Critical patient vital sign data detection modeling analysis system and life integrated machine

Abstract

The application relates to the technical field of medical data processing and vital sign monitoring, and discloses a critical patient vital sign data detection modeling analysis system and a life integrated machine. The critical patient vital sign data detection modeling analysis system comprises the following modules: an initial baseline risk index is calculated by a main control module, and an artifact verification instruction is generated; a maximum aligned cross-correlation coefficient of an airway pressure sequence and a driving current sequence is calculated by a breathing module based on the artifact verification instruction; a pathological disturbance is determined and a respiratory compliance sequence is extracted by a monitoring module according to the maximum aligned cross-correlation coefficient; a waveform sequence data is collected by a deep anesthesia module, and a dynamic physiological disturbance degradation coefficient is mapped in combination with the respiratory compliance sequence; a target control module deduces drug concentration data to generate a dynamic early warning trigger threshold, and compares the initial baseline risk index to trigger a pre-warning action. The technical scheme of airway pressure sequence variance ratio monitoring combined with valve driving current cross-correlation optimization is adopted, so that the technical effect of accurately stripping physical artifacts and locking real pathological disturbances is achieved.

Description

Technical Field

[0001] This invention relates to the field of medical data processing and vital sign monitoring technology, specifically a machine learning-based system for detecting, modeling, and analyzing vital sign data of critically injured patients. Background Technology

[0002] Critically injured patients are extremely unstable during emergency transport and intensive care. Vital signs data are the direct basis for assessing the progression of their injuries. Medical staff rely on real-time vital signs monitoring equipment to assist in clinical intervention decisions. Timely and accurate capture of dynamic changes in various physiological parameters is crucial for seizing the right opportunity for rescue. Vital signs data monitoring and analysis systems have become an indispensable basic medical device in the emergency medical process.

[0003] Conventional vital sign monitoring systems primarily employ a mechanism of independent acquisition of single-channel physiological signals and comparison with fixed thresholds. These devices can intuitively present absolute values ​​of basic indicators such as heart rate, airway pressure, and blood oxygen saturation. In clinical applications, medical personnel can quickly obtain the current macroscopic physiological state of the injured person through the terminal interface. The independent parameter monitoring mode significantly reduces the computational load on the device's main control module. Simultaneously, the fixed alarm upper and lower limit settings facilitate standardized emergency operating procedures across different medical institutions, and the devices are easy to operate and have extremely high adoption rates.

[0004] Emergency transport environments are often accompanied by severe jolting. Mechanical vibrations are directly transmitted along the breathing tube to the front-end sensors. Conventional systems cannot distinguish between such physical waveform disturbances and actual pathological abnormalities, easily triggering invalid alarms frequently during transport. The human body has a strong stress compensation mechanism in the early stages of trauma. Independent monitoring of single parameters is insufficient to detect subtle abnormal linkages between organ systems. When a slight mismatch between cerebral blood oxygen supply and demand occurs in the early stages of decreased respiratory compliance, traditional algorithms cannot capture this hidden signal phase characteristic, easily missing the early intervention window. Anesthesia or analgesics are often administered during emergency treatment. Pharmacokinetic processes themselves cause normal fluctuations in vital signs. Existing static alarm thresholds do not incorporate the nonlinear effects of drug intervention and completely ignore the individual patient's genetic risk baseline. Using a uniform, rigid boundary to assess critically ill individuals in complex intervention states prevents the early warning mechanism from playing a personalized preventative role. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a machine learning-based system for detecting, modeling, and analyzing vital signs data of critically injured patients. This system solves the problems in existing technologies, such as the susceptibility of vital sign monitoring to false alarms caused by physical disturbances during emergency transport, the difficulty in capturing the hidden deterioration trend of multiple systems during the compensatory period of patients when monitoring single parameters independently, and the inability of static and fixed alarm thresholds to take into account individual risk baseline differences and the impact of clinical drug interventions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a machine learning-based system for detecting, modeling, and analyzing vital signs data of critically injured patients, comprising: The main control module calculates the initial baseline risk index and generates artifact verification instructions; The breathing module receives the artifact verification command, collects the high-frequency sequence of airway pressure and the drive current sequence, calculates and outputs the maximum alignment cross-correlation coefficient to the main control module; When the main control module determines a pathological disturbance based on the maximum alignment cross-correlation coefficient, the monitoring module extracts the respiratory compliance sequence and sends it to the main control module. The anesthesia module collects waveform sequence data based on the pathological disturbance and sends it to the main control module, which then maps the respiratory compliance sequence to the waveform sequence data into a dynamic physiological disturbance degradation coefficient. The target control module sends the estimated drug concentration data of the effect room to the main control module. The main control module uses the estimated drug concentration data of the effect room to generate a dynamic early warning trigger threshold, and fuses the initial baseline risk index with the dynamic physiological disturbance degradation coefficient to trigger an early warning action by comparing the dynamic early warning trigger threshold.

[0007] Preferably, the main control module calculates the initial baseline risk index and generates artifact verification instructions, including: Obtain data on the methylation level of specific CpG sites in wounded personnel samples; The specific CpG site methylation level data is input into a support vector machine baseline classifier to calculate the initial baseline risk index. A time window is set to monitor the variance ratio of the high-frequency airway pressure sequence. When the variance ratio exceeds the fluctuation mutation threshold, the artifact verification command is generated and sent to the respiratory module.

[0008] Preferably, the step of inputting the specific CpG site methylation level data into a support vector machine baseline classifier to calculate the initial baseline risk index includes: Read the methylation ratio values ​​of multiple specific CpG sites in whole blood samples and perform standardized preprocessing; The methylation ratio values ​​are arranged in a preset feature dimension order to construct a feature vector; The support vector machine baseline classifier is invoked to perform hyperplane mapping on the feature vectors and output the initial baseline risk index.

[0009] Preferably, the step of receiving the artifact verification command, acquiring the high-frequency airway pressure sequence and the driving current sequence, calculating and outputting the maximum alignment cross-correlation coefficient to the main control module includes: The high-frequency sequence of airway pressure and the drive current sequence including delay redundancy are extracted in a local annular buffer. By introducing a time delay search step size variable, the discrete cross-correlation function between the high-frequency sequence of airway pressure and the shifted driving current sequence is calculated. Within the set acoustic delay search interval, the value of the time delay search step size variable is continuously changed to perform the search, and the extreme value is extracted as the maximum alignment cross-correlation coefficient and sent to the main control module.

[0010] Preferably, the step of extracting the respiratory compliance sequence and sending it to the main control module when the main control module determines that it is a pathological disturbance based on the maximum alignment cross-correlation coefficient includes: The received maximum alignment cross-correlation coefficient is compared with a preset correlation determination threshold; When the maximum alignment cross-correlation coefficient is greater than the correlation determination threshold, it is determined to be a physical artifact and the state machine is locked. When the maximum alignment cross-correlation coefficient is less than or equal to the correlation determination threshold, it is determined to be a pathological disturbance and the state machine lock is released. The respiratory compliance sequence is then extracted and sent to the main control module.

[0011] Preferably, the step of mapping the respiratory compliance sequence and the waveform sequence data to a dynamic physiological perturbation degradation coefficient by the main control module includes: The instantaneous phase of the electroencephalogram (EEG) signal and the instantaneous phase of the near-infrared cerebral blood oxygenation pulsation waveform are extracted from the waveform sequence data, respectively. The phase unwinding algorithm is invoked to restore the continuously monotonically increasing absolute physical phase sequence and the phase offset is calculated within the set allowable error time window; The first time derivative of the respiratory compliance sequence is extracted, and the first time derivative and the phase shift are input into the ridge regression model to map and output the dynamic physiological perturbation degradation coefficient.

