0080] After the light source driver receives this judgment result, the originally set LED near-infrared channel drive current is 10mA. At this time, according to the step-by-step increase + feedback observation strategy, the system increases the light supply by 1.5mA every 0.5 seconds for three times until it reaches 15mA. At the same time, the system records the waveform changes during the light supply period and continuously calculates Π(t) to observe its response trend. For example, the integral value of Π(t) is recalculated 2 seconds after the fill light (t=12s), which is 0.31. It has increased by about 0.1 compared with the initial state, and the slope is If it is positive, it means that the fill light behavior is effective. According to the control strategy, the system keeps the LED current unchanged at this time and enters a stable output state.

[0081] If Π(t) is found to be stagnant and fluctuating in the subsequent operation, and fluctuates back and forth within ±0.01 in the short term, that is, The system maintains the current light intensity and pulse frequency and activates the micro-thermal management module to limit the LED temperature rise. If continued monitoring reveals that Π(t) continues to rise, reaching 0.56 at the 15th second, exceeding the threshold and continuing to rise, the system determines that tissue perfusion has significantly improved. Based on the energy-saving strategy, the LED current is reduced from 15mA to 12mA, and the duty cycle is adjusted from 40% to 30% to save power and prevent tissue overexposure.

[0082] Furthermore, during dynamic fill-in illumination, the system records the tissue's physiological response to light and extracts its morphological structure within high-response bands, including the onset time of the shoulder peak, the emergence of the second-order waveform, and the steep increase in slope. This information is used to guide the subsequent signal enhancement module in selecting a signal reconstruction template. When a fill-in illumination event is confirmed to have resulted in structural changes, the system marks it as a valid response segment and prioritizes it in the SpO2 estimation algorithm, eliminating low- or no-response segments to further enhance the reliability of the estimation.

[0083] Throughout this process, the comprehensive scoring function Π(t) provides a dynamic, continuous, and interpretable means of perfusion assessment, completely different from the traditional method based on a single-point AC / DC ratio hard trigger. Through a nonlinear integral structure, this invention integrates three physiological level indicators: rising power, volume response, and microcirculatory rhythm, enabling the device to respond to real physiological changes. Simultaneously, its control logic, through continuous trend judgment feedback, makes the LED light source both energy-efficient and physiologically adaptable, truly forming a dynamic physiological response-driven light compensation system. This is particularly suitable for people like Mr. Wang who suffer from chronic hypoperfusion, enabling continuous blood oxygen monitoring at home with medical-grade accuracy.

