A clinical trial blood oxygen collection calibration method based on multi-modal fusion

Abstract

The application discloses a clinical test blood oxygen acquisition calibration method based on multi-modal fusion and relates to the technical field of medical signal processing. Dual-wavelength photoelectric plethysmogram signals and inertial, ambient light, contact pressure and surface quality auxiliary modalities are synchronously acquired and time-aligned; motion pollution, environmental interference and contact stability intermediate states are constructed based on the aligned data to generate window credibility; whether to enter a calibration process is judged according to the window credibility, a heart beat level alternating component and a direct current component are extracted in a credible window to construct a robust ratio, and dynamic calibration is performed in combination with perfusion, contact stability, ambient light and surface quality; in a low credibility window, parameter updating is frozen, and a reference blood oxygen value or a degraded blood oxygen value and a quality mark are output. The stability, traceability and cross-condition consistency of blood oxygen calibration in a continuous monitoring scene are improved.

Description

Technical Field

[0001] This invention belongs to the field of medical signal processing technology, specifically relating to a clinical trial blood oxygen acquisition calibration method based on multimodal fusion. Background Technology

[0002] With the increasing digitalization of clinical trial data collection, blood oxygen saturation, as an important physiological indicator reflecting the respiratory and circulatory status of subjects, has been widely used for drug safety monitoring, anesthesia recovery assessment, observation of respiratory diseases, and continuous vital sign recording in multicenter clinical studies. Current blood oxygenation detection is typically based on pulse oximetry, which extracts the pulsating and DC components from the photoplethysmography signal by exploiting the difference in absorption of different wavelengths of light in oxyhemoglobin and deoxyhemoglobin, and then constructs a ratio to map it to an estimated blood oxygen saturation value.

[0003] However, clinical trial scenarios differ significantly from routine resting monitoring scenarios. Subjects at the bedside may turn over, raise their hands, speak, exhibit slight tremors, or undergo nursing procedures, leading to motion artifacts overlapping with the heart rate frequency in the photoplethysmography signal. Changes in ward lighting, probe contact pressure, peripheral perfusion fluctuations, skin color differences, and local temperature variations also collectively affect the optical path propagation and reception process, making the traditional "single-channel filtering + fixed calibration curve" processing framework prone to baseline drift, AC component distortion, and ratio shift. The fundamental reason is that optical propagation, tissue scattering, contact state, and motion state in a real clinical environment are coupled and change; relying solely on filtering within a single-mode waveform domain is insufficient to simultaneously maintain pulse morphology, suppress non-stationary interference, and maintain the consistency of calibration mapping.

[0004] To address the aforementioned issues, existing technologies have explored certain approaches. The first category of solutions revolves around primary signal quality assessment and posterior compensation. For example, patent CN202510502083.2 introduces a signal quality index, Bayesian filtering, and a time-series prediction module to dynamically correct blood oxygenation results under low signal-to-noise ratio conditions; patent CN202511080860.5 constructs a blood oxygenation calibration model using hand-reflective optical paths and facial video information to improve the accuracy of non-contact monitoring. A common characteristic of these technologies is the construction of compensation relationships around the primary observation link, with auxiliary modes primarily undertaking result mapping enhancement or compensation roles, generally approaching a posterior logic of "correcting results from results."

[0005] The second category of solutions focuses on multispectral or imaging reconstruction. For example, patent CN202511266846.4 obtains lung tissue oxygenation imaging results by simultaneously acquiring white light images and two near-infrared wavelength images, extracting multiple scattering components, and performing depth mismatch correction; patent CN202210786161.2 uses the reconstruction of RGB images to hyperspectral images to achieve tissue oxygenation assessment. This type of technology is mainly aimed at tissue imaging scenarios with well-defined spatial structures and imaging conditions, focusing on the reconstruction of spatial oxygenation distribution, rather than the stability of wearing the device, motion artifact separation, and online calibration stability in bedside continuous acquisition.

[0006] The existing technologies described above still have the following shortcomings in the calibration of pulse oximetry data acquisition in continuous clinical trials. First, multi-source interference is not explicitly decoupled; motion artifacts, ambient light variations, contact pressure fluctuations, and perfusion variations are not modeled separately in different projection modes in the dual-wavelength channels, easily leading to the aliasing of calibration and noise biases. Second, the calibration process lacks explicit intermediate states and causal chains, making it difficult to specify when calibration should be allowed, when parameter updates should be frozen, and when only reference or degraded values ​​should be output, thus failing to meet the requirements of traceability and auditability of anomalies in clinical trials. Third, the objects and time scales of imaging reconstruction technologies do not match those of bedside continuous acquisition, making it difficult to directly solve the problems of real-time artifact separation, quality grading, and online parameter updates for wearable pulse oximetry terminals operating continuously at the second or sub-second level.

[0007] Therefore, it is still necessary to provide a blood oxygen acquisition calibration method for continuous monitoring scenarios in clinical trials, which can separate and characterize motion contamination, environmental interference, and contact stability. Based on this, a window confidence level can be generated to further determine whether to perform calibration, whether to update parameters, and how to handle abnormal degradation output, thereby improving the stability, traceability, and cross-condition consistency of blood oxygen calibration under continuous monitoring conditions. Summary of the Invention

[0008] The purpose of this invention is to address the problems in existing technologies, such as undecoupled multi-source interference, lack of explicit state modeling in the calibration process, and insufficient online update mechanisms in continuous monitoring scenarios of clinical trials, by providing a multimodal fusion-based calibration method for clinical trial blood oxygen acquisition. This method is designed for bedside continuous acquisition environments. It establishes a multimodal observation system based on dual-wavelength photoplethysmography signals and auxiliary modal signals, and explicitly constructs intermediate state chains between motion contamination, environmental interference, contact stability, and calibration reliability states. This determines whether the current time window is suitable for entering the calibration process, and whether parameter updates are allowed or degraded output should be performed.

[0009] To achieve the above objectives, the technical solution provided by the present invention includes:

[0010] This invention provides a multimodal fusion-based clinical trial blood oxygen acquisition calibration method. This method simultaneously acquires the subject's red photoplethysmography (PPG) signal, infrared PPG signal, and at least one auxiliary modal signal. The auxiliary modal signal includes at least one of a three-axis inertial signal, ambient light signal, probe contact pressure signal, and skin surface image quality features. A unified timestamp is appended to each modal signal in the auxiliary modal signal, and each modal signal is aligned to the time axis of the red and infrared PPG signals. Based on the aligned dual-wavelength PPG signal and modal signal, an intermediate state reflecting motion contamination, environmental interference, and contact stability is constructed. Motion contamination is jointly characterized by morphological synchronization deviation and inertial fluctuations of the dual-wavelength PPG signal; environmental interference is jointly characterized by changes in the dual-channel DC baseline and ambient light fluctuations; and contact stability is jointly characterized by probe pressure deviation and pulse perfusion changes. A current time window is generated based on the intermediate state. The system determines the reliability of the time window and decides whether the current time window is allowed to enter the blood oxygen calibration process. When the window reliability meets the reliability condition, the AC and DC components of the red and infrared channels are extracted at the heart rate granularity to construct the robust ratio. Dynamic calibration is then performed in conjunction with perfusion information, contact stability information, ambient light level, and surface quality information to obtain the blood oxygen calibration value for the current time window. When the window reliability does not meet the reliability condition, the calibration parameter update is frozen, and a reference blood oxygen value or a degraded blood oxygen value is output according to the interval where the window reliability is located. At the same time, a quality mark or alarm information corresponding to the output is generated.

