Respiration intensive care method and system

Abstract

The invention discloses a respiratory intensive care method and system, and the method comprises the steps: deploying an acoustic sensor on a respirator loop, and synchronously obtaining the ventilation parameters and optional temperature and humidity data of a respirator; preprocessing the acquired multi-modal data, and converting an acoustic signal into a time-frequency spectrogram; the preprocessed data are input into a multi-modal fusion fault recognition model, the model comprises a convolutional neural network branch used for extracting acoustic features and a recurrent neural network branch used for capturing time sequence dynamics, and the probability of a specific fault precursor is output by combining the features of all the branches through a fusion layer; and finally, performing early warning decision according to the probability. The invention also comprises a system for implementing the method. According to the invention, through deep fusion of acoustic, pneumatic and thermodynamic data, accurate prediction of fault precursor is realized, and an earlier time window is provided for clinical intervention, so that the monitoring safety is improved.

Description

Technical Field

[0001] This application relates to the field of real-time monitoring and electronic data processing, specifically to a respiratory intensive care method and system. Background Technology

[0002] Acute respiratory distress syndrome (ARDS) is a common critical illness in the field of intensive care, and mechanical ventilation is the core life support method. Clinical studies have confirmed that driving pressure (ΔP) is the respiratory mechanics parameter most closely related to the prognosis of ARDS patients, and maintaining driving pressure within a safe threshold is key to implementing lung-protective ventilation strategies and reducing the risk of ventilator-associated lung injury (VILI).

[0003] For elderly ARDS patients, due to their unique physiological and pathological characteristics, their lungs exhibit extremely high sensitivity to even minor changes in driving pressure; any unexpected pressure increase can lead to fatal consequences. However, existing respiratory monitoring technologies have significant limitations. Ventilator safety systems primarily rely on passive alarm mechanisms based on fixed thresholds; for example, the system only triggers an alarm when parameters such as airway pressure or tidal volume clearly exceed preset limits.

[0004] This passive alarm mode has a monitoring blind spot between the early stages of a fault and the actual triggering of the alarm. Before the alarm is triggered, a series of minor, progressive mechanical faults may have already occurred in the ventilator circuit, such as humidifier condensation buildup in the tubing (rain-out), minor leaks at tubing connections, or incomplete kinking of the tubing. While these minor faults are insufficient to immediately trigger a conventional alarm, they continuously and insidiously increase airway resistance or alter ventilation efficiency, causing the actual driving pressure applied to the patient's lungs to gradually rise unnoticed. The dynamic changes in temperature and humidity within the circuit are key physical factors contributing to phenomena such as "rain-out," but these are not effectively utilized by current monitoring technologies. For elderly ARDS patients, this undetected pressure shift poses a significant safety hazard.

[0005] Furthermore, current artificial intelligence technologies applied in intensive care primarily focus on predicting clinical outcomes (such as mortality risk) or optimizing weaning timing based on patient physiological data. These technologies generally assume that the ventilator and its circuit system are stable and reliable ideal components, neglecting the proactive and predictive monitoring of the physical integrity of the treatment tool itself. Therefore, existing technologies cannot proactively identify minute mechanical malfunction precursors that could adversely affect the driving pressure in elderly ARDS patients before traditional alarms are triggered, leaving a technological gap. Summary of the Invention

[0006] The technical problem this invention aims to solve is to provide a respiratory intensive care method and system that addresses the shortcomings of existing technologies in ensuring ventilation safety for specific vulnerable patient groups (especially elderly patients with acute respiratory distress syndrome (ARDS) who are extremely sensitive to changes in driving pressure). This invention aims to achieve early and accurate identification and warning of any minor mechanical malfunction precursors that may cause unexpected deviations in driving pressure by proactively and with high sensitivity continuously monitoring the physical state of the ventilator itself, before traditional alarm thresholds are exceeded. This elevates the safety of lung-protective ventilation strategies to a new level.

[0007] To address the aforementioned technical problems, the present invention provides a respiratory intensive care method, which includes the following steps: Step S1: Data Acquisition. At least one acoustic sensor is deployed at a preset location in the ventilator circuit to synchronously acquire real-time ventilation parameter data output by the ventilator device. The acoustic sensor is used to continuously acquire acoustic signals of gas flow in the ventilator circuit, forming a raw acoustic data stream; the real-time ventilation parameter data includes at least one or more time-series data of airway pressure, gas flow rate, and tidal volume.

[0008] Preferably, to achieve preliminary fault location and more refined feature capture, at least one acoustic sensor is deployed in each of the inspiratory and expiratory phase tubing of the ventilator circuit. More preferably, the acoustic sensor is a high-sensitivity contact microphone or a piezoelectric accelerometer to minimize environmental noise interference and directly capture vibration signals from the tubing walls.

[0009] Preferably, to construct a more comprehensive monitoring dimension, the present invention may also deploy temperature sensors and humidity sensors at the inspiratory and expiratory ends of the ventilator circuit, respectively, to synchronously collect real-time temperature and humidity data of the gas in the circuit, forming an environmental status data stream. This data stream can provide key thermodynamic characteristics for accurately identifying fault precursors closely related to temperature and humidity changes (such as water accumulation in pipelines).

[0010] Step S2: Data Preprocessing. The raw acoustic data stream and real-time ventilation parameter data collected in Step S1, as well as optional environmental condition data, are processed to generate a data format suitable for input to the artificial intelligence model.