[0012] Preferably, the step of generating a dynamic early warning trigger threshold by the main control module using the effect room estimated drug concentration data includes: Extract the preset initial safety threshold; A nonlinear monotonically increasing function is used in conjunction with the estimated drug concentration data in the effect room to perform a subtraction compensation operation on the initial safety critical threshold. The numerical result of the subtraction compensation operation is used as the dynamic early warning trigger threshold.

[0013] Preferably, the step of performing a subtraction compensation operation on the initial safety threshold using a nonlinear monotonically increasing function combined with the effect-room predicted drug concentration data includes: Obtain the set maximum permissible compensatory downregulation range and the drug concentration constant that produces the half maximum pharmacological effect; The Hill coefficient power of the predicted drug concentration data in the effect room is calculated as the numerator, and the sum of the Hill coefficient power of the drug concentration constant that produces the half maximum pharmacological effect and the numerator is calculated as the denominator. The numerator is divided by the denominator to obtain the proportionality coefficient. Multiply the maximum allowable compensation reduction by the proportional coefficient to obtain the reduction product, and subtract the reduction product from the initial safety threshold.

[0014] Preferably, the step of fusing the initial baseline risk index with the dynamic physiological disturbance degradation coefficient and triggering the warning action based on the dynamic warning trigger threshold includes: Obtain the preset positive constant bias term; The initial baseline risk index is added to the positive constant bias term to generate the bias risk base; The exponential degradation value is calculated by using the natural constant as the base and the dynamic physiological disturbance degradation coefficient as the exponent. Multiply the bias risk base value by the exponential degradation value to output a comprehensive early warning risk value; When the comprehensive early warning risk value is continuously greater than or equal to the dynamic early warning trigger threshold for a preset number of judgment cycles, the terminal is driven to output the early warning action.

[0015] Preferably, the step of adding the initial baseline risk index to the positive constant bias term to generate the bias risk base includes: The initial baseline risk index is retrieved from the fixed address space of static random access memory; Extract the positive constant bias term set to prevent underflow of risk values ​​in low baseline populations; The addition operation is performed to combine the initial baseline risk index with the positive constant bias term, outputting the bias risk base, and then proceeding to the multiplication operation step.

[0016] The second aspect of this application discloses a life-saving integrated machine equipped with the aforementioned system for detecting, modeling, and analyzing vital signs data of critically injured patients.

[0017] Furthermore, the integrated life support system also includes conventional medical modules such as a respiratory module, a monitoring module, anesthesia module, and an infusion module; The breathing module is used to support the wounded soldier's breathing. The monitoring module is used to monitor the vital signs of the injured. The anesthesia module is used to anesthetize the wounded. The infusion module is used to administer intravenous fluids to the injured.

[0018] This invention provides a machine learning-based system for detecting, modeling, and analyzing vital signs data of critically injured patients. It offers the following advantages: 1. This invention employs a technical solution combining airway pressure sequence variance ratio monitoring with valve drive current cross-correlation optimization, achieving the technical effect of accurately removing physical artifacts and locking onto real pathological disturbances. Compared to existing technologies that rely on a single channel fixed threshold to filter waveform interference, this invention overcomes the shortcomings of false alarms easily triggered by mechanical vibration or transportation bumps during patient evacuation.

[0019] 2. This invention employs a technique that maps the first derivative of respiratory compliance over time to EEG and cerebral oxygenation phase shifts using a ridge regression model, resulting in a degradation coefficient. This achieves the technical effect of real-time quantification of the occult deterioration trend in multi-organ involvement. Compared to existing technologies that only independently monitor the absolute values ​​of various physiological parameters or simply superimpose them linearly, this invention overcomes the limitation of failing to capture subtle abnormal features when the injured person is in the multi-system compensation phase.

[0020] 3. This invention employs a nonlinear compensatory early warning threshold for effect-room drug concentration and utilizes a multiplicative degradation model to fuse baseline and real-time risk, achieving a dynamic early warning effect that matches the individual patient's tolerance and anesthesia / analgesia status. Compared to existing technologies that use uniform static physiological alarm limits for all patients, this invention overcomes the shortcomings of ineffective alarms caused by normal physiological fluctuations during drug intervention and the inability to consider individual underlying risk differences. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the system architecture according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 3 This is a schematic diagram of waveform cross-correlation analysis for physical artifact verification performed by the respiratory module in an embodiment of the present invention; Figure 4 This is a time-series evolution curve of the main control module performing cross-determination of comprehensive early warning risk value and dynamic early warning trigger threshold in an embodiment of the present invention.

[0022] Among them, 10 is the main control module; 20 is the respiratory module; 30 is the monitoring module; 40 is the anesthesia depth module; and 50 is the target control module. Detailed Implementation

[0023] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Reference Figure 1 The present invention provides a machine learning-based system for detecting, modeling and analyzing vital signs data of critically injured patients. The system includes a main control module 10, a respiratory module 20, a monitoring module 30, an anesthesia depth module 40 and a target control module 50.

[0025] The main control module 10 serves as the system's data processing center, internally equipped with a memory and processor, and externally connected to a multi-touch display screen. The main control module 10 establishes communication connections with the respiratory module 20, monitoring module 30, anesthesia depth module 40, and target control module 50 via the controller area network bus and Ethernet network communication protocol. It is used to receive heterogeneous data collected by each hardware module, perform calculations for the machine learning early warning model, and issue human-machine interaction and alarm commands.

[0026] The breathing module 20 is internally configured with a microcontroller, pneumatic components, and an analog-to-digital converter (ADC). The microcontroller integrates a hardware timer for synchronously triggering data acquisition actions from different ADC channels. The breathing module 20 performs low-level calculations of respiratory mechanics parameters and acquires the drive current sequence of the pneumatic components to perform local edge computing tasks for physical artifact verification.

[0027] The monitoring module 30 is equipped with a peripheral physiological signal sensing unit, which is responsible for the real-time acquisition of routine peripheral vital signs signals, filtering and preprocessing, and transmission of characteristic parameters such as respiratory compliance and airway resistance.

[0028] The anesthesia module 40 is equipped with a dual-channel acquisition front-end for EEG and near-infrared brain blood oxygenation, which performs high-frequency synchronous acquisition of central nervous system state signals and sends waveform sequence data to the main control module 10.

[0029] The target control module 50 is equipped with an electric infusion pump and a pharmacokinetic calculation unit. It receives control commands to execute the infusion of anesthetic drugs, performs real-time simulations, and outputs the estimated effect-site drug concentration data to the main control module 10.