[0011] In this invention, to facilitate the establishment of synchronization relationships between modes with different sampling frequencies, window-level multimodal input data can be constructed first. This can be represented as: , in, Indicates the first A multimodal input tensor with a time window; Indicates the first Red light photoplethysmography feature sequence within a time window; Indicates the first Infrared photoplethysmography feature sequence within a time window; This represents the sequence of inertial features aligned with this time window; This represents the ambient light feature sequence aligned with this time window; This represents the contact pressure characteristic sequence aligned with this time window; This represents the low-dimensional feature sequence of image quality aligned with this time window. When image modalities are not deployed... The input tensor is empty, and its corresponding weights are reset to zero in subsequent fusion. The purpose of this input tensor is to establish a synchronization relationship between observations of different modes on a unified time axis, providing a basis for subsequent decomposition of interference sources.

[0012] Furthermore, aligning each auxiliary modal signal to the time axis of the red photoplethysmography (PPG) signal and the infrared PPG signal includes: performing saturation sample removal, outlier code removal, smoothing, and separating AC candidate components from DC components on the red PPG signal and the infrared PPG signal; extracting the acceleration modulus, angular velocity modulus, and their changes from the triaxial inertial signal; extracting the de-peaked slow-varying trend features from the ambient light signal and the probe contact pressure signal; resampling the auxiliary modal signals to the PPG signal time axis according to a unified timestamp, wherein continuous numerical signals are aligned using interpolation and category state signals are aligned using state preservation; when the missing proportion of any auxiliary modality in the current window exceeds a preset threshold, the auxiliary modality is marked as a low-confidence modality, and the corresponding weight of the auxiliary modality is downgraded during subsequent intermediate state fusion.

[0013] Furthermore, the skin surface image quality features are low-dimensional quality features extracted from the skin surface image around the probe acquired by the miniature camera. These low-dimensional quality features include at least one of regional average brightness, skin color stability, and occlusion ratio. The skin surface image quality features only participate in intermediate state generation and dynamic calibration as feature values, without transmitting or storing the original image. When no image modality is deployed, the corresponding features of the image modality are set to null, and their corresponding weight in window confidence generation is reset to zero. This allows for the use of skin surface conditions to assist in identifying occlusion, displacement, and local reflection changes, while also reducing privacy risks and terminal computational burden.

[0014] Furthermore, the construction of the intermediate state reflecting motion contamination, environmental interference, and contact stability includes: generating a synchronization artifact coupling index based on the difference in first-order changes between the red light photoplethysmography signal and the infrared photoplethysmography signal within the current window, combined with the fluctuation intensity of inertial characteristics within the current window, to characterize the motion contamination state; generating an environmental incident interference index based on the changes in the red light DC baseline and the infrared DC baseline between the current window and the previous window, as well as the fluctuation of ambient light characteristics, to characterize the environmental interference state; generating a contact stability index based on the deviation of the average probe pressure in the current window from the reference pressure and the deviation of the pulse perfusion index in the current window from the reference perfusion index, to characterize the contact stability state; and fusing the synchronization artifact coupling index, the environmental incident interference index, the contact stability index, and waveform quality characteristics to obtain the intermediate state.

[0015] In one alternative implementation, the synchronization artifact coupling index Determined according to the following formula: , in, Indicates the first Synchronization artifact coupling index for each time window; This indicates the number of sampling points within the current time window; Indicates the first The red light photoplethysmography signal in the time window is at the... First-order difference values ​​at each sampling point; Indicates the first The infrared photoplethysmography signal in the time window is at the... First-order difference values ​​at each sampling point; This indicates a smoothing term used to prevent the denominator from being zero; Indicates the first The first time window Inertial feature magnitude at each sampling point; Indicates the first The mean of the inertial characteristic modulus within a time window; Indicates the first The standard deviation of the inertial feature magnitude within a time window; This represents the inertial coupling weight. In the above formula, the first term measures the inconsistency in the morphological changes of the dual-wavelength waveform, and the second term introduces actual motion evidence, thus enabling motion artifacts to be distinguished from the differences in the true pulse morphology. When When approaching zero, through To maintain computational stability, when window length degradation makes it impossible to stably compute differences, only the inertial stability term can be retained.

[0016] Furthermore, the environmental incident interference index It can be determined according to the following formula: , in, Indicates the first Environmental incident interference index for each time window; Indicates the first Average value of DC component of red light channel within a time window; Indicates the first Average value of DC component of infrared channel within a time window; This represents the average value of the DC component of the red light channel within the previous time window; This represents the average DC component of the infrared channel within the previous time window. Indicates the first Within the first time window The ambient light characteristic values ​​corresponding to each sampling point; Indicates the first Mean values ​​of ambient light characteristics within a time window; Indicates the first Standard deviation of ambient light characteristics within a time window; Indicates ambient light weight; The meaning is the same as described above. The first part of the formula characterizes the degree of change in the dual-channel DC baseline between adjacent windows, while the second part introduces evidence of ambient light fluctuations to identify baseline drift caused by changes in illumination rather than by actual physiological changes. When there is no preceding value in the first window, the DC mean of the previous window can be set equal to the DC mean of the current window, making this term initially zero; when the ambient light mode is missing, the second term can be set to zero.

[0017] Furthermore, the contact stability index Determined according to the following formula: , in, Indicates the first Contact stability index for each time window; Indicates the first Average probe pressure over a time window; This represents the average reference pressure obtained during the initial stabilization phase. The parameter representing the fluctuation of the reference pressure; Indicates the first Pulse perfusion index for each time window; This represents the reference perfusion index obtained during the initial stabilization phase; This represents a smoothing term to prevent the denominator from being zero. The index characterizes whether the probe fit is within a stable range by the combined attenuation of pressure deviation and perfusion deviation. Specifically, if the average reference pressure is not obtained during the initial stabilization phase, a preset pressure reference value is used as a temporary reference value, and updated to an individualized reference value after the first time window of stabilization conditions is met. When the reference perfusion index is below a preset lower limit, the index terms related to the pulse perfusion index are suspended, and the contact stability index is determined solely based on the degree of pressure deviation.

[0018] Furthermore, the credibility of the window It can be obtained by fusing the above intermediate states, and the fusion form can be expressed as: , in, Indicates the first Window reliability for each time window; Represents the Sigmoid function; , , and These represent the weights of synchronization artifacts, environmental interference, contact stability, and overall quality, respectively. It represents a comprehensive quality score composed of waveform morphological stability, peak spacing variability, and optional image quality features; This indicates a bias term. The confidence level is not based on the blood oxygen value itself, but rather on a risk assessment of whether the current window can serve as a valid calibration sample. When image modalities are missing, It consists only of non-image quality features, and the corresponding weights are normalized.