[0011] Preferably, step S2 includes the following sub-steps: Sub-step S201: Acoustic data processing. The continuous raw acoustic data stream is segmented into short audio frames with fixed durations (e.g., 1-2 seconds); each audio frame undergoes a short-time Fourier transform (STFT) or Mel frequency transform to convert it from a one-dimensional time-domain signal into a two-dimensional time-spectrum or Mel spectrum. This process visualizes the characteristics of the audio signal in both the time and frequency domains, facilitating subsequent image feature extraction.

[0012] Sub-step S202: Ventilation parameter data processing. The synchronously acquired real-time ventilation parameter time series data is aligned and segmented according to the same time window as the audio frame, forming a one-to-one corresponding ventilation parameter sequence segment. These sequence segments can be normalized to eliminate the influence of dimensions.

[0013] Sub-step S203: Environmental state data processing. If environmental state data has been collected, the synchronously acquired time-series data such as temperature and humidity will be aligned and segmented according to the same time window as the audio frame to form a one-to-one corresponding environmental state sequence segment, which can then be normalized.

[0014] Sub-step S204: Noise Suppression. During or after the generation of the time-of-flight spectrum, algorithms such as spectral subtraction, Wiener filtering, or deep learning-based denoising autoencoders are used to suppress environmental noise or background equipment noise in order to enhance the signal-to-noise ratio of the fault characteristic signal.

[0015] Step S3: Individualized baseline calibration and fault precursor identification. The preprocessed data is input into a pre-trained multimodal fusion fault identification model with online adaptive learning capabilities. The model outputs the probability that a specific mechanical fault precursor exists in the ventilator circuit at the current moment.

[0016] Preferably, when this method is first applied to a specific patient and device combination, an initial individualized baseline calibration sub-step is performed: in the initial stage of confirming that the ventilator circuit is in normal working order, the system not only collects normal operation data for a period of time (e.g., 5-10 minutes), but also simultaneously records or prompts the operator to input the patient's initial core respiratory mechanics parameters (e.g., lung compliance (Crs) and airway resistance (Rrs) measured by the end-inspiratory breath-hold maneuver). Based on this, the model learns and constructs a human-machine system health baseline coupled with the patient's individual physiological characteristics. This baseline reflects the "normal" acoustic and aerodynamic behavior pattern of the device under specific lung compliance and resistance conditions. This ensures that subsequent anomaly detection is based on changes in the device's own state, rather than misjudgments caused by changes in the patient's own lung mechanics, thereby greatly improving the sensitivity and specificity of subsequent anomaly detection.

[0017] The multimodal fusion fault identification model, within a unified framework, can simultaneously learn and fuse acoustic features, ventilator dynamics features, and optional thermodynamic features. Preferably, the model includes multiple parallel feature extraction branches: The first feature extraction branch is an acoustic feature extractor based on a convolutional neural network (CNN). This branch receives the temporal spectrogram as input and automatically extracts high-dimensional spatial features such as local and global texture features and harmonic structure features from the temporal spectrogram related to different faults (such as the specific low-frequency sound "gurgling" produced by water accumulation, the "hissing" sound produced by micro-leakage, the turbulent sound produced by blockage, or the abnormal vibration sound of components such as valves inside the ventilator).

[0018] The second feature extraction branch is a ventilation parameter temporal feature extractor based on a recurrent neural network (RNN, such as a long short-term memory network LSTM or a gated recurrent unit GRU). This branch receives ventilation parameter sequence fragments as input to capture subtle but time-dependent dynamic change patterns in the waveforms of parameters such as airway pressure and flow rate under the influence of minor faults.

[0019] The third feature extraction branch is a ventilation environment state temporal feature extractor based on a recurrent neural network. This branch receives environmental state sequence segments as input and captures temperature and humidity drift or fluctuation patterns associated with specific faults (especially water accumulation in pipes).

[0020] In a more preferred embodiment, the model of the present invention may further include a fourth feature extraction branch, which is a real-time physiological parameter calculation and feature extractor. This branch continuously receives real-time airway pressure and flow rate data and calculates in real-time derived indicators reflecting the patient's lung mechanical state, such as dynamic driving pressure per breath, stress index, or airway resistance change trends. This branch uses the time series of these derived physiological indicators as input to extract features that reflect the patient's response to subtle changes in the device.

[0021] After extracting the feature vectors from each branch, a fusion layer (e.g., splicing operation, attention mechanism, or tensor fusion) effectively combines acoustic features, temporal features of ventilation parameters, optional temporal features of environmental states, and optional real-time physiological response features to form a fused feature vector containing more comprehensive information. Finally, this fused feature vector is passed through a fully connected layer and a Softmax activation function to output a probability distribution, corresponding to the probability values ​​of preset fault categories such as "circuit normal", "circuit water accumulation", "circuit minor leakage", "circuit partial blockage", "filter blockage", and "internal component malfunction".

[0022] As attachedFigure 4 The implementation details of this model are illustrated below. The first feature extraction branch can specifically employ a lightweight convolutional neural network, such as MobileNetV2 or ResNet-18, to meet the needs of real-time computing on bedside devices. This CNN model receives a 128xN Mel spectrogram and outputs, for example, a 512-dimensional acoustic feature vector. The second and third feature extraction branches can both employ a two-layer Long Short-Term Memory (LSTM) network with 128 hidden units per layer. They receive ventilation parameter sequences and environmental state sequences, respectively, and output two 128-dimensional temporal feature vectors. In the fusion layer, this embodiment preferably employs an attention-based fusion mechanism. This mechanism can dynamically assign weights to features from different modalities. For example, when judging "loop water accumulation," the model automatically assigns higher weights to acoustic features (gurgling sounds) and thermodynamic features (temperature and humidity changes) through learning. When judging "minor loop leaks," combined with the input from the fourth branch, the model might detect that when a tiny "hissing" feature appears, the patient's dynamic driving pressure also begins to show a weak but continuous upward trend, thus greatly increasing the confidence in judging minor leak faults. Finally, the weighted and fused feature vectors are passed through a fully connected layer, and the final probability of each fault category is calculated by the Softmax function. It can be understood that in another embodiment, the fusion layer could also be a simple concatenation operation, directly concatenating all feature vectors into a longer vector before subsequent processing.

