Multi-part sensing and machine vision fused vital sign monitoring method for large-area burn patient

Abstract

The invention discloses a multi-part sensing and machine vision fused vital sign monitoring method for a large-area burn patient, relates to the technical field of medical monitoring, and is used for solving the problems of monitoring blind areas and shock early warning lag in a large-area burn scene. The method comprises the following steps: firstly, combining waveform statistical characteristics and body temperature change, and accurately dividing a hypoperfusion channel and a fall-off channel; then, eliminating eschar dressing interference based on a three-dimensional mask and chromaticity features, and extracting high-signal-to-noise-ratio non-contact signs; then, a self-adaptive weighting matrix is constructed according to the channel state, and the visual weight is dynamically improved to compensate a contact blind area; finally, by analyzing the cross covariance of the heart rate and the perfusion index, haemodynamic discrete features are captured, shock compensation period advanced early warning is achieved, and continuity and accuracy of monitoring data are guaranteed.

Description

Technical Field

[0001] This invention relates to the field of medical monitoring technology, specifically to a vital signs monitoring method that integrates multi-site sensing and machine vision for patients with extensive burns. Background Technology

[0002] Extensive burns are a critical and severe condition. Patients are highly susceptible to hypovolemic shock in the early post-injury period. During this time, accurate and continuous monitoring of vital signs is crucial for guiding fluid resuscitation, assessing the effectiveness of anti-shock treatment, and detecting sudden changes in the patient's condition. Because burn patients have damaged skin barriers and experience massive fluid exudation, their bodies are in an extremely unstable physiological state. Any abnormal fluctuations or failures in monitoring data can affect medical decisions. Therefore, establishing a highly reliable vital sign monitoring system is of great significance for saving patients' lives.

[0003] Current clinical monitoring primarily relies on contact sensors to collect physiological parameters, but this approach has significant limitations when dealing with patients with extensive burns. Firstly, patients often have severe wounds or thick dressings covering their limbs and trunk, making it difficult to find suitable placement for conventional clip-on or patch-type probes. Furthermore, wound exudate can interfere with the stability of sensor-skin contact. Secondly, burn patients in shock often experience severe microcirculatory disturbances and cold, clammy extremities, resulting in weak or absent peripheral pulse wave signals, causing contact monitoring devices to fail to read data or generate frequent false alarms. In addition, relying solely on contact methods cannot automatically compensate for probe detachment or signal quality degradation using other modalities, leading to monitoring blind spots in the system. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a vital sign monitoring method that integrates multi-site sensing and machine vision for patients with extensive burns, thus solving the problems mentioned above.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a vital sign monitoring method integrating multi-site sensing and machine vision for patients with large-area burns, comprising the following steps: S1: Simultaneously acquiring photoplethysmography (PPG), body surface temperature, and a color video stream containing depth information from multiple anatomical test points of the patient; calculating waveform statistical features to generate a signal quality index; classifying channels with signal quality indices below an effective threshold based on the change characteristics of the corresponding body surface temperature data; marking channels with continuously decreasing or normal temperatures as low-perfusion channels or detachment channels, respectively; and outputting a state classification vector and a blood perfusion index sequence; S2: Constructing a three-dimensional background mask using depth information to remove video background noise; performing pixel-level segmentation based on the chromaticity features of burn eschar and medical dressings; calculating the gray-level variance in the remaining region and selecting the connected component with the smallest variance to define... S3: For the effective signal extraction region, spatial averaging is performed on the pixel data within this region to obtain the original color channel signal. Non-contact heart rate and respiratory rate are extracted through blind source separation operation; S4: The contact pulse rate and non-contact heart rate are time-aligned. An adaptive weighting matrix is ​​constructed based on the channel state classification vector and signal quality index. The weight coefficients of each channel are dynamically adjusted using the normalized signal quality index. When a low-perfusion channel or a detached channel is identified, its weight is automatically reduced and the weight of the non-contact heart rate is increased. Weighted fusion is performed on the multi-source data to output a fused heart rate sequence; S5: A time sliding window is constructed. The fused heart rate sequence and the blood perfusion index sequence are input simultaneously. The cross-covariance of the two sets of sequences is calculated within the sliding window. When the cross-covariance is less than the preset negative correlation threshold and the variance of the fused heart rate sequence is continuously greater than the fluctuation threshold, a hemodynamic discrete early warning signal is generated.

[0006] Furthermore, the specific process of simultaneously acquiring photoplethysmography (PPG), body surface temperature, and color video stream containing depth information from multiple anatomical test points of the patient, and calculating the signal quality index by waveform statistical features is as follows: Bandpass filtering is performed on the acquired PPG signal to suppress high-frequency noise; a detrending algorithm is used to eliminate baseline drift; a time segment of preset length is extracted; and the skewness coefficient, kurtosis coefficient, and Shannon entropy of the signal amplitude within the time segment are calculated respectively. Interval normalization mapping is performed on the skewness coefficient, kurtosis coefficient, and Shannon entropy; a preset feature weight vector is applied to linearly weighted sum the three normalized feature values; and the weighted sum is output as the signal quality index for the corresponding time segment.

[0007] Furthermore, based on the changing characteristics of the corresponding body surface temperature data, channels with signal quality indices below the effective threshold are classified into state attributes. Channels with continuously decreasing or normal temperatures are marked as low-perfusion channels or detached channels, respectively. The specific process of outputting the state classification vector and blood perfusion index sequence is as follows: Identify the peak and trough amplitudes of the photoplethysmography (PPG) signal, extract the AC and DC components from the waveform, and calculate the ratio of the AC to DC components to generate the blood perfusion index sequence; calculate the first-order difference of the time series of body surface temperature data. When the signal quality index of the channel to be determined is below the effective threshold, if the first-order difference corresponding to the channel to be determined is negative and its absolute value is greater than the preset temperature change rate threshold, the channel to be determined is marked as a low-perfusion channel in the state classification vector; if the absolute value of the first-order difference corresponding to the channel to be determined is less than the preset temperature fluctuation tolerance, the channel to be determined is marked as a detached channel in the state classification vector.

[0008] Furthermore, a 3D background mask is constructed using depth information to remove video background noise. Pixel-level segmentation is performed based on the chromaticity features of burn eschar and medical dressings. The specific process of calculating the gray-level variance in the remaining region and selecting the connected component with the smallest variance as the effective signal extraction region is as follows: The depth image is mapped to the color video stream coordinate system using a coordinate transformation matrix. A depth truncation threshold is set to generate a binary depth mask. The depth mask is applied to the color video stream to perform a logical AND operation to retain near-field pixels. The retained pixels are converted to the YCrCb chromaticity space. A first filtering threshold interval is set according to the chromaticity distribution of burn eschar, and a second filtering threshold interval is set according to the chromaticity distribution of medical dressings. Pixels whose chromaticity components fall within the first or second filtering threshold interval are removed. Morphological opening operations are performed on the binary mask composed of the remaining pixels to eliminate isolated noise. The gray-level variance of each connected component in the luminance channel is calculated, and the connected component with the smallest gray-level variance is marked as the effective signal extraction region.