[0030] See attached document Figure 2 This invention provides a machine learning-based method for detecting, modeling, and analyzing vital signs data of critically injured patients, comprising the following steps: S100: Obtain methylation level data of specific CpG sites in the wounded samples, calculate and store the static initial baseline risk index through the support vector machine baseline classifier; S200, with each hardware module working in parallel, synchronously acquires peripheral physiological signal sequences, central nervous system state signal sequences, effect room estimated drug concentration data, and high-frequency sequences of driving current and airway pressure of the underlying pneumatic components. S300, the main control module 10 sets a multi-scale time window to monitor the variance ratio of the airway pressure sequence. When the variance ratio exceeds the threshold, it sends an artifact verification command to the respiratory module 20. The respiratory module 20 performs a delay optimization cross-correlation calculation locally. The main control module 10 decouples the physical artifacts and locks the state machine based on the judgment result of the maximum aligned cross-correlation coefficient. S400, when the disturbance is determined to be a real pathological disturbance, the main control module 10 extracts the time first derivative of the respiratory compliance sequence and the phase offset of the central nervous system waveform sequence, and maps the feature parameters to the dynamic physiological disturbance degradation coefficient through the ridge regression model. S500, the main control module 10 uses the effect room to estimate the drug concentration to perform nonlinear compensation on the set safety threshold to generate a dynamic early warning trigger threshold. It integrates the initial baseline risk index with the dynamic physiological disturbance degradation coefficient to generate a comprehensive early warning risk value. It then compares the comprehensive early warning risk value with the dynamic early warning trigger threshold according to a fixed calculation cycle to trigger an early warning action.

[0031] In this embodiment, the main control module 10 executes step S100, which involves acquiring casualty sample data and anchoring static baseline risk. This step specifically includes the following sub-steps: S101, the main control module 10 reads the sequencing value vector of methylation levels of specific CpG sites in the whole blood sample of the wounded soldier through the data interface. In a preferred embodiment, the hardware communication interface of the main control module 10 receives a formatted data packet from an external sequencing device. This data packet contains the methylation ratios of multiple CpG sites in the promoter region of a specific gene of the wounded soldier. To ensure dimensional consistency and numerical stability in subsequent model calculations, the processor inside the main control module 10 parses the data packet, extracts a predetermined number of methylation ratio values ​​for specific CpG sites, and performs Z-score normalization preprocessing on these values ​​to eliminate the dimensional differences in absolute values ​​caused by individual basal metabolism. After the above normalization preprocessing, the main control module 10 arranges the methylation ratio values ​​according to a preset feature dimension order and constructs a methylation feature vector with a fixed dimension in memory. It should be noted that the specific biochemical extraction and analysis process for whole blood sample collection and specific CpG site methylation sequencing can be achieved by those skilled in the art using existing gene sequencing technology and methylation microarray chip technology. The biochemical sequencing principle is a well-known technology in this field and will not be elaborated here.

[0032] S102, the main control module 10 calls the support vector machine baseline classifier in the internal memory to process the methylation feature vector. The hyperplane mapping calculation is performed. During this calculation, to ensure the Support Vector Machine (SVM) baseline classifier has accurate risk prediction capabilities, the system of this invention has undergone a complete offline model training process before leaving the factory. Specifically, during the offline training phase, a large amount of clinical sample data from past trauma patients is collected as the training set. Samples diagnosed with traumatic lung injury are labeled as 1, and samples without this complication are labeled as 0. The training system extracts the methylation features of these historical samples as input, and uses a hinge loss function combined with a linear kernel function to iteratively optimize the SVM model to find a high-dimensional hyperplane that maximizes the differentiation between the two types of samples. To convert the geometric distance calculated by this hyperplane into the probability values ​​of disease states required for actual business operations, after training, the Platt scaling method is further used, and the logistic regression function is used to calibrate the model output probability.

[0033] After completing the offline training described above, the parameters of the converged support vector machine baseline classifier model are permanently stored in the non-volatile memory of the main control module 10. In this embodiment, the main control module 10 will extract the generated methylation feature vector. The data is substituted into the hyperplane mapping function for calculation. From a general principle perspective, this calculation process essentially assesses the physical distance of the current wounded soldier's genetic characteristics relative to the historical boundary between healthy and diseased groups in the multidimensional data space, and then maps it into a continuous disease tendency index. The specific calculation formula executed by the main control module 10 is as follows: ; in, To calculate the initial baseline risk index for the output; It is the base of the natural logarithm; The hyperplane normal vector, i.e. the weight coefficient vector, is obtained by training the baseline classifier of the support vector machine through an offline dataset. Its dimension is strictly consistent with the dimension of the input feature vector. The input is the methylation feature vector; This is the intercept bias term of the support vector machine baseline classifier, which is determined by the support vector boundaries during training.

[0034] S103, the main control module 10 outputs an initial baseline risk index in the range of 0 to 1. and the initial baseline risk index The initial baseline risk index is stored in memory as a static fundamental constant for subsequent system calculations. In the specific application scenario of this embodiment, the initial baseline risk index... Numerically, this quantifies the inherent susceptibility of injured individuals to traumatic lung injury at the molecular biological level. The closer the value is to 1, the higher the physiological risk of compensatory respiratory failure at the underlying genetic level. As a preferred storage and retrieval mechanism, the main control module 10 will calculate the initial baseline risk index. The data is written into a fixed address space within the internal high-speed static random access memory. During a single monitoring cycle of the casualty access system, the system, based on the biological prerequisite that the casualty's gene expression is stable in the short term, sets the initial baseline risk index... The initial baseline risk index is locked as a static constant. During subsequent execution of dynamic physiological feature fusion calculation instructions, the main control module 10 directly retrieves the initial baseline risk index from the fixed address space of the high-speed static random access memory. It participates in the calculation, no longer repeatedly triggering the external data reading and support vector machine classifier mapping calculation tasks, thereby significantly saving the computing power overhead of the underlying controller and improving the overall real-time response performance of the system.

[0035] In this embodiment, after the system enters the steady-state operation phase, the data aggregation step S200 of multi-module distributed high-frequency acquisition and communication is executed. This step specifically includes the following sub-steps: S201, the monitoring module 30 and the anesthesia depth module 40 synchronously acquire peripheral circulation parameters and central nervous system waveform sequences according to their respective set hardware sampling rates. At the specific physical sensing implementation level of the system, the monitoring module 30 acquires the patient's electrocardiogram, blood pressure, and blood oxygen saturation characteristics through non-invasive or invasive sensors, and extracts peripheral circulation characteristic parameters such as heart rate variability and perfusion index in real time within its internal microprocessor. Simultaneously, the respiratory module 20 continuously acquires airway physical quantities using built-in flow and pressure sensors, and calculates respiratory compliance and airway resistance parameters in real time within its local microcontroller. Based on the application requirement of synchronously monitoring the central nervous system status, the anesthesia depth module 40 uses a dual-channel sensor attached to the patient's head to acquire high-frequency electroencephalogram (EEG) signals and near-infrared cerebral oxygenation signals. As a preferred data alignment mechanism, to ensure the consistency of the time dimension of multi-source heterogeneous data in subsequent joint analysis, both the monitoring module 30 and the anesthesia depth module 40 call their internal real-time clocks to add a globally unified timestamp to the data packets after acquiring the underlying electrical signals and completing analog-to-digital conversion. After completing the aforementioned time alignment operation, the monitoring module 30 and the anesthesia depth module 40, according to a preset communication transmission cycle, package and send the timestamped physiological characteristic data to the main control module 10 via the controller local area network bus. It should be noted that the design of filtering and amplification circuits for conventional physiological signals and the algorithms for extracting basic characteristic parameters can be implemented using existing standard signal processing schemes for medical monitoring instruments by those skilled in the art. The signal conditioning and conventional calculations are well-known technologies in this field and will not be elaborated upon here.

