A monitoring and early warning system for changes in the position of endotracheal tubes in ICU nursing

Abstract

This invention provides a monitoring and early warning system for endotracheal intubation position changes in ICU nursing, relating to the field of image data processing technology. This invention addresses the problems of lag and limited diagnostic dimensions in existing endotracheal intubation monitoring by constructing a cross-modal fusion framework guided by temporal spectral dynamic features and adaptively enhanced spatial thermal imaging. Through deep temporal spectral analysis of airflow signals and the introduction of change rate and stability factors, early quantitative capture of airway dysfunction precursors is achieved. Using these functional features as input, a two-level cascaded correlation mapping model generates and applies nonlinear local image correction parameters in real time, thereby enhancing visualization on thermal images and improving monitoring sensitivity. This enables early warning of airway events caused by functional instability. Through spatiotemporal joint analysis, acute risks and chronic states are effectively distinguished, reducing the clinical false alarm rate.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, specifically to an ICU nursing endotracheal tube position change monitoring and early warning system. Background Technology

[0002] In the field of modern critical care medicine, the monitoring of patients' vital signs is evolving from traditional threshold alarms based on isolated parameters to predictive risk assessments that integrate multimodal information. Especially in invasive mechanical ventilation, ensuring the stability and patency of artificial airways (such as endotracheal intubation) is fundamental to maintaining patient life support. Therefore, developing advanced monitoring systems that can integrate multidimensional data, provide early warnings of potential airway risks, and offer accurate diagnostic information has become a key technological direction for improving the quality of critical care and patient safety.

[0003] Currently, technologies for monitoring changes in endotracheal tube position still face challenges in terms of sensitivity, specificity, and the completeness of diagnostic information. Specifically, existing technologies have the following limitations: Existing technologies, such as the scheme disclosed in Chinese patent application CN119379637A, rely on comparing the currently acquired airway internal images with pre-stored standard images for similarity. This type of method is essentially a "static snapshot"-style structural morphological analysis. Its limitation is that it can only trigger an alarm after a significant, structurally observable physical displacement has occurred at the intubation site. However, in clinical practice, many critical airway events (such as sputum obstruction or tube wall compression) first manifest as "functional instability" in airflow dynamics before evolving into structural displacement. Existing technologies lack the ability to perceive this functional precursor information, resulting in an inherent time lag in their monitoring; it is a "post-event" response rather than a "pre-event" warning.

[0004] Existing technologies typically rely on a single source of information for decision-making. While purely image-based methods can provide spatial location information, they lack sensitivity to functional changes and are easily affected by non-specific visual factors such as secretions and tissue edema. On the other hand, monitoring methods based solely on functional signals (such as airway pressure and end-tidal carbon dioxide) can sensitively detect abnormalities in ventilation function but cannot provide the specific spatial location of the abnormal event. This disconnect between functional and spatial information creates a technological dilemma, limiting the rapid and accurate diagnosis of the root causes of complex airway events. Summary of the Invention

[0005] The purpose of this invention is to provide an endotracheal tube position change monitoring and early warning system for ICU nursing, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: An ICU nursing endotracheal tube position change monitoring and early warning system specifically includes: The first data acquisition module is used to acquire first data characterizing the stability of airflow function in endotracheal intubation, wherein the first data is the time-domain spectral characteristics of the airflow pressure or flow rate signal. The second data acquisition module is used to acquire second data characterizing the spatial distribution of airflow around the tip of the endotracheal tube. The second data is a sequence of original thermal image frames acquired by a thermal imaging sensor. The image processing module is configured as follows: Based on the changes in the first data, the target region of interest in the original thermal image frame sequence is identified; Furthermore, for the target region of interest, a set of nonlinear image correction parameters are generated and applied in real time to locally enhance the dynamic range of the image in the target region of interest, thereby generating an enhanced thermal image; The early warning module is used to generate a diagnostic early warning signal based on visual features presented in the enhanced thermal image that are not visible in the original thermal image frame sequence.

[0007] Compared with existing technologies, the advantages of this invention are: by introducing time-domain spectral characteristics analysis of airflow pressure or flow signals for dynamic characteristics analysis; and by calculating and continuously tracking the high-frequency harmonic energy ratio of the respiratory cycle, the airflow turbulence index, and their respective rates of change and stability factors, the dynamic evolution of airway status from stable to unstable can be captured from a functional perspective. This allows for the identification of the nascent stage of an event before structural displacement occurs, thus enabling the monitoring mode to have predictive capabilities.

[0008] This design employs a spectrum-guided adaptive thermal image dynamic range enhancer. This design serves as a bridge connecting functional and spatial information. Instead of passively analyzing the raw thermal image, the system utilizes the aforementioned multi-dimensional temporal spectral features as a "diagnostic probe." These spectral features are transformed in real-time into two sets of key instructions through a built-in correlation mapping model: one set is the location signal of the region of interest (ROI), which tells the imaging system the area to focus on; the other set is nonlinear image correction parameters, which guide the imaging system on how to observe. Specifically, the generation of these image correction parameters uses a two-stage cascaded processing, generating basic and dynamic enhancement terms based on the severity and dynamism of the event. This mechanism enables the system to adaptively, locally, and nonlinearly enhance the dynamic range of the thermal image in key regions according to the degree of functional instability, thereby "making explicit" changes in the topological structure of the heat flux field caused by subtle airflow variations that are invisible in the original thermal image. This achieves a technical closed loop where sensitive signals guide imaging and imaging results verify signals. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the overall system application process of the present invention; Figure 2 This is a schematic diagram comparing the original thermal image (left) and the enhanced thermal image (right) in an embodiment of the present invention; Figure 3 This is a schematic diagram of the execution logic from the first data acquisition module to the image processing module of the present invention; Figure 4 This is a schematic diagram of the execution logic of the early warning module of the present invention. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0011] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various elements, but unless otherwise stated, these elements are not limited by these terms. These terms are used only to distinguish one element from another.

[0012] Example 1: Please see Figures 1 to 4 The present invention provides a technical solution: An ICU nursing endotracheal tube position change monitoring and early warning system includes: The first data acquisition module is used to acquire first data characterizing the stability of airflow function in endotracheal intubation. The first data is the time-domain spectral characteristics of the airflow pressure or flow rate signal. The second data acquisition module is used to acquire second data characterizing the spatial distribution of airflow around the tip of the endotracheal tube. The second data is a sequence of original thermal image frames acquired by a thermal imaging sensor. The image processing module is configured as follows: Based on the changes in the first data, the target region of interest (ROI) in the original thermal image frame sequence is identified. Furthermore, for the target region of interest, a set of nonlinear image correction parameters are generated and applied in real time to locally enhance the dynamic range of the image in the target region of interest, thereby generating an enhanced thermal image; The early warning module is used to generate diagnostic early warning signals based on visual features presented in the enhanced thermal images that are not visible in the original thermal image frame sequence.

[0013] Further explanation: The first data acquisition module is configured as follows: A fast Fourier transform is performed on the pressure or flow rate signal representing the first data to calculate the high-frequency harmonic energy ratio of the breathing cycle and the airflow turbulence index within a preset high-frequency band, and the calculation results are used as time-domain spectral features. The second data acquisition module includes: A miniature uncooled infrared focal plane array sensor, custom-designed for the tip of an endotracheal tube, is used to acquire raw thermal image frame sequences to dynamically image the topology of the airflow heat flux field formed around the tip of the endotracheal tube.

[0014] Further explanation: The first data acquisition module is also configured as follows: Based on the historical high-frequency harmonic energy ratio and historical airflow turbulence index values ​​within a preset time window, the rate of change and stability factor of the historical high-frequency harmonic energy ratio and the rate of change and stability factor of the historical airflow turbulence index are calculated respectively. These four time-dimensional dynamic parameters, together with the high-frequency harmonic energy ratio and airflow turbulence index of the respiratory cycle, are used as inputs to the image processing module.

[0015] Further explanation: The image processing module includes a spectrum-guided adaptive thermal image dynamic range enhancer with a built-in pre-trained association mapping model; The correlation mapping model is configured to receive the high-frequency harmonic energy ratio of the respiratory cycle and the airflow turbulence index, as well as four time-dimensional dynamic parameters as inputs, and output the localization signal of the target region of interest and nonlinear image correction parameters.

[0016] Further explanation: The association mapping model is also configured to perform a two-stage cascaded process to generate non-linear image correction parameters: The image processing module is also configured to perform two-stage cascaded fusion processing: Level 1 processing: Based on the high-frequency harmonic energy ratio of the respiratory cycle, the airflow turbulence index, and four time-dimensional dynamic parameters, an event severity factor and an event dynamic factor are generated. Second-level processing: Based on the event severity factor and the event dynamic factor, nonlinear image correction parameters are generated.

[0017] Further explanation: In the second-level processing, the image processing module is also configured as follows: Based on the event severity factor, determine the basic enhancement terms in the image correction parameters; Furthermore, based on the event dynamics factor, the dynamic enhancement term in the image correction parameters is determined; The nonlinear image correction parameters consist of a combination of basic enhancement terms and dynamic enhancement terms.

[0018] Further explanation: In the second-level processing, the image processing module is also configured as follows: Based on the event dynamics factor, the size of the target region of interest (ROI) is adaptively adjusted.