[0009] Furthermore, spatial averaging is performed on the pixel data within the effective signal extraction area to obtain the original color channel signals. The specific process of extracting non-contact heart rate and respiratory rate through blind source separation is as follows: The spatial mean of all pixels within the effective signal extraction area in the red, green, and blue channels is calculated frame by frame to construct a three-channel time-series signal. Zero-mean unit variance standardization is performed on the three-channel time-series signal. The standardized signal is demixed into independent source signal components using an independent component analysis algorithm. Fast Fourier transform is performed on each source signal component to obtain the power spectral density. The frequency corresponding to the maximum power spectral density within the preset heart rate frequency range is searched as the non-contact heart rate, and the frequency corresponding to the maximum power spectral density within the preset respiratory frequency range is searched as the respiratory rate.

[0010] Furthermore, the specific process of aligning contact pulse rate and non-contact heart rate data in time and constructing an adaptive weighted matrix based on channel state classification vectors and signal quality indices is as follows: Extract the data timestamps of each contact and non-contact channel; select the time axis of the contact photoplethysmography (PPG) signal with the highest sampling frequency as the master clock reference; perform linear interpolation resampling based on timestamp matching on the non-contact heart rate data to establish a point-to-point mapping relationship between multi-source data in the same time dimension; construct a fusion weight vector with dimensions adapted to the sum of the number of all channels; call the signal quality index of each contact channel to fill the corresponding element position of the fusion weight vector as the initial weight of the contact channel; and initialize the weight elements of the non-contact channel to a preset baseline confidence value.

[0011] Furthermore, the weight coefficients of each channel are dynamically adjusted using a normalized signal quality index. When a low-perfusion or detached channel is identified, its weight is automatically reduced while the weight of the non-contact heart rate is increased. The specific process of performing weighted fusion on multi-source data to output a fused heart rate sequence is as follows: Iteratively read the element labels in the channel state classification vector. When the label of the contact channel to be processed is a detached channel, the corresponding weight coefficient in the fusion weight vector is set to a logic zero value. When the label of the contact channel to be processed is a low-perfusion channel, a preset attenuation factor is applied to the corresponding signal quality index to perform a multiplication operation to update the weight coefficients. Calculate the sum of the updated weight coefficients of all contact channels, calculate the difference between the unit scalar 1 and the sum of the weight coefficients to obtain the residual weight margin, and add the residual weight margin to the non-contact channel weight coefficients. Perform L1 norm normalization on the fusion weight vector, calculate the vector dot product of the heart rate value of each channel and the normalized fusion weight vector, and output the fused heart rate sequence.

[0012] Furthermore, a time-sliding window is constructed, and the fused heart rate sequence and blood perfusion index sequence are input synchronously. The specific process of calculating the cross-covariance of the two sets of sequences within the sliding window is as follows: A fixed-capacity synchronous data buffer queue is established as a time-sliding window, and the time-aligned fused heart rate data and low-perfusion channel blood perfusion index data are stored in real time according to the first-in-first-out principle; the arithmetic mean of the fused heart rate data and blood perfusion index data in the buffer queue are calculated respectively, and a mean vector is obtained by tensor product operation between the all-1 vector and the arithmetic mean; the mean vector is generated by subtracting the mean vector from the original data vector in the buffer queue; the centralized fused heart rate vector and the centralized blood perfusion vector are generated; the dot product of the centralized fused heart rate vector and the centralized blood perfusion vector is calculated, and the dot product result is multiplied by the reciprocal of the buffer queue capacity value to output the real-time cross-covariance.

[0013] Furthermore, when the cross-covariance is less than a preset negative correlation threshold and the variance of the fused heart rate sequence is continuously greater than the fluctuation threshold, the specific process for generating a hemodynamic discrete early warning signal is as follows: Initialize the early warning judgment accumulator, compare the real-time cross-covariance with the preset negative correlation threshold in real time, and synchronously calculate the variance of the fused heart rate sequence within the sliding window and compare it with the preset fluctuation threshold; only when both the real-time cross-covariance is less than the negative correlation threshold and the variance of the fused heart rate sequence is greater than the fluctuation threshold are satisfied, perform an incremental operation on the early warning judgment accumulator; if either condition is not satisfied, perform a zeroing and reset operation on the early warning judgment accumulator; when the accumulated value of the early warning judgment accumulator exceeds the preset time confirmation threshold, trigger the alarm circuit to output a logic high-level signal as a hemodynamic discrete early warning signal.

[0014] The present invention has the following beneficial effects:

[0015] (1) A vital sign monitoring method integrating multi-site sensing and machine vision for patients with large-area burns. By simultaneously acquiring data from multiple anatomical points and analyzing waveform statistical features, combined with the characteristics of body surface temperature changes, it achieves accurate differentiation between hypoperfusion and probe physical detachment states, avoiding false alarms caused by cold and wet extremities during the shock period, and effectively improving the confidence of contact monitoring in complex physiological states. At the same time, the three-dimensional background mask constructed using depth information and color features and pixel-level segmentation technology can automatically remove eschar, dressings and background noise interference in the complex body surface environment of patients with large-area burns. By locking the effective area with the smallest gray-scale variance for non-contact signal extraction, it ensures that high-precision visual vital sign data can still be obtained when contact methods are limited.

[0016] (2) A vital sign monitoring method integrating multi-site sensing and machine vision for patients with extensive burns utilizes an adaptive weighted matrix to fuse contact and non-contact data from multiple sources. The weighting coefficients are dynamically adjusted based on channel status classification results. When a contact probe fails due to low perfusion or detachment, non-contact visual data is automatically and seamlessly compensated, eliminating blind spots in single monitoring modalities and ensuring the continuity and stability of monitoring data. Furthermore, by constructing a time-sliding window to calculate the cross-covariance between fused heart rate and perfusion index, the method can sensitively capture the negative correlation divergence trend between compensatory increases in heart rate and a sustained decrease in peripheral perfusion. This allows for the generation of hemodynamic discrete early warning signals before a sharp drop in blood pressure, providing clinicians with a valuable window for anti-shock treatment.