[0019] Furthermore, when the window confidence level meets the confidence condition, the AC and DC components of the red and infrared channels are extracted at the heartbeat granularity to construct a robust ratio, including: locating the effective heartbeat cycle within the current time window and calculating the AC amplitude and DC baseline of each effective heartbeat in the red and infrared channels respectively; constructing the ratio of the red AC / DC ratio to the infrared AC / DC ratio for each effective heartbeat; aggregating the ratios of all effective heartbeats within the current time window using the median to obtain the robust ratio; when the number of effective heartbeats within the current time window is less than a preset number, or the DC component of any effective heartbeat is lower than the minimum effective illuminance threshold, the robust ratio is not output, and the current time window is determined to be an unusable time window for calibration.

[0020] In one alternative implementation, the ratio of the robustness ratios Determined according to the following formula: ,

[0021] in, Indicates the first The ratio of robustness ratios for each time window; Indicates the first The number of valid heartbeats detected within a time window; Indicates the first The first time window The AC amplitude of an effective heartbeat in the red light channel; Indicates the first The first time window The DC baseline of an effective heartbeat in the red light channel; Indicates the first The first time window The AC amplitude of an effective heartbeat in the infrared channel; Indicates the first The first time window The DC baseline of an effective heartbeat in the infrared channel; Represents the median operator; This indicates a smoothing term. This method uses median aggregation of heartbeat granularity instead of a full-window average, thereby reducing the bias of local abnormal heartbeats on the construction of the full-window ratio. When No output when the number is less than the preset number. .

[0022] Further, the step of performing dynamic calibration to obtain the blood oxygen calibration value for the current time window includes: constructing a calibration feature vector that includes at least the robustness ratio, artifact index smoothing value, environmental disturbance index smoothing value, contact stability index, perfusion index, ambient light baseline level, and surface quality stability index; inputting the calibration feature vector into a confidence-gated dynamic calibration model to obtain the blood oxygen calibration value for the current time window, wherein the dynamic calibration model applies conservative corrections to time windows with low confidence; when the window confidence is lower than the parameter update lower limit or the robustness ratio is missing, calibration parameter updates are prohibited, and the process directly switches to reference output or degraded output.

[0023] In one alternative implementation, the calibration feature vector It can be represented as: , in, Indicates the first The calibration feature vector for each time window; This represents the ratio of robustness ratios; This represents the window-level artifact exponential smoothing value; This represents the smoothed value of the window-level environmental disturbance index; Indicates the contact stability index; Indicates the interpeak variation index; Indicates the perfusion index; Indicates the ambient light baseline level; This represents an indicator of skin surface reflectance stability or image quality stability. This feature vector is used as input for dynamic model calibration, rather than being directly used as the final display value.

[0024] Furthermore, the dynamic calibration model for the confidence gating can take the following form: , in, Indicates the first Blood oxygen calibration values ​​for each time window; Indicates the intercept parameter; to These represent the regression coefficients corresponding to the ratio of robustness ratios, infusion index, contact stability index, ambient light baseline level, and surface quality stability index, respectively. Indicates the coefficient of the low-confidence penalty term; Indicates the first The confidence level of a time window. In this formula, Reflecting the main optical absorptivity relationship, and These respectively reflect the explainable deviations of the calibration from the infusion and contact conditions. and It reflects environmental and skin surface quality factors, and This constraint is used to express that the model output should be more conservative when the window confidence is lower. When it exceeds the physical range, it can be cropped to the preset effective blood oxygen range; when Below the confidence limit or If the model is missing, it will not be executed, and the process will switch to a degenerate output flow.

[0025] Furthermore, the calibration parameter update employs a selective online update mechanism, including: statistically analyzing the average window confidence, average environmental interference level, and average contact stability level across multiple recent consecutive time windows; allowing individualized calibration parameters to be updated only when the average window confidence is not lower than a confidence update threshold, the average environmental interference level is not higher than an environmental update threshold, and the average contact stability level is not lower than a contact update threshold; the individualized calibration parameter update only includes fine-tuning of the intercept parameter and principal coefficient parameters, without freely updating all calibration parameters; the fine-tuning of the individualized calibration parameters is achieved using recursive least squares or an equivalent low-complexity recursive parameter estimation method. By limiting the online update conditions and the range of updated parameters, the risk of long-term calibration curve drift caused by erroneous updates within abnormal windows can be reduced.

[0026] Further, the step of outputting a reference blood oxygen value or a degraded blood oxygen value according to the interval of the window confidence level, and simultaneously generating a quality label or alarm information corresponding to the output, includes: when the window confidence level is not lower than a first confidence threshold, outputting a calibrable blood oxygen value and recording the evidence chain corresponding to the calibrable blood oxygen value, wherein the evidence chain includes at least the motion contamination state, environmental interference state, contact stability state, robustness ratio ratio, and current parameter version number; when the window confidence level is lower than the first confidence threshold but not lower than a second confidence threshold, outputting a reference blood oxygen value with an attached medium-risk label, and simultaneously blocking calibration parameter updates; when the window confidence level is lower than the second confidence threshold, outputting the blood oxygen value maintenance result corresponding to the most recent high confidence time window, or outputting a short-term amplitude-limited prediction result based on the most recent high confidence time window, and generating an alarm information indicating that the current acquisition conditions do not meet the calibration requirements.

[0027] In one optional implementation, the short-time amplitude-limiting prediction result Determined according to the following formula: , in, Indicates the first Degraded output values ​​for each time window; This indicates the blood oxygen value corresponding to the most recent high-confidence time window; This indicates the blood oxygen value corresponding to the previous high-confidence time window before the most recent high-confidence time window; The index representing the most recent high-confidence time window; Indicates the attenuation coefficient; This represents a limiting function used to restrict input variations within preset upper and lower bounds. This formula is used to maintain short-term output continuity during low-confidence phases, preventing unreasonable large jumps when acquisition conditions are not met; if the historical high-confidence window is insufficient, the most recent available value can be directly retained.

[0028] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the above-described clinical trial blood oxygen acquisition calibration method based on multimodal fusion.

[0029] Compared with existing technologies, this invention, through the collaborative design of multimodal state decomposition, credibility gating calibration, and abnormal degradation output, can form a closed-loop processing logic of "technical problem - state identification - calibration execution - safe output" in continuous acquisition scenarios in clinical trials, thereby achieving the following beneficial effects.

[0030] First, this invention addresses the problem of motion artifacts being mixed with real pulse components in existing technologies by constructing a synchronization artifact coupling index. By combining evidence of dual-wavelength waveform morphology synchronization and inertial fluctuations, motion contamination is first identified at the state level, and then the window confidence level determines whether the current window is allowed to enter the calibration process. This reduces the probability of a semi-contaminated window being mistakenly included in the calibration process, and mechanistically suppresses calibration drift caused by motion artifacts. Compared with a single-mode fixed threshold filtering scheme, this invention can effectively reduce the mean square error of blood oxygen and the error update rate of parameters in motion interference scenarios.

[0031] Secondly, this invention addresses the problem that the ratio of ambient light fluctuations and changes in contact state jointly affect the DC component, causing a systematic shift in the ratio. This is achieved by constructing separate ambient incident interference indices. and contact stability index This invention separates and models DC baseline changes, ambient light fluctuations, pressure deviations, and perfusion deviations, then incorporates ambient light levels and contact status into the dynamic calibration model. This distinguishes DC offset-type interferences from actual physiological changes, reducing continuous output instability caused by factors such as lighting switching and probe loosening. Under mixed conditions of lighting switching and probe loosening, this invention can reduce the median absolute error of blood oxygen estimation and the number of continuous output jumps.