[0019] Figure 2 The comparison between the left and right parts demonstrates the core image processing effect of this invention. The left part 100 represents the original thermal image acquired by the second data acquisition module, showing the heat flux distribution around the tip of the endotracheal tube. In this original image, due to the limited dynamic range, the weak thermal anomaly signal 101 (local temperature difference caused by a small amount of sputum accumulation), representing early pathological changes, has low contrast with the surrounding background tissue, making it difficult to directly identify by the naked eye or conventional algorithms. The right part 200 represents the enhanced thermal image generated after processing by the image processing module of this embodiment. The image processing module identifies a target region of interest (ROI) 102 on the original image, marked by a dashed circle. The module performs a local, non-linear dynamic range enhancement operation only on the pixels within the ROI 102. The result is that the image area outside 102 remains unchanged, while the weak thermal anomaly signal inside 102 is significantly enhanced, forming a clear feature region 201 with high contrast. This process effectively highlights diagnostic information that is buried in the background, improving the ability to detect early signs of endotracheal intubation-related complications.

[0020] The following is a detailed description of the implementation of the above content: The core technical feature of this invention lies in the introduction of analysis of the dynamic characteristics of time-domain spectral features. By calculating and fusing the rate of change and historical stability of these features, a multi-dimensional "event feature vector" is constructed. This vector can distinguish the dynamic attributes and severity of unstable events. Based on this feature vector, a two-level cascaded dynamic adaptive fusion mechanism is designed. The first-level fusion mechanism qualitatively and quantitatively characterizes the event attributes and generates intermediate state factors; the second-level fusion mechanism uses these state factors to dynamically and nonlinearly generate local correction parameters for thermal images. This enables the fusion mechanism not only to "see" the precursors of subclinical events, but also to predictively classify the root causes and development trends of the precursors, thereby elevating the monitoring capability from "event detection" to "prospective diagnosis." The technical solution provided by this invention will be clearly and completely described below with reference to a preferred specific embodiment. In this embodiment, all algorithms are implemented as Python scripts. At the beginning of script execution, the data loading module is configured to read a locally stored, structured spreadsheet file, which predefines all configurable runtime parameters in this embodiment. All subsequent calculation steps will query these parameters from the data structures loaded into the program's working memory.

[0021] The high-frequency harmonic energy ratio of the respiratory cycle, its parameter symbol is: This parameter characterizes the proportion of high-frequency components in the airflow signal relative to the fundamental respiratory energy. Its physical meaning lies in quantifying the intensity of non-laminar or turbulent components caused by minor disturbances within the airway (including sputum and airway wall contact).

[0022] The calculation of this parameter is based on Fourier transform theory in the field of signal processing. The calculation logic includes: a) The raw data comes from a high-frequency response MEMS pressure sensor integrated into the ventilator tubing. In this embodiment, the Honeywell HSC series or equivalent products are used, with key performance indicators of: sampling rate not less than 500Hz, pressure response range covering ±50cmH2O, and response time less than 1 millisecond. The acquired raw signal is a one-dimensional time-series voltage value, which has a linear relationship with airway pressure.

[0023] b) During the data loading phase, the program loads the gain and bias coefficients of the specific sensor according to the calibration file path specified in the "Running Configuration Parameters". Before processing each batch of raw voltage values, a calibration operation must be performed: the bias coefficient value is subtracted from the raw voltage value, and then the result is multiplied by the gain coefficient value to obtain a pressure value sequence in cmH2O; the pressure value sequence corresponds to 1024 points, corresponding to 2 seconds of data. c) To eliminate the effects of baseline pressure fluctuations, a high-pass filter is applied to the calibrated pressure value sequence. Specifically, a second-order Butterworth high-pass filter is used, with its cutoff frequency set to 0.05 Hz. This value is specified in the spreadsheet file of the operating configuration parameters to filter out slow drifts below the normal respiratory rate.

[0024] To filter out high-frequency electronic noise introduced by the sensor itself and the environment, a low-pass filter is then applied to the signal. Specifically, a fourth-order Butterworth low-pass filter is used, with its cutoff frequency set to 220Hz, a value specified in the spreadsheet file of the operating configuration parameters.

[0025] d) For the preprocessed pressure signal sequence with N sampling points, a Hanning window function is applied for point-by-point multiplication to smooth the signal ends and reduce spectral leakage. Subsequently, a Fast Fourier Transform is performed on the windowed signal sequence to obtain a frequency domain sequence containing N / 2+1 complex numbers.

[0026] e) For each complex number in the frequency domain sequence output by the Fast Fourier Transform (FFT), calculate the square of its modulus to obtain the power spectral density; read the lower and upper limits of the fundamental frequency breathing band and the lower and upper limits of the high-frequency harmonic band from the running configuration parameters; the program calculates the index range of the FFT output sequence corresponding to these frequencies. Summate all power spectral density values ​​within the fundamental frequency band index range and the high-frequency band index range respectively to obtain the total fundamental frequency energy and the total high-frequency energy. In this embodiment, the lower and upper limits of the fundamental frequency breathing band are 0.1Hz and 2Hz, respectively; the lower and upper limits of the high-frequency harmonic band are 50Hz and 200Hz, respectively. Calculate the ratio of total high-frequency energy to total fundamental frequency energy; if the total fundamental frequency energy is below a preset threshold, then the total fundamental frequency energy is considered to be the threshold. Finally, take the logarithm to base 10 of this ratio and multiply the result by 10 to obtain the final high-frequency harmonic energy ratio for the respiratory cycle. .

[0027] The airflow turbulence index, whose parameter symbol is This parameter characterizes the disorder or complexity of the airflow power spectrum. Its physical meaning lies in quantifying the degree of deviation from airflow stability. Highly ordered laminar flow corresponds to a low exponent, while turbulent airflow corresponds to a high exponent. In this invention, the energy distribution of the power spectrum at different frequencies is analogized to a probability distribution. Turbulent flow, characterized by a flat and broad power spectrum, has high uncertainty and therefore a high entropy value; while steady laminar flow, characterized by a power spectrum with energy concentrated at a few frequency points, has low uncertainty and a low entropy value. The calculation logic is as follows: the only input is the aforementioned ratio of high-frequency harmonic energy in the breathing cycle. The power spectral density sequence generated during the calculation process, covering frequencies from DC to the Nyquist frequency.

[0028] The entire input power spectral density sequence is summed to obtain the total energy across the entire frequency band. Each power spectral density value in the sequence is divided by the total energy across the entire frequency band to obtain a new sequence in which each element has a value between 0 and 1, and the sum of all elements is 1. This new sequence is considered the probability distribution sequence of the energy.

[0029] Iterate through each probability value in the probability distribution sequence; for each non-zero probability value, calculate the product of the probability value and its logarithm to the base of the natural constant e; sum all the calculated products.

[0030] Taking the negative of the sum, the result is the final airflow turbulence index. The value is a dimensionless positive number. In this embodiment, after probabilistic processing, the probability value at the selected frequency point A is 0.5, the probability value at frequency point B is 0.5, and the probability value for the rest is 0. Therefore, the term for frequency point A is calculated as 0.5 multiplied by ln(0.5); the term for frequency point B is calculated as 0.5 multiplied by ln(0.5). These two terms are added together, and the sum is negative to obtain the airflow turbulence index. .

[0031] HHER rate of change, the parameter sign is This parameter characterizes the ratio of high-frequency harmonic energy during the respiratory cycle. The trend and rate of change over a recent period are used to determine the suddenness of functional instability events.

[0032] The calculation of this parameter is based on linear regression analysis in statistics; the module maintains and stores the high-frequency harmonic energy ratios of the past N1 respiratory cycles. The module performs a univariate linear regression calculation on the data points in this queue, using time as the independent variable and the ratio of high-frequency harmonic energy during the respiratory cycle. Using the variable as the dependent variable, calculate the slope of the regression line. This slope is the rate of change of the HHER. In this embodiment, N1 is configured as 10 by an external file, corresponding to 20 seconds; HHER stability factor, its parameter symbol is: This parameter characterizes the ratio of high-frequency harmonic energy during the respiratory cycle. The degree of fluctuation over a recent period is used to determine whether the functional instability is persistent or intermittent. This parameter is calculated based on the standard deviation in statistics; the calculation logic utilizes the aforementioned ratio of high-frequency harmonic energies stored over the past N1 respiratory cycles. The queue of values ​​is used to calculate the standard deviation of all values ​​in the queue, and then the reciprocal of this deviation is taken to obtain the HHER stability factor. To avoid division by zero, when the standard deviation is zero, it is set to a small positive number 1e-9.

[0033] The rate of change of ATI, denoted as ATI stability factor, denoted as The method for determining these two parameters is related to the rate of change of HHER. and HHER stability factor The same, except the input data source for the calculation is replaced with the airflow turbulence index. .

[0034] The specific methods for determining the HHER and ATI change rates are as follows: The raw data consists of the ratio of high-frequency harmonic energy in the respiratory cycle, calculated at fixed time intervals of "every 2 seconds". or airflow turbulence index A historical value sequence; a fixed-length First-In-First-Out (FIFO) queue is maintained in memory, the length of which is specified as 10 by the runtime configuration parameters. Each time a new respiratory cycle's high-frequency harmonic energy ratio is reached... or airflow turbulence index Once the value is calculated, it is pushed to the end of the queue, while the oldest value at the head of the queue is removed.