[0036] S202, the target control module 50 runs the internally embedded pharmacokinetic model to extrapolate the current effect-cell predicted drug concentration. The concentration value is then transmitted to the main control module 10 via a communication bus. In this embodiment, the target control module 50 is internally equipped with a control unit and a precision infusion actuator. The control unit continuously calculates the drug infusion rate required to maintain the current depth of anesthesia based on the target concentration command set by the clinician and the patient's age, weight, and vital signs, and drives the actuator accordingly. During the actual drug infusion cycle, the target control module 50 utilizes a built-in three-compartment pharmacokinetic model to continuously calculate the transport rate constant of the anesthetic drug between the central compartment and the effect compartment, thereby deriving the estimated drug concentration in the effect compartment that diffuses from plasma to the target site and produces a hysteresis effect. As a preferred embodiment, the effect chamber predicts the drug concentration. The physical unit is set as μg / mL or ng / mL, and its specific value characterizes the actual pharmacological state of the patient's central nervous system under drug inhibition, providing a reliable calculation basis for the dynamic compensation of the system's subsequent warning threshold. For the specific mathematical derivation of the pharmacokinetic model and the electromechanical closed-loop control of the infusion pump, those skilled in the art can use existing Marsh or Schnider models. Their derivation and control principles are well-known technologies in the field and will not be elaborated upon here.

[0037] S203, the respiratory module 20, based on a synchronous constraint mechanism using a shared clock, synchronously triggers the first and second analog-to-digital conversion (ADC) channels within the internal microcontroller to perform signal conversion operations using a unique hardware timer. In a preferred embodiment of the invention, to completely eliminate high-frequency signal phase integration errors caused by clock domain drift at the underlying hardware level, the respiratory module 20 avoids using an external independent ADC chip driven by a separate crystal oscillator. Instead, it directly configures the dual ADC peripherals integrated within the main microcontroller. The microcontroller configures an internal general-purpose timer to generate a fixed-frequency update event interrupt. This update event is directly routed as a hardware trigger signal to the trigger terminals of the first and second ADC channels. Specifically, in terms of channel signal allocation, the first ADC channel is responsible for acquiring the analog airway pressure signal at the patient end of the respiratory circuit, and the second ADC channel is responsible for synchronously acquiring the analog drive current signal of the piezoelectric proportional valve controlling the fresh gas flow. This underlying hardware routing and register-level configuration method, which uses synchronous triggering based on the same source clock, strictly controls the sampling time delay between the two independent physical channels to the nanosecond level, thus providing an indispensable reference time anchor for the system to subsequently eliminate physical artifacts caused by pipeline acoustic delay.

[0038] S204, the respiratory module 20 acquires the high-frequency airway pressure sequence and the piezoelectric proportional valve drive current sequence, and overwrites and stores the two sets of data into a circular buffer defined in the local static random access memory. As a core hardware caching mechanism of this invention, to prevent high-frequency sampling data from directly occupying the main processor's computing resources and causing congestion on the system's main communication bus, the respiratory module 20 is internally configured with a direct memory access controller. The direct memory access controller automatically stores the airway pressure sequence continuously generated by the first analog-to-digital conversion channel in the background bus matrix. and the drive current sequence continuously generated by the second analog-to-digital conversion channel The data is transferred to static random access memory at a preset sampling rate of 1 kHz. In this embodiment, the airway pressure sequence... The physical unit is cmH2O, and the driving current sequence is... The physical unit of measurement is mA. The system pre-allocates a contiguous memory address space in static random access memory to form a circular buffer. To ensure that the data in this buffer can fully cover the low-frequency respiratory cycle characteristics of critically ill patients, the capacity of the circular buffer is allocated according to the above-set sampling rate and the preset maximum time window span (e.g., covering at least one complete 10-second respiratory action span), that is, at least 10,000 discrete sampling points are allocated and stored for a single physical channel. When the latest sampling point data is written to the end of the physical address of the circular buffer, the target address pointer of the direct memory access controller will automatically wrap back to the beginning address of the buffer to perform the data overwrite operation. Through this edge-side high-frequency caching mechanism, the respiratory module 20 can always maintain a sliding high-frequency data window of a fixed time length in local memory. This mechanism allows the system to avoid uploading a massive amount of low-level electromechanical signals to the main control module 10, thereby realizing the local residence of data for edge verification calculation tasks at the physical network topology level.

[0039] In this embodiment, after the system completes the synchronous aggregation of multi-source data, it executes the artifact cross-decoupling step S300 based on edge computing and physical linkage. This step specifically includes the following sub-steps: S301, the main control module 10 opens short-time windows and long-time windows in parallel on the continuous airway pressure sequence, and calculates the ratio of the variance of the short-time window to the variance of the long-time window. As a preferred implementation, to support real-time low-frequency monitoring at the main control end, the respiratory module 20 synchronously performs equally spaced downsampling operations on the high-frequency airway pressure sequence in the background, and sends the downsampled normal airway pressure sequence to the receiving buffer of the main control module 10 in real time via the controller area network bus. After acquiring the sequence, to accurately capture transient abnormal fluctuations in airway pressure, the processor inside the main control module 10 sets a short-time window spanning 0.5 seconds to 1 second and a long-time window spanning 5 seconds to 10 seconds on the time axis. In the specific calculation process, within one calculation cycle, the main control module 10 synchronously calculates the first variance of the airway pressure data within the short-time window and the second variance of the same airway pressure data within the long-time window. Based on this, the main control module 10 calculates the dimensionless variance ratio parameter by dividing the first variance by the second variance. The calculation mechanism of calculating the variance ratio using both long and short time windows can filter out the interference of slow drift in the patient's basic respiratory rhythm while highlighting the characteristics of sudden abnormal high-frequency airflow disturbances.

[0040] S302, the main control module 10 determines whether the variance ratio is greater than the set fluctuation mutation threshold, and if the condition is met, it pauses the feature update action and sends an artifact verification command to the respiratory module 20. In the application scenario of this embodiment, the main control module 10 pre-stores a fluctuation mutation threshold based on clinical statistics in its memory. The value range of this threshold is typically set to 3.0 to 5.0. When the main control module 10 finds that the currently calculated variance ratio is greater than the fluctuation mutation threshold, the system determines that a significant pressure mutation has occurred in the current respiratory circuit. To prevent such unconfirmed mutation data from being incorrectly input into the subsequent risk prediction model as a pathological feature and causing false alarms, the main control module 10 immediately enters a state machine locking mode, pausing the feature update action of the current calculation cycle. As a coordinated action to deal with this mutation state, the main control module 10 extracts the global timestamp of the moment the mutation occurs and sends an artifact verification command carrying the timestamp to the underlying respiratory module 20 via the controller area network bus.

[0041] S303, the breathing module 20 extracts data from its local circular buffer, introduces a time delay search step size variable, and calculates the discrete cross-correlation function between the high-frequency airway pressure sequence and the shifted piezoelectric proportional valve drive current sequence. Upon receiving the artifact verification command, the breathing module 20 addresses the circular buffer in its local static random access memory according to the timestamp in the command. From the perspective of the system's underlying physical linkage principle, when the breathing tubing is subjected to external pressure or mechanical disturbances such as bending, the airway pressure will generate transient spikes. At this time, the ventilator's underlying closed-loop PID controller will rapidly adjust the opening of the piezoelectric proportional valve to maintain the set pressure, causing the drive current to generate compensatory fluctuations with similar waveforms. Due to the inherent delays in the acoustic transmission of gas in the tubing and the action of the mechanical valve, these two waveforms exhibit slight translational misalignment on the time axis.