[0017] (3) Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0018] Figure 1This is a flowchart of a vital sign monitoring method that integrates multi-site sensing and machine vision for patients with large-area burns, according to the present invention. Detailed Implementation

[0019] This application provides a vital sign monitoring method that integrates multi-site sensing and machine vision for patients with large-area burns. This method solves the problems of vital sign monitoring failure or false alarms caused by the inability to wear probes due to wound coverage and peripheral circulatory disorders during the shock period in patients with large-area burns.

[0020] The overall approach of the scheme in this application is as follows: First, by combining multi-site contact signals and body surface temperature characteristics, the physical detachment of sensors or physiological hypoperfusion is intelligently identified. At the same time, machine vision technology is used to extract non-contact vital signs after removing eschar and dressing interference. Then, the fusion weights of multimodal data are automatically adjusted according to the channel status, and visual signals are used for dynamic compensation when contact signals are damaged. Finally, by analyzing the cross-covariance changes of fused heart rate and blood perfusion index over time, the divergence trend of the two is identified to achieve early warning of burn shock risk.

[0021] Please see Figure 1This invention provides a technical solution: a method for monitoring vital signs of patients with extensive burns by fusing multi-site sensing and machine vision, comprising the following steps: S1: Simultaneously acquiring photoplethysmography (PPG), body surface temperature, and a color video stream containing depth information from multiple anatomical test points of the patient; calculating waveform statistical features to generate a signal quality index; classifying channels with signal quality indices below an effective threshold based on the change characteristics of the corresponding body surface temperature data; marking channels with continuously decreasing or normal temperatures as low-perfusion channels or detachment channels, respectively; and outputting a state classification vector and a blood perfusion index sequence; S2: Constructing a three-dimensional background mask using depth information to remove video background noise; performing pixel-level segmentation based on the chromaticity features of burn eschar and medical dressings; calculating the gray-level variance in the remaining region and selecting the connected component with the smallest variance as the effective signal. The extraction area is defined as follows: spatial averaging is performed on the pixel data within the area to obtain the original color channel signal. Non-contact heart rate and respiratory rate are extracted through blind source separation operation; S3: contact pulse rate and non-contact heart rate are time-aligned. An adaptive weighting matrix is ​​constructed based on the channel state classification vector and signal quality index. The weight coefficients of each channel are dynamically adjusted using the normalized signal quality index. When a low-perfusion channel or a detached channel is identified, its weight is automatically reduced and the weight of non-contact heart rate is increased. Weighted fusion is performed on the multi-source data to output a fused heart rate sequence; S4: a time sliding window is constructed. The fused heart rate sequence and blood perfusion index sequence are input simultaneously. The cross-covariance of the two sequences is calculated within the sliding window. When the cross-covariance is less than the preset negative correlation threshold and the variance of the fused heart rate sequence is continuously greater than the fluctuation threshold, a hemodynamic discrete early warning signal is generated.

[0022] In this implementation plan, step S1 is mainly used to establish a multi-source data foundation and intelligently predict the working status of sensors. During this process, the system not only simultaneously collects contact-type photoplethysmography (PPG) and temperature data, but also acquires non-contact depth video streams. Waveform statistical characteristics refer to quantifying the regularity of a signal by calculating statistical indicators such as skewness and kurtosis, thereby generating a quantitative indicator called the signal quality index. The core of this step is the judgment logic based on changes in body surface temperature. It utilizes the physiological characteristics of cold and wet extremities during burn shock, identifying low-quality signals accompanied by temperature decreases as low-perfusion channels (i.e., weakened physiological signals), while identifying low-quality signals with normal temperature as detachment channels (i.e., physical contact failure). The technical role of this step is to accurately distinguish between equipment failure and disease deterioration, providing reliable decision labels for subsequent data fusion strategies, while retaining the blood perfusion index sequence of low-perfusion channels as a key parameter for subsequent shock early warning. Step S2 is mainly used to extract high-confidence non-contact vital signs in the complex burn surface environment using machine vision technology. This step utilizes depth information to construct a 3D background mask, separating the patient's body from the background environment using distance information, effectively removing ambient light and shadow interference. Pixel-level segmentation based on chromaticity features involves setting a threshold in the color space to automatically identify and remove dark areas of burn eschar and white areas of medical dressings. From the remaining areas, the most uniform skin tone and least affected by illumination are selected as effective signal extraction regions by calculating the grayscale variance. Subsequently, spatial averaging is used to convert image pixels into time-series signals, and blind source separation is applied. This algorithm separates independent source signals from mixed signals, capable of resolving pure pulse wave components from RGB signals containing motion artifacts. This step addresses the challenge of monitoring large-area burn patients lacking intact skin contact points, achieving high-precision extraction of non-contact signals. Step S3 primarily performs intelligent decision-making and fusion of multimodal data, ensuring the continuity and accuracy of monitoring data. First, contact pulse rate and non-contact heart rate at different sampling rates are time-aligned, making them comparable on the same time axis. The adaptive weighting matrix is ​​the core algorithm tool in this step. It dynamically assigns trust weights to each sensor based on the channel state classification vector output in step S1. When the system detects that a contact channel is in a state of low perfusion or detachment, the algorithm automatically reduces or even zeros the weight coefficient of that channel and simultaneously increases the weight ratio of the non-contact visual heart rate sensor. The technical advantage of this mechanism is that it eliminates the monitoring blind spot caused by the failure of a single sensor. Especially when burn patients experience fluid exudation or limb ischemia leading to the failure of contact probes, it can seamlessly switch to machine vision monitoring mode, ensuring that the final output fused heart rate sequence remains stable and reliable. Step S4 is mainly used for early warning of critical states based on correlation analysis of multidimensional physiological parameters. Constructing a time sliding window refers to extracting the latest data stream for real-time analysis at fixed time intervals.Within this window, the cross-covariance is calculated. This metric measures the correlation between the fused heart rate sequence and the perfusion index sequence in terms of their changing trends. When the cross-covariance is less than a preset negative correlation threshold, it means that the heart rate is compensatorily increasing while peripheral blood perfusion is continuously decreasing, which is a typical characteristic of the compensatory phase of shock. Simultaneously, the variance of the fused heart rate sequence is used to determine the severity of heart rate fluctuations. When both conditions are met, a hemodynamic discrete early warning signal is generated. The technical advantage of this step lies in overcoming the lag of traditional alarms that rely solely on numerical over-limits. By capturing the divergent trends of the waxing and waning of physiological parameters, it can detect the risk of shock before the blood pressure of patients with extensive burns completely collapses, buying valuable time for clinical rescue.