[0035] The input for univariate linear regression is all the values ​​in the historical data queue, along with the timestamp sequence corresponding to each value. Standard least-squares linear regression is performed; specifically, the average of the time series and the parameter value sequence are calculated separately. Then, the covariance of the time series and the parameter value sequence is calculated. The variance of the time series is calculated. Dividing the covariance by the variance of the time series yields the slope of the regression line, which is defined as the rate of change parameter HHER or ATI.

[0036] The specific methods for determining the HHER stability factor and the ATI stability factor are as follows: In calculating the reciprocal of the standard deviation, the input is all the values ​​in the historical data queue; the arithmetic mean of all values ​​in the queue is calculated. For each value in the queue, the square of its difference from the mean is calculated. All these squared differences are summed to obtain the total sum of squares. The total sum of squares is divided by the length of the queue to obtain the variance. The square root of the variance is taken to obtain the standard deviation; the reciprocal of this standard deviation is calculated; to prevent a division-by-zero error when the standard deviation is zero (i.e., all historical values ​​are exactly the same), the calculated standard deviation is first compared with the smallest positive number specified by the runtime configuration parameters, and the larger of the two is taken before calculating the reciprocal. This reciprocal value is the HHER stability factor or ATI stability factor.

[0037] The above six parameters These parameters have different dimensions and numerical ranges. To fuse them, each parameter must be normalized before being input into the subsequent fusion model, mapping it to a closed interval [0,1]. This process is based on the sigmoid function in mathematics.

[0038] For each input parameter value to be normalized, the module reads the two configuration coefficients corresponding to that parameter from an external configuration file: the center point and the scaling factor. First, it calculates the difference between the input value and the center point, then multiplies this difference by the scaling factor, and finally negates the result. Next, it calculates the negative power of the natural constant e. It then adds 1 to this power. Finally, it divides 1 by this sum to obtain the final normalized parameter value within the [0,1] interval. This ensures that the output is close to 0.5 when the parameter is within the normal range, and asymptotically approaches 0 or 1 in extreme anomalies.

[0039] In this embodiment, the core computational process of the fusion mechanism is designed as a periodically executed sequence, and its specific steps are broken down as follows: 1.1) At the start of each computation cycle, the fusion mechanism first executes the data loading module, reading and updating all algorithm parameters from the local spreadsheet file into memory. To ensure accurate alignment between the high-frequency pressure signal and the low-frequency image frame, a hardware-level clock synchronization controller acquires and unifies the absolute timing markers of the two data streams. As a concrete implementation example, the hardware-level clock synchronization controller uses a general-purpose microcontroller with an independent real-time clock and hardware interrupt pins. This microcontroller generates a reference clock signal through a timer and synchronously sends this clock signal to the sampling trigger of the pressure sensor and the frame capture trigger of the infrared sensor, generating an aligned global timestamp at the physical level. The first data acquisition module acquires the latest 2-second airflow pressure signal based on this global timestamp; the second data acquisition module acquires the latest frame of raw thermal image based on this synchronized global timestamp. This image is generated by a 32×32 pixel miniature uncooled infrared focal plane array sensor customized at the end of the endotracheal cannula.

[0040] 1.2) The first data acquisition module calculates the high-frequency harmonic energy ratio of the current respiratory cycle. and airflow turbulence index Instantaneous value.

[0041] The module updates the queue storing historical data and calculates... Four time-dimensional dynamic parameters; for these six core parameters, the defined S-shaped function normalization model is applied to obtain six normalized input factors in the interval [0,1].

[0042] 1.3) The image processing module inputs these six normalized input factors into its built-in "two-level cascaded dynamic adaptive fusion mechanism," which is the core of this invention. This fusion mechanism performs a series of internal calculations and ultimately outputs three key control parameters: the center coordinates and radius of the target region of interest (ROI); and the gamma correction coefficient. Brightness adjustment factor .

[0043] 1.4) The image processing module receives the control parameters output in step 1.3); on the input raw thermal image frame, it locates the specified Region of Interest (ROI). For each pixel within the ROI region, the module obtains its original grayscale value, uses this grayscale value as the base, and applies a gamma correction factor. The result is then exponentially multiplied by a brightness adjustment factor. This yields new pixel grayscale values. Pixels outside the ROI region remain unchanged.

[0044] After the above processing, the fusion mechanism generates and outputs an "enhanced thermal image". This "enhanced thermal image" will be sent to the display unit and the early warning module for further processing.

[0045] Furthermore, the technical elements of this fusion mechanism consist of six normalized input factors. These elements collectively constitute a six-dimensional event feature vector, describing the intensity, complexity, suddenness, and duration of airway dysfunction events. This invention reveals that instantaneous values ​​alone cannot effectively distinguish between airway events of different natures. Both a violent cough and a severe intubation displacement can produce high instantaneous values, but the former has a high rate of change and low stability, while the latter has a relatively low rate of change but stabilizes at a high level once it occurs. Therefore, there is a nonlinear, strongly coupled intrinsic correlation between instantaneous values ​​and dynamic characteristic parameters; this correlation is key to distinguishing the root causes of events.

[0046] Traditional linear weighting methods cannot handle this non-linear coupling relationship; simply summing the six factors by weight can confuse the characteristics of different event types, resulting in a consistent enhancement effect in the output and failing to achieve diagnostic classification. Fixed threshold rules are even more rigid and cannot adapt to individual patient differences and disease evolution.

[0047] To address this limitation, this invention designs a two-level cascaded dynamic adaptive fusion mechanism that deeply integrates the design concepts of "dynamic adaptive weights" and "multi-level cascaded models".

[0048] Level 1: Event Attribute Representation Layer. This layer's function is to "interpret" the input vector, transforming it into the following two intermediate state factors with clear physical meaning: The first "intermediate state factor" is the event severity factor, denoted as... The severity factor of this event is mainly determined by the normalized ratio of high-frequency harmonic energies in the respiratory cycle. and airflow turbulence index The instantaneous value is determined by a weighted sum of the two values. The weights here are not fixed, but are determined by a normalized stability factor. Dynamic adjustments are made. Specifically, when the stability factor is high, it indicates that the state is stable, and the weight of the instantaneous value will increase accordingly; conversely, it will decrease. The weight is determined as follows: Determine an optimal set of baseline weights and a dynamic adjustment function to ensure that the calculated event severity factor... Maximize the correlation between the event severity and the clinician's subjective rating of the event severity.

[0049] The dataset used to train the machine learning model contains 50,000 samples, each with a six-dimensional feature vector and a corresponding clinical event label. Clinical event labels include no event, mild sputum, severe sputum blockage, and tube displacement. Three senior ICU physicians were invited to anonymously rate the severity of each sample on a scale of 1 to 10, and the average score was taken as the true severity value for that sample. .

[0050] Event Severity Factor The calculation logic is defined as: calculating the ratio of high-frequency harmonic energy in the respiratory cycle. The weight is determined by adding the HHER stability factor to the baseline weight value. The relevant gain terms are constructed; secondly, the airflow turbulence index is calculated in the same manner. The weights; finally, the normalized ratio of high-frequency harmonic energies in the respiratory cycle. and airflow turbulence index The severity factor is obtained by multiplying each factor by its corresponding dynamic weight and then summing them. This process involves four parameters to be calibrated: the ratio of high-frequency harmonic energy during the respiratory cycle. The baseline weights and HHER stability factors Gain coefficient, airflow turbulence index The baseline weights and ATI stability factors Gain coefficient.

[0051] The calculated event severity factor Sequence and Severity Truth Value The Pearson correlation coefficient between the sequences is used as the objective function.

[0052] The Particle Swarm Optimization (PSO) algorithm is used to search for the four-parameter combination that maximizes the objective function within a predefined parameter space.

[0053] Through the above optimization experiments, a set of preferred parameters was obtained: the ratio of high-frequency harmonic energy in the respiratory cycle. The baseline weight is 0.6, with a value range of 0.5-0.7, for the HHER stability factor. The gain coefficient is 0.3, and its value ranges from 0.2 to 0.4. The airflow turbulence index... The baseline weight is 0.4, with a value range of 0.3-0.5, and is the ATI stability factor. The gain coefficient is 0.2, and its value ranges from 0.1 to 0.3.

[0054] The second "intermediate state factor" is the event dynamic factor, denoted as... The dynamic factors of this event Mainly composed of normalized rate of change factor The decision was also obtained through a weighted summation.

[0055] For the second-level processing: construct an image correction parameter generation layer; this layer receives the event severity factor output from the first level. and event dynamics factors As input, it generates the final image correction parameters.