[0023] Preferably, the model has online adaptive learning capabilities, and can use data that the system continuously judges as "normal circuit" as new samples during operation to dynamically fine-tune the individualized baseline to adapt to changes in the patient's condition or environment and maintain the accuracy of long-term monitoring.

[0024] Step S4: Early Warning Decision and Generation. Based on the probability of fault precursors output in Step S3, determine whether to trigger an early warning.

[0025] This step includes setting a probability threshold (e.g., 85%) and a duration threshold (e.g., three consecutive identification cycles). When the model outputs a probability for a specific fault precursor that is consistently higher than its corresponding probability threshold for the set duration, the occurrence of that fault precursor event is confirmed. Subsequently, the system generates and issues a warning signal that not only indicates the risk but also clearly identifies the type of fault precursor and provides guidance for intervention.

[0026] In a preferred embodiment, the early warning decision logic is dynamically adaptive. The system allows clinicians to pre-define a "patient vulnerability level" (e.g., divided into "low," "medium," and "high") based on the patient's specific circumstances (such as ARDS severity, age, underlying diseases, etc.). The system automatically adjusts the early warning threshold according to this level. For example, for a patient assessed as having a "high" vulnerability level, the system might lower the probability threshold for "micro-leakage in the circuit" from the default 85% to 75%, and shorten the duration threshold from 3 identification cycles to 2. This means that for the most vulnerable patients, the system will operate with the highest sensitivity, ensuring that any potential risks are detected as early as possible. This adaptive early warning strategy based on patient risk stratification enables monitoring resources to be more intelligently allocated to patients who need them most, achieving the goal of personalized precision medicine.

[0027] This embodiment provides a quantitative explanation of the early warning decision-making logic. The probability threshold is not a single fixed value, but is associated with the type of fault. For example, the probability threshold for an emergency risk such as "partial circuit blockage" can be set to a higher 90%, while the probability threshold for a relatively mild risk such as "minor circuit leakage" can be set to 80%. The duration threshold is set to three consecutive recognition cycles, with each recognition cycle being 2 seconds. This means that the system will only officially trigger an early warning when the probability of the model reporting the same fault precursor for 6 consecutive seconds is higher than the threshold. The principle behind this design is that by setting a time window, false alarms caused by instantaneous data fluctuations due to sudden coughing or changes in body position can be effectively filtered out, avoiding the clinical pain point of "alarm fatigue" and improving the reliability and clinical credibility of the early warning.

[0028] To achieve the above method, the present invention also provides a respiratory intensive care system, the system comprising: A data acquisition module includes at least one acoustic sensor deployed on a ventilator circuit and means for connecting to a ventilator data interface, the data acquisition module being configured to perform step S1 of the aforementioned method. Preferably, the module further includes a temperature sensor and a humidity sensor deployed on the ventilator circuit.

[0029] A processing module is electrically connected to the data acquisition module. This processing module has a built-in processor and memory. The memory stores computer program instructions, which, when executed by the processor, will execute steps S2, S3, and S4 of the aforementioned method. Specifically, this processing module is configured to preprocess the acquired data, run a multimodal fusion fault identification model with adaptive learning capabilities for calculation, and execute early warning decision logic based on the model output.

[0030] A user interaction and alarm module, electrically connected to the processing module, is used to output specific fault precursor warning signals through sound, light, or by sending information to the central monitoring system after receiving a warning decision command from the processing module. Preferably, the module also includes a display interface for visually presenting the fault precursor information according to the instructions.

[0031] The technical solution provided by this invention innovatively extends the monitoring focus from the patient's biological body to the physical circuitry of the treatment device itself, analyzing both as an interactive human-machine coupled system. By utilizing an adaptive, multimodal artificial intelligence model that deeply integrates acoustic, aerodynamic, and thermodynamic data, this invention achieves accurate prediction of the precursors of minute mechanical faults in areas where traditional technologies are ineffective. Its core value lies in the fact that it no longer passively waits for faults to significantly impact the patient's physiological parameters, but rather issues warnings minutes or even longer before the fault has a substantial adverse effect on the driving pressures of vulnerable patients such as elderly patients with ARDS. This proactive risk management model gains a decisive time window for clinical intervention, ensuring that the core treatment strategy of lung-protective ventilation can be implemented within its designed safe trajectory, thereby significantly improving the life safety of critically ill patients and possessing significant clinical application value. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Appendix Figure 1 This is a schematic diagram of the structure of a respiratory intensive care system provided in an embodiment of the present invention; Appendix Figure 2 This is a flowchart of a respiratory intensive care method provided in an embodiment of the present invention; Appendix Figure 3 This is a comparison diagram of the time-frequency domain characteristics of acoustic signals corresponding to different fault types in embodiments of the present invention; Appendix Figure 4 This is a detailed architectural diagram of the multimodal fusion neural network in an embodiment of the present invention; Appendix Figure 5 This is a schematic diagram of the process of individualized baseline calibration and online adaptive learning in an embodiment of the present invention. Detailed Implementation

[0034] 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 of the invention and are not intended to limit the scope of protection of the invention.