[0032] Third, this invention addresses the common problem in existing technologies of lacking a system-level causal chain that defines "when calibration is possible, when to freeze updates, and when to output only reference or degraded values." It establishes a window confidence level, a parameter freezing mechanism, and a tiered output mechanism. Each calibrable output can be linked to a chain of evidence including motion contamination status, environmental disturbance status, contact stability status, robustness ratios, and parameter version numbers, thus ensuring the traceability of the calibration process. For low-confidence windows, continuous operation is maintained and anomaly propagation is reduced through reference or degraded output mechanisms. Compared to manual artifact identification combined with conventional calibration procedures, this invention reduces the proportion of low-confidence windows erroneously participating in calibration and the cross-center average bias.

[0033] Fourth, this invention addresses the issue of calibration parameters being easily skewed by abnormal windows during long-term online operation under conditions of low perfusion and individual differences. It employs a selective online update strategy, allowing fine-tuning of the intercept and principal coefficient parameters only under continuous window conditions with high average reliability, low environmental interference, and good contact stability, rather than freely updating all parameters. This preserves individualized adaptation capabilities while controlling the risk of instability during the online learning process. In scenarios involving mixed skin color differences and low perfusion, this invention can reduce the 95th percentile absolute error. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings adopt a flowchart form that is strongly related to the method flow of the present invention, and are used to illustrate the logical relationship between the core processing links of the present invention. The flowchart in the accompanying drawings does not constitute a limitation on the scope of protection of the present invention.

[0035] Figure 1 The flowchart for interference state decomposition in this embodiment of the invention shows the process of calculating the synchronization artifact coupling index, environmental incident interference index, and contact stability index based on dual-wavelength photoplethysmography sequence, inertial sequence, ambient light sequence, and pressure sequence, and further generating window confidence.

[0036] Figure 2 This is a flowchart of the reliable segment selection and ratio construction in an embodiment of the present invention. It shows the process of filtering time windows based on window reliability, and performing heartbeat segmentation, AC component and DC component extraction, and robust ratio construction within the reliable window.

[0037] Figure 3 The flowchart shown in this embodiment of the invention illustrates a dynamic calibration model that uses calibration feature vectors and window confidence input confidence gating, and selectively updates individualized calibration parameters based on continuous window conditions.

[0038] Figure 4 The flowchart for abnormal degradation and result output in this embodiment of the invention shows the processing flow of executing calibrable output, reference output or degradation-preserving output according to the window confidence interval, and generating quality marks, alarm information and evidence chains. Detailed Implementation

[0039] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0040] Example 1: Overall Flowchart of a Clinical Trial Blood Oxygen Acquisition Calibration Method Based on Multimodal Fusion

[0041] like Figures 1 to 4 As shown in this embodiment, a clinical trial blood oxygen acquisition calibration method based on multimodal fusion is provided, comprising six collaborative steps: multimodal synchronous acquisition, heterogeneous data preprocessing and time alignment, interference source state decomposition, reliable segment selection and ratio construction, dynamic calibration, and abnormal degradation output. This method is executed by the processor of the clinical trial acquisition terminal. The processor reads program instructions and a parameter dictionary from memory and interacts with the red light transmitting and receiving module, the infrared transmitting and receiving module, the inertial sensor, the ambient light sensor, the pressure sensor, and the display or upload module. Optionally, it can also interact with a miniature camera to obtain low-dimensional quality features of the skin surface image around the probe.

[0042] Step S1: Perform multimodal synchronous acquisition and initialization reference establishment to obtain a dual-wavelength photoplethysmography sequence and an auxiliary mode sequence with a unified timestamp.

[0043] The main body executing this step is the data acquisition terminal processor. The input is the raw sampling signals from each sensor after the subject has finished wearing the device. The purpose of processing is to generate synchronous multimodal data that can be used for subsequent state decomposition and to establish individualized reference pressure and reference perfusion under stable conditions. The output is a sequence of red photoplethysmography (RPG) recordings, infrared photoplethysmography (IPG) recordings, inertial sequences, ambient light sequences, pressure sequences, and optional image quality feature sequences with uniform timestamps, and generates initial reference parameters.

[0044] S101, the processor controls the red light emitter, infrared emitter, and corresponding photodetector to work alternately, simultaneously acquiring red and infrared photoplethysmography (PPG) signals; simultaneously, it acquires three-axis inertial signals from a three-axis accelerometer and a three-axis gyroscope, obtains the ambient light intensity from an ambient light sensor, and acquires the probe contact pressure from a pressure sensor. If the device is equipped with a miniature camera, it simultaneously acquires images of the skin surface around the probe and extracts at least one low-dimensional quality feature locally, including average brightness of the area, skin color stability, and occlusion ratio. The original image frame, after low-dimensional feature extraction, is no longer used as input for subsequent processes, thereby reducing privacy exposure risks and transmission burden.

[0045] S102. The processor adds a unified timestamp to each modal sampling point. The unified timestamp is used to represent the same moment or time period to which each modal sample belongs, so that subsequent resampling and alignment processes can be synchronized with the time axis of the dual-wavelength photoplethysmography signal as a reference. After this processing, sensor data with different sampling frequencies can be mapped to a unified time reference, and subsequent state calculations no longer rely on manual alignment.

[0046] S103. After the device is worn, the processor enters the initialization and stabilization phase, evaluating whether the current wearing state meets the conditions for establishing a stable reference through multiple consecutive time windows. The purpose of this phase is not to immediately output the blood oxygen calibration value, but to first find a window where the probe fit and perfusion are relatively stable, in order to establish the individualized mean reference pressure and reference perfusion index. If a window that meets the stabilization conditions is not obtained during the initialization phase, the processor prompts for readjustment of the probe fit.

[0047] S104. When a time window meeting the stability conditions first appears during the initialization stabilization phase, the processor aggregates the pressure averages of that window and its adjacent stable windows to obtain a reference pressure average, aggregates the perfusion indices in the corresponding windows to obtain a reference perfusion index, and records the reference pressure fluctuation parameters. If a reference pressure average is not obtained during the initialization phase, a preset pressure reference value is used as a temporary reference value, and it is updated to an individualized reference value after the time window meeting the stability conditions for the first time appears. If the reference perfusion index is lower than a preset lower limit, the use of perfusion-related terms is suspended in subsequent contact stability index calculations.

[0048] Step S2: Perform heterogeneous data preprocessing and time alignment to obtain window-level multimodal input data.

[0049] The main body executing this step is the data acquisition terminal processor. The input is the original time series of each modality obtained in step S1. The processing aims to remove obviously invalid samples, extract basic features suitable for state estimation, and align different frequency modes to a unified time axis. The output is window-level multimodal input data and confidence labels for each modality.

[0050] S201, the processor performs saturation sample removal and abnormal code removal on the red photoplethysmography (PPG) signal and the infrared PPG signal. Saturation samples indicate sample points where the receiver has reached the upper or lower quantization limit and cannot reflect the true optical changes; abnormal codes indicate invalid coded values ​​generated by analog-to-digital conversion errors or sensor link malfunctions. Removing these invalid samples can reduce the distortion of subsequent AC and DC component estimations.