[0056] For adaptive adjustment of the size of the Region of Interest (ROI) based on event dynamics factors, specifically including ROI localization and radius determination: the association mapping model receives the original six input factors and outputs the center coordinates of the ROI. The initial search radius of the ROI is determined by the event dynamics factors. Adjustments are made: Higher dynamism indicates a more intense event, requiring a larger initial search radius to capture a wider range of impacts; specifically, the "ROI base radius" is obtained from the loaded runtime configuration parameters. This parameter defines the ROI base radius when the system is completely silent, i.e. The minimum ROI size when it is 0 ensures a basic monitoring range. Obtain the "ROI dynamic radius scaling factor" from the runtime configuration parameters. This parameter defines the "sensitivity" or "gain" of the ROI radius to event dynamics factors, determining how the radius scales. The degree of change. The "event dynamics factor" of the current computation cycle is received from the first-level processing layer of the fusion model. The value has been normalized to the [0,1] interval. The product of the "Event Dynamics Factor" and the "ROI Dynamic Radius Scaling Factor" is calculated to obtain the dynamic increment value of the radius. This "Dynamic Increment Value of Radius" is added to the "ROI Base Radius," and the sum is the final output ROI radius for the current image frame. In this embodiment, the "ROI Base Radius" is preferably 10 pixels; the "ROI Dynamic Radius Scaling Factor" is preferably 20 pixels.

[0057] The construction of the correlation mapping model is explained as follows: The core objective of the correlation mapping model is to predict the most likely thermodynamic location of the instability in the tracheal carina region, i.e., the center coordinates of the output ROI, based on the functional instability characteristics of the airflow.

[0058] This embodiment employs a feed-forward neural network. Its structure consists of an input layer with 6 neurons, two hidden layers each containing 32 neurons, and an output layer with 2 neurons (corresponding to the X and Y coordinates respectively). Each hidden layer uses a modified linear unit (ReLU) as the activation function, while the output layer uses a linear activation function. The dataset was constructed by combining computational fluid dynamics (CFD) simulations with clinically labeled data.

[0059] Using software such as ANSYS-Fluent, a detailed 3D model of the trachea and carina was constructed. Tens of thousands of simulation scenarios were generated by simulating varying sizes of sputum adhesion or minute deformations of the trachea wall at different locations. For each scenario, the corresponding airflow pressure signal and heat flux distribution map at the trachea inlet were calculated. Then, a six-dimensional event feature vector was calculated from the pressure signal, and the center points of abnormal areas were manually marked from the heat flux map, forming (feature vector, coordinate) data pairs.

[0060] Clinical Data Supplement: Simultaneous pressure and thermal imaging data were collected one minute prior to the occurrence of real clinical events (including sputum blockage and displacement) confirmed by bronchoscopy. Two senior ICU physicians independently retrospectively annotated the thermal images, and the average position of the annotated points was taken as the true value.

[0061] Dataset Size and Split: A total of 50,000 data samples were randomly split into three groups: 80% (training set), 10% (validation set), and 10% (test set). The Mean-Squared Error (MSE) loss function was used to calculate the squared Euclidean distance between the model's predicted coordinates and the true coordinates. The Adam optimizer was used with an initial learning rate of 0.001. The total number of training epochs was 100, and the batch size was 64. During training, if the validation set loss did not decrease for five consecutive epochs, training was terminated early to prevent overfitting.

[0062] After training, the model's weights and bias parameters are saved as separate files. During the fusion mechanism, this model file is loaded into memory. The image processing module takes the real-time six-dimensional feature vector as input and performs a single forward propagation through the model to obtain the center coordinates of the ROI.

[0063] Furthermore, gamma correction coefficient The determination method is as follows: it is composed of a baseline state item, a basic item, and a dynamic item. Specifically, the preset gamma baseline value is 1 to represent the unenhanced silent state; the basic item is determined by the basic weight coefficient and the event severity factor. The product of these factors determines the strength of the enhanced baseline; the dynamic term is determined by the dynamic weighting coefficient and the event dynamics factor. The product of these factors is used to provide an additional, brief, enhancing "pulse" during sudden events to highlight rapid changes. The final calculation logic is as follows: ;in Characterizing the basic weight coefficients, Characterizes dynamic weighting coefficients.

[0064] Brightness adjustment factor The method for determining this value is as follows: it is used to compensate for the overall brightness change that gamma correction may cause, and its value is determined by using a brightness adjustment lookup table (LUT) based on the gamma correction coefficient. The values ​​are determined to maintain visual comfort in the ROI area. The brightness adjustment lookup table (LUT) is generated as follows: A set of optimal values ​​(gamma correction coefficients, brightness adjustment factor) is determined. The mapping relationship ensures that the average brightness and visual recognizability of the ROI region of the enhanced thermal image remain within the optimal range under different enhancement intensities.

[0065] One hundred representative raw thermal images were selected, covering different background temperatures and noise levels. These were processed in a DICOM-compliant medical imaging reading room environment using a calibrated professional medical monitor. Ten engineers or physicians with experience in thermal imaging interpretation were invited as evaluators. For each sample image, the program applied a gamma correction factor to a pre-defined central region. The value increases from 1 (no enhancement) in steps of 0.1 to 4 (maximum enhancement).

[0066] In each gamma correction coefficient At this value, evaluators can adjust the brightness adjustment factor in real time using a slider. The value (range 0.5-1.5) is used until they deem the area to have "the clearest detail and no obvious overexposure or underexposure." Each evaluator records the value for each gamma correction factor. Brightness adjustment factor selected at the value Value. For each gamma correction coefficient Step size points are used to collect all brightness adjustment factors selected by the evaluators. The values ​​are calculated by removing outliers and then averaging them. These (gamma correction coefficients, average brightness adjustment factor) are used... The data pairs are used to generate a smooth brightness adjustment lookup table (LUT) containing 1024 data points through a cubic spline interpolation algorithm.

[0067] This embodiment demonstrates the beneficial technical effects of the two-level cascaded fusion mechanism: By decoupling the "severity" and "dynamics" of an event, the final image enhancement effect is no longer a simple change in "brightness," but rather a visual encoding with "diagnostic information." An enhanced image dominated by "basic factors" and exhibiting continuous high brightness strongly suggests a chronic, serious problem (e.g., phlegm blockage); while an image showing brief, intense "pulsating" enhancement points to a sudden event (e.g., displacement caused by a change in body position).

[0068] Furthermore: when the pressure signal is stable and undisturbed, all six input factors, after normalization, approach a baseline of 0 or 0.5. At this point, the event severity factor... and event dynamics factors All of them approach zero, and the final output gamma correction coefficient The brightness adjustment factor is 1. A value of 1 is equivalent to no enhancement operations, and the fusion mechanism remains stably in a "silent" state.

[0069] When a catastrophic event causes all input factors to reach their saturation limits, the event severity factor... and event dynamics factors It will reach its design limit, and the final output gamma correction coefficient will be... The ROI radius will also be limited to the maximum safe value preset in the external configuration file to prevent invalid or misleading image output.

[0070] The operational configuration parameters defined in this invention, such as the aforementioned filter cutoff frequency, historical data queue length, and normalized function coefficients, are all pre-configured and stored in a structured external data carrier. This data carrier is logically a collection of key-value pairs, and its physical implementation can be a file stored in the device's local file fusion mechanism or a database table entry stored on a cloud server.

[0071] When the fusion mechanism or method of this invention is activated or runs within a preset period, it first executes a parameter loading program. This program is configured to access the aforementioned external data carrier, retrieve and extract a set of currently effective running configuration parameter values ​​through predefined keys, and load them into the program's working memory data structure to guide the computational logic of all subsequent algorithm modules.

[0072] In a preferred embodiment, the structured external data carrier is a local text file stored in comma-separated values ​​(CSV) format. The parameter loader then parses the contents of this CSV file, converting it into an internal hash table or dictionary data structure, where the file's column headers serve as keys and the corresponding row values ​​as values. This approach allows technical personnel without a programming background to easily view and modify the configuration using a common text editor or spreadsheet software.

[0073] Further explanation: The early warning module is further configured as follows: Image segmentation is performed on the enhanced thermal image to identify and separate the heat flux regions corresponding to the left and right bronchi, respectively; Based on the heat flux regions of the left and right bronchi, the heat flux symmetry coefficient of the two lungs, which represents the symmetrical state at the current moment, is calculated.

[0074] Further explanation: The early warning module is further configured as follows: Based on the symmetry coefficient of bilateral lung heat flux, the instantaneous event severity, which characterizes the severity of the current event, is calculated. Maintain and store a symmetrical time series of the symmetry coefficients of the heat flux of both lungs at multiple consecutive moments within a preset time window; Based on the changing trends and fluctuations of symmetrical time series, the dynamic deterioration risk degree, which characterizes the risk of event development, is calculated.

[0075] Further explanation: The early warning module is further configured as follows: The severity of an instantaneous event is combined with the risk of dynamic deterioration to generate a final alert level; Based on the final alarm level and image segmentation results, a diagnostic warning signal containing event location information and suggested alarm level is generated.

[0076] Further explanation: The method of integrating the instantaneous event severity and the dynamic deterioration risk is as follows: the dynamic deterioration risk is used as a dynamic gain factor to nonlinearly amplify the instantaneous event severity in order to determine the final alarm level.

[0077] Further explanation: Continuously monitor the validity of the data streams from the first data acquisition module and the second data acquisition module; When any data stream interruption is detected, the system automatically switches to a single-modal monitoring mode driven by the remaining valid data streams and generates a system status degradation notification. The continuous monitoring method is as follows: maintain a heartbeat timer for each data stream, and determine that the data stream is interrupted when no valid data is received within a preset time window.

[0078] The following is a detailed implementation description of the above content: The core technology of this embodiment lies in the design of a two-level cascaded dynamic risk assessment model, which is implemented as follows: Level 1: Segment and extract features from the current enhanced thermal image frame to calculate the instantaneous event severity that characterizes the airway patency at the current moment.