[0035] The core design concept of the methods and systems disclosed in this invention is rooted in a deep understanding of the clinical needs of specific critically ill patient groups. Taking elderly patients with acute respiratory distress syndrome (ARDS) as an example, their lung tissue has poor compliance and high fragility. For these patients, lung-protective ventilation strategies require strict control of the driving pressure (ΔP) within a safe threshold (e.g., ≤15 cmH2O), as any slight, undetected pressure increase can cause irreversible lung damage. However, traditional monitoring systems are passive, only issuing alarms when damage has already occurred or is occurring (i.e., after parameters exceed the threshold). This invention aims to resolve this fundamental contradiction by shifting the monitoring focus forward, from monitoring abnormal physiological parameters to predicting precursors of equipment malfunction, thereby providing proactive safety assurance for elderly patients with ARDS and other patients with highly unstable lung mechanics. The following embodiments will illustrate in detail how this goal is achieved. Example 1

[0036] See attached document Figure 2 This embodiment provides a respiratory intensive care method. This method aims to achieve early and accurate identification and warning of minor mechanical fault precursors affecting drive pressure before traditional alarm thresholds are exceeded by analyzing continuous, multimodal sensor data from the ventilator circuit. (Appendix) Figure 2 The core steps of the method are shown.

[0037] Step S1: Data Acquisition. At least one acoustic sensor is deployed at a preset location in the ventilator circuit to synchronously acquire real-time ventilation parameter data output by the ventilator device. The acoustic sensor is used to continuously acquire acoustic signals of gas flow in the ventilator circuit, forming a raw acoustic data stream; the real-time ventilation parameter data includes at least one or more time-series data of airway pressure, gas flow rate, and tidal volume.

[0038] In one specific implementation of this embodiment, the details of step S1 have been optimized. The acoustic sensor can specifically be a microelectromechanical system (MEMS) microphone, which has the advantages of small size, low power consumption, and high sensitivity, making it suitable for integration onto the surface of the tubing. The sensor's sampling rate is set to 44.1 kHz, sufficient to capture high-frequency hissing sounds of micro-leakage or subtle sounds of liquid flow within the tubing. Data acquisition is accomplished through a dedicated data acquisition card, which receives acoustic data via the I2S (Inter-IC Sound) bus protocol and synchronously reads ventilation parameters from the digital communication port of the ventilator (e.g., mainstream brands such as Dräger and Hamilton) via an RS-232 or Ethernet interface following Modbus or HL7 protocols. This ensures strict alignment of the acoustic data and ventilation parameter data in terms of timestamps, providing a high-quality data foundation for subsequent multimodal fusion analysis.

[0039] Preferably, to achieve preliminary fault location and more refined feature capture, at least one acoustic sensor is deployed in each of the inspiratory and expiratory phase tubing of the ventilator circuit. More preferably, the acoustic sensor is a high-sensitivity contact microphone or a piezoelectric accelerometer to minimize environmental noise interference and directly capture vibration signals from the tubing walls.

[0040] In another embodiment, the selection and deployment of the sensor are further elaborated. A piezoelectric accelerometer is used as a contact acoustic sensor, and medical-grade double-sided tape is used to tightly attach it to the midpoint of the outer wall of the inspiratory and expiratory tubing. The principle behind this deployment is that when water accumulation, micro-leakage, or blockage occurs in the circuit, the turbulence state of the airflow changes, causing unique high-frequency vibrations in the tubing wall that can be captured by the piezoelectric sensor. Compared to microphones that capture airborne sound, this method can more effectively filter out background noise propagated through the air, such as nurses' conversations and alarms from other equipment in the ICU environment, thereby significantly improving the signal-to-noise ratio of the target signal. It is understood that in special medical environments with strong electromagnetic interference, the acoustic sensor can also be a fiber optic acoustic sensor to ensure the stability of signal acquisition.

[0041] Preferably, to construct a more comprehensive monitoring dimension, the present invention may also deploy temperature sensors and humidity sensors at the inspiratory and expiratory ends of the ventilator circuit, respectively, to synchronously collect real-time temperature and humidity data of the gas in the circuit, forming an environmental status data stream. This data stream can provide key thermodynamic characteristics for accurately identifying fault precursors closely related to temperature and humidity changes (such as water accumulation in pipelines).

[0042] To illustrate this preferred approach, consider the following example: Elderly ARDS patients often require the use of heated humidifiers to maintain airway humidification. When the warm, moist air from the humidifier flows through the relatively cool ventilator tubing, condensation can easily form at low points in the tubing, a phenomenon known as "rain-out," if the ambient temperature is low or the tubing is improperly laid. By deploying high-precision digital temperature and humidity sensors (such as the SHT3x series) at the inspiratory end (near the humidifier outlet) and the expiratory end (near the patient's Y-tube), the system can monitor the temperature and humidity differences between the two ends of the tubing in real time. When an abnormal drop in temperature or persistent humidity saturation is detected at the expiratory end, combined with the "gurgling" sound captured by the acoustic sensor, the model can determine with high confidence that a precursor to tubing water accumulation has occurred. This demonstrates the crucial role of integrating thermodynamic features in improving the accuracy of specific fault identification.

[0043] Step S2: Data Preprocessing. The raw acoustic data stream and real-time ventilation parameter data collected in Step S1, as well as optional environmental condition data, are processed to generate a data format suitable for input to the artificial intelligence model.

[0044] The purpose of this step is to transform raw, multi-source, heterogeneous sensor data into a structured, standardized format that highlights key features, so that it can be used as input for subsequent artificial intelligence models.