[0023] Specifically, the process of simultaneously acquiring photoplethysmography (PPG), body surface temperature, and color video stream containing depth information from multiple anatomical test points of the patient, and calculating the signal quality index by waveform statistical features is as follows: Bandpass filtering is performed on the acquired PPG signal to suppress high-frequency noise; a detrending algorithm is used to eliminate baseline drift; a time segment of preset length is extracted; and the skewness coefficient, kurtosis coefficient, and Shannon entropy of the signal amplitude within the time segment are calculated respectively. Interval normalization mapping is performed on the skewness coefficient, kurtosis coefficient, and Shannon entropy; a preset feature weight vector is applied to linearly weighted sum the three normalized feature values; and the weighted sum is output as the signal quality index for the corresponding time segment.

[0024] In this implementation scheme, the photoplethysmography (PPG) pulse wave signal undergoes preprocessing and feature extraction. The system first uses bandpass filtering to remove high-frequency noise introduced by power line interference and low-frequency baseline drift caused by human respiratory movements, ensuring that subsequent analysis is based on pure blood flow pulsation components. Then, within a pre-defined time segment, three waveform statistical features—skewness coefficient, kurtosis coefficient, and Shannon entropy—are calculated. Skewness coefficient characterizes the asymmetry of the waveform probability distribution, reflecting whether the pulse wave morphology is distorted by motion artifacts; kurtosis coefficient characterizes the steepness of the waveform probability distribution, with effective pulse waves typically exhibiting a specific kurtosis range; and Shannon entropy measures the complexity and disorder of the signal, with noisy signals usually exhibiting higher entropy values. To eliminate the influence of different dimensional features on the evaluation results, the above features are interval-normalized and mapped, and a linear weighted summation algorithm is used to generate a signal quality index. The formula for calculating this signal quality index is as follows: ;In the formula, in the formula, represents the signal quality index of the calculated time segment; k represents the feature number involved in the calculation, with a value range of 1 to 3, corresponding to the skewness coefficient, kurtosis coefficient and Shannon entropy, respectively. This represents the weight component in the preset feature weight vector corresponding to the k-th feature, where the sum of the three weight components is 1. This weight component is predetermined by the analytic hierarchy process based on the contribution of each feature to the signal quality. This represents the k-th original statistical feature value calculated within the current time segment; This represents the minimum boundary of the feature in historical statistical data; This indicates the maximum boundary of the feature in historical statistical data. This step, through the weighted fusion of multidimensional statistical features, can comprehensively quantify the reliability of the signal from three dimensions: waveform symmetry, steepness, and disorder.

[0025] Specifically, based on the changing characteristics of the corresponding body surface temperature data, channels with signal quality indices below the effective threshold are classified into state attributes. Channels with continuously decreasing temperatures or maintaining normal temperatures are marked as low-perfusion channels or detached channels, respectively. The specific process for outputting the state classification vector and blood perfusion index sequence is as follows: Identify the peak and trough amplitudes of the photoplethysmography (PPG) signal, extract the AC and DC components from the waveform, and calculate the ratio of the AC to DC components to generate the blood perfusion index sequence; calculate the first-order difference of the time series of body surface temperature data. When the signal quality index of the channel to be determined is below the effective threshold, if the first-order difference corresponding to the channel to be determined is negative and its absolute value is greater than the preset temperature change rate threshold, the channel to be determined is marked as a low-perfusion channel in the state classification vector; if the absolute value of the first-order difference corresponding to the channel to be determined is less than the preset temperature fluctuation tolerance, the channel to be determined is marked as a detached channel in the state classification vector.

[0026] In this implementation scheme, abnormal channels are accurately classified based on the correlation between physiological parameters and physical states. First, time-domain feature analysis is performed on the photoplethysmography (PPG) signal to identify the peak and trough amplitudes within the signal period. The trough amplitude is used as the DC component representing the absorption of light by venous blood and surrounding tissues, while the difference between the peak and trough amplitudes is used as the AC component representing the change in light absorption caused by arterial blood pulsation. The perfusion index is generated by calculating the ratio of the AC to DC components. This index directly reflects the peripheral blood flow filling at the monitored site. The formula for calculating the perfusion index is as follows: ; This indicates the blood perfusion index of the current monitoring channel. This represents the amplitude of the extracted pulse wave AC component. This represents the amplitude of the extracted DC component of the pulse wave. This represents the peak light intensity value of the photoplethysmography signal identified within the sampling period. This represents the trough intensity of the photoplethysmography (PPG) signal detected within the sampling period. The system performs a secondary assessment of low-quality signal channels based on the changing trend of body surface temperature data. The first-order difference of body surface temperature is calculated to quantify the rate of temperature change over time. When the signal quality index of a channel is below the effective threshold, it indicates that the data from that channel is unreliable. In this case, the rate of temperature change is used to distinguish between physiological shock and physical detachment. If the first-order difference corresponding to the channel to be assessed is negative and its absolute value is greater than the preset temperature change rate threshold, it indicates that the temperature at that site is rapidly decreasing, consistent with the physiological characteristics of cold, wet extremities and microcirculatory contraction during burn shock, and is therefore marked as a low-perfusion channel. Conversely, if the absolute value of the first-order difference is less than the preset temperature fluctuation tolerance, it indicates that the temperature at that site remains within a stable range, and the signal deterioration is more likely due to probe loosening or detachment, and is therefore marked as a detached channel. Through the above logic, the system can effectively eliminate invalid detached data while retaining clinically valuable low-perfusion data.

[0027] Specifically, the process of constructing a 3D background mask using depth information to remove video background noise, performing pixel-level segmentation based on the chromaticity features of burn eschar and medical dressings, and calculating the gray-level variance in the remaining region and selecting the connected component with the smallest variance as the effective signal extraction region is as follows: The depth image is mapped to the color video stream coordinate system using a coordinate transformation matrix; a depth truncation threshold is set to generate a binary depth mask; the depth mask is applied to the color video stream to perform a logical AND operation to retain near-field pixels; the retained pixels are converted to the YCrCb chromaticity space; a first filtering threshold interval is set based on the chromaticity distribution of burn eschar, and a second filtering threshold interval is set based on the chromaticity distribution of medical dressings; pixels whose chromaticity components fall within the first or second filtering threshold interval are removed; morphological opening operations are performed on the binary mask composed of the remaining pixels to eliminate isolated noise; the gray-level variance of each connected component in the luminance channel is calculated; and the connected component with the smallest gray-level variance is marked as the effective signal extraction region.