[0042] To effectively avoid the array out-of-bounds algorithm dead zone caused by data shift addressing, the total length of data extracted by the breathing module 20 within the buffer is set to... ,in To calculate the core window length for cross-correlation, This is the maximum number of delay sampling points set. The microcontroller's truncation length is... The high-frequency sequence of airway pressure and the high-frequency sequence of piezoelectric proportional valve drive current containing delay redundancy are used. A time delay search step size variable is introduced, and a normalized discrete cross-correlation function is calculated on the translated sequence. The specific calculation formulas performed by the respiratory module 20 during this process are as follows: ; in, For the set time delay parameter The normalized cross-correlation coefficient is such that the numerator and denominator of this parameter have the same physical dimensions (i.e., the product of pressure and current), making the final calculation result a dimensionless pure number. The total number of discrete sampling points in the core matching calculation window; This is the current sampling point sequence index; For the first Airway pressure values ​​at each sampling point; The arithmetic mean of the extracted high-frequency airway pressure sequence; The time delay search step size variable represents the number of sampling points shifted in the sequence; To translate backwards The piezoelectric proportional valve drive current value after each sampling point. This is the arithmetic mean of the extracted high-frequency drive current sequence.

[0043] S304, the breathing module 20 extracts the maximum alignment cross-correlation number within the set acoustic delay search interval. The scalar value is then transmitted back to the main control module 10 via the bus. To avoid wasting computing power and creating logical dead zones due to blind global search, this embodiment pre-determines a reasonable time delay search boundary range based on the typical physical length of the system's pneumatic pipeline (e.g., 1.5 to 2.0 meters) and the velocity of sound at room temperature (approximately 340 m / s), superimposed with the mechanical response time of the electromechanical proportional valve. As a preferred data optimization method, the time span corresponding to this acoustic delay search range is set to 0 ms to 50 ms. The breathing module 20 continuously changes the time delay search step size variable only within this set acoustic delay search range. The value of the interval is calculated, and the discrete cross-correlation function mentioned above is called multiple times to generate a series of cross-correlation coefficients. After completing the traversal search of this interval, the microcontroller of the breathing module 20 extracts the extreme value with the largest absolute value from this series of cross-correlation coefficients through a comparison algorithm, and defines it as the maximum aligned cross-correlation coefficient. Finally, the breathing module 20 will calculate the maximum alignment cross-correlation coefficient. As an independent scalar data packet, it is transmitted back to the receive buffer of the main control module 10 via the bus.

[0044] S305, Main control module 10 compares the maximum alignment cross-relation number A correlation threshold is used to perform state machine locking and decoupling determination. In this embodiment, the main control module 10 internally presets a correlation threshold, which is used to define the mathematical boundary between physical deformation artifacts and real pathological perturbations, and its preferred value range is 0.75 to 0.85. The main control module 10 parses the received maximum alignment cross-correlation coefficient. And compare the numerical value with the correlation determination threshold. When When the value exceeds the correlation threshold, it indicates a high degree of physical time delay following isomorphism between the sudden change waveform of airway pressure and the closed-loop compensation current waveform of the piezoelectric proportional valve. Based on this characteristic, the main control module 10 confirms that the current pressure fluctuation is a physical artifact caused by external mechanical deformation such as pipeline pressure or kinking. Under this determination, the main control module 10 maintains the state machine locked, keeping the comprehensive early warning risk value of the previous cycle as the output of the current cycle, thereby achieving complete decoupling of physical artifacts and preventing false alarms. To prevent continuous physical interference from causing long-term paralysis of the system's physiological monitoring function, the main control module 10 is equipped with a watchdog timer. When the state machine is continuously locked for more than the set safe reset time (e.g., continuously locked for 15 seconds), the main control module 10 will forcibly unlock and prioritize broadcasting a high-priority device-level alarm command for pipeline blockage or physical fault to the bus. If the comparison result shows... If the value is less than or equal to the correlation determination threshold, it indicates that the pressure fluctuation cannot be explained by the physical closed-loop behavior of the electromechanical system. The main control module 10 then confirms that the fluctuation is a real pathological disturbance of the injured person, and then releases the state machine lock, allowing the system to enter the subsequent dynamic physiological parameter extraction and early warning feature update calculation steps.

[0045] In a specific embodiment of the present invention, when the system confirms through physical artifact cross-decoupling that the current data fluctuation belongs to a real pathological disturbance, the main control module 10 then unlocks the state machine and executes the dynamic physiological disturbance degradation coefficient extraction step S400, which specifically includes the following sub-steps: S401, the main control module 10 calculates the time first derivative of the long-term respiratory compliance sequence for data intervals confirmed as pathological disturbances. In exploring the specific physical mechanisms underlying the compensatory capacity of the respiratory system in wounded soldiers, an increase in airway resistance or a depletion of alveolar surfactant directly reflects changes in lung elasticity. Therefore, the main control module 10 retrieves the respiratory compliance time series continuously calculated and uploaded by the respiratory module 20. To quantify the dynamic deterioration trend of the lung's mechanical properties, the microprocessor inside the main control module 10 performs a differential operation on the discrete respiratory compliance sequence within a set long time window (e.g., a span of 5 to 10 seconds). This involves subtracting the mean compliance at the start of the time window from the mean compliance at the current moment, and then dividing by the time span of that window to calculate the first-order time derivative of the respiratory compliance sequence. In a preferred embodiment of the present invention, the time first derivative... The standard physical dimension of this value is mL / (cmH2O·s). An increase in the absolute value within the negative range directly indicates an abnormal increase in alveolar elastic recoil force and a rate of worsening airway obstruction. For the basic measurements of conventional respiratory compliance parameters and the discrete sequence time difference algorithm, those skilled in the art can utilize existing clinical respiratory mechanics monitoring techniques. The basic algorithms are well-known in the field and will not be elaborated upon here.

[0046] S402, the main control module 10 extracts the EEG signal envelope and near-infrared cerebral blood oxygenation pulsation waveform uploaded by the anesthesia depth module 40, and calculates the phase offset between the two within the allowable error window. Based on the physiological mechanism that peripheral ventilation deterioration is often accompanied by central nervous system microcirculatory blood perfusion impairment, the electrophysiological metabolic activity of brain tissue is highly dependent on real-time blood oxygen delivery. When oxygen supply is delayed, a temporal decoupling mismatch occurs between the pulsating waveform representing physical blood oxygen supply and the EEG waveform representing neuronal aerobic discharge activity. Based on the above-mentioned physical-biochemical linkage principle, the main control module 10 synchronously processes the neurophysiological waveform sequence uploaded in real time by the anesthesia module 40. The main control module 10 uses Hilbert transform to extract the instantaneous phase of the low-frequency envelope waveform of the EEG signal and the instantaneous phase of the near-infrared brain blood oxygen pulsating waveform. To effectively avoid the phase jump dead zone caused by the limitation of the inverse trigonometric function range, after calculating the original instantaneous phase, the main control module 10 forcibly calls the internal phase dewinding algorithm, and by detecting jumps exceeding π between adjacent sampling points and compensating for integer multiples of 2π, restores the absolute physical phase sequence that continuously and monotonically increases with time. After completing the unwinding operation, the main control module 10 calculates the phase offset between the two sets of waveform sequences within a set allowable error time window. In this embodiment, the phase offset... The physical unit is radians (rad). Phase offset. The increase in the numerical value reflects, at the underlying mechanism, the pathological time lag between the firing activity of central neurons and the oxygen supply response of microvessels, providing a cross-confirmation dimension for the deterioration of the body's overall compensatory capacity.

