A method for processing ventilator data
By collecting and structuring multimodal data from ventilators in real time, and combining this with adjustments and evaluations by medical staff, the problems of time sequence and operational correspondence in traditional ventilator data recording methods have been solved. This enables data support for intelligent decision-making systems and improves the integrity and usability of the data.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI SONGJIANG DISTRICT CENTRAL HOSPITAL
- Filing Date
- 2026-03-24
- Publication Date
- 2026-06-02
AI Technical Summary
Traditional ventilator data recording methods lack timeliness and operational correspondence, cannot meet the needs of multimodal, high-dimensional time-series data analysis, cannot support the training of intelligent algorithms, and lack expert feedback information from medical personnel.
Multimodal data is collected in real time through the ventilator interface. Combined with the step-by-step adjustment operations and evaluations of medical staff, the data is preprocessed and structured to construct a training dataset for artificial intelligence models, including time-series correspondence, preprocessing, and quantitative feedback label generation.
This has enabled the transformation from single-point-of-time parameter snapshots to full-process structured datasets, improving data integrity and availability, supporting the training of various artificial intelligence algorithms such as reinforcement learning and imitation learning, and promoting the development of intelligent decision-making systems for ventilators.
Smart Images
Figure CN122132839A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical data processing technology, and specifically relates to a method for processing ventilator data. Background Technology
[0002] In clinical practice, medical staff need to frequently adjust ventilator parameters based on patients' real-time physiological feedback. This adjustment process contains a wealth of clinical decision-making information. However, traditional ventilator data recording methods have significant shortcomings. Existing data records are mostly intermittent, single-point-in-time parameter snapshots, recording only the ventilator settings and monitoring values at a specific moment. This fails to fully reflect the complete decision-making process of medical staff from observation and evaluation to step-by-step adjustments. The data records lack temporal sequence and operational correspondence, making it impossible to reconstruct the adjustment logic and decision-making path. Traditional data records do not include the subjective judgment information of medical staff on the effectiveness of each adjustment step. In actual adjustment, medical staff assess the changing trends of respiratory parameters after each operation to determine whether the current adjustment is effective. This type of labeled data with expert feedback is crucial for training artificial intelligence models, but existing collection methods completely ignore this critical information.
[0003] The data types output by ventilators encompass high-frequency waveform data (waveforms of airway pressure, flow rate, and tidal volume changes over time), discrete monitoring parameters (blood oxygen saturation, end-expiratory carbon dioxide partial pressure, etc.), and setting parameters (PEEP, inhaled oxygen concentration, etc.). However, traditional methods often only record the values of some setting parameters, failing to meet the needs of multimodal, high-dimensional time-series data analysis. Furthermore, due to the lack of structured, end-process regulation data, existing data cannot support the training of autonomous decision-making systems for ventilators based on intelligent algorithms such as reinforcement learning and temporal reasoning. Reinforcement learning algorithms require training data to be organized in a Markov decision process format of "state-action-reward-next state," a structure completely absent in traditional data.
[0004] Therefore, there is an urgent need for a method that can comprehensively, continuously, and structurally collect and process data throughout the entire process of ventilator regulation, so as to lay a data foundation for the development of intelligent ventilator decision-making systems. Summary of the Invention
[0005] This invention aims to collect multimodal respiratory parameters in real time from the ventilator interface using dedicated software, and combine this with the step-by-step adjustment operations and subjective evaluations of medical personnel. Through systematic data preprocessing and structured encapsulation, a structured training dataset for training artificial intelligence models is constructed.
[0006] To achieve the above objectives, the present invention provides a ventilator data processing method, the method comprising the following steps: Step S1: Collect multimodal raw data output by the ventilator in real time through the ventilator's data transmission interface. The multimodal raw data includes respiratory waveform data, monitoring parameter data, and setting parameter data.
[0007] Step S2: In response to the step-by-step adjustment operation of the ventilator by medical staff, data acquisition is triggered after each adjustment operation is performed, the temporal correspondence between each adjustment operation and the acquired data is established, and a temporal operation data sequence with operation timestamp and operation step identifier is generated.
[0008] Step S3: Receive the effect evaluation information of the respiratory parameter changes after each adjustment operation from medical staff, convert the effect evaluation information into a quantitative feedback label, and associate the quantitative feedback label with the time sequence operation data of the corresponding step.
[0009] Step S4: Preprocess the multimodal raw data. The preprocessing includes: transmission error detection and correction, data segment time sequence reconstruction, missing parameter calculation and completion, and waveform reconstruction.
[0010] Step S5: Pair the preprocessed data according to adjacent operation steps in the time-series operation data sequence to construct state transition data pairs. Each state transition data pair is associated with corresponding operation action information and quantization feedback label.