[0084] During Mr. Wang's continuous use, the accuracy of blood oxygen saturation estimation in low perfusion states was further improved. Based on the traditional ratio method, a response weighting mechanism was introduced. By analyzing the morphological changes in the photoplethysmography waveform before and after the light supplement, the physiological response of the tissue to light stimulation was quantified. This response degree was used as a weighting factor to correct the output of the traditional ratio method, thereby offsetting signal fluctuations or abnormal ratio shifts caused by insufficient perfusion. During data acquisition, the device continuously acquired reflected light signals at two wavelengths, 660nm and 880nm, at a sampling rate of 250Hz. The light source control module accurately recorded the timestamps of the start and end of the LED light supplement, enabling the system to accurately segment the PPG waveforms before the light supplement (e.g., the first 2 seconds), during the light supplement (e.g., the first 1-3 seconds), and after the light supplement (e.g., the last 3 seconds). The system then preprocessed these waveform segments, including first-order difference detrending, wavelet packet denoising, and amplitude normalization, to ensure that each waveform segment had the same data scale and comparability. The system then extracts multi-dimensional features such as slope, amplitude, and morphological structure (such as the appearance of a shoulder peak and a bimodal structure) from the waveforms before and after the fill light, and calculates the changes in morphological similarity between the waveforms, including indicators such as waveform mean square error, changes in the direction of morphological gradients, and frequency domain energy redistribution. It then uses a trained lightweight model to map these difference values ​​to a response weight value between 0 and 1. This model is a small neural network consisting of two fully connected hidden layers. The input is the difference vector of waveform features before and after the fill light, and the output is the response weight value. The model uses approximately 10,000 sets of manually annotated data during training. The data labels are set based on whether the tissue has a significant response (for example, whether it is a true response band label given by empirical perfusion indicators or human experts). The loss function is weighted binary cross entropy, and the optimizer is Adam. Data enhancement strategies are added during training to improve the model's generalization ability under low signal-to-noise ratios. During operation, during a test conducted by Mr. Wang, after the supplemental light treatment, the system calculated that the rising slope of the near-infrared waveform increased by 38%, the peak-to-valley amplitude increased by 22%, the probability of the shoulder peak appearance increased to 80%, and the morphological similarity index increased from 0.42 to 0.74. The response weight output by the model was 0.82, indicating that the tissue responded significantly to the supplemental light stimulation. Traditional ratio methods usually estimate SpO2 only based on the AC / DC ratio of the red light and near-infrared waveforms. However, in low perfusion conditions, this ratio can be distorted or underestimated due to insufficient light absorption or flat waveforms. This traditional ratio R is multiplied by the response weight and fed into a correction function for correction. The higher the weight, the more reliable the waveform quality and the closer the result is to the actual blood oxygen level. Conversely, a low weight corresponds to a larger ratio deviation and a stronger correction.Taking this case as an example, the SpO2 estimated by the traditional ratio method was 91%, but after response weight correction, the system estimated value was 93.4%, which was highly close to Mr. Wang's historical normal blood oxygen value and consistent with the 93% measured by the fingerclip oximeter. The error was controlled within ±0.5%, which is far better than the conventional solution under the non-response correction strategy.

[0085] Example 2:

[0086] Combined with attachment Figure 5 As shown in Example 1, during Mr. Wang's long-term fingertip blood oxygen monitoring, due to his chronic diabetes combined with peripheral microvascular sclerosis, the local perfusion of his fingers fluctuated frequently and was in a low perfusion state most of the time. The dynamic light compensation control system embedded in the device will regularly emit light stimulation at a preset rhythm, triggering LED fill light behavior every 30 seconds, and dividing this behavior into three time windows: before fill light (first 3 seconds), during fill light (lasting 2 seconds), and after fill light (lasting 5 seconds). These three waveforms are analyzed separately to determine whether the tissue has a real blood flow response to the light stimulation.

[0087] During one monitoring session, the system triggered the 20th round of fill light at the 10th minute. The LED driver emitted a 16mA current into the near-infrared channel, generating a 2-second-wide square wave fill light pulse and simultaneously recording the corresponding waveform. The system extracted the peak-to-valley amplitude difference of the PPG waveform during the 3-second period before fill light, finding it to be 0.072V. The calculated average phase delay of the main pulse cycle was 0.35 seconds, and the morphological features showed a smooth waveform profile with no obvious shoulder peaks or second-order rising components. During the mid-fill light period, the amplitude increased to 0.094V, the phase advanced to 0.29 seconds, and the waveform's left shoulder showed a sharp rise. The morphological recognition module identified two reflection-like peaks, signaling a high-response signal. Within 5 seconds after fill light, the amplitude dropped slightly to 0.085V, but remained higher than the pre-fill light level. The phase remained at 0.30 seconds, and the morphological structure retained a clear, sharp rise before the peak. The standard deviation of the waveform's micro-fluctuation period decreased to 0.08 seconds, indicating that capillary circulation was activated after fill light.

[0088] The system then compares the three waveforms for amplitude, phase, and morphology, calculating a response score using an internally constructed difference mapping function. For example, the amplitude response score is (0.094-0.072) / 0.072=0.305, a 30.5% improvement; the phase shortening response ratio is (0.35-0.29) / 0.35=0.171, a 17.1% shortening. The morphological changes are compared using a trained neural network model encoding vector similarity, increasing from 0.41 to 0.74. The overall response score is 0.78 (on a scale of 0-1). The system sets a physiological response threshold of 0.65, thus determining this fill light response as a significant response.