[0051] S202. The processor performs smoothing processing on the discarded dual-wavelength photoplethysmography signal and separates the AC candidate component and DC component. The AC candidate component is used for heartbeat localization and valid heartbeat determination, while the DC component is used for calculating the environmental incident interference index and constructing the heartbeat-level AC-DC ratio. The bandpass processing result here is only used as a candidate waveform and is not directly used as the final calibration basis to avoid mistaking the artifact filtering result for the true blood oxygenation characterization.

[0052] S203. The processor extracts the acceleration modulus, angular velocity modulus, and their changes from the triaxial inertial signal. The acceleration and angular velocity moduli reflect the relative motion intensity between the probe and the tissue, and their changes are used to enhance the perception of abrupt changes in motion. The processed inertial features will serve as external evidence of motion contamination.

[0053] S204. The processor performs peak removal processing on the ambient light and pressure signals while retaining the slow-varying trend term. The ambient light slow-varying trend is used to identify the impact of lighting switching, screen flickering, or occlusion changes on the dual-channel DC baseline; the pressure slow-varying trend is used to identify probe over-looseness, over-tightness, or localized pressure changes. If the original auxiliary modes have short-term spikes, this processing can reduce the impact of occasional spikes on state estimation. S205. The processor resamples the auxiliary modes to the time axis of the dual-wavelength photoplethysmography signal according to a unified timestamp, forming window-level multimodal input data. Interpolation is used for alignment of continuous numerical signals, and state preservation is used for alignment of categorical state signals. After resampling, each auxiliary mode can establish a correspondence with the dual-wavelength waveform in the same window and at the same sampling position, which is beneficial for subsequent sample-level or window-level correlation analysis.

[0054] S206. The processor calculates the missing proportion of each auxiliary modality within the current time window. For gaps where the missing proportion is less than a preset threshold, local interpolation is used for completion. When the missing proportion of any auxiliary modality within the current window exceeds the preset threshold, the auxiliary modality is marked as a low-confidence modality, and its corresponding weight is reduced during subsequent intermediate state fusion. If a modality is missing for the entire window, other modalities are retained and execution continues without interrupting the entire process. This degradation process allows the system to still output usable results when some sensor links are unavailable.

[0055] Step S3: Decompose the state of the interference source to generate the motion pollution state, environmental interference state, contact stability state, and window confidence.

[0056] The main body executing this step is the data acquisition terminal processor. The input is the window-level multimodal input data obtained in step S2. The purpose of the processing is not to directly map the multimodal input to blood oxygen value, but to explicitly estimate the intermediate states that affect calibration stability. The outputs are the synchronization artifact coupling index, environmental incident interference index, contact stability index, and window confidence level.

[0057] S301. The processor reads the first-order change information of the red light waveform and infrared waveform within the current time window, and calculates the synchronization artifact coupling index according to the aforementioned formula, based on the inertial characteristic modulus within the current window. The consistency of the changes in the two-wavelength derivatives is used to measure whether the true pulse morphology is dominant; the inertial fluctuation term is used to supplement motion evidence. If the inertial feature fluctuation is small in the current window, the judgment mainly relies on the synchronicity of the two-wavelength morphology; if the inertial mode output is abnormal or unavailable, a conservative estimate of motion contamination can be obtained based on historical weighted average or by retaining only the waveform synchronization term.

[0058] S302. The processor reads the average red DC baseline value, the average infrared DC baseline value, and the ambient light characteristics of the current time window from the previous time window, and calculates the ambient incident interference index according to the aforementioned formula. If the current window is the first window, the difference term between adjacent windows is initialized to zero; if the ambient light mode is missing, the dual-channel DC baseline change term is retained, and the ambient light fluctuation term is set to zero. After this processing, the system can distinguish between short-term synchronous abrupt changes in the dual-channel DC baseline and real physiological changes.

[0059] S303. The processor reads the average pressure of the current time window, the average reference pressure obtained during the initialization phase, the current window perfusion index, and the reference perfusion index, and calculates the contact stability index according to the aforementioned formula. If the reference injection index is lower than the preset lower limit, the injection-related index items are suspended, and the contact stability index is calculated only based on the degree of pressure deviation. This avoids the injection item from unreasonably amplifying the contact stability estimate under extreme low injection conditions.

[0060] S304, processor combination , , and the reliability of the waveform quality feature generation window If an image modality is deployed, its image quality features are included in the overall quality score; if an image modality is not deployed, its corresponding weights are reset to zero, and the weights of other quality items are renormalized. If an auxiliary modality is marked as a low-confidence modality, its weight is simultaneously reduced during fusion. After this processing, It reflects whether the current window is suitable as a calibration sample, rather than the blood oxygen level of the current window.

[0061] Step S4: Perform trusted fragment selection and robustness ratio construction.

[0062] The main body executing this step is the acquisition terminal processor. The inputs are the window confidence level obtained in step S3 and the dual-wavelength photoplethysmography waveform. The processing objective is to construct robust ratios only for windows that meet the confidence criteria and to prepare features for the dynamic calibration model. The outputs are the robust ratios, calibration feature vectors, or uncalibrable window markers.

[0063] S401. The processor compares the current window confidence level with a preset confidence threshold. If the window confidence level is lower than the confidence threshold, the current window will not enter the heart rate ratio construction process, but will be directly recorded as a time window that cannot be used for calibration, and will be transferred to the subsequent reference output or degraded output process. This gating mechanism moves the "whether calibration is allowed" decision to the beginning of the calibration model, which can reduce the probability of abnormal windows contaminating parameters.

[0064] S402. When the current window meets the confidence criteria, the processor locates the valid heartbeat cycle within the window. Valid heartbeat location can be achieved based on the peak-trough relationship of the AC candidate components, the zero-crossing information of the derivative, and the peak spacing constraint. If the detected heartbeat cycle does not meet the preset heart rate range or has an abnormal waveform, the corresponding heartbeat is not considered a valid heartbeat for subsequent calculations. The output of this step is the set of valid heartbeats within the current window.

[0065] S403. For each valid heartbeat, the processor calculates the AC amplitude and DC baseline in the red light channel, and the AC amplitude and DC baseline in the infrared channel. If the DC component corresponding to any valid heartbeat is lower than the minimum effective illuminance threshold, the heartbeat is discarded to avoid constructing an unstable AC-DC ratio under insufficient illuminance conditions.

[0066] S404. The processor calculates the ratio of red light AC / DC to infrared AC / DC for all valid heartbeats, and calculates the ratio for the current window according to the aforementioned robust ratio formula. If the number of valid heartbeats in the current window is less than the preset number, no output will be generated. The current window is then determined to be unusable for calibration. By using median aggregation, the influence of local abnormal heartbeats on the overall window results is suppressed, allowing subsequent calibration models to operate on a more stable optical ratio basis.

[0067] S405, The processor constructs the calibration feature vector for the current window. (Except for...) In addition, it includes artifact index smoothing value, environmental disturbance index smoothing value, contact stability index, interpeak variation index, perfusion index, ambient light baseline level, and surface quality stability index. If an image mode is missing, the surface quality stability index can be replaced by a non-image quality index or left blank, and will be handled by the model side in the default manner.