[0045] Preferably, step S2 includes the following sub-steps: Sub-step S201: Acoustic data processing. The continuous raw acoustic data stream is segmented into short audio frames with fixed durations (e.g., 1-2 seconds); each audio frame undergoes a short-time Fourier transform (STFT) or Mel frequency transform to convert it from a one-dimensional time-domain signal into a two-dimensional time-spectrum or Mel spectrum. This process visualizes the characteristics of the audio signal in both the time and frequency domains, facilitating subsequent image feature extraction.

[0046] In a specific implementation of this embodiment, sub-step S201 is further decomposed and quantized. First, a Hannian window is used as the window function to window the continuous audio stream, with a window length of 1024 sampling points and a frame shift (Hop Length) of 512 sampling points. Then, a Fast Fourier Transform (FFT) is performed on each windowed audio frame. Finally, the obtained spectral energy is passed through a Mel filter bank containing 128 filters and logarithmic operations are performed to generate a 128×N (N is the number of time frames) Mel spectrogram. The Mel spectrogram is chosen because its frequency scale simulates the auditory characteristics of the human ear, providing higher resolution for low-frequency components and more finely characterizing abnormal harmonics and formant changes in breathing sounds, thereby improving the sensitivity of fault identification.

[0047] As attached Figure 3 As shown, different types of mechanical faults exhibit distinctly different characteristics in the time-frequency domain: the acoustic spectrum under normal operating conditions shows a regular periodic distribution; water accumulation faults in the circuit manifest as significant energy accumulation in the low-frequency range of 50-200Hz, corresponding to a "gurgling" water sound; while micro-leakage faults display a continuous energy band in the high-frequency range of 800-1500Hz, corresponding to a "hissing" leaking sound. This difference in frequency domain characteristics provides a reliable physical basis for subsequent fault classification.

[0048] Sub-step S202: Ventilation parameter data processing. The synchronously acquired real-time ventilation parameter time series data is aligned and segmented according to the same time window as the audio frame, forming a one-to-one corresponding ventilation parameter sequence segment. These sequence segments can be normalized to eliminate the influence of dimensions.

[0049] Specifically, the alignment of ventilation parameter sequence segments is based on the precise timestamps recorded in S1. For example, a 2-second audio frame corresponds to a 2-second ventilation parameter sequence segment. Data within this segment is normalized using Z-Score standardization, which involves subtracting the mean of the segment from each data point and then dividing by its standard deviation. This method transforms parameters in different physical units (such as cmH2O for pressure and L / min for flow rate) into dimensionless values ​​that follow a standard normal distribution. This avoids parameters with large numerical ranges (such as pressure) from having an excessively disproportionate impact on model weights during model training, thus accelerating model convergence and improving accuracy.

[0050] Sub-step S203: Environmental state data processing. If environmental state data has been collected, the synchronously acquired time-series data such as temperature and humidity will be aligned and segmented according to the same time window as the audio frame to form a one-to-one corresponding environmental state sequence segment, which can then be normalized.

[0051] This step is processed similarly to sub-step S202. It also performs alignment and segmentation based on timestamps, and can use Z-Score or Min-Max normalization methods to process temperature and humidity data. This allows thermodynamic features to be fused with features from other modalities on the same numerical scale.

[0052] Sub-step S204: Noise Suppression. During or after the generation of the time-of-flight spectrum, algorithms such as spectral subtraction, Wiener filtering, or deep learning-based denoising autoencoders are used to suppress environmental noise or background equipment noise in order to enhance the signal-to-noise ratio of the fault characteristic signal.

[0053] In a preferred implementation, this invention employs a deep learning-based denoising autoencoder scheme. During the model training phase, a large amount of clean acoustic signals from normal breathing circuits recorded in a shielded environment, as well as pure noise signals from an ICU environment, are collected. Then, the clean signals and noise signals are artificially synthesized into a noisy signal, which is used to train a denoising autoencoder-decoder model. This model learns to reconstruct the spectrum of the clean signal from the spectrum of the noisy signal. In practical applications, after generating the spectrum of all collected acoustic data, a "clean" spectrum is first output through this pre-trained denoising model before being fed into the subsequent fault identification model. This effectively filters out interference from irrelevant signals such as monitor beeps and footsteps, allowing the fault identification model to focus more on analyzing acoustic features directly related to the circuit state.

[0054] Step S3: Individualized baseline calibration and fault precursor identification. The preprocessed data is input into a pre-trained multimodal fusion fault identification model with online adaptive learning capabilities. The model outputs the probability that a specific mechanical fault precursor exists in the ventilator circuit at the current moment.

[0055] This step is the core of the method of this invention. It not only utilizes the generalization ability of the pre-trained model, but also achieves a high degree of adaptability to specific patients and scenarios through individualized calibration and online learning.

[0056] Preferably, when this method is first applied to a specific patient and device combination, an initial individualized baseline calibration sub-step is performed: in the initial stage of confirming that the ventilator circuit is in normal working order, the system not only collects normal operation data for a period of time (e.g., 5-10 minutes), but also simultaneously records or prompts the operator to input the patient's initial core respiratory mechanics parameters (e.g., lung compliance (Crs) and airway resistance (Rrs) measured by the end-inspiratory breath-hold maneuver). Based on this, the model learns and constructs a human-machine system health baseline coupled with the patient's individual physiological characteristics. This baseline reflects the "normal" acoustic and aerodynamic behavior pattern of the device under specific lung compliance and resistance conditions. This ensures that subsequent anomaly detection is based on changes in the device's own state, rather than misjudgments caused by changes in the patient's own lung mechanics, thereby greatly improving the sensitivity and specificity of subsequent anomaly detection.