[0079] Level 2: The instantaneous event severity output from Level 1 is used as the latest data point of the time series for maintenance, and the dynamic deterioration risk degree, which characterizes the development trend and urgency of the event, is calculated based on the rate of change and volatility of the time series.

[0080] The final alert level is generated by combining the severity of the instantaneous event with the risk of dynamic deterioration through a nonlinear fusion mechanism.

[0081] In this embodiment, all configurable operating parameters are predefined and stored in a structured external data carrier. Preferably, this data carrier is a spreadsheet file, which is read by a data loading module within a Python program script and loaded into the program's working memory upon system startup, thereby achieving technical decoupling between the core algorithm logic and specific application strategies. The following are the key parameters upon which the early warning module depends for operation: The symmetry coefficient of heat flux in both lungs, with the parameter symbol being... This parameter quantifies the symmetry of heat flux distribution in the left and right bronchial regions; its physical meaning is a dimensionless similarity measure. Its value range is normalized to the interval [0,1], where 1 represents complete symmetry and 0 represents complete asymmetry. The specific implementation steps are as follows: Acquire a single-frame enhanced thermal image output by the image processing module. This enhanced thermal image is a two-dimensional matrix, where the value of each pixel represents the relative heat flux intensity at the corresponding physical location.

[0082] The early warning module receives the enhanced thermal image and applies a label-based watershed segmentation algorithm. The algorithm's specific execution logic is as follows: it performs a distance transformation on the image and finds local maxima in the transformation result as initial label points or "seed points." Then, using these label points as starting points, it performs region growing on the image gradient map, ultimately segmenting the image into multiple regions. In this embodiment, the algorithm is configured to specifically identify and separate two high-heat-flux regions corresponding to the left and right bronchi, respectively, denoted as the left region. and the right side area .

[0083] To achieve robust description of the region shape, the module targets the segmented left-side region. and the right side area The module independently calculates a set of mathematical features that remain invariant under translation, rotation, and scaling transformations of the image. In a preferred embodiment, the features used are Hu-moment-invariants. The module calculates a complete set of invariant moments consisting of seven independent values ​​for each of the two regions, forming two seven-dimensional feature vectors, denoted as the left feature vector. and the right-hand feature vector .

[0084] The module calculates the left-hand feature vector. and the right-hand feature vector The Euclidean distance between the two regions is used to quantify the difference in shape. A larger distance indicates a greater shape difference and poorer symmetry. In this embodiment, to convert this distance value into an intuitive and easily calculated symmetry coefficient within the range [0,1], the module uses a negative exponential function for normalization. The calculation logic is as follows: Obtain the calculated Euclidean distance value, denoted as D; calculate the natural constant e raised to the power of -D; this result is the final symmetry coefficient of the two-lung heat flux. This function ensures the symmetry coefficient of the heat flux of the two lungs when the Euclidean distance D is 0, representing complete symmetry. The coefficient of symmetry of heat flux in the lungs is 1; as the Euclidean distance D approaches infinity, the coefficient of symmetry of heat flux in the lungs is 1. It approaches 0.

[0085] The threshold for determining symmetry deterioration, with parameter symbol as follows: This parameter is used to determine the symmetry coefficient of current bilateral lung heat flux. This refers to the static threshold determining whether a value has entered a critical abnormal range. It is determined through statistical analysis of thermal image data from a large number of healthy individuals and patients known to have unilateral ventilatory dysfunction, selecting a bilateral lung thermal flux symmetry coefficient that can distinguish between the two groups with 95% confidence. The value is used as the threshold. In a preferred embodiment, this symmetry deterioration determination threshold... It was set to 0.8.

[0086] Symmetric time series, whose parameter notation is This parameter is a first-in, first-out (FIFO) queue used to store the symmetry coefficients of the bilateral lung heat flux calculated from the most recent N2 consecutive sampling times. These values ​​form the data foundation for time-domain analysis.

[0087] The window length for time series analysis, its parameter symbol is: This parameter defines the symmetry of the time series. The length of the queue refers to the number of historical data points used for time-domain analysis. It is determined based on the typical duration of clinical respiratory events, aiming to ensure the window covers the dynamic changes of the event without being excessively long and sluggish. In a preferred embodiment, this value is set to 50, corresponding to 5 seconds of data at an image processing frequency of 10 Hz.

[0088] The rate of change of symmetry, with parameter sign as This parameter is used to quantify the symmetry coefficient of heat flux in both lungs. The trend and rate of change within the time window are physically represented by the first derivative of symmetry with time. Its range is normalized to the [0,1] interval, with larger values ​​indicating a faster rate of decrease. The determination method is detailed below: Its core principle is the application of least squares linear regression in statistics, aiming to find a straight line that best fits a set of two-dimensional data points and extract the slope of this line as a measure of the rate of change. The calculation logic is as follows: The input is a symmetric time series. A set of data points. This set of data includes the time series analysis window length. There are 1 data pairs, each indexed by a time index (from 1 to 1). ) and the corresponding bilateral lung heat flux symmetry coefficient Value composition.

[0089] Specifically, the average value of the time-indexed sequence and the symmetry coefficient of the bilateral lung heat flux are calculated. The average of the value sequence.

[0090] The module calculates the symmetry coefficient of the time index sequence and the heat flux of the two lungs. The covariance of the value sequence, and the variance of the time index sequence itself.

[0091] Divide the covariance calculated in the previous step by the variance. The quotient is the slope of the best-fit line, denoted as Slope.

[0092] The slope value is judged and normalized: if the slope is positive or zero, it indicates that the symmetry is improving or remaining unchanged, then the rate of change of symmetry is normalized. The value is assigned to 0; if the slope is negative, indicating that the symmetry is deteriorating, the module takes its absolute value and maps it to the interval (0,1) using the logistic function. This mapping ensures that the larger the absolute value of the slope, the faster the deterioration and the faster the rate of symmetry change. The closer the value is to 1, the better. The logistic growth rate is denoted as k; it characterizes the sensitivity of the deterioration rate to the nonlinear amplification of the system's early warning risk output. The logistic offset is denoted as... Its characterization system triggers a critical point boundary for the rate of increase in the dynamic deterioration risk rating. The aforementioned parameters are pre-set to ensure a good dynamic range of output within the clinical deterioration rate range.

[0093] The key parameters (growth rate and offset) in the aforementioned logistic function were determined through offline analysis of a large amount of historical clinical data. During the analysis, clinical experts labeled events with different rates of deterioration, and then optimization algorithms were used to find the function parameters that best distinguish the severity of these events.

[0094] To determine the growth rate k and offset in the logistic function This invention employs the following offline calibration method: A historical monitoring dataset containing at least 100 ICU patients, with a total duration exceeding 2000 hours, and which has undergone ethical approval and data anonymization, is prepared. This dataset contains a synchronous data stream consistent with the first and second data acquisition modules of this invention. A panel of experts consisting of three senior ICU physicians evaluates the symmetry coefficients of thermal flux in all bilateral lungs within the dataset. Events showing a significant decline were reviewed, and those marked as "slowly deteriorating" were determined based on the absolute value of the slope, which represents the rate of decline. "Medium-speed deterioration" ) and "rapid deterioration" There are three levels.

[0095] The optimization objective is set as follows: to find a set of parameters. This allows the logistic function to transform the absolute value of the slope of these three types of events, resulting in a symmetric rate of change. The output values ​​achieve optimal class discrimination. Specifically, by executing a grid search optimization algorithm, a set of parameters is found within a predefined parameter space that maximizes the distance between the mean values ​​of the three classes of output values; in this example, the predefined parameter space is... In a preferred embodiment, the preferred value of the growth rate k determined by the above method is 100, and the offset is... The preferred value is 0.1.

[0096] Symmetric volatility factor, its parameter sign is This parameter is used to quantify the symmetry coefficient of heat flux in both lungs. The degree of instability or oscillation within a time window. Its physical meaning is the dispersion of the time series. Its value range is normalized to the [0,1] interval; a larger value indicates more severe fluctuations. The determination method is detailed below: it applies the standard deviation used in descriptive statistics to measure the dispersion of data; the calculation logic is as follows: The input is a symmetric time series. Time series analysis window length Symmetry coefficient of heat flux in both lungs value.

[0097] Calculate the length of this time series analysis window. Symmetry coefficient of heat flux in both lungs The arithmetic mean of the values.

[0098] For each bilateral lung heat flux symmetry coefficient in the sequence The variance of the data series is calculated by taking the square of the difference between the variance and the arithmetic mean. The average of all these squared differences is then calculated. The square root of this variance is calculated to obtain the standard deviation, denoted as StdDev. To normalize this standard deviation to the [0,1] interval, the module applies a preset scaling function. This function maps the standard deviation, which represents "high volatility" as determined by clinical data, to 1 and 0 to 0, using linear interpolation in between. Any standard deviation greater than this "high volatility" threshold is capped at 1. This normalized value is the final symmetry volatility factor. In this embodiment, the standard deviation of "high volatility" is set to 0.3. Transient event severity, its parameter symbol is This parameter is the output of the first level, representing the severity of the event reflected in the current single-frame image. Its value range is [0,1], with larger values ​​indicating a more severe current state. Its calculation logic is a conditional linear mapping based on a threshold. The input is the bilateral lung heat flux symmetry coefficient. And the threshold for symmetry deterioration The output is the severity score in the range [0,1]. This logic ensures that the severity score is only calculated when the symmetry deteriorates to a certain extent, effectively filtering out minor fluctuations within the normal range.