[0068] Step S5: Perform dynamic calibration and selective online updates for the trustworthiness gating.

[0069] The main body executing this step is the data acquisition terminal processor. The inputs are the calibration feature vector and window confidence obtained in step S4. The processing objective is to output the current window's blood oxygen calibration value and, when conditions are met, to perform limited fine-tuning of individualized parameters. The outputs are the blood oxygen calibration value and the updated parameter status.

[0070] S501, the processor determines whether the current window simultaneously meets the dynamic calibration calculation conditions, namely, the window confidence reaches the calibration execution condition and the robustness ratio is not missing. If either condition is not met, the dynamic calibration formula is not executed, and the process switches to step S6. This avoids generating seemingly continuous but unfounded calibration values ​​when core features are missing.

[0071] S502. For windows that meet the conditions, the processor calls the aforementioned confidence-gated dynamic calibration model to calculate the current window's blood oxygen calibration value based on the calibration feature vector and window confidence. If the model output exceeds the preset effective blood oxygen range, amplitude limiting and pruning are performed. This is because the model explicitly includes... When the window confidence level reaches the threshold but is still below the high confidence level, the model output will automatically become more conservative, thereby reducing the over-correction caused by the boundary window.

[0072] S503: The processor calculates the average window confidence, average environmental interference level, and average contact stability level for the most recent consecutive time windows. Individualized calibration parameter updates are only allowed if the average window confidence is not lower than the confidence update threshold, the average environmental interference level is not higher than the environmental update threshold, and the average contact stability level is not lower than the contact update threshold. If any condition is not met, the parameters are frozen, and the current parameter version continues to be output.

[0073] S504. When the online update conditions are met, the processor uses recursive least squares or an equivalent low-complexity recursive parameter estimation method to fine-tune the intercept and principal coefficient parameters, instead of freely updating all parameters. This adapts to the slow changes in individual skin reflectivity and adhesion differences while controlling the model's degrees of freedom, reducing the risk of instability during online learning. After the update is complete, the processor synchronously refreshes the parameter version number for use in subsequent evidence chain recordings.

[0074] Step S6: Perform abnormal degradation and output the results.

[0075] This step is executed by the data acquisition terminal processor. Inputs include the current window confidence level, dynamic calibration results, or the most recent high-confidence output history. The processing objective is to output calibrable values, reference values, or degradation values ​​based on the confidence level, and to generate quality markers and alarm information. Outputs are the final displayed or uploaded blood oxygenation results and their quality status.

[0076] S601. When the confidence level of the current window is not lower than the first confidence threshold, and the current window has obtained a valid calibration value, the processor outputs a calibrable blood oxygen value. Simultaneously, it records the evidence chain corresponding to this output. The evidence chain includes at least the motion contamination state, environmental disturbance state, contact stability state, robustness ratio, and the current parameter version number. This evidence chain supports subsequent review, auditing, and multi-center consistency tracking.

[0077] S602. When the confidence level of the current window is lower than the first confidence threshold but not lower than the second confidence threshold, the processor outputs a reference blood oxygen value with a medium-risk label, and simultaneously blocks calibration parameter updates. At this time, the system still provides continuous reference readings, but explicitly indicates that the current window is not used as a basis for parameter learning. This strategy allows the system to maintain continuous readings under medium-risk conditions while avoiding incorporating edge-quality segments into the parameter evolution process.

[0078] S603. When the current window's confidence level is lower than the second confidence threshold, the processor enters a degraded output mode. It prioritizes outputting the blood oxygen value retention result corresponding to the most recent high-confidence time window; if there are enough historical high-confidence windows, it can generate degraded blood oxygen values ​​based on the aforementioned short-term amplitude-limiting prediction method using the most recent high-confidence window history. If the prediction result exceeds a preset physical range, pruning is performed. If there are insufficient historical high-confidence windows, only the most recent available value is retained, and extrapolation is not performed.

[0079] S604. The processor synchronously generates quality markers or alarm messages for the reference output or degraded output. The alarm messages at least indicate that the current acquisition conditions do not meet the calibration requirements, and if necessary, prompt the system to readjust the probe or remain stationary. After this processing, the system can maintain continuous output capability even in scenarios such as short-term displacement, strong light flashing, or momentary subject movement, while avoiding misinterpreting low-confidence output as high-quality calibration results.

[0080] In one optional implementation, when the method of the present invention involves remote data uploading or result transmission between a terminal and a host computer, the output result frame may include at least one of the following: result number, device number, generation timestamp, window start and end time identifier, blood oxygen value, quality grade, parameter version number, and integrity check code. The receiving end can determine whether the result is within a valid receiving window based on the timestamp and integrity check code, and confirm whether truncation, tampering, or out-of-order transmission has occurred during transmission. This field is designed to support the traceability of clinical trial data and is not limited to a single implementation.

[0081] Example 2: Finger clip continuous pulse oximetry calibration in bedside drug clinical trials

[0082] This embodiment addresses a bedside continuous monitoring scenario in a Phase I or Phase II drug clinical trial ward. Blood oxygen saturation data is continuously collected from subjects before and after drug administration, but situations such as turning over, talking, raising hands, nurse checks, changes in ward lighting, and slight probe loosening may occur. Traditional single-mode waveform filtering typically only coarsely removes obviously abnormal segments, making it difficult to prevent semi-contaminated windows from entering the calibration process and causing continuous curve drift.

[0083] This embodiment uses a finger-clamp terminal, which includes a dual-wavelength red and infrared light transmitting and receiving module, a six-axis inertial sensor, an ambient light sensor, and a pressure sensor. The red light wavelength is... The infrared wavelength is The sampling frequency of the photoplethysmography signal is The inertial data sampling frequency is The sampling frequency for both ambient light and pressure data is [missing information]. The length of the sliding window is The sliding step size is The confidence threshold is The number of parameter update windows is... The degradation output threshold is The normal operating pressure range is... The normal range for ambient light is .

[0084] In this embodiment, after the device is worn, it first enters... Initialization stabilization phase. During this phase, the processor searches for at least continuous... A credibility greater than A time window is established to set a reference mean pressure, reference perfusion index, and initial DC baseline. If these conditions are not met, a probe readjustment is prompted. After this step, the contact stabilization state has an individualized reference, rather than relying on a factory-standard threshold.

[0085] After entering the continuous acquisition phase, taking a certain time window as an example, the processor obtains... Red light photoplethysmography sample, Point infrared photoplethysmography sample, Point inertial sample, Point ambient light samples and Point pressure samples are used. If a subject raises their hand to speak within the window, the inertial characteristic modulus fluctuation increases, and the difference between the red light and infrared derivatives increases, resulting in a higher synchronization artifact coupling index calculated by the processor. Simultaneously, if the ward lighting changes, causing an increase in the average ambient light and slight probe loosening, the environmental incident interference index increases, and the contact stability index decreases. After fusion, if the window's reliability drops below the reliability threshold, the processor no longer allows that window to enter calibration updates, only outputting a reference blood oxygen value with a medium-risk label. This mechanism effectively blocks abnormal window contamination parameters.