[0057] One clinical application example is when a 78-year-old female patient diagnosed with ARDS was first connected to the system. After confirming that the ventilator tubing was properly connected and the parameters were set correctly, the responsible nurse clicked the "Start Baseline Calibration" button on the system interface. The system then began collecting and analyzing data over the next 10 minutes, simultaneously recording the patient's initial Crs as 35 mL / cmH2O and Rrs as 12 cmH2O / L / s. The "individualized health baseline" learned and constructed by the model is specific to this human-machine coupled system with "Crs=35, Rrs=12". All subsequent real-time data will be compared with this dynamically updated baseline. This greatly reduces false alarms caused by physiological differences between patients and provides crucial evidence for distinguishing between equipment malfunctions and changes in patient condition.

[0058] The multimodal fusion fault identification model, within a unified framework, can simultaneously learn and fuse acoustic features, ventilator dynamics features, and optional thermodynamic features. Preferably, the model includes multiple parallel feature extraction branches: The first feature extraction branch is an acoustic feature extractor based on a convolutional neural network (CNN). This branch receives the temporal spectrogram as input and automatically extracts high-dimensional spatial features such as local and global texture features and harmonic structure features from the temporal spectrogram related to different faults (such as the "gurgling" sound caused by water accumulation, the "hissing" sound caused by micro-leakage, the turbulent sound caused by blockage, or the abnormal vibration sound of components such as valves inside the ventilator).

[0059] The second feature extraction branch is a ventilation parameter temporal feature extractor based on a recurrent neural network (RNN, such as a long short-term memory network LSTM or a gated recurrent unit GRU). This branch receives ventilation parameter sequence fragments as input to capture subtle but time-dependent dynamic change patterns in the waveforms of parameters such as airway pressure and flow rate under the influence of minor faults.

[0060] The third feature extraction branch is a ventilation environment state temporal feature extractor based on a recurrent neural network. This branch receives environmental state sequence segments as input and captures temperature and humidity drift or fluctuation patterns associated with specific faults (especially water accumulation in pipes).

[0061] In a more preferred embodiment, the model of the present invention may further include a fourth feature extraction branch, which is a real-time physiological parameter calculation and feature extractor. This branch continuously receives real-time airway pressure and flow rate data and calculates in real-time derived indicators reflecting the patient's lung mechanical state, such as dynamic driving pressure per breath, stress index, or airway resistance change trends. This branch uses the time series of these derived physiological indicators as input to extract features that reflect the patient's response to subtle changes in the device.

[0062] After extracting the feature vectors from each branch, a fusion layer (e.g., splicing operation, attention mechanism, or tensor fusion) effectively combines acoustic features, temporal features of ventilation parameters, optional temporal features of environmental states, and optional real-time physiological response features to form a fused feature vector containing more comprehensive information. Finally, this fused feature vector is passed through a fully connected layer and a Softmax activation function to output a probability distribution, corresponding to the probability values ​​of preset fault categories such as "circuit normal", "circuit water accumulation", "circuit minor leakage", "circuit partial blockage", "filter blockage", and "internal component malfunction".

[0063] As attached Figure 4The implementation details of this model are illustrated below. The first feature extraction branch can specifically employ a lightweight convolutional neural network, such as MobileNetV2 or ResNet-18, to meet the needs of real-time computing on bedside devices. This CNN model receives a 128xN Mel spectrogram and outputs, for example, a 512-dimensional acoustic feature vector. The second and third feature extraction branches can both employ a two-layer Long Short-Term Memory (LSTM) network with 128 hidden units per layer. They receive ventilation parameter sequences and environmental state sequences, respectively, and output two 128-dimensional temporal feature vectors. In the fusion layer, this embodiment preferably employs an attention-based fusion mechanism. This mechanism can dynamically assign weights to features from different modalities. For example, when judging "loop water accumulation," the model automatically assigns higher weights to acoustic features (gurgling sounds) and thermodynamic features (temperature and humidity changes) through learning. When judging "minor loop leaks," combined with the input from the fourth branch, the model might detect that when a tiny "hissing" feature appears, the patient's dynamic driving pressure also begins to show a weak but continuous upward trend, thus greatly increasing the confidence in judging minor leak faults. Finally, the weighted and fused feature vectors are passed through a fully connected layer, and the final probability of each fault category is calculated by the Softmax function. It can be understood that in another embodiment, the fusion layer could also be a simple concatenation operation, directly concatenating all feature vectors into a longer vector before subsequent processing.

[0064] Preferably, the model has online adaptive learning capabilities, and can use data that the system continuously judges as "normal circuit" as new samples during operation to dynamically fine-tune the individualized baseline to adapt to changes in the patient's condition or environment and maintain the accuracy of long-term monitoring.

[0065] The principle behind online adaptive learning here is that the system continuously fine-tunes model parameters using data samples that are highly confidently considered "normal" at a low learning rate. For example, when the model outputs a "normal loop" probability higher than 99% for more than one hour consecutively, the system adds the data features collected during this period to the baseline database and performs a small update to the model weights. The technical effect is that it allows the monitoring system to adapt to slow changes in the patient's lung condition (such as gradual improvement or deterioration of lung compliance), avoiding performance degradation due to an outdated baseline and ensuring robustness in long-term monitoring. (See attached image) Figure 5 As shown in the figure, this diagram illustrates the complete adaptive learning process from initial baseline calibration, real-time monitoring, high-confidence data selection to dynamic updating of model parameters.

[0066] Step S4: Early Warning Decision and Generation. Based on the probability of fault precursors output in Step S3, determine whether to trigger an early warning.