[0028] In this implementation, this step first utilizes the joint constraints of depth information and chromaticity features to accurately locate skin areas suitable for physiological parameter extraction in the complex burn ward scene. First, a coordinate transformation matrix is ​​used to map the depth image acquired by the depth camera to the RGB coordinate system of the color camera, achieving spatial alignment. A depth truncation threshold is set, determined by measuring the vertical distance from the camera to the bed plane. The system retains depth pixels smaller than this distance, generating a binary depth mask, and using logical AND operations to remove personnel movement or equipment interference from the far-field background in the color video stream. Subsequently, the retained near-field pixels are converted from the RGB color space to the YCrCb chromaticity space, which can separate the luminance component Y from the chromaticity components Cr and Cb, reducing the impact of changes in light intensity on skin color recognition. The system determines a first filtering threshold interval based on a pre-constructed histogram of burn eschar chromaticity distribution, and a second filtering threshold interval based on the histogram of chromaticity distribution of commonly used medical gauze and dressings. The image is traversed pixel-by-pixel, removing pixels falling into the above two intervals, thereby eliminating interference from non-skin areas. Morphological opening operations are performed on the binary mask composed of the remaining pixels. This involves first performing erosion to remove small, isolated noise points, and then performing dilation to restore the region's morphology, thus obtaining several independent connected components. To ensure the highest signal-to-noise ratio of the extracted pulse wave signal, the system needs to select the region with the most uniform illumination and the smoothest surface. Therefore, the system calculates the gray-level variance of each connected component in the brightness channel and selects the connected component with the smallest variance as the effective signal extraction region. The formula for calculating the gray-level variance is as follows: ;In the formula, This represents the gray-level variance of the m-th connected component. The smaller the value, the more uniform the illumination distribution within that region. This represents the total number of pixels contained in the m-th connected component; (u,v) represents the two-dimensional pixel coordinate index of the image; Represents the set of pixel coordinates of the m-th connected component; This represents the grayscale value of the Y channel of the pixel with coordinates (u,v) in the YCrCb space. This represents the arithmetic mean of the grayscale values ​​of all pixels within the m-th connected region. Using this formula, the system can automatically avoid areas with strong edge shadows or reflections, and pinpoint the optimal region of interest for the skin.

[0029] Specifically, spatial averaging is performed on pixel data within the effective signal extraction area to obtain the original color channel signal. The specific process of extracting non-contact heart rate and respiratory rate through blind source separation is as follows: The spatial mean of all pixels in the red, green, and blue channels within the effective signal extraction area is calculated frame by frame to construct a three-channel time-series signal. Zero-mean unit variance standardization is performed on the three-channel time-series signal. The standardized signal is demixed into independent source signal components using an independent component analysis algorithm. Fast Fourier transform is performed on each source signal component to obtain the power spectral density. The frequency corresponding to the maximum power spectral density within the preset heart rate frequency range is searched as the non-contact heart rate, and the frequency corresponding to the maximum power spectral density within the preset respiratory frequency range is searched as the respiratory rate.

[0030] In this implementation, since the original color channel signal is a linear mixture of blood volume change signals caused by heartbeats, chest rise and fall signals caused by respiratory movements, and ambient light interference signals, direct analysis is insufficient to obtain an accurate heart rate. Therefore, this embodiment employs the Independent Component Analysis (ICA) algorithm in blind source separation computation. Under the assumption of statistical independence of source signals, it demixes the independent source signal components from the mixed observation signal. Before performing ICA, the three-channel time-series signal is first subjected to zero-mean, unit-variance standardization to eliminate signal amplitude differences and meet the whitening requirements of the algorithm. The source signal estimation formula based on ICA is as follows: ;In the formula, This represents the source signal vector estimated at time t, which contains the separated independent components, such as pulse wave components, respiratory components, and noise components. denoted as the demixing matrix, which is obtained by the FastICA algorithm through iterative solution by maximizing non-Gaussianity and is used to back-project the mixed signal back into the source signal space; This represents the standardized observation signal vector after whitening, obtained by centering and eigenvalue decomposition of the original red, green, and blue three-channel mean signals. This step successfully separates the weak pulse wave signal that was previously submerged in motion and light noise. Finally, a Fast Fourier Transform is performed on each separated source signal component to convert the time-domain signal into a power spectral density in the frequency domain. The power spectral density directly reflects the energy distribution of the signal at different frequencies. The system searches for the maximum peak value of the power spectral density within a preset heart rate frequency range of 0.7 Hz to 4.0 Hz; the frequency corresponding to this peak value is the non-contact heart rate. Similarly, it searches for the frequency corresponding to the maximum peak value within a preset respiratory frequency range of 0.1 Hz to 0.5 Hz as the respiratory rate. This process achieves the final mapping from the pure source signal to clinical physiological indicators.

[0031] Specifically, the process of aligning contact pulse rate and non-contact heart rate in time and constructing an adaptive weighted matrix based on channel state classification vectors and signal quality indices is as follows: Extract the data timestamps of each contact and non-contact channel; select the time axis of the contact photoplethysmography (PPG) signal with the highest sampling frequency as the master clock reference; perform linear interpolation resampling based on timestamp matching on the non-contact heart rate data to establish a point-to-point mapping relationship between multi-source data in the same time dimension; construct a fusion weight vector with dimensions adapted to the sum of the number of all channels; call the signal quality index of each contact channel to fill the corresponding element position of the fusion weight vector as the initial weight of the contact channel; and initialize the weight elements of the non-contact channel to a preset baseline confidence value.

[0032] In this implementation, this step aims to address the data fusion obstacle caused by inconsistent sampling frequencies of multimodal sensors and to complete the initialization preparation before fusion. First, data cleaning and timing alignment are performed, with the system extracting timestamp information from the header of each channel's data packet. Since photoplethysmography (PPG) sensors typically have high sampling rates (e.g., 125Hz or higher), while non-contact cameras have relatively low frame rates (e.g., 30fps), direct fusion would lead to phase misalignment. Therefore, the contact signal time axis with the highest sampling frequency is selected as the master clock reference, and linear interpolation resampling based on timestamp matching is performed on the non-contact heart rate data. This process uses two adjacent visual heart rate data points to estimate the value corresponding to the contact signal sampling time through a linear equation, thereby establishing a point-to-point mapping relationship between multi-source data in the same time dimension, ensuring that each subsequent frame of data is strictly synchronized in physical time. Subsequently, the system constructs an initial fusion weight vector. The length of this vector is equal to the sum of the number of all enabled contact probes and the number of non-contact visual channels. To provide an initial basis for subsequent weighted fusion, the system calls the signal quality indices of each contact channel calculated in the previous steps and directly fills them into the corresponding contact element positions in the weight vector. For non-contact channels, due to the lack of waveform statistical characteristics, the system initializes their weight elements to a preset baseline confidence value. The baseline confidence value is usually preset based on the camera's hardware parameters and ambient light intensity, for example, set to 0.5 under standard lighting conditions, to characterize the basic reliability of the visual signal in an interference-free state. This step completes the data alignment and parameter initialization of the fusion model, laying the foundation for subsequent dynamic adjustments.