[0086] After the subject returns to rest, the processor locates the effective heartbeat within a new high-confidence window, constructs robust ratios and calibration features such as perfusion index, ambient light level, and surface quality indicators, and then calls the dynamic calibration model to output the blood oxygen calibration value. If recently... If the average confidence level, average environmental interference level, and average contact stability level of each window simultaneously meet the online update conditions, then the intercept parameter and principal coefficient parameter can be fine-tuned to adapt to the current individual fit status and reflection characteristics.

[0087] When a subject's short-term intense activity causes the confidence level to fall below the degradation threshold for multiple consecutive windows, the system stops calibration output and instead maintains the most recent high confidence value or performs short-term amplitude-limited prediction, prompting the system to readjust the probe or remain stationary. This degradation mechanism allows the system to maintain short-term continuous output even when acquisition conditions temporarily do not meet calibration requirements, and avoids unreasonable large jumps.

[0088] In the exemplary simulation results, a digital twin acquisition environment was constructed that included minor movements, lighting changes, probe loosening, and low perfusion fluctuations. The simulation duration was [duration missing]. , total simulation The study included [number] subjects. The control group used single-modal acquisition, fixed threshold filtering, and conventional ratio mapping. Results showed that this invention effectively reduces the mean square error of blood oxygenation, median absolute error, parameter error update rate, and number of consecutive output jumps. These results indicate that this invention reduces the risk of semi-contaminated segments mistakenly entering the calibration process through confidence-based pre-gating and state decomposition mechanisms.

[0089] Example 3: Consistency Calibration of Attached Oxygen Monitoring in a Multicenter Clinical Trial

[0090] This embodiment is designed for a multi-center clinical trial scenario. Although different trial centers may use the same batch of adhesive pulse oximeters, there are differences in ward lighting conditions, nursing habits, subject skin color, and peripheral perfusion conditions. Traditional procedures typically rely on manual annotation of obvious artifacts followed by a standardized calibration curve, which can easily lead to increased inter-center bias.

[0091] This embodiment uses an adhesive finger base terminal and adds an optional low-resolution skin surface image quality module, extracting only three low-dimensional indicators: average brightness, occlusion ratio, and reflection stability. The photoplethysmography signal sampling frequency is... The inertial data sampling frequency is The update frequency of ambient light, pressure, and image quality features is: Window length is Step size is The calibration output cycle is The confidence threshold is Online update conditions are still based on the most recent Window statistics; image occlusion alarm threshold is... .

[0092] This embodiment first performs multi-center pre-training. The processor or training server imports window-level samples from multiple test centers. Each sample contains a calibration feature vector, window confidence level, and reference blood oxygen label. During training, a confidence-weighted strategy is used, and stratified sampling is performed by center to avoid excessive sample size in one center, which could lead to model bias. The obtained basic parameters are then distributed to each center terminal as the factory calibration model.

[0093] If, after a subject wears the device, the ward lighting is too strong, the ambient light baseline is consistently high, and peripheral perfusion is low, then a traditional uniform fixed curve might misinterpret a decrease in AC amplitude as a decrease in blood oxygenation. This embodiment allows only fine-tuning of the intercept and principal coefficient parameters within a continuous high-confidence window, and uses perfusion index, ambient light level, and surface quality indicators for joint correction, thereby reducing the impact of central environmental differences and individual low perfusion differences on the consistency of results.

[0094] During online operation, if a window is obstructed by nursing operations, causing the image occlusion ratio to exceed the alarm threshold, and simultaneously, localized ambient light flicker increases the environmental incident interference index, but the inertial change is not significant and the pressure remains stable, the system can identify this window as a non-motion, low-confidence window, outputting only the reference blood oxygen value without triggering parameter updates. Once the occlusion of the next window is resolved, the surface quality indicators recover, the confidence level rises, and the system immediately resumes normal calibration output. This process demonstrates that the present invention can not only determine whether an anomaly occurs but also use state combinations to reflect the source of the anomaly, thereby enabling more targeted processing strategies.

[0095] In low-perfusion nighttime scenarios, the decrease in temperature at the subject's fingertips reduces the AC component. Traditional fixed-threshold methods easily misclassify such windows as artifacts and directly reject them; however, in this embodiment, because the synchronization artifact coupling index and environmental incident interference index are kept low, and the contact stability index is kept high, the window is still allowed to enter the calibration process under credible conditions, and the conservatism of the output is controlled by the perfusion index, thereby reducing the misjudgment of "low perfusion equals low quality".

[0096] In the exemplary simulation results provided in this invention, the construction A multi-condition joint simulation set at multiple test centers, with each center simulating... Continuous data acquisition was performed, including data on skin color differences, illumination differences, low perfusion differences, and localized occlusion events. The control group consisted of manual artifact identification, conventional frequency domain filtering, and a unified calibration curve. Results showed that this invention reduces cross-center mean bias, the proportion of erroneous calibration entries in the low confidence window, and the 95th percentile absolute error. This indicates that this invention can improve the consistency and auditability of blood oxygen acquisition calibration under multi-center clinical trial conditions.

[0097] Example 4: Electronic Equipment

[0098] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of a multimodal fusion-based clinical trial blood oxygen acquisition and calibration method described in any of the embodiments one to three.

[0099] The memory stores program instructions, a parameter dictionary, window cache data, historical high-reliability output records, and parameter version information. The processor invokes program instructions to perform multimodal synchronous acquisition control, preprocessing and alignment, state decomposition, reliable segment selection, dynamic calibration, online parameter updates, and abnormal degradation output. If necessary, the electronic device may also include a data acquisition interface connected to sensors, a display interface connected to a display terminal, and a communication interface connected to a host computer or clinical trial data platform.

[0100] In one implementation, the processor, when executing a computer program, forms an acquisition function module, a state decomposition function module, a reliability determination function module, a dynamic calibration function module, and a result output function module. These function modules are logical functional modules formed when the processor executes the program and do not constitute additional limitations on the structure of the independent device.

[0101] Terms such as "first" and "second" are used only to distinguish different objects or different processing states, and do not indicate an actual order, primary or secondary relationship, or quantity limitation. Terms such as "including" and "containing" are open-ended expressions and do not exclude the existence of other elements or steps not explicitly listed.

[0102] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the principles and spirit of the present invention should be included within the scope of protection of the present invention.

Claims

1. A clinical trial blood oxygen acquisition calibration method based on multimodal fusion, characterized in that, include: Simultaneously acquire the subject's red light photoplethysmography signal, infrared photoplethysmography signal, and at least one auxiliary modal signal. The auxiliary modal signal includes at least one of triaxial inertial signal, ambient light signal, probe contact pressure signal, and skin surface image quality features. A uniform timestamp is added to each modal signal in the auxiliary modal signal. Each modal signal is aligned to the time axis of the red light photoplethysmography signal and the infrared photoplethysmography signal. Based on the aligned dual-wavelength photoplethysmography signal and modal signals, an intermediate state reflecting motion contamination, environmental interference, and contact stability is constructed. The motion contamination is characterized by the morphological synchronization deviation and inertial fluctuation of the dual-wavelength photoplethysmography signal. The environmental interference is characterized by the change of the dual-channel DC baseline and the fluctuation of ambient light. The contact stability is characterized by the change of probe pressure and the change of pulse perfusion. The window confidence level of the current time window is generated based on the intermediate state, and it is determined whether the current time window is allowed to enter the blood oxygen calibration process based on the window confidence level. When the window confidence level meets the confidence condition, the AC and DC components of the red light channel and infrared channel are extracted at the heartbeat granularity, the robust ratio is constructed, and dynamic calibration is performed in combination with perfusion information, contact stability information, ambient light level and surface quality information to obtain the blood oxygen calibration value of the current time window. When the confidence level of the window does not meet the confidence condition, the calibration parameter update is frozen, and a reference blood oxygen value or a degraded blood oxygen value is output according to the interval where the confidence level of the window is located. At the same time, a quality mark or alarm information corresponding to the output is generated.