[0067] This step includes setting a probability threshold (e.g., 85%) and a duration threshold (e.g., three consecutive identification cycles). When the model outputs a probability for a specific fault precursor that is consistently higher than its corresponding probability threshold for the set duration, the occurrence of that fault precursor event is confirmed. Subsequently, the system generates and issues a warning signal that not only indicates the risk but also clearly identifies the type of fault precursor and provides guidance for intervention.

[0068] In a preferred embodiment, the early warning decision logic is dynamically adaptive. The system allows clinicians to pre-define a "patient vulnerability level" (e.g., divided into "low," "medium," and "high") based on the patient's specific circumstances (such as ARDS severity, age, underlying diseases, etc.). The system automatically adjusts the early warning threshold according to this level. For example, for a patient assessed as having a "high" vulnerability level, the system might lower the probability threshold for "micro-leakage in the circuit" from the default 85% to 75%, and shorten the duration threshold from 3 identification cycles to 2. This means that for the most vulnerable patients, the system will operate with the highest sensitivity, ensuring that any potential risks are detected as early as possible. This adaptive early warning strategy based on patient risk stratification enables monitoring resources to be more intelligently allocated to patients who need them most, achieving the goal of personalized precision medicine.

[0069] This embodiment provides a quantitative explanation of the early warning decision-making logic. The probability threshold is not a single fixed value, but is associated with the type of fault. For example, the probability threshold for an emergency risk such as "partial circuit blockage" can be set to a higher 90%, while the probability threshold for a relatively mild risk such as "minor circuit leakage" can be set to 80%. The duration threshold is set to three consecutive recognition cycles, with each recognition cycle being 2 seconds. This means that the system will only officially trigger an early warning when the probability of the model reporting the same fault precursor for 6 consecutive seconds is higher than the threshold. The principle behind this design is that by setting a time window, false alarms caused by instantaneous data fluctuations due to sudden coughing or changes in body position can be effectively filtered out, avoiding the clinical pain point of "alarm fatigue" and improving the reliability and clinical credibility of the early warning.

[0070] Preferably, the warning signal is visualized on the monitoring system interface. Abnormal frequency bands in the acoustic time-spectrum graph are highlighted, or the real-time ventilation parameter waveform is superimposed and compared with the theoretical normal waveform generated by the system based on historical data, and preliminary fault location guidance is provided in text form (e.g., "Warning: Increased risk of water accumulation in the expiratory tubing, please check the humidifier and the low point of the tubing").

[0071] Here's a concrete application example: When the system detects early signs of "expiratory loop fluid accumulation," a window pops up on the central monitoring station's display interface. The left side of the window displays the real-time Mel-spectrum graph of the acoustic signal collected by the expiratory acoustic sensor, highlighting the abnormal low-frequency energy region (e.g., the 50-200Hz band) corresponding to the "gurgling" sound with a red box. The right side of the window displays both the real-time airway pressure waveform (solid red line) and the theoretical normal pressure waveform generated by the system based on an individualized baseline (dashed blue line) in superimposed curve form, allowing doctors to visually see the subtle distortions in the pressure curve. The bottom of the window displays clear text information: "Warning [Level 2]: Early signs of expiratory loop fluid accumulation detected (94% confidence level). Recommended action: Immediately check and drain fluid from the low point of the expiratory loop." This multi-dimensional information presentation not only informs of the risk but also explains the basis for the judgment and provides clear action guidelines, greatly improving the efficiency and accuracy of clinical intervention. Example 2

[0072] To achieve the above method, this embodiment also provides a respiratory intensive care system. (See attached document.) Figure 1 The system includes: a data acquisition module 10, a processing module 20, and a user interaction and alarm module 30.

[0073] A data acquisition module 10 includes at least one acoustic sensor 11 deployed on a ventilator circuit and a device 12 for connecting to a ventilator data interface. The data acquisition module is configured to perform step S1 of the aforementioned method. Preferably, the module further includes a temperature sensor and a humidity sensor 13 deployed on the ventilator circuit.

[0074] Specifically, the data acquisition module 10 is a physical unit with integrated clips on its housing for securing it to the ventilator tubing. An acoustic sensor 11 (e.g., a piezoelectric sensor) is built into the inner wall of the clips to ensure good contact with the tubing. The device 12 is a multi-functional interface box providing multiple physical interfaces such as RS-232, USB, and RJ45, and includes a built-in protocol conversion chip to ensure compatibility with data output formats from different brands of ventilators. Temperature / humidity sensors 13 are integrated into two probes that can be inserted into the tubing sampling ports.

[0075] A processing module 20 is electrically connected to the data acquisition module 10. This processing module 20 has a built-in processor 21 and a memory 22. The memory 22 stores computer program instructions. When the processor 21 executes these instructions, it performs steps S2, S3, and S4 of the aforementioned method. Specifically, this processing module is configured to preprocess the acquired data, run a multimodal fusion fault identification model with adaptive learning capabilities for calculation, and execute early warning decision logic based on the model output.

[0076] In this embodiment, the processing module 20 can be a compact edge computing device (e.g., a device based on the NVIDIA Jetson series or Intel NUC) placed at the patient's bedside. The processor 21 is specifically a system-on-a-chip (SoC) containing a CPU and an embedded GPU, giving it sufficient computing power to perform data preprocessing and deep learning model inference locally in real time. The memory 22 is non-volatile memory (such as SSD), which stores the weight parameters of the pre-trained multimodal fusion fault identification model, the denoising model parameters, and the computer program executing all method steps. By performing the main computation at the edge, the dependence on the hospital's central network bandwidth can be reduced, and the core monitoring and early warning functions can still operate reliably even in the extreme case of network outages.