[0089] Based on this judgment, the system uses the PPG waveforms during and after the fill light as the blood oxygen estimation signal source for this cycle, while retaining the signal before the fill light as a control baseline to establish contextual differences. The system calculates the standard AC / DC ratio of the waveforms in the two time periods as 0.043 for red light and 0.027 for near-infrared light, with a traditional ratio R value of 1.59, resulting in an initial estimate of 91.7%. The response score of 0.78 is then used as a weighting factor, and the traditional R value is fine-tuned using a curve fitting and input into the response correction model, resulting in a corrected blood oxygen saturation of 93.1%, which is consistent with Mr. Wang's clinical blood oxygen measurement value at rest.

[0090] On the contrary, in the system's previous 17th round of fill light event, there was no obvious difference in the waveform after the fill light and before the fill light, the amplitude only increased by 2.1%, the phase almost did not change, the morphological feature score rose slightly from 0.43 to 0.46, and the total response score was 0.19, which was far below the threshold. Based on this, the system judged this fill light as an invalid response, and the entire waveform signal was downgraded and retained only as data feature extraction and training samples, and was not used for real-time SpO2 estimation.

[0091] Through the process of segmented extraction → feature analysis → scoring judgment → dynamic preferential input, the waveform segments that are truly relevant to physiological perfusion changes are efficiently screened out, and LED power control and valuation decisions are dynamically guided, completely getting rid of the error accumulation of full-segment averaging or single-factor judgment of signal intensity in traditional solutions. This ensures that even for patients with chronic hypoperfusion such as Mr. Wang, stable, accurate, and energy-controlled continuous blood oxygen monitoring can still be achieved, providing a highly reliable solution for home chronic disease management and remote vital sign identification.

[0092] The amplitude and phase characteristics of the pulse waveform are directly used to determine whether the tissue has produced a change in reflection enhancement and response speed in response to the fill-in light behavior, thereby inferring the true physiological response of the tissue's local blood flow to light stimulation, deciding whether to use this segment of data for blood oxygen estimation, and dynamically adjusting the light compensation strategy accordingly. During the data acquisition phase, the device's LED emits modulated pulse light of two wavelengths, 660nm and 880nm. The fingertip reflection signal is collected by a close-fitting photodetector. The sampling frequency of each wavelength signal is 250Hz. Each fill-in light behavior of the system lasts for 2 seconds, and 3 seconds of data are collected before and after the fill-in light, forming a complete three-segment time window data block. The original PPG waveform undergoes data preprocessing, including bandpass filtering to remove baseline drift (filter range 0.5Hz-5Hz), wavelet denoising (DB4 three-layer decomposition), and normalization to ensure that the peak and valley characteristics are comparable in each segment. Amplitude feature extraction involves the system using a cycle detection algorithm to identify local extreme points within each complete cardiac cycle. The system then measures the amplitude difference between each pair of adjacent peak and valley values ​​and normalizes this difference to the DC baseline within that cycle to obtain a standardized amplitude value. During Mr. Wang's 21st light-infusion cycle, the near-infrared channel amplitudes were 0.064V, 0.066V, and 0.062V in the three cycles before light-infusion. These amplitudes increased to 0.081V and 0.087V in the two cycles during light-infusion, representing an amplitude increase of 25%-40%. The system interprets this result as significant evidence of enhanced light reflection and labels it as a high-amplitude response. Phase feature extraction involves temporally locating the onset of each pulse, calculating the RR interval for each cycle, and determining the relative positional change of the phase starting point. The system uses the first-order derivative zero-crossing method to detect the waveform slope reversal point, thereby determining the heartbeat starting point. Subtracting this from the starting point of the previous cycle yields the phase delay or advance. For example, the average onset time delay before Mr. Wang's fill light was 370ms, while during the fill light, due to the increase in local capillary blood flow velocity, the onset time of the waveform after reflection enhancement was advanced to 312ms. The system measured a phase advance value of 58ms and a phase change ratio of 15.7%. This shows that the tissue reaction speed is accelerated, which is a direct manifestation of the local perfusion enhancement triggered by the fill light. The above two features are encoded as input vectors and sent to the response recognition module, which is a lightweight multi-layer perceptron (MLP). The network structure includes an input layer (two nodes: amplitude change ratio and phase advance value), two hidden layers (8 nodes in each layer, activation function is Swish), and the output layer is a Sigmoid node output response score (ranging from 0 to 1). The model's training data comes from more than 18,000 light-filling cycle samples collected from real wearers. Doctors label the samples as responsive or non-responsive based on changes in waveform morphology and the degree of perfusion improvement. Weighted binary cross entropy is used as the loss function during training, the optimizer is Adam, the number of training rounds is 100, and the validation set accuracy reaches 92.7%.The model is ultimately deployed on the terminal chip, with a response time of no more than 15ms, which can meet the real-time judgment needs of edge devices.