2. The clinical trial blood oxygen acquisition calibration method based on multimodal fusion according to claim 1, characterized in that, Aligning each auxiliary mode signal to the time axis of the red photoplethysmography signal and the infrared photoplethysmography signal includes: The red photoplethysmography signal and the infrared photoplethysmography signal are subjected to saturation sample removal, outlier code removal, smoothing, and separation of AC candidate components and DC components. The acceleration modulus, angular velocity modulus and their changes are extracted from the triaxial inertial signal, and the slow-change trend characteristics after peak removal are extracted from the ambient light signal and the probe contact pressure signal. The auxiliary modal signals are resampled to the time axis of the photoplethysmography signal according to a unified timestamp to form window-level multimodal input data, wherein continuous numerical signals are aligned by interpolation and category state signals are aligned by state preservation. When the proportion of missing auxiliary modalities in the current window exceeds a preset threshold, the auxiliary modality is marked as a low-confidence modality, and the corresponding weight of the auxiliary modality is reduced during subsequent intermediate state fusion.

3. A clinical trial blood oxygen acquisition calibration method based on multimodal fusion according to claim 1 or 2, characterized in that, The skin surface image quality features are low-dimensional quality features extracted from the skin surface image around the probe acquired by the miniature camera. The low-dimensional quality features include at least one of regional average brightness, skin color stability, and occlusion ratio. The skin surface image quality features only participate in the intermediate state generation and dynamic calibration in the form of feature values, and do not transmit or store the original image; When no image modality is deployed, the corresponding feature of the image modality is set to null, and its corresponding weight in window confidence generation is reset to zero.

4. The clinical trial blood oxygen acquisition calibration method based on multimodal fusion according to claim 1, characterized in that, The construction of the intermediate state reflecting motion pollution, environmental disturbance, and contact stability includes: Based on the difference in first-order changes between the red photoplethysmography signal and the infrared photoplethysmography signal within the current window, and combined with the fluctuation intensity of the inertial characteristics within the current window, a synchronization artifact coupling index is generated to characterize the motion contamination state. An environmental incident interference index is generated based on the changes in the red light DC baseline and infrared DC baseline between the current window and the previous window, as well as the fluctuations in ambient light characteristics, to characterize the environmental interference state. A contact stability index is generated based on the deviation of the average probe pressure in the current window from the reference pressure and the deviation of the pulse perfusion index in the current window from the reference perfusion index, in order to characterize the contact stability state. The intermediate state is obtained by fusing the synchronization artifact coupling index, the environmental incident interference index, the contact stability index, and the waveform quality characteristics.

5. The clinical trial blood oxygen acquisition calibration method based on multimodal fusion according to claim 4, characterized in that, The contact stability index Determined according to the following formula: ， in, Indicates the first Contact stability index for each time window; Indicates the first Average probe pressure over a time window; This represents the average reference pressure obtained during the initial stabilization phase. The parameter representing the fluctuation of the reference pressure; Indicates the first Pulse perfusion index for each time window; This represents the reference perfusion index obtained during the initial stable phase; This indicates a smoothing term to prevent the denominator from being zero; When the average reference pressure is not obtained during the initial stabilization phase, a preset pressure reference value is first used as a temporary reference value, and then updated to an individualized reference value after the time window for first meeting the stabilization conditions appears. When the reference perfusion index is lower than the preset lower limit, the use of index items related to the pulse perfusion index is suspended, and the contact stability index is determined only based on the degree of pressure deviation.

6. The clinical trial blood oxygen acquisition calibration method based on multimodal fusion according to claim 1, characterized in that, When the window confidence level meets the confidence condition, the AC and DC components of the red and infrared channels are extracted at the heartbeat granularity to construct a robust ratio, including: Within the current time window, locate the effective heartbeat cycle and calculate the AC amplitude and DC baseline of each effective heartbeat in the red light channel and infrared channel respectively; For each valid heartbeat, the ratio of red light AC / DC ratio to infrared AC / DC ratio is constructed; The ratio of the robust ratios is obtained by aggregating the median values ​​of all valid heartbeats within the current time window.

7. A clinical trial blood oxygen acquisition calibration method based on multimodal fusion according to claim 1 or 6, characterized in that, The process of performing dynamic calibration to obtain the blood oxygen calibration value for the current time window includes: Construct a calibration feature vector that includes at least the ratio of robustness ratios, artifact index smoothing value, environmental disturbance index smoothing value, contact stability index, perfusion index, ambient light baseline level, and surface quality stability index; The calibration feature vector is input into the confidence-gated dynamic calibration model to obtain the blood oxygen calibration value for the current time window.

8. The clinical trial blood oxygen acquisition calibration method based on multimodal fusion according to claim 7, characterized in that, The calibration parameter updates employ a selective online update mechanism, including: The average window reliability, average environmental disturbance level, and average contact stability level of the most recent consecutive time windows were statistically analyzed. Individualized calibration parameters may be updated only when the average window confidence is not lower than the confidence update threshold, the average environmental interference level is not higher than the environmental update threshold, and the average contact stability level is not lower than the contact update threshold. The individualized calibration parameter update only includes fine-tuning of the intercept parameter and principal coefficient parameter, without freely updating all calibration parameters; The fine-tuning of the individualized calibration parameters is achieved using recursive least squares or an equivalent low-complexity recursive parameter estimation method.

9. The clinical trial blood oxygen acquisition calibration method based on multimodal fusion according to claim 1, characterized in that, The step involves outputting a reference blood oxygen value or a degraded blood oxygen value according to the range of the window's confidence level, and simultaneously generating a quality marker or alarm information corresponding to the output, including: When the confidence level of the window is not lower than the first confidence threshold, the calibrable blood oxygen value is output and the evidence chain corresponding to the calibrable blood oxygen value is recorded. The evidence chain includes at least the motion pollution state, environmental disturbance state, contact stability state, robustness ratio ratio and current parameter version number. When the confidence level of the window is lower than the first confidence threshold and not lower than the second confidence threshold, the reference blood oxygen value is output and a medium-risk quality mark is attached, while the calibration parameter update is blocked. When the confidence level of the window is lower than the second confidence threshold, the blood oxygen value maintenance result corresponding to the most recent high confidence time window is output, or the short-term amplitude limiting prediction result based on the most recent high confidence time window is output, and an alarm message is generated indicating that the current acquisition conditions do not meet the calibration requirements.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of a clinical trial blood oxygen acquisition calibration method based on multimodal fusion as described in any one of claims 1-9.

Images