[0099] The dynamic deterioration risk level, whose parameter symbol is: This parameter is the second-level output, characterizing the deterioration trend and risk of the event over time. Its value range is [0,1], with larger values ​​indicating a higher risk of deterioration. Its calculation logic is based on the rate of change of symmetry. and symmetry fluctuation factor The weighted geometric mean is designed to emphasize the combined effect of two risk factors, "rapid change" and "violent fluctuation," and amplifies risk signals more effectively than the arithmetic mean.

[0100] The final alarm level, its parameter symbol is: This parameter is the final output of the entire early warning module. It is a continuous value in the range [0,1] and is used to drive specific alarm devices. The larger the value, the higher the alarm level.

[0101] The data stream heartbeat monitoring time window, its parameter symbol is: This parameter is used for degradation strategies, defining the maximum allowable period of no data for determining whether a data stream is interrupted. It is determined based on the nominal update frequency of the system data stream, and is 2 to 3 times the data update cycle. For example, if the nominal update frequency of the HHER / ATI value is 10Hz and the cycle is 100ms, then the data stream heartbeat monitoring time window... The preferred setting is 300ms.

[0102] The early warning module is configured to periodically execute the following calculation process, performing the data flow monitoring process in parallel.

[0103] 2.0) At the beginning of each computation cycle, acquire the latest frame of enhanced thermal image output by the image processing module and access the symmetric time series stored in the working memory. .

[0104] Execute Level 1 transient event severity calculate; 2.1) Before performing watershed segmentation on the enhanced thermal image, the real-time yaw angle features of the endotracheal tube tip are obtained through the spatial attitude perception module. As a specific general engineering implementation example, the spatial attitude perception module uses an inertial measurement unit (IMU). The image processing module reads the gravitational acceleration vector output by the IMU, calculates the yaw angle of the current tube relative to the coronal plane of the human body, and uses this yaw angle to perform an inverse two-dimensional rotation affine transformation on the two-dimensional matrix of the enhanced thermal image. After this reference coordinate correction, a label-based watershed segmentation algorithm is applied to the enhanced thermal image to obtain the heat flux regions of the left and right bronchi, thereby ensuring that the segmented left and right regions of the image are physically accurate and always correspond to the patient's anatomical left and right bronchi.

[0105] 2.2) Calculate the symmetry coefficient of the heat flux of both lungs at the current moment based on these two regions. .

[0106] 2.3) The latest bilateral lung heat flux symmetry coefficient Values ​​are pushed into symmetric time series The oldest value at the head of the queue is removed from the end of the queue to maintain the queue length as the time series analysis window length. .

[0107] 2.4) Module execution logic judgment: If the current bilateral lung heat flux symmetry coefficient Value greater than the symmetry deterioration threshold The severity of the instantaneous event will be determined. The value is assigned to 0; otherwise, the symmetry coefficient of the bilateral lung heat flux is adjusted using a linear mapping function. Value from By mapping the interval back to the [0,1] interval, the instantaneous event severity can be obtained. .

[0108] Perform the second level of calculation for dynamic deterioration risk: 3.1) Based on the updated symmetry deterioration threshold Queues, calculate the rate of symmetric change .

[0109] 3.2) Based on the updated symmetry time series Queue, calculate the symmetry fluctuation factor .

[0110] 3.3) The rate of change of symmetry and symmetry fluctuation factor Weighted fusion is performed to calculate the dynamic deterioration risk level. In a preferred embodiment, this fusion is accomplished by calculating a weighted geometric mean of the two parameters to emphasize the risk when both are present simultaneously; the risk level is dynamically deteriorating. The specific calculation logic is as follows: Obtaining the rate of change of symmetry and their corresponding weight parameters and symmetric volatility factor and their corresponding weight parameters .in, Used to quantify the importance of factors related to the rate of change. Used to quantify the importance of fluctuation factors, where the sum of the two is 1. Calculate the rate of change for symmetry separately. of Powers and symmetric fluctuation factors of The result of the two exponentiation operations is multiplied together, and the product is the final dynamic deterioration risk level. .

[0111] This embodiment is for determining weight parameters. and To determine the optimal value, a regression analysis based on historical data was performed. In this analysis, clinically confirmed "emergency intervention events" (including those requiring emergency reintubation) were used as the target variable, and the symmetrical rate of change over the period preceding the event was used as the metric. and symmetry fluctuation factor As independent variables, the contribution of the two independent variables to the target variable is analyzed using a logistic regression model. The analysis results show that the predictive power of the rate of change factor is significantly higher than that of the fluctuation factor. Based on this analysis, in a preferred embodiment, the weighting parameters... The preferred value is 0.7, and the weighting parameter is... The preferred value is 0.3.

[0112] 4.1) Determine the severity of the instantaneous events output by the first level. Dynamic deterioration risk level of the second-level output Perform nonlinear fusion to calculate the final alarm level. 4.2) Execute logical judgment: If the final alarm level If the threshold exceeds the preset minimum threshold for triggering an alarm, the module generates a structured diagnostic warning signal. This signal consists of two parts: event location information, namely the centroid coordinates and circumscribed rectangle of the asymmetric region determined by the watershed algorithm; and a suggested alarm level, which is the final alarm level. The values ​​are mapped to three discrete levels: "low", "medium", and "high", as shown in the table below: Table 1: Key Operational Configuration Parameters for the Early Warning Module

[0113] Continuous monitoring: A timer is continuously maintained in the background for both the first and second data acquisition modules.

[0114] Heartbeat Reset: The timer for the corresponding data stream is reset each time a valid data packet containing HHER / ATI values ​​or a new thermal image frame is successfully received.

[0115] 3.1) If the timer value of the pressure signal stream exceeds the heartbeat monitoring time window of the data stream... The module immediately triggers a system-level degradation mode switch: it bypasses all spectrum-guided local adaptive image enhancement functions that depend on HHER / ATI values, and instructs the image processing module to execute a global compensation mechanism, i.e., performing conventional global histogram equalization on the original thermal image to passively and holistically increase the matrix dynamic contrast; simultaneously, the warning module's calculation process will only perform symmetry analysis based on the thermal image after this global equalization, ensuring that the visual extraction capability of the baseline can still be preserved by brightening even in the absence of precise target localization features. The module also generates a system status degradation notification of "Spectrum monitoring failure, switch to pure thermal imaging mode".

[0116] 3.2) If the timer value of the thermal image signal stream exceeds the heartbeat monitoring time window of the data stream... The module also triggers a mode switch: it stops the image analysis process of all image processing and early warning modules and instructs the backup algorithm module, which detects abnormal waveforms solely based on historical HHER / ATI values, to begin operation. Simultaneously, the module generates a system status degradation notification: "Thermal imaging monitoring failed; switching to pure spectrum analysis mode."

[0117] When the early warning module detects an interruption in the thermal image signal stream and switches to single-mode monitoring mode, its backup abnormal waveform detection algorithm module activates. The specific workflow of this backup module is as follows: During normal dual-mode operation, this backup module continuously receives historical time series of HHER and ATI values ​​and maintains the autoregressive integral moving average model ARIMA(p,d,q) online. The autoregressive order of the model is denoted as p; it represents the backtracking time span window by which the model uses past historical abnormal fluctuation values ​​to predict the current signal trend. The difference order of the model is denoted as d; it represents the number of mathematical derivatives required for the system to transform the non-stationary airflow time series into a stationary data series. The moving average order of the model is denoted as q; it represents the noise-resistant smoothing range by which the system uses past prediction residuals to correct the current prediction value. The various order parameters (p,d,q) of this model are periodically and automatically optimized according to the Akaike Information Criterion (AIC) to best fit the normal data fluctuation pattern of the most recent 5-minute period.

[0118] When mode switching is triggered, for each newly received HHER or ATI value, the module performs the following steps: Using the established ARIMA model, based on previous data points, predict the value at the current time point and its 99% prediction confidence interval. Calculate the difference between the currently received value and the model's predicted value, i.e., the prediction residual.

[0119] If the absolute value of the prediction residual exceeds the boundary of the prediction confidence interval given by the model, an abnormal fluctuation event is determined to have occurred. If the cumulative number of such abnormal fluctuation events exceeds 3 within a preset time window of 10 seconds, the module will generate and output a diagnostic warning signal, which reads "Thermal imaging monitoring failed, and the spectral data showed significant abnormalities."

[0120] If both data streams return to normal, the system will automatically switch back to dual-modal monitoring mode.

[0121] Further explanation: The severity of instantaneous events is determined using a conventional linear weighted average method. and dynamic deterioration risk Adding them together fails to effectively address critical clinical scenarios: when the severity of transient events... High but with a risk of dynamic deterioration When the value is 0, indicating a long-term stable state, linear weighting will still output a high alert; conversely, when the severity of the instantaneous event is high... The value is small but the risk of dynamic deterioration is high. When the intensity is high, the linear weighting is based on the severity of the instantaneous event. The low weight of the signal makes it impossible to issue high-level alerts in a timely manner.