[0093] In Mr. Wang's current fill-light cycle, the model output a response score of 0.84, far exceeding the set response threshold of 0.6. The system then marked this waveform segment as high-confidence data, allowing it to participate in blood oxygen estimation. The system also fed the amplitude boost ratio and phase advance value as feedback signals into the light compensation scheduling module, maintaining the current LED power and fill-light rhythm. In contrast, in the 18th fill-light cycle, the amplitude changed by less than 10%, the phase remained unchanged, or even slightly delayed, resulting in a score of only 0.22. The system downgraded this data segment and excluded it from the SpO2 calculation process.

[0094] During Mr. Wang's 22nd consecutive low perfusion blood oxygen monitoring cycle, the system initiated a standard LED dynamic light compensation operation. The device output 880nm near-infrared pulsed light with a 12mA drive current for 2 seconds. The system divided the light-filling behavior into three time periods: the pre-light-filling window (3 seconds), the mid-light-filling window (2 seconds), and the post-light-filling window (5 seconds). The system extracted three core morphological features of the photoelectric capacitance waveform collected in each window: waveform contour characteristics, rising edge slope, and micro-fluctuation structure. The system then constructed a composite waveform function based on these morphological factors and performed an integral difference function calculation to determine whether the light-filling behavior triggered real microcirculatory physiological changes.

[0095] In the pre-fill window, the system detected that Mr. Wang's near-infrared waveform showed a slightly flat-top shape, with a waveform symmetry of 0.61 (calculated based on the normalized left and right area ratio), no obvious shoulder peak, a steep rise lasting about 84ms, a rising edge slope of 0.011V / ms, and irregular micro-fluctuation periods with a standard deviation of 0.15s. The waveform morphology function χ constructed by the system based on this is b The (τ) value is 0.48 (normalized to the unit interval). During the fill-in period, the system observed significant changes in the waveform profile. The symmetry was improved to 0.78, the left shoulder formation time was advanced by 18ms, the steep rise was shortened to 61ms, the slope increased to 0.017V / ms, the micro-wave structure showed periodic activation, the autocorrelation value increased from 0.42 to 0.69, and the standard deviation decreased to 0.07s. The corresponding morphological function χ s (τ) reaches 0.77. After the fill light, the waveform in the window drops slightly, but the outline is still clear, the symmetry remains at 0.74, the slope drops slightly to 0.015V / ms, and the micro-wave structure still maintains good periodicity. r (τ)=0.69.

[0096] At this point, the system performs the following integral difference function calculation:

[0097]

[0098] The integration interval was set from t1 = 0s (beginning before the fill light) to t3 = 10s (ending after the fill light), with sampling every 0.5 seconds, generating 20 integration points. The physiological credibility weight function Ω(τ) was generated by the rhythm index (based on heartbeat stability and microfluctuation synchronization), with a typical value set between 0.8 and 1.2 for each integration point. After calculation, the total difference value ΔΨ = 0.183 was obtained for this waveform response after integration, while the system's preset response threshold ε was 0.12 (the threshold range is generally set between 0.10 and 0.15, and the actual value can be dynamically adjusted to match the individual waveform noise characteristics).