[0011] Step S6: The state transition data pairs and their associated operation action information and quantization feedback labels are encapsulated in a structured manner according to a preset data pattern, and output as a structured training dataset for training artificial intelligence models.
[0012] Furthermore, in the multimodal raw data: The respiratory waveform data includes airway pressure-time waveform data, airflow velocity-time waveform data, and tidal volume-time waveform data; wherein, the airway pressure-time waveform data is a waveform signal reflecting the continuous change of airway pressure in the patient's breathing circuit over time, the airflow velocity-time waveform data is a waveform signal reflecting the change of gas flow velocity over time during inspiration and expiration, and the tidal volume-time waveform data is a waveform signal reflecting the cumulative change of the volume of gas entering and leaving the lungs over time in a single respiratory cycle.
[0013] Furthermore, the monitoring parameters include blood oxygen saturation, end-tidal carbon dioxide partial pressure, measured respiratory rate, measured tidal volume, measured minute ventilation, peak airway pressure, and plateau airway pressure.
[0014] Furthermore, the setting parameter data includes breathing mode, set respiratory rate, set tidal volume, positive end-expiratory pressure, inhaled oxygen concentration, inspiratory pressure support level, and inspiratory-to-expiratory ratio.
[0015] Furthermore, the time-series reassembly of the data fragments includes: The discrete data segments output by the ventilator at first preset time intervals are sequentially concatenated along the time axis to form a continuous data segment covering a second preset time window. The concatenation relationship is represented as follows: ;in, The number of data segments required for splicing represents the number of discrete data segments contained within a complete time window; The first preset time interval is the time interval between two consecutive data outputs from the ventilator, ranging from 3 to 10 seconds. The second preset time window is the total time length covered by the continuous data segments after splicing, and the value ranges from 30 to 120 seconds.
[0016] During the splicing process, boundary consistency is checked at the junctions of adjacent data segments. When a value jump at the junction is detected to exceed a preset threshold, interpolation smoothing is used to eliminate splicing artifacts.
[0017] Furthermore, the missing parameter calculation and completion includes: The target parameters that the ventilator does not directly output in the current operating mode are determined based on the current operating mode of the ventilator; the values of the target parameters are calculated using a respiratory mechanics model based on the associated parameters collected in the multimodal raw data.
[0018] Furthermore, the respiratory mechanics model includes: Calculate respiratory rate based on respiratory cycle: ;in, The calculated respiratory rate is expressed in breaths per minute, representing the number of times the patient breathes per minute. The duration of a single respiratory cycle, expressed in seconds, represents the time elapsed from the start of one inhalation to the start of the next.
[0019] Dynamic lung compliance is calculated based on tidal volume and pressure difference: ;in, Dynamic lung compliance reflects the elastic expansion capacity of lung tissue under dynamic respiratory conditions; Tidal volume is the volume of gas that enters the lungs during a single respiratory cycle. Peak airway pressure, which is the highest pressure value in the airway during the inspiratory phase; Positive end-expiratory pressure (PEEP) is the positive pressure maintained in the airway at the end of expiration.
[0020] Calculate airway resistance based on inspiratory flow rate and pressure difference: ;in, Airway resistance reflects the magnitude of the frictional resistance encountered by gas as it passes through the respiratory tract; The plateau pressure is the static pressure in the airway at the end of inspiration when the patient holds their breath, reflecting the pressure within the alveoli. This is the peak inlet velocity, which is the maximum inlet velocity of the gas in the inlet phase.
[0021] Furthermore, the continuous waveform reconstruction of the discrete numerical sequence includes: extracting airway pressure sample values, airflow velocity sample values, and tidal volume sample values stored in the form of a numerical list from the multimodal raw data; determining the sampling time corresponding to each sample value according to the transmission protocol of the data transmission interface; and reconstructing the discrete sampling points into a continuous waveform curve using a piecewise linear interpolation method with the sampling time as the horizontal axis and the sample value as the vertical axis. The interpolation expression is as follows: ;in, For any time The waveform signal value obtained by interpolation at that location; and For two adjacent sampling times, represent the th sampling time, respectively. The and the first The time coordinates corresponding to each sampling point; and At the sampling time respectively and The actual signal value collected at the location.
[0022] Furthermore, after piecewise linear interpolation reconstruction, the method further includes: performing bandpass filtering on the reconstructed waveform curve. The bandpass filtering is used to remove high-frequency noise interference and low-frequency baseline drift components from the signal, retain the effective signal components that match the frequency range of respiratory physiological activities, and output the filtered standard respiratory waveform data for use in subsequent data processing steps.
[0023] Furthermore, the method