[0047] S403, the main control module 10 calculates the first derivative of time. With phase offset In the ridge regression model preset by the input system, orthogonal dimension feature fusion is performed by combining internal weight parameters. Considering that the two feature variables mentioned above belong to completely different physical dimensions, direct linear operation lacks physical meaning. Therefore, as a preferred data alignment mechanism, before performing model fusion, the main control module 10 first calls the historical mean and standard deviation of the healthy group recorded in the memory to align the input data. and Perform standardization preprocessing to transform it into dimensionless standardized feature variables. and .

[0048] After unifying the dimensions of the multi-source data, the main control module 10 combines these two standardized features into a vector and inputs it into the ridge regression model stored in the internal non-volatile memory. To ensure that the model can accurately output the deterioration coefficient and support the technical solution of the relevant claims, this invention collected multimodal waveform segments from the critical care clinical database as training samples before system deployment. The labels of these samples are clearly defined as clinical scores of physiological deterioration in the range of 0 to 10, labeled by professional physicians. Given the strong collinearity between respiratory mechanics features and neuroischemic features in the process of organ failure, a mean squared error loss function with a regularization term (i.e., ...) was used in the offline training phase. ,in The ridge regression model is iteratively converged using a regularization hyperparameter.

[0049] After offline training, the main control module 10 stores the optimal weight parameters of the ridge regression model, and these weights guide the online inference calculation. In this embodiment, the specific linear mapping calculation formula executed by the main control module 10 is as follows: ; in, The output is the dynamic physiological perturbation degradation coefficient, which is a dimensionless pure number. This is the standardized dimensionless first time derivative; The dimensionless phase offset after standardization; and The first and second partial regression coefficients in the ridge regression model are trained and solidified using an offline dataset. Based on physiological mechanisms, it is known that respiratory compliance decreases as it deteriorates (i.e., the first derivative becomes negative). The first partial regression coefficients are derived from the model's autonomous convergence using historical samples. This is typically represented by a negative value, thus correctly converting the negative input characteristics into a positive deterioration gain in terms of physical logic. This is the intercept constant term of the ridge regression model.

[0050] S404, the main control module 10 calculates the continuously changing dynamic physiological disturbance degradation coefficient. This is used to characterize the rate of deterioration of current vital signs. The main control module 10 outputs this continuously changing degradation coefficient by running the aforementioned ridge regression algorithm. This degradation coefficient numerically quantifies the acceleration of pathological deterioration in the injured person across both respiratory mechanics and central nervous system dimensions. To prevent numerical overflow and damage to the warning level caused by subsequent nonlinear exponential scaling calculations, while preserving the model's dynamic resolution across the entire clinical score range, the main control module 10 calls a range normalization function before outputting the final degradation coefficient. This linearly compresses the original clinical score range of 0 to 10 to a sensitive index range of 0 to 2. After linear scaling, boundary protection logic is applied: when abnormal fluctuations cause the scaling value to be less than 0, a value of 0 is assigned; when it is greater than 2, a value of 2 is assigned. This smoothly and completely preserves the deterioration cascade gradient of the original data. In specific application scenarios, as a key intermediate feature variable for the system's subsequent dynamic risk prediction, the main control module 10 calculates the dynamic physiological disturbance degradation coefficient. The data is continuously written to the internal high-speed random access memory. This data will be used to perform nonlinear compensation on the system's static baseline risk index in a time-series manner, and will serve as the core judgment basis for the dynamic triggering early warning mechanism.

[0051] In a specific embodiment of the present invention, after completing the multidimensional feature fusion and degradation coefficient extraction, the system enters the pharmacokinetic mapping and adaptive early warning tolerance triggering step S500, which specifically includes the following sub-steps: S501, the main control module 10 uses a multiplicative degradation model to convert the initial baseline risk index... Degradation coefficient with dynamic physiological disturbance The comprehensive early warning risk value at the current moment is calculated. In specific clinical applications, patients typically have underlying physiological damage upon accessing the system. Therefore, the main control module 10 pre-stores an initial baseline risk index in its memory, derived from the patient's clinical score (e.g., SOFA score) upon admission. To integrate this static baseline pathological data with the dynamically extracted deterioration features in real time, the main control module 10 calls its internal microprocessor to execute a multiplicative degradation algorithm. As a preferred mathematical model of this invention, the main control module 10 employs an exponential base scaling mechanism, the specific calculation formula of which is as follows: ; in, The comprehensive early warning risk value output at the current moment has the same dimension as the baseline index and is a dimensionless pure number. The initial baseline risk index preset for the system; The positive constant bias term, set to prevent the risk value of low baseline populations from overflowing, is preferably set to a value range of 0.5 to 1.0, to ensure that wounded patients with low genetic susceptibility still have a sufficient base to exceed the alarm threshold when extreme physiological deterioration occurs. This is the dimensionless dynamic physiological perturbation degradation coefficient calculated in the previous stage; It is an exponential function with the natural constant as its base. Through this multiplicative degradation model, when the body deteriorates rapidly, the exponential term can nonlinearly amplify the basic risk, thus significantly highlighting the acute deterioration trend of the wounded's vital signs at the numerical calculation level.

[0052] S502, Main Control Module 10 Extracts Effect Chamber Predicted Drug Concentration The system utilizes a normalized, nonlinearly monotonically increasing function to perform a subtraction compensation operation on the initial safety threshold. From the perspective of clinical anesthetic pharmacology, higher concentrations of analgesic and sedative drugs substantially inhibit the central nervous system's autonomic compensatory reflexes to hypoxia and hypotension. This drug inhibition phenomenon leads to a false steady state in routine vital sign monitoring indicators, i.e., a masking effect. To counteract this masking effect at the system control level, the main control module 10 obtains the current estimated drug concentration in the effect room in real time from the target control module 50 via the controller local area network bus. Based on this, the main control module 10 introduces the classic Hill equation from pharmacodynamics as a nonlinear monotonically increasing function to dynamically correct the preset initial safety threshold downwards. The specific subtraction compensation calculation formula is as follows: ; in, The dynamic early warning trigger threshold is generated after compensation calculation, and it is related to... and They all have the same dimensionless risk index; This is an initial safety threshold set by the system for conscious injured persons. In a specific embodiment of the present invention, this initial safety threshold... The cutoff point for specificity and sensitivity is not arbitrarily assigned, but rather determined by calculating the Youden index using the receiver operating characteristic curve (ROC curve) based on a pre-collected large-scale database of conscious critically ill patients. This cutoff is tailored to match the scaled-down data volume of the comprehensive early warning risk value. The preferred value range is set between 1.5 and 3.0. The maximum allowable reduction in compensation is set to limit the physical boundary of threshold correction to prevent system overreaction; Estimate the drug concentration in the effect room for the current input, in μg / mL; The drug concentration constant that produces the half maximum pharmacological effect has the same dimension as μg / mL, so that the fractional term in the formula becomes a dimensionless proportionality coefficient in the range of 0 to 1. The Hill coefficient is used to adjust the steepness of the concentration-compensation effect curve. To avoid the system falling into a logic dead zone of continuous false alarms due to excessive threshold reduction, this embodiment imposes strict limits on the maximum allowable compensation reduction range. The range of values ​​is to Between these points, ensure that the dynamic early warning trigger threshold is met. Always remain positive. This is a preferred data configuration for commonly used sedative drugs such as propofol. The value is usually set between 2.0 and 4.0 μg / mL. The values ​​are set from 1.5 to 3.0, and these parameters are pre-programmed into the pharmacokinetic parameter library of the main control module 10.