[0122] To address this limitation, this invention designs an adaptive gain fusion mechanism based on dynamic degradation risk. The core idea of ​​this mechanism is to integrate the dynamic degradation risk... As an amplification factor, it dynamically and non-linearly adjusts the severity of instantaneous events. Impact on the final alert level.

[0123] Perform the following steps to determine the final alarm level. Get the severity of an instantaneous event. and dynamic deterioration risk Next, the dynamic gain factor is calculated, which is the dynamic deterioration risk level. The exponential function is specifically defined as follows: taking the basic gain constant as the base, and dynamically deteriorating the risk level. Perform a power operation on the exponent; combine the calculated dynamic gain factor with the instantaneous event severity. The multiplication is used to extract the first alarm feedforward term; simultaneously, the dynamic deterioration risk level is extracted. The first alarm feedforward term is used as the second alarm feedforward term. The first and second alarm feedforward terms are summed. This summation result is processed by a peak-shaving function represented by the "tanh function" to ensure that the final output value is strictly limited to the interval [0,1], thus obtaining the final alarm level. .

[0124] In a preferred embodiment, the preferred value of the base gain constant is e. In other embodiments, this value is also adjusted within a reasonable range of [2,4] to accommodate the specific sensitivity requirements of different monitoring scenarios.

[0125] When the risk level deteriorates dynamically When the dynamic gain factor approaches 0, it approaches 1, and the final alarm level... Approaching the instantaneous event severity The system accurately reflects the severity of the current situation and will not trigger excessive alarms.

[0126] When the risk level deteriorates dynamically As the value approaches 1, the dynamic gain factor increases exponentially, even with the severity of instantaneous events. Although its value itself is not large, its product with the gain factor will quickly saturate to near 1, triggering the highest level of alarm. This makes the system highly sensitive to acute deterioration events, achieving risk "prediction" rather than a delayed "response".

[0127] When the symmetry coefficient of heat flux in both lungs When perfectly symmetrical and the duration is 1, the severity of the instantaneous event Constantly 0, rate of change of symmetry and symmetry fluctuation factor Also zero, leading to a dynamic deterioration risk level. A value of 0 indicates the final alert level. The value is stable at 0, and the system outputs no alarms.

[0128] When the symmetry coefficient of heat flux in both lungs When the instantaneous event severity drops and remains at 0, the instantaneous event severity is... Immediately saturates to 1. In the early stages of change, the rate of symmetric change... The risk of reaching its peak and causing dynamic deterioration will be high. It also reached its peak, leading to the final alarm level. It rapidly saturates to 1. Subsequently, due to the symmetry coefficient of heat flux in both lungs... Stabilized at 0, rate of change of symmetry and symmetry fluctuation factor It will drop to 0, leading to a dynamic deterioration of the risk level. The alert level drops to 0, at which point it becomes the final alert level. It will fall back to the severity of the instantaneous event. The determined high-level stability value. This behavior represents the clinical logic of "acute event occurrence → highest emergency alarm → transition to continuous severe state alarm".

[0129] The following are detailed implementation instructions for the above content: Final Alert Level It is a dimensionless value normalized to the interval [0,1]; when the final alarm level The smaller the time, the more accurately the system judges the current endotracheal tube position and the more symmetrical and stable the ventilation of both lungs.

[0130] When the final alarm level As the value approaches 0, it indicates that the system correctly identifies the current endotracheal tube position and that bilateral lung ventilation is symmetrical and stable. When the final alarm level... The closer the value is to 1, the more severe and rapidly deteriorating the asymmetry in bilateral lung ventilation is detected by the characterization system, indicating a greater likelihood of acute and dangerous deviation of the endotracheal intubation position, and a greater need for the highest priority clinical intervention.

[0131] The six time-domain spectral features input to the image processing module are: the high-frequency harmonic energy ratio of the respiratory cycle, the airflow turbulence index, and their respective rates of change and stability factors.

[0132] The event severity factor is positively correlated with the high-frequency harmonic energy ratio of the respiratory cycle and the airflow turbulence index. The physical logic is that the higher the static values ​​of these two indices, the more severe the airflow obstruction or turbulence, which constitutes the basic severity of the event.

[0133] The event dynamics factor is positively correlated with four time-dimensional dynamic parameters (rate of change and stability factor). The underlying logic is that these dynamic parameters characterize the changing trends of airflow states. A high rate of change or a low stability factor indicates an evolving, unstable airflow event. This dynamic instability is key to triggering image enhancement. Spectral features indicating acute deterioration trends, including but not limited to high rates of change in the airflow turbulence index, will, through two-stage cascaded processing, significantly increase the event dynamics factor, thereby increasing the weight of the dynamic enhancement term and adaptively expanding the ROI size.

[0134] For the final alert level Impact analysis: Instantaneous event severity input to the early warning module and dynamic deterioration risk Instantaneous event severity Symmetry coefficient of heat flux with both lungs Monotonically negatively correlated. Symmetry coefficient of bilateral lung heat flux. The lower the value, the more asymmetrical the ventilation between the two lungs, and the more severe the current condition.

[0135] Dynamic deterioration risk level The rate of change and volatility of the symmetry time series are positively correlated. This reflects the clinical reality: stable but mild asymmetry (including but not limited to that caused by the patient’s inherent anatomy) and sudden and worsening asymmetry (including but not limited to intubation slipping into a unilateral main bronchus) have vastly different levels of clinical urgency.

[0136] Final Alert Level Severity of instantaneous events and dynamic deterioration risk All are positively correlated, and due to the non-linear amplification of the fusion method, the risk of dynamic deterioration is high. For the final alert level The influence exhibits an exponential relationship. When other parameters remain constant, the dynamic deterioration risk... The increase will lead to the final alert level. The growth rate is much faster than the severity of instantaneous events. The linear increase in [something] enables this invention to accurately identify the most dangerous "acute exacerbation" events in clinical practice.

[0137] To verify the beneficial effects of the technical solution of this invention, four typical clinical scenarios were designed for comparative experiments. The experiments aimed to demonstrate the significant advancements of this invention in two core innovative aspects: 1) spectrum-guided adaptive image enhancement capabilities; and 2) the ability to distinguish between the severity and urgency of events through spatiotemporal joint early warning. See the table below for details. Table 2: Comparison of Monitoring Data in Typical Clinical Scenarios

[0138] Standard System - Alarm Output: This parameter is used to compare with existing technologies that rely solely on thermal image symmetry for threshold judgment. Its calculation logic is as follows: when the "bi-lung heat flux symmetry coefficient" is lower than a fixed, relatively sensitive threshold of 0.60, a high-level alarm of 0.95 is output; otherwise, no alarm is output (0).

[0139] "Overall Enhancement Strength" is a dimensionless parameter normalized to the [0,1] interval. Its design purpose is to quantify the overall intensity of the nonlinear image correction applied by the "Spectrum-Guided Adaptive Thermal Image Dynamic Range Intensifier." The calculation of this parameter is the final quantization output of the two-stage cascaded processing in the image processing module. The specific calculation logic is as follows: The image processing module first obtains two core intermediate variables generated by its internal first-level processing: Event Severity Factor: A numerical value that characterizes the severity of static anomalies reflected by the current airflow spectrum characteristics.

[0140] Event dynamics factor: A numerical value that characterizes the time-series variation trend and stability of the current airflow spectrum characteristics.

[0141] To achieve differentiated responses to events of different natures, the system has two built-in preset weight coefficients for adjusting the enhancement strategy: Base weighting coefficient: Used to adjust the contribution of the "event severity factor" to the final enhancement strength. In a preferred embodiment, this coefficient is set to 0.3.

[0142] Dynamic weighting coefficient: Used to adjust the contribution of the "event dynamics factor" to the final enhancement strength. In a preferred embodiment, this coefficient is set to 0.7.

[0143] The dynamic weighting coefficient is assigned a higher weight, the technical logic of which is that the "event dynamics factor," which characterizes the dynamically changing and unstable airflow state, indicates an ongoing acute event. Its clinical risk and the need for early visual detail observation are higher than those of long-term, stable abnormal states. Therefore, the system is designed to be more sensitive to "dynamic" signals.

[0144] The image processing module performs the following calculation steps to determine the final "overall enhancement strength": The obtained "event severity factor" is multiplied by a preset "base weight coefficient" to obtain the "base enhancement component". The obtained "event dynamism factor" is multiplied by a preset "dynamism weight coefficient" to obtain the "dynamism enhancement component". The "base enhancement component" and "dynamism enhancement component" obtained in the first two steps are added together. The sum obtained from this addition operation is the final output "overall enhancement strength" referenced in the experimental data table.

[0145] Comparing Scenario 1 and Scenario 2: In Scenario 2 (slow sputum blockage), due to the increase in the airflow turbulence index and its rate of change, the overall enhancement intensity of this invention (mean 0.695) is much higher than in Scenario 1 (mean 0.055). This active image enhancement driven by spectral features enables the system to capture subtle changes in heat flux caused by sputum blockage, thereby calculating a slight decrease in the bilateral lung heat flux symmetry coefficient (mean 0.845) and generating a low-level warning signal, with a final alarm level of mean 0.245.