[0099] Since ΔΨ=0.183>ε=0.12 in this fill-light behavior, the system determines that this behavior is an effective physiological response fill-light. The system immediately marks the waveform in this time period as a high-confidence segment and enters the main path of blood oxygen estimation, while keeping the current LED power and pulse rhythm unchanged to continue the effective response strategy. Conversely, if ΔΨ is lower than 0.08, for example, in Mr. Wang's 19th round of fill-light behavior, the system found that the waveform in the fill-light only had a slight increase in slope, no significant change in symmetry, and the micro-fluctuation structure was still in a disordered state. The resulting integral difference was only 0.06. Based on this, the system determined that this was an invalid response fill-light. Subsequently, in the next round, the LED current was reduced to 10mA and the duty cycle was adjusted from 40% to 25% to avoid the adverse effects of excessive fill-light on tissue heat load or energy consumption.

[0100] In addition, each morphological sub-function in the integral difference function is generated through training. The system uses a large amount of real user sample data (more than 25,000), manually annotates the response and non-response waveforms, and trains the distribution function of each morphological feature in the three-dimensional feature space. The weighted fusion function χ(τ) is constructed, and the model is fitted and solved in real time using polynomial approximation on the edge device, so that even on low-power chips, a complete morphological response calculation can be completed once per second, truly achieving edge intelligence + physiological reality.

[0101] During Mr. Wang's 23rd round of continuous blood oxygen monitoring, the system successfully stimulated a microcirculatory response in his fingertip tissue through dynamic LED illumination. To further improve the accuracy of blood oxygen estimation in low perfusion states, a response confidence scoring mechanism based on the differences in three waveform responses was designed. This mechanism weighted each waveform segment collected in each round in the SpO2 estimation, distinguishing between strong, weak, and no-response segments. This maximizes the use of data with true physiological reference value while filtering out data that could be affected by noise, optical reflection failure, or physiological inactivity. During operation, each LED illumination compensation action is clearly calibrated by the system with a start and end time, divided into three windows: before illumination compensation (e.g., t1 to t2), during illumination compensation (t2 to t3), and after illumination compensation (t3 to t4). Within each window, the system extracts multidimensional features from the PPG reflectance waveform, including three primary axes: amplitude, phase, and morphology, resulting in nine secondary features, including peak-to-valley amplitude difference, slope change rate, symmetry index, and microfluctuation period stability. The system compares the differences between these three data segments dimension by dimension to form a response difference vector Δ, where each dimension represents the degree of significant change in a certain type of physiological parameter before, during, and after the supplemental lighting. Subsequently, the system uses a lightweight multi-input scoring network to construct a response credibility scoring function, inputs the Δ vector into the model, and outputs a normalized weight value R between 0 and 1. s (ResponseScore). The model structure is a three-layer perceptron network: the input layer contains 9 nodes, the two hidden layers contain 16 and 8 neurons respectively, the activation function adopts a mixed structure of Tanh and Swish, and the output layer is a Sigmoid node. The training data comes from more than 37,000 dynamic light compensation sample segments, which are labeled by experts (whether there is a real tissue response) to form supervised samples. The system uses weighted cross entropy as the loss function for training, and introduces the real SpO2 reference value as an auxiliary label to improve the model's ability to identify error risk segments. After training, the model is quantized and compressed and deployed on the MCU, and the calculation delay is less than 30ms. In Mr. Wang's fill light data this time, the average amplitude of the window before fill light was 0.066V, the slope was 0.011V / ms, and the waveform symmetry was 0.58. During fill light, they increased to 0.083V, 0.016V / ms, and 0.75, respectively. After fill light, they remained at a similar high level. The system calculated the response credibility score R through the feature difference vector Δ s=0.87, the system believes that the fill light response is strong and the waveform reliability is high, and the waveforms in the fill light and after the fill light are input into the main channel of blood oxygen estimation, and a weight coefficient of 0.87 is assigned to them. In contrast, in the 21st round of fill light, the amplitude only increased by 7%, the slope remained almost unchanged, the symmetry improvement was less than 0.1, the Δ vector difference was extremely small, and the final response score was only 0.22. The system marked this waveform as low-confidence data, and the weight coefficient was reduced to 0.22 or directly eliminated when participating in SpO2 estimation. This response scoring mechanism ultimately acts on the fusion calculation model of blood oxygen estimation, that is, the SpO2 estimation result output by the traditional ratio method will be based on the current response score R s Perform linear or nonlinear weighting, or participate in the regression weight allocation of the machine learning valuation model as sample importance. A test of the system after Mr. Wang wore it for 30 minutes showed that the mean square error of the valuation after introducing the response score was reduced by 38% compared to the original ratio method, and the error dropped by more than 50% during the low perfusion period (PI < 0.3). In addition, in terms of device power consumption, since the fill light rhythm is only continued in the high-response segment and the power is reduced or the fill light frequency is reduced in the low-response segment, the overall energy saving rate is increased by more than 25%, significantly extending battery life in wearable long-term working environments.