[0053] S503, the main control module 10 generates a dynamic early warning trigger threshold through compensation calculation. This implements a control logic that increases the system's sensitivity to minor physiological abnormalities as the depth of anesthesia deepens. Based on the mathematical characteristics of the above formula, when the target control module (50) extrapolates the predicted drug concentration in the effect room... As the value gradually increases, the output value of the nonlinear term in the latter half of the compensation calculation formula will monotonically approximate the maximum allowable reduction in compensation. This calculation process determines the dynamic early warning trigger threshold. The trigger threshold decreases non-linearly as the depth of anesthesia increases. Because the overall baseline of the trigger threshold is automatically lowered, the activation energy required by the system to determine risk is correspondingly reduced. This process is specifically implemented in the specification, using mathematical control, to achieve the adaptive adjustment function of the warning tolerance outlined in the claims. Through this specific technical means, the system can trigger the alarm mechanism with high sensitivity in advance if a patient under anesthesia suppression experiences a genuine pathological disturbance, effectively preventing the risk of delayed treatment caused by the drug masking effect.

[0054] S504, the main control module 10 performs a rolling comparison according to a fixed period. and The size, in consecutive greater than or equal to When the preset number of judgment cycles is reached, the multi-touch display and the sound unit are driven to execute an alarm command. In the system's real-time monitoring main loop, the logic comparison unit inside the main control module 10 synchronously reads the updated comprehensive early warning risk value at fixed time intervals. With dynamic early warning trigger threshold To prevent false alarms caused by transient electromagnetic interference during signal transmission or a single abnormal sampling point, as a preferred embodiment of this invention, the main control module 10 allocates a status counter in memory. When the comparison result is displayed... Greater than or equal to When the status counter increments, it increments by one; if the comparison result falls back, the status counter is reset to zero. Only when the accumulated value of the status counter reaches the preset number of judgment cycles (e.g., 5 consecutive operation cycles) does the main control module 10 finally confirm that the hardware-level alarm trigger condition has been met. After the trigger condition is met, the main control module 10 immediately sends an interface redraw command to the multi-touch display screen through its internal high-definition multimedia interface, rendering a high-risk visual warning waveform in a preset color; simultaneously, it drives the sound unit through pulse width modulation peripherals to output a continuous alarm prompt tone conforming to medical equipment standards. For conventional multi-touch display rendering drive and buzzer sound control mechanisms, those skilled in the art can use existing medical electrical equipment alarm standards to implement them. The hardware-level drive and interface communication are well-known technologies in the field and will not be described in detail here.

[0055] To further clarify the collaborative working process of the technical solution described in this invention, a specific working scenario example will be used for illustration below. In this embodiment, a traumatically ill patient is connected to the machine learning-based vital sign data detection, modeling, and analysis system for critically injured patients of this invention.

[0056] First, in the initial stage of the casualty access system, the main control module 10 acquires specific CpG site methylation sequencing data from the casualty's whole blood sample via an interface. The main control module 10 then extracts feature vectors. The initial baseline risk index is then calculated and fed into the internal support vector machine baseline classifier. In this embodiment, the calculation is as follows: The value is 0.65, which is stored in memory by the main control module 10 as a static constant within a single monitoring cycle.

[0057] Subsequently, the system enters the continuous monitoring phase. The respiratory module 20, monitoring module 30, anesthesia depth module 40, and target control module 50 operate synchronously and in parallel. During a certain monitoring period, the patient's breathing tubing is subjected to mechanical compression during medical personnel's operations. The respiratory module 20 detects pressure fluctuations in the locally acquired high-frequency airway pressure sequence and calculates that the variance ratio of the short-term window to the long-term window exceeds the set fluctuation mutation threshold. The main control module 10 then sends an artifact verification command carrying a timestamp to the respiratory module 20.

[0058] like Figure 3 As shown in the figure, the horizontal axis represents the time axis of discrete sampling points, and the vertical axis represents the normalized amplitude of airway pressure and piezoelectric proportional valve drive current, respectively. The respiratory module 20 extracts the airway pressure sequence before and after the abrupt change in its local annular buffer. And extract the drive current sequence containing delay redundancy. Within the set acoustic delay search interval, the breathing module 20 introduces a time delay search step size variable. Normalized discrete cross-correlation calculations are performed on the two sets of sequences. Figure 3 The text shows when the variable At a specific value, the airway pressure spike waveform and the closed-loop compensating current waveform achieve the highest degree of alignment in time phase. Under this alignment, the respiratory module 20 calculates the maximum alignment cross-correlation coefficient. The value is 0.92. The main control module 10 receives this value and compares it with the preset correlation threshold of 0.80. Because... If the pressure exceeds the threshold, the main control module 10 confirms that the pressure fluctuation is a physical artifact caused by external mechanical deformation. It then keeps the state machine locked and does not input the fluctuation data into the subsequent early warning model, thus avoiding false alarms from the system.

[0059] After the monitoring process entered the next stage, due to trauma complications, the injured person experienced a pathological worsening of decreased respiratory compliance and impaired central microcirculatory perfusion. At this point, the system again detected that the variance ratio of airway pressure fluctuations exceeded the standard. The respiratory module 20 returned the data after local optimization calculation. The value is 0.35. The main control module 10 determines that this value is less than the correlation determination threshold, confirms that a real pathological disturbance has occurred, and then releases the state machine lock.

[0060] After the state machine is unlocked, the main control module 10 extracts the first-order time derivative of the respiratory compliance sequence within the set time window. The phase shift between the EEG waveform and the near-infrared cerebral blood oxygenation waveform was calculated using Hilbert transform and phase unwinding algorithm. The main control module 10 inputs the two sets of standardized features into the internal ridge regression model, and combines them with the partial regression coefficients fixed by offline training to calculate the dynamic physiological disturbance degradation coefficient, which characterizes the rate of deterioration of the body. During this stage, the system continuously outputs... The values ​​show a clear upward trend.

[0061] like Figure 4 As shown in the figure, the horizontal axis represents the system running time, and the vertical axis represents the dimensionless risk index value. The figure contains two main evolution curves: one is the dynamic early warning trigger threshold adjusted downwards based on drug compensation. The other curve represents the comprehensive early warning risk value, reflecting the real-time condition of the injured. curve.

[0062] In the specific operation of this embodiment, the target control module 50 continuously infuses sedative drugs into the wounded, calculates and outputs the estimated drug concentration in the effect room to the main control module 10. .like Figure 4 As shown, with As the value increases, the main control module 10 uses the Hill equation to adjust the initial safety threshold. Perform a nonlinear subtraction compensation operation. This operation process makes... Figure 4 In The curve exhibits a step-like correction characteristic that decreases over time, which, in terms of control logic, achieves an adaptive reduction of the alarm threshold under anesthesia inhibition.

[0063] At the same time, the main control module 10 will statically set the initial baseline risk index. Degradation coefficient of dynamic physiological perturbation output in real time Substitute into the multiplicative degradation model. For example... Figure 4 As shown, with The increase is obtained by scaling the exponential term. The curve exhibits a non-linear, accelerating upward trend. Figure 4 The point at which the two curves intersect, and the continuously rising comprehensive early warning risk value. The dynamic early warning trigger threshold for downward compensation has been exceeded. The main control module 10 performs rolling comparisons at fixed intervals and confirms the status by checking the status counter. After continuously reaching the preset number of judgment cycles, the system immediately drives the external multi-touch display screen to render a high-risk warning waveform and controls the sound unit to output an alarm prompt sound.