[0033] Specifically, the weight coefficients of each channel are dynamically adjusted using a normalized signal quality index. When a low-perfusion or detached channel is identified, its weight is automatically reduced while the weight of the non-contact heart rate is increased. The specific process of performing weighted fusion on multi-source data to output a fused heart rate sequence is as follows: Iteratively read the element labels in the channel state classification vector. When the label of the contact channel to be processed is a detached channel, the corresponding weight coefficient in the fusion weight vector is set to a logic zero value. When the label of the contact channel to be processed is a low-perfusion channel, a preset attenuation factor is applied to the corresponding signal quality index to perform a multiplication operation to update the weight coefficients. Calculate the sum of the updated weight coefficients of all contact channels, calculate the difference between the unit scalar 1 and the sum of the weight coefficients to obtain the residual weight margin, and add the residual weight margin to the non-contact channel weight coefficients. Perform L1 norm normalization on the fusion weight vector, calculate the vector dot product of the heart rate value of each channel and the normalized fusion weight vector, and output the fused heart rate sequence.

[0034] In this implementation plan, this step is the core of achieving blind-spot-free monitoring. It utilizes an intelligent weight scheduling strategy to automatically elevate the decision-making role of machine vision when contact sensors fail. The system iteratively reads the element labels in the channel status classification vector and corrects the weights of contact channels one by one. When a label indicates a channel is physically detached, it means the sensor has completely left the human body, and the data is pure noise. Therefore, the weight coefficient of this channel is forcibly set to logic zero, completely cutting off the input of erroneous data. When a label indicates a low-perfusion channel, it means the signal exists but the signal-to-noise ratio is reduced. At this time, a preset attenuation factor is applied to penalize the signal quality index, preventing low-quality signals from dominating the fusion result. The system performs the core residual weight transfer operation. It calculates the sum of the corrected weights of all contact channels, treating the unit scalar 1 as the system's total expected confidence level for contact monitoring, and calculates the difference between the two to obtain the residual weight margin. This margin represents the share of trust "lost" by the contact sensor due to detachment or low perfusion. The system adds this lost trust to the weights of non-contact channels, thereby achieving automatic weight compensation. Finally, the weight vector is normalized using the L1 norm to ensure that the sum of all weights is 1, and a vector dot product is performed to output the final result. The calculation formula for multi-source data weighted fusion is as follows: ;In the formula, This represents the final fused heart rate sequence value calculated at time t; The L1 norm of the final fused weight vector is the sum of the absolute values ​​of all elements in the vector, used to normalize the denominator; N represents the number of channels in the contact sensor; M represents the number of channels in the non-contact vision sensor. This represents the heart rate value of the nth channel at time t; This represents the weighting coefficient of the nth channel after dynamic adjustment and residual compensation. Specifically, the weighting coefficient for the contactless channel... The calculation logic is as follows: ;In the formula, This represents the preset baseline confidence level value for the contactless channel; This represents the signal quality index of the i-th contact channel; This represents the state decay factor of the i-th contact channel. It is 1 when the state is normal, a preset decay value (e.g., 0.5) when the perfusion is low, and 0 when the channel is detached. Through the above calculation, the system cleverly uses mathematical logic to implement an adaptive mechanism that increases the weight of machine vision as the number of contact sensor failures increases, ensuring the continuity and robustness of monitoring data in extreme burn scenarios.

[0035] Specifically, a time-sliding window is constructed, and the fused heart rate sequence and blood perfusion index sequence are input synchronously. The specific process of calculating the cross-covariance of the two sets of sequences within the sliding window is as follows: A fixed-capacity synchronous data buffer queue is established as a time-sliding window, and the time-aligned fused heart rate data and low-perfusion channel blood perfusion index data are stored in real time according to the first-in-first-out principle; the arithmetic mean of the fused heart rate data and blood perfusion index data in the buffer queue are calculated respectively, and a mean vector is obtained by tensor product operation between the all-1 vector and the arithmetic mean; the mean vector is generated by subtracting the mean vector from the original data vector in the buffer queue; the centralized fused heart rate vector and the centralized blood perfusion vector are generated; the dot product of the centralized fused heart rate vector and the centralized blood perfusion vector is calculated, and the dot product result is multiplied by the reciprocal of the buffer queue capacity value to output the real-time cross-covariance.

[0036] In this implementation plan, this step aims to quantify the dynamic correlation between heart rate changes and peripheral perfusion changes using mathematical statistical methods, thereby capturing the physiological compensatory characteristics in the early stages of shock. The system first establishes a first-in, first-out queue as a data buffer container. The capacity of this container determines the observation duration of the time sliding window, for example, set to 30 to 60 seconds, to ensure coverage of the complete respiratory modulation cycle and vasomotor response cycle. After real-time data storage, the system performs centering processing on the data within the window, a necessary prerequisite for calculating the covariance. Centering, or zero-mean normalization, is achieved by calculating the arithmetic mean of the data within the window and constructing an all-1 vector. A mean vector with all elements equal to this arithmetic mean is then generated using tensor product operations. This mean vector is then subtracted from the original data vector. This operation eliminates baseline numerical differences between different patients or under different physiological states, retaining only the fluctuation component of the data over time. Subsequently, the dot product of the two centered vectors is calculated and normalized to obtain the real-time cross-covariance. The formula for calculating the real-time cross-covariance is as follows: ;In the formula, This represents the calculated real-time cross-covariance value within the current sliding window; This represents the capacity of the buffer queue, i.e., the total number of sampling points contained in the sliding window; This represents a column vector consisting of the fused heart rate data stored within the sliding window; Representing vectors The arithmetic mean of all elements in the set; Representing dimensions and The same column vector of all 1s is used to construct the mean vector through tensor product operations; This represents a column vector consisting of blood perfusion index data for low-perfusion channels stored within the sliding window; Representing vectors The formula is the arithmetic mean of all elements in the formula. When the heart rate is increasing and the perfusion index is decreasing, the signs of the two are opposite after centering, and the dot product is significantly negative, thus numerically reflecting the body's shock compensation mechanism.