[0102] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting low perfusion blood oxygen saturation based on dynamic light compensation, characterized in that The following steps are involved: An optical characterization of red blood cell density in tissues is constructed based on the combined characteristic relationship between multi-wavelength reflection intensity and fluctuation amplitude. The current perfusion level is inferred from the morphological characteristics of the reflection waveform, including pulse edge slope, peak-to-valley ratio, and micro-fluctuation period. To establish optical indices of perfusion; The time series modeling method used is to analyze the changing trend of light reflection dynamics over a period of time; The LED driving strategy is dynamically adjusted to form a physiological rhythm-like lighting behavior with three modes: step-by-step ramping, pulse stimulation, and cycle maintenance. In the time windows before, during, and after the onset of fill light, the differences in amplitude, phase, and morphology of the plethysmographic waveforms of the three light segments are calculated. Dynamic morphological similarity discrimination is introduced to quantify whether each fill light segment truly elicits a tissue response and distinguish true physiological fluctuations from noise disturbances. For low-response fill light, the fill light intensity is reduced as a feedback mechanism. On the contrary, the key deformation features of high-response fill light are extracted for signal reconstruction; The pulsating component and baseline component of each wavelength signal are separated. Combined with the weights set for the perfusion optical indicators, the pulsating information of different wavelengths is ratio-converted. A pre-calibrated conversion model is called to output the blood oxygen saturation value. At the same time, a result quality index based on the waveform morphology similarity score is given as the final detection result.

2. The method for detecting low perfusion blood oxygen saturation based on dynamic light compensation according to claim 1, characterized in that The method for establishing perfusion optical indicators comprises: s1. Collect multi-wavelength light reflection signals from the target detection site and obtain the pulse photoplethysmography waveform; S2. Extracting morphological features of the reflected waveform, including: a) The slope of the rising edge of the pulse, b) the ratio of the pulse peak to the trough amplitude, c) Repeatability and similarity of micro-fluctuation cycles; S3. constructing an optical perfusion state index based on the waveform morphology characteristics to determine whether the current tissue is in a low perfusion state; when the perfusion state index is lower than a preset threshold, triggering a dynamic light compensation control module to adjust the incident light power; S4. During the light compensation process, continuously monitor the waveform morphology changes, determine the light compensation effect, and adjust the light output accordingly; calculate the blood oxygen saturation estimate based on the reflection signal optimized by light compensation.

3. The method for detecting low perfusion blood oxygen saturation based on dynamic light compensation according to claim 2, characterized in that The extraction of the pulse edge slope includes analyzing the time and amplitude change rate of the rising edge of the photoplethysmography waveform to determine the blood propulsion speed and the tissue light response speed.