[0064] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A machine learning-based system for detecting, modeling, and analyzing vital signs data of critically injured patients, characterized in that: include: The main control module calculates the initial baseline risk index and generates artifact verification instructions; The breathing module receives the artifact verification command, collects the high-frequency sequence of airway pressure and the drive current sequence, calculates and outputs the maximum alignment cross-correlation coefficient to the main control module; When the main control module determines that a pathological disturbance has occurred based on the maximum alignment cross-correlation coefficient, the monitoring module extracts the respiratory compliance sequence and sends it to the main control module. The anesthesia module collects waveform sequence data based on the pathological disturbance and sends it to the main control module, which then maps the respiratory compliance sequence to the waveform sequence data into a dynamic physiological disturbance degradation coefficient. The target control module sends the estimated drug concentration data of the effect room to the main control module. The main control module uses the estimated drug concentration data of the effect room to generate a dynamic early warning trigger threshold, and fuses the initial baseline risk index with the dynamic physiological disturbance degradation coefficient to trigger an early warning action by comparing the dynamic early warning trigger threshold.

2. The machine learning-based vital sign data detection, modeling, and analysis system for critically injured patients according to claim 1, characterized in that, The main control module calculates the initial baseline risk index and generates artifact verification instructions, including: Obtain data on the methylation level of specific CpG sites in wounded personnel samples; The specific CpG site methylation level data is input into a support vector machine baseline classifier to calculate the initial baseline risk index. A time window is set to monitor the variance ratio of the high-frequency airway pressure sequence. When the variance ratio exceeds the fluctuation mutation threshold, the artifact verification command is generated and sent to the respiratory module.

3. The machine learning-based vital sign data detection, modeling, and analysis system for critically injured patients according to claim 2, characterized in that, The step of inputting the specific CpG site methylation level data into a support vector machine baseline classifier to calculate the initial baseline risk index includes: Read the methylation ratio values ​​of multiple specific CpG sites in whole blood samples and perform standardized preprocessing; The methylation ratio values ​​are arranged in a preset feature dimension order to construct a feature vector; The support vector machine baseline classifier is invoked to perform hyperplane mapping on the feature vectors and output the initial baseline risk index.

4. The machine learning-based vital sign data detection, modeling, and analysis system for critically injured patients according to claim 1, characterized in that, The step of receiving the artifact verification command, acquiring the high-frequency sequence of airway pressure and the driving current sequence, calculating and outputting the maximum alignment cross-correlation coefficient to the main control module includes: The high-frequency sequence of airway pressure and the drive current sequence including delay redundancy are extracted in a local annular buffer. By introducing a time delay search step size variable, the discrete cross-correlation function between the high-frequency sequence of airway pressure and the shifted driving current sequence is calculated. Within the set acoustic delay search interval, the value of the time delay search step size variable is continuously changed to perform the search, and the extreme value is extracted as the maximum alignment cross-correlation coefficient and sent to the main control module.

5. The machine learning-based vital sign data detection, modeling, and analysis system for critically injured patients according to claim 1, characterized in that, The step of extracting the respiratory compliance sequence and sending it to the main control module when the main control module determines that it is a pathological perturbation based on the maximum alignment cross-correlation coefficient includes: The received maximum alignment cross-correlation coefficient is compared with a preset correlation determination threshold; When the maximum alignment cross-correlation coefficient is greater than the correlation determination threshold, it is determined to be a physical artifact and the state machine is locked. When the maximum alignment cross-correlation coefficient is less than or equal to the correlation determination threshold, it is determined to be a pathological disturbance and the state machine lock is released. The respiratory compliance sequence is then extracted and sent to the main control module.

6. The machine learning-based vital sign data detection, modeling, and analysis system for critically injured patients according to claim 1, characterized in that, The process by which the main control module maps the respiratory compliance sequence to the waveform sequence data into a dynamic physiological perturbation degradation coefficient includes: The instantaneous phase of the electroencephalogram (EEG) signal and the instantaneous phase of the near-infrared cerebral blood oxygenation pulsation waveform are extracted from the waveform sequence data, respectively. The phase unwinding algorithm is invoked to restore the continuously monotonically increasing absolute physical phase sequence and the phase offset is calculated within the set allowable error time window; The first time derivative of the respiratory compliance sequence is extracted, and the first time derivative and the phase shift are input into the ridge regression model to map and output the dynamic physiological perturbation degradation coefficient.

7. The machine learning-based vital sign data detection, modeling, and analysis system for critically injured patients according to claim 1, characterized in that, The dynamic early warning trigger threshold generated by the main control module using the effect-room estimated drug concentration data includes: Extract the preset initial safety threshold; A nonlinear monotonically increasing function is used in conjunction with the estimated drug concentration data in the effect room to perform a subtraction compensation operation on the initial safety critical threshold; The numerical result of the subtraction compensation operation is used as the dynamic early warning trigger threshold.

8. The machine learning-based vital sign data detection, modeling, and analysis system for critically injured patients according to claim 7, characterized in that, The step of performing a subtraction compensation operation on the initial safety threshold using a nonlinear monotonically increasing function combined with the effect-room predicted drug concentration data includes: Obtain the set maximum permissible compensatory downregulation range and the drug concentration constant that produces the half maximum pharmacological effect; The Hill coefficient power of the predicted drug concentration data in the effect room is calculated as the numerator, and the sum of the Hill coefficient power of the drug concentration constant that produces the half maximum pharmacological effect and the numerator is calculated as the denominator. The numerator is divided by the denominator to obtain the proportionality coefficient. Multiply the maximum allowable compensation reduction by the proportional coefficient to obtain the reduction product, and subtract the reduction product from the initial safety threshold.

9. The machine learning-based vital sign data detection, modeling, and analysis system for critically injured patients according to claim 1, characterized in that, The step of fusing and comparing the initial baseline risk index with the dynamic physiological disturbance degradation coefficient to trigger a warning action based on the dynamic warning trigger threshold includes: Obtain the preset positive constant bias term; The initial baseline risk index is added to the positive constant bias term to generate the bias risk base; The exponential degradation value is calculated by using the natural constant as the base and the dynamic physiological disturbance degradation coefficient as the exponent. Multiply the bias risk base value by the exponential degradation value to output a comprehensive early warning risk value; When the comprehensive early warning risk value is continuously greater than or equal to the dynamic early warning trigger threshold for a preset number of judgment cycles, the terminal is driven to output the early warning action.

10. The machine learning-based vital sign data detection, modeling, and analysis system for critically injured patients according to claim 9, characterized in that, The step of adding the initial baseline risk index to the positive constant bias term to generate the bias risk base includes: The initial baseline risk index is retrieved from the fixed address space of static random access memory; Extract the positive constant bias term set to prevent underflow of risk values ​​in low baseline populations; The addition operation is performed to combine the initial baseline risk index with the positive constant bias term, outputting the bias risk base, and then proceeding to the multiplication operation step.

11. The life-integrated machine of the machine learning-based vital sign data detection, modeling and analysis system for critically injured patients as described in any one of claims 1-10.

12. The life-integrated machine according to claim 11, characterized in that, It also includes conventional medical modules such as a respiratory module, a monitoring module, anesthesia module, and infusion module; The breathing module is used to support the wounded soldier's breathing. The monitoring module is used to monitor the vital signs of the injured. The anesthesia module is used to anesthetize the wounded. The infusion module is used to administer intravenous fluids to the injured.

Images