[0146] In contrast, conventional systems, lacking this adaptive enhancement capability, cannot identify such subtle changes from raw thermal images. Their calculated symmetry coefficients remain within the normal range, resulting in a consistently zero alarm output. This demonstrates that this invention, guided by cross-modal data, can elevate detection sensitivity to levels unattainable by conventional pure thermal imaging techniques, enabling early detection of potential risks.

[0147] Comparing Scenario 3 and Scenario 4: The instantaneous event severity of Scenario 4 (chronic problem) (mean 0.79) and Scenario 3 (acute problem) (mean 0.835) are in a highly comparable range. Under the premise of "control variables" with almost identical "static severity," the following comparative analysis is conducted: Because conventional systems rely solely on static symmetry thresholds, they detected severe ventilation asymmetry in both scenarios three and four (symmetry coefficients were both far below 0.60), resulting in indistinguishable, identical high-level alarms (0.95). This exposes the fundamental flaw in existing technologies' inability to determine the level of risk urgency.

[0148] However, by introducing an analysis of the time dimension, this invention makes the following judgment: In Scenario 3, the high dynamic deterioration risk (mean 0.965) serves as a dynamic gain factor, non-linearly amplifying the final alarm level to 0.99, thus issuing the highest priority emergency alarm.

[0149] In scenario four, although the static severity is also high, the low dynamic deterioration risk (mean 0.025) failed to trigger the gain amplification mechanism, so the final alarm level was reasonably suppressed to a medium level (mean 0.50).

[0150] Assuming the instantaneous event severity is roughly the same, the final alert level for scenario three is 98% higher than that for scenario four. This 98% figure was obtained by calculating the average of the final alert levels for both scenarios: 0.99 for scenario three and 0.50 after adjustment for scenario four. Next, the difference was calculated: 0.99 minus 0.50, resulting in 0.49. Then, this difference was divided by the average of scenario four (0.50), yielding a result of 0.98. Finally, this result was multiplied by 100%.

[0151] This data comparison, with strict variable control, demonstrates that the present invention, by integrating dynamic and static risk indicators, can effectively distinguish between events with the same apparent severity but vastly different clinical urgency. This capability directly addresses the "alarm fatigue" problem commonly found in existing technologies, caused by the inability to differentiate risk levels.

[0152] Based on the above experimental data analysis and combined with expert experience in ICU clinical workflows, the final alarm level of this invention is determined. The output range is divided into practical applications as follows to transform quantitative output into specific, executable clinical operation instructions. See the table example below for details: Table 3: Mapping Table of Final Alert Levels and Clinically Recommended Practices

[0153] The computational logic involved in this application can be constructed using algorithms such as regression analysis in machine learning. A mathematical model is established by analyzing the inherent trends and interrelationships of the collected parameters. This process can be implemented using specialized computational tools (such as Python's Scikit-learn library or the R language environment). To ensure the effectiveness and accuracy of the model, techniques such as cross-validation will be used to evaluate model performance, and iterative optimization will be performed based on continuous feedback to ensure that the computational content truly reflects the inherent laws of objective data. Throughout all computational processes, to eliminate the influence of different physical dimensions and ensure that data is compared and analyzed on the same scale, the input parameters in each formula are dimensionless. The dimensionless techniques used include, but are not limited to, min-max-normalization or Z-score standardization. It should be emphasized that the foregoing embodiments are merely illustrative of preferred implementations of the present invention and are not intended to limit the scope of protection of the present invention. Those skilled in the art can make various modifications, equivalent substitutions, or improvements based on the technical solutions, spirit, and principles disclosed in this invention. Any such changes that do not depart from the scope of the technical solutions of this invention should fall within the protection scope of the claims of this invention. Furthermore, the technical features of the various embodiments described in the specification can be combined arbitrarily without causing technical contradictions. This application also provides a computer-readable storage medium storing computer program instructions thereon.

Claims

1. A monitoring and early warning system for changes in the position of endotracheal tubes used in ICU nursing, characterized in that, Specifically, it includes: The first data acquisition module is used to acquire first data characterizing the stability of airflow function in endotracheal intubation, wherein the first data is the time-domain spectral characteristics of the airflow pressure or flow rate signal. The second data acquisition module is used to acquire second data characterizing the spatial distribution of airflow around the tip of the endotracheal tube. The second data is a sequence of original thermal image frames acquired by a thermal imaging sensor. The image processing module is configured as follows: Based on the changes in the first data, the target region of interest in the original thermal image frame sequence is identified; Furthermore, for the target region of interest, a set of nonlinear image correction parameters are generated and applied in real time to locally enhance the dynamic range of the image in the target region of interest, thereby generating an enhanced thermal image; The early warning module is used to generate a diagnostic early warning signal based on visual features presented in the enhanced thermal image that are not visible in the original thermal image frame sequence.

2. The ICU endotracheal tube position change monitoring and early warning system according to claim 1, characterized in that: The first data acquisition module is configured as follows: The first data representing the pressure or flow rate of the airflow is subjected to a fast Fourier transform to calculate the high-frequency harmonic energy ratio of the breathing cycle and the airflow turbulence index within a preset high-frequency band, and the calculation result is used as the time-domain spectral feature. The second data acquisition module includes a miniature uncooled infrared focal plane array sensor customized at the tip of the endotracheal tube, used to acquire the original thermal image frame sequence to dynamically image the topology map of the airflow heat flux field formed around the tip of the endotracheal tube.

3. The ICU endotracheal tube position change monitoring and early warning system according to claim 2, characterized in that: The first data acquisition module is also configured to: Based on the historical high-frequency harmonic energy ratio and historical airflow turbulence index value within a preset time window, the rate of change and stability factor of the historical high-frequency harmonic energy ratio and the rate of change and stability factor of the historical airflow turbulence index are calculated respectively, and these rates of change and stability factors are constructed into four time-dimensional dynamic parameters. The four time-dimensional dynamic parameters, along with the high-frequency harmonic energy ratio of the respiratory cycle and the airflow turbulence index, are input to the image processing module.

4. The ICU endotracheal tube position change monitoring and early warning system according to claim 3, characterized in that: The image processing module includes a spectrum-guided adaptive thermal image dynamic range enhancer characterized by a built-in pre-trained correlation mapping model. The correlation mapping model is configured to receive the high-frequency harmonic energy ratio of the respiratory cycle and the airflow turbulence index, as well as the four time-dimensional dynamic parameters, as inputs, and output the localization signal of the target region of interest and the nonlinear image correction parameters.

5. The ICU endotracheal tube position change monitoring and early warning system according to claim 4, characterized in that: The association mapping model is also configured to perform a two-stage cascaded fusion process to generate the nonlinear image correction parameters: First-level processing: Based on the high-frequency harmonic energy ratio of the respiratory cycle and the airflow turbulence index, as well as the four time-dimensional dynamic parameters, an event severity factor and an event dynamic factor are generated; Second-level processing: Based on the event severity factor and the event dynamics factor, generate the nonlinear image correction parameters.

6. The ICU nursing endotracheal tube position change monitoring and early warning system according to claim 5, characterized in that: In the second-level processing, the image processing module is further configured to: Based on the event severity factor, determine the basic enhancement terms in the image correction parameters; Furthermore, based on the event dynamism factor, the dynamic enhancement term in the image correction parameters is determined; The nonlinear image correction parameters are composed of the basic enhancement term and the dynamic enhancement term; In the second-level processing, the image processing module is further configured to: The size of the target region of interest is adaptively adjusted based on the event dynamics factor.

7. The ICU endotracheal tube position change monitoring and early warning system according to claim 6, characterized in that: The early warning module is further configured as follows: Image segmentation is performed on the enhanced thermal image to identify and separate the heat flux regions corresponding to the left and right bronchi, respectively; Based on the heat flux regions of the left and right bronchi, calculate the symmetry coefficient of the heat flux of the two lungs, which characterizes the symmetrical state at the current moment.

8. The ICU nursing endotracheal tube position change monitoring and early warning system according to claim 7, characterized in that: The early warning module is further configured as follows: Based on the symmetry coefficient of the bilateral lung heat flux, the instantaneous event severity, which characterizes the severity of the current event, is calculated; Maintain and store a symmetric time series of the symmetry coefficients of the heat flux of the two lungs at multiple consecutive moments within a preset time window; Based on the changing trends and fluctuations of the symmetrical time series, the dynamic deterioration risk degree, which characterizes the risk of event development, is calculated.

9. The ICU endotracheal tube position change monitoring and early warning system according to claim 8, characterized in that: The early warning module is further configured as follows: The instantaneous event severity is fused with the dynamic deterioration risk level to generate the final alert level; Based on the final alarm level and the image segmentation results, a diagnostic warning signal containing event location information and a suggested alarm level is generated.

10. The endotracheal tube position change monitoring and early warning system for ICU nursing according to claim 9, characterized in that: The method of fusing the instantaneous event severity and the dynamic deterioration risk is as follows: the dynamic deterioration risk is used as a dynamic gain factor to nonlinearly amplify the instantaneous event severity in order to determine the final alarm level; Continuously monitor the validity of the data streams from the first data acquisition module and the second data acquisition module; When any data stream interruption is detected, the system automatically switches to a single-modal monitoring mode driven by the remaining valid data streams and generates a system status degradation notification. The continuous monitoring method is as follows: maintain a heartbeat timer for each data stream, and determine that the data stream is interrupted when no valid data is received within a preset time window.

Images