[0037] Specifically, when the cross-covariance is less than a preset negative correlation threshold and the variance of the fused heart rate sequence is continuously greater than the fluctuation threshold, the specific process for generating a hemodynamic discrete early warning signal is as follows: Initialize the early warning judgment accumulator, compare the real-time cross-covariance with the preset negative correlation threshold in real time, and synchronously calculate the variance of the fused heart rate sequence within the sliding window and compare it with the preset fluctuation threshold; only when both the real-time cross-covariance is less than the negative correlation threshold and the variance of the fused heart rate sequence is greater than the fluctuation threshold are satisfied, perform an incremental operation on the early warning judgment accumulator; if either condition is not satisfied, perform a zeroing and reset operation on the early warning judgment accumulator; when the accumulated value of the early warning judgment accumulator exceeds the preset time confirmation threshold, trigger the alarm circuit to output a logic high-level signal as a hemodynamic discrete early warning signal.

[0038] In this implementation scheme, this step achieves accurate alarm for discrete hemodynamic events through multidimensional constraints and temporal persistence verification. The system first calculates the variance of the fused heart rate sequence to assess the stability of heart rate fluctuations. Variance reflects heart rate variability and agitation; patients in the early stages of shock often exhibit increased heart rate fluctuations. Subsequently, an early warning accumulator mechanism is introduced, which is a time-domain filtering strategy. The system monitors two concurrent conditions in real time: first, whether the real-time cross-covariance variance is less than a preset negative correlation threshold, indicating whether heart rate and perfusion show a significant negative correlation divergence; second, whether the heart rate variance is consistently greater than a fluctuation threshold, indicating whether the physiological state is unstable. The accumulator only performs an increment operation when both conditions are met simultaneously; otherwise, it is immediately reset to zero. The technical function of this mechanism is to filter out transient data jumps caused by body movement, coughing, etc., ensuring that the alarm signal is triggered only when the abnormal physiological state persists, thereby greatly reducing the false alarm rate. The generation logic formula for the discrete hemodynamic early warning signal is as follows: ; ;In the formula, This indicates the current count value of the accumulator at time t for the early warning determination. This represents the count value of the accumulator at the previous sampling time. This represents the preset negative correlation threshold, which is usually set to a negative number to define the degree of association between increased heart rate and decreased perfusion. This represents the variance of the fused heart rate sequence within the current sliding window; This indicates a preset fluctuation threshold, used to define the boundaries of heart rate instability; This indicates the logic state of the final output hemodynamic discrete warning signal, where 1 represents a high-level alarm and 0 represents normal operation. Indicates the system's sampling interval time; This represents the preset time confirmation threshold, i.e., the minimum duration the abnormal state must last. This step, through logical AND operations and time integration effects, achieves specific identification of the compensatory period of shock in large-area burns.

[0039] In summary, this application has at least the following effects:

[0040] A vital sign monitoring method integrating multi-site sensing and machine vision for patients with extensive burns is proposed. This method utilizes heterogeneous fusion of multi-anatomical point contact sensing and machine vision, combined with a state attribute classification mechanism based on waveform statistical features and body temperature gradients. This effectively overcomes monitoring blind spots and data distortion caused by wound coverage and the cold, wet extremities during the shock phase in burn patients. Depth information and color segmentation technology are used to automatically remove eschar and dressing interference. When contact channels fail due to exudation or low perfusion, an adaptive weighted matrix seamlessly switches to a non-contact visual monitoring mode, ensuring the continuity and robustness of monitoring data. Simultaneously, a hemodynamic discrete early warning model based on temporal cross-covariance analysis is constructed to keenly capture early pathological features such as compensatory heart rate elevation and persistent peripheral perfusion contraction, achieving advanced early warning for the compensatory phase of burn shock and significantly improving the success rate of clinical treatment.

[0041] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0042] This invention is described with reference to flowchart illustrations and / or block diagrams of systems, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0043] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0044] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0045] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0046] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for monitoring vital signs in patients with extensive burns by fusing multi-site sensing and machine vision, characterized in that, Includes the following steps: S1: Simultaneously acquire photoplethysmography pulse waves, body surface temperature, and color video stream containing depth information from multiple anatomical test points of the patient. Calculate waveform statistical features to generate a signal quality index. Combine the change characteristics of the corresponding body surface temperature data to classify the state attributes of channels with signal quality indices below the effective threshold. Mark channels with continuously decreasing or normal temperature as low perfusion channels or detached channels, respectively. Output state classification vector and blood perfusion index sequence. S2: Construct a 3D background mask using depth information to remove video background noise. Perform pixel-level segmentation based on the color features of burn eschar and medical dressing. Calculate the gray-level variance in the remaining area and select the connected region with the smallest variance as the effective signal extraction region. Perform spatial averaging on the pixel data in this region to obtain the original color channel signal. Extract non-contact heart rate and respiratory rate through blind source separation operation. S3: Time-series alignment of contact pulse rate and non-contact heart rate, construction of adaptive weighting matrix based on channel state classification vector and signal quality index, dynamic adjustment of weight coefficient of each channel using normalized signal quality index, automatic reduction of weight of low perfusion channel or detached channel and increase of non-contact heart rate weight when low perfusion channel or detached channel is identified, and weighted fusion of multi-source data to output fused heart rate sequence. S4: Construct a time sliding window, synchronously input the fused heart rate sequence and the blood perfusion index sequence, calculate the cross-covariance of the two sequences within the sliding window, and generate a hemodynamic discrete early warning signal when the cross-covariance is less than the preset negative correlation threshold and the variance of the fused heart rate sequence is continuously greater than the fluctuation threshold.

2. The vital sign monitoring method for patients with extensive burns, which integrates multi-site sensing and machine vision, as described in claim 1, is characterized in that: The specific process of simultaneously acquiring photoplethysmography pulse waves, body surface temperature, and color video streams containing depth information from multiple anatomical test points on the patient, and calculating waveform statistical features to generate a signal quality index is as follows: Bandpass filtering was performed on the acquired photoplethysmography pulse wave signal to suppress high-frequency noise. Baseline drift was eliminated by detrending algorithm. Time segments of preset length were extracted, and the skewness coefficient, kurtosis coefficient and Shannon entropy of the signal amplitude within the time segments were calculated respectively. The skewness coefficient, kurtosis coefficient, and Shannon entropy are subjected to interval normalization mapping. The three normalized eigenvalues ​​are linearly weighted and summed using a preset feature weight vector. The weighted summation result is output as the signal quality index of the corresponding time segment.