[0077] A user interaction and alarm module 30, electrically connected to the processing module 20, is used to output specific fault precursor warning signals through sound, light, or by sending information to the central monitoring system after receiving a warning decision command from the processing module 20. Preferably, the module also includes a display interface 31 for visually presenting the fault precursor information according to the command.

[0078] The user interaction and alarm module 30 can be integrated into the chassis of the processing module 20 or it can be a standalone flat panel display device. It includes a high-decibel buzzer and a high-brightness LED indicator for audible and visual alarms. Simultaneously, it connects to the hospital's monitoring network via a separate Ethernet interface, enabling it to send structured alarm information (compliant with the HL7 standard) to the central monitoring workstation or push it to nurses' mobile terminals via IoT protocols (such as MQTT). The display interface 31 is a touchscreen LCD that clearly displays detailed information about the warnings and suggested actions to clinical staff in a visual manner as described in Example 1.

[0079] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A respiratory intensive care method, characterized in that, The method includes the following steps: Step S1: Data acquisition. At least one acoustic sensor is deployed on the ventilator circuit and the real-time ventilation parameter data output by the ventilator device is acquired synchronously. The acoustic sensor is used to continuously acquire the acoustic signal of gas flow in the ventilator circuit to form a raw acoustic data stream. The real-time ventilation parameter data includes at least one or more time series data of airway pressure, gas flow rate and tidal volume. Step S2: Data preprocessing, processing the raw acoustic data stream and real-time ventilation parameter data collected in step S1 to generate a data format suitable for input to the artificial intelligence model; Step S3: Fault precursor identification. The preprocessed data is input into a multimodal fusion fault identification model. The model outputs the probability that the ventilator circuit has one or more preset types of mechanical fault precursors at the current moment. The preset types of mechanical fault precursors include at least one of the following: circuit water accumulation, circuit micro-leakage, circuit partial blockage, filter blockage, or internal component abnormality. Step S4: Early warning decision and generation. Based on the probability of the fault precursor output in step S3, determine whether to trigger an early warning.

2. The respiratory intensive care method according to claim 1, characterized in that, In step S1, at least one acoustic sensor is deployed on the inspiratory phase line and the expiratory phase line of the ventilator circuit, respectively.

3. A respiratory intensive care method according to claim 1 or 2, characterized in that, In step S1, a temperature sensor and a humidity sensor are also deployed on the ventilator circuit to synchronously collect real-time temperature and humidity data of the gas in the circuit, forming an environmental state data stream; and in step S3, the multimodal fusion fault identification model also integrates the thermodynamic features in the environmental state data stream.

4. The respiratory intensive care method according to claim 1, characterized in that, The data preprocessing in step S2 includes: dividing the continuous original acoustic data stream into short audio frames, and performing time-frequency domain transformation on each audio frame to convert it from a one-dimensional time-domain signal into a two-dimensional time-spectrum or Mel-spectrum for feature extraction by the multimodal fusion fault identification model.

5. A respiratory intensive care method according to claim 1, characterized in that, The multimodal fusion fault identification model includes multiple parallel feature extraction branches, and the feature extraction branches include at least: The first feature extraction branch is an acoustic feature extractor based on a convolutional neural network, which is used to receive the time-spectrum map generated according to the original acoustic data stream as input and extract the high-dimensional spatial features therein; The second feature extraction branch is a ventilation parameter time-series feature extractor based on a recurrent neural network, which is used to receive ventilation parameter sequence segments aligned with the time window of the time spectrum as input and capture their time-dependent dynamic change patterns. And a fusion layer for combining the features extracted by the first and second feature extraction branches to provide the probability of the fault precursor for subsequent output.

6. A respiratory intensive care method according to claims 3 and 5, characterized in that, The feature extraction branch of the multimodal fusion fault identification model also includes: The third feature extraction branch is a ventilation environment state time series feature extractor based on a recurrent neural network, which is used to receive environmental state sequence segments aligned with the time window of the time spectrum as input and capture temperature and humidity drift or fluctuation patterns related to specific faults. Furthermore, the fusion layer uses an attention mechanism to weightedly combine the features extracted from the first, second, and third feature extraction branches.

7. A respiratory intensive care method according to claim 1, characterized in that, Before step S3, there is an initial individualized baseline calibration sub-step: in the initial stage of confirming that the ventilator circuit is in normal operation, normal operation data for a period of time is collected, and combined with the patient's initial respiratory mechanics parameters, the model learns and constructs a human-machine system health baseline that couples the patient's individual physiological characteristics.

8. A respiratory intensive care method according to claim 7, characterized in that, The multimodal fusion fault identification model has online adaptive learning capabilities, and can use data that the system continuously judges as being in a normal circuit state as new samples during operation to dynamically fine-tune the individualized baseline.

9. A respiratory intensive care method according to claim 1, characterized in that, In step S4, the probability threshold or duration threshold used to determine whether an early warning is triggered is dynamically adjusted based on a preset patient risk level.

10. A respiratory intensive care system, characterized in that, The system includes: A data acquisition module includes at least one acoustic sensor deployed on a ventilator circuit and means for connecting to a ventilator data interface, the data acquisition module being configured to perform step S1 as described in claim 1; A processing module is electrically connected to the data acquisition module. The processing module has a built-in processor and a memory. The memory stores computer program instructions. When the processor executes these instructions, it is configured to perform steps S2, S3 and S4 as described in claim 1. A user interaction and alarm module, electrically connected to the processing module, is used to output specific fault precursor warning signals after receiving the warning decision instruction issued by the processing module.

Images