4. The method for detecting low perfusion blood oxygen saturation based on dynamic light compensation according to claim 3, characterized in that The amplitude ratio of the peak value to the valley value is used to represent the amplitude of tissue volume change during the cardiac cycle. The smaller the amplitude, the weaker the perfusion. The micro-fluctuation period is a secondary fluctuation outside the main pulse waveform and is used to judge the capillary return or vascular elasticity in the tissue.

5. The method for detecting low perfusion blood oxygen saturation based on dynamic light compensation according to claim 4, characterized in that The optical perfusion status index is a comprehensive score calculated based on the weighted calculation of multiple waveform features, which is used to reflect the real-time tissue perfusion quality; the dynamic light compensation control module includes a light source driver and a feedback controller, which is used to dynamically adjust the power and pulse rhythm of the LED light source according to the optical perfusion status index.

6. The method for detecting low perfusion blood oxygen saturation based on dynamic light compensation according to claim 5, characterized in that The blood oxygen saturation calculation process introduces a response weight based on the waveform morphology changes before and after light compensation, which is used to correct the estimation deviation of the traditional ratio method.

7. The method for detecting low perfusion blood oxygen saturation based on dynamic light compensation according to claim 1, characterized in that The method for respectively calculating the response differences in amplitude, phase, and morphology of the capacitance waveforms of the three light segments includes: Dynamic light compensation signals are sent to the test tissue at a preset rhythm; each light compensation behavior is divided into three time windows: before, during, and after light compensation; and multi-dimensional features of amplitude, phase, and morphology are extracted from the photoplethysmographic waveforms collected in each time window. The characteristics of the three waveforms are compared to determine the differences in the physiological responses of tissues to the supplemental lighting behavior; the waveforms within the supplemental lighting period with physiological response characteristics are used for blood oxygen saturation estimation, and the remaining waveforms are downgraded or discarded.

8. The method for detecting low perfusion blood oxygen saturation based on dynamic light compensation according to claim 7, characterized in that The amplitude feature includes the amplitude change between the peak and valley values ​​of the pulse waveform, which is used to determine whether the supplemental light causes enhanced tissue reflection; the phase feature includes the advance or delay of the pulse wave start time relative to the reference point of the previous cycle, which is used to evaluate the tissue's response speed to the supplemental light.

9. The method for detecting low perfusion blood oxygen saturation based on dynamic light compensation according to claim 8, characterized in that The morphological characteristics include changes in the contour of the pulse waveform, the slope of the rising edge, and the appearance and disappearance of micro-wave structures. By constructing an integral function and quantifying the differences in the three waveforms, we can determine whether the fill light causes physiological changes in tissue microcirculation: in: ΔΨ: defined waveform response morphology difference index, used to measure the significance of physiological response; t1, t3: respectively represent the start time of the window before fill light and the end time of the window after fill light; T: any time point in the integration interval; χ s (τ): The composite morphological function of the waveform in the fill light window, which includes the fusion of the contour factor, slope factor, and micro-fluctuation factor; χ b (τ), χ r (τ): composite waveform morphology value at the corresponding time points before and after fill light; Ω(τ): physiological credibility weight function; |·|: represents the absolute morphology offset between the three segments, indicating the net change of the waveform caused by fill light; When ΔΨ exceeds the system's preset response threshold ε, the fill light behavior is determined to be an effective physiological response fill light; otherwise, it is considered to be invalid fill light or optical over-fill light, and the LED intensity in the next fill light cycle will be reduced or the fill light duration will be shortened.

10. The method for detecting low perfusion blood oxygen saturation based on dynamic light compensation according to claim 9, characterized in that The response differences of the three waveforms are used to construct a response credibility score by weighting different waveforms in the blood oxygen saturation calculation; each waveform segment collected in each round is weighted in the SpO2 estimation to distinguish between strong response segments, weak response segments and no response segments.

Images