3. The vital sign monitoring method for patients with large-area burns, which integrates multi-site sensing and machine vision, as described in claim 1, is characterized in that: Based on the changing characteristics of corresponding body surface temperature data, channels with signal quality indices below the effective threshold are classified into state attributes. Channels with continuously decreasing temperatures or maintaining normal temperatures are marked as low-perfusion channels or detached channels, respectively. The specific process of outputting state classification vectors and blood perfusion index sequences is as follows: Identify the peak and trough amplitudes of the photoplethysmography (PPG) signal, extract the AC and DC components from the waveform, and calculate the ratio of the AC to DC components to generate a blood perfusion index sequence. Calculate the first-order difference of the time series of body surface temperature data. When the signal quality index of the channel to be determined is lower than the effective threshold, if the first-order difference of the channel to be determined is negative and the absolute value is greater than the preset temperature change rate threshold, the channel to be determined is marked as a low perfusion channel in the state classification vector. If the absolute value of the first-order difference corresponding to the channel to be judged is less than the preset temperature fluctuation tolerance, the channel to be judged will be marked as a detachment channel in the state classification vector.

4. The vital sign monitoring method for patients with extensive burns, which integrates multi-site sensing and machine vision, as described in claim 1, is characterized in that: The process of constructing a 3D background mask using depth information to remove video background noise, performing pixel-level segmentation based on the chromaticity features of burn eschar and medical dressings, calculating the gray-level variance in the remaining region, and selecting the connected component with the smallest variance as the effective signal extraction region is as follows: The depth image is mapped to the coordinate system of the color video stream using a coordinate transformation matrix. A depth truncation threshold is set to generate a binary depth mask. The depth mask is then applied to perform a logical AND operation on the color video stream to preserve near-field pixels. The retained pixels are converted to the YCrCb color space. A first filtering threshold range is set according to the color distribution of burn eschar, and a second filtering threshold range is set according to the color distribution of medical dressing. Pixels whose color components fall into the first filtering threshold range or the second filtering threshold range are removed. Morphological opening operations are performed on the binary mask formed by the remaining pixels to eliminate isolated noise. The gray-level variance of each connected component in the brightness channel is calculated, and the connected component with the smallest gray-level variance is selected as the effective signal extraction region.

5. The vital sign monitoring method for patients with large-area burns, which integrates multi-site sensing and machine vision according to claim 4, is characterized in that: The process of performing spatial averaging on pixel data within the effective signal extraction area to obtain the original color channel signal, and extracting non-contact heart rate and respiratory rate through blind source separation operation is as follows: The spatial mean of all pixels in the red, green and blue channels within the effective signal extraction area is calculated frame by frame to construct a three-channel time-series signal. Zero-mean unit variance standardization is performed on the three-channel time-series signal, and the standardized signal is demixed into independent source signal components through independent component analysis algorithm. Perform a fast Fourier transform on each source signal component to obtain the power spectral density. Search for the frequency corresponding to the maximum power spectral density within a preset heart rate frequency range as the non-contact heart rate, and search for the frequency corresponding to the maximum power spectral density within a preset respiratory frequency range as the respiratory rate.

6. The vital sign monitoring method for patients with extensive burns, which integrates multi-site sensing and machine vision, as described in claim 1, is characterized in that: The specific process of time-series alignment of contact pulse rate and non-contact heart rate, and construction of an adaptive weighted matrix based on channel state classification vector and signal quality index is as follows: Extract the timestamps of the data from each contact and non-contact channel, select the time axis of the contact photoplethysmography pulse wave signal with the highest sampling frequency as the master clock reference, perform linear interpolation resampling based on timestamp matching on the non-contact heart rate data, and establish a point-to-point mapping relationship between multi-source data in the same time dimension. Construct a fusion weight vector with dimensions adapted to the total number of all channels. Call the signal quality index of each contact channel and fill the corresponding element position of the fusion weight vector as the initial weight of the contact channel. Initialize the weight elements of the non-contact channels to the preset baseline confidence value.

7. The vital sign monitoring method for patients with extensive burns, which integrates multi-site sensing and machine vision, as described in claim 6, is characterized in that: The weighting coefficients of each channel are dynamically adjusted using a normalized signal quality index. When a low-perfusion or detached channel is identified, its weight is automatically reduced while the weight of the non-contact heart rate is increased. The specific process of performing weighted fusion on multi-source data to output a fused heart rate sequence is as follows: Iteratively read the element labels in the channel state classification vector. When the label of the contact channel to be processed is the detached channel, set the corresponding weight coefficient in the fusion weight vector to a logical zero value. When the label of the contact channel to be processed is a low-perfusion channel, the corresponding signal quality index is multiplied by the preset attenuation factor to update the weight coefficient. Calculate the sum of the updated weight coefficients for all contact channels, calculate the difference between the unit scalar 1 and the sum of the weight coefficients to obtain the residual weight margin, and then add the residual weight margin to the weight coefficients of the non-contact channels. The L1 norm normalization process is performed on the fusion weight vector, the vector dot product of the heart rate value of each channel and the normalized fusion weight vector is calculated, and the fusion heart rate sequence is output.

8. The vital sign monitoring method for patients with large-area burns, which integrates multi-site sensing and machine vision, as described in claim 1, is characterized in that: The specific process of constructing a time-sliding window, simultaneously inputting fused heart rate and blood perfusion index sequences, and calculating the cross-covariance of the two sequences within the sliding window is as follows: A fixed-capacity synchronous data buffer queue is established as a time sliding window, and time-aligned fused heart rate data and low-perfusion channel blood flow perfusion index data are stored in real time according to the first-in-first-out principle. The arithmetic mean of the fused heart rate data and blood perfusion index data in the buffer queue are calculated separately. The mean vector is obtained by tensor product operation between the all-1 vector and the arithmetic mean. The mean vector is then subtracted from the original data vector in the buffer queue to generate the centered fused heart rate vector and the centered blood perfusion vector. Calculate the dot product of the centralized fused heart rate vector and the centralized blood perfusion vector, multiply the dot product result by the reciprocal of the buffer queue capacity value, and output the real-time cross-covariance.

9. A vital sign monitoring method for patients with extensive burns, integrating multi-site sensing and machine vision, as described in claim 8, characterized in that: When the cross-covariance is less than the preset negative correlation threshold and the variance of the fused heart rate sequence is consistently greater than the fluctuation threshold, the specific process for generating a hemodynamic discrete early warning signal is as follows: Initialize the early warning judgment accumulator, compare the real-time cross-covariance with the preset negative correlation threshold in real time, and synchronously calculate the variance of the fused heart rate sequence within the sliding window and compare it with the preset fluctuation threshold. Incremental operation is performed on the warning judgment accumulator only when both the real-time cross-covariance variance is less than the negative correlation threshold and the variance of the fused heart rate sequence is greater than the fluctuation threshold. If any condition is not met, the warning judgment accumulator is cleared and reset. When the accumulated value of the warning judgment accumulator exceeds the preset time confirmation threshold, the alarm circuit is triggered to output a logic high-level signal as a hemodynamic discrete warning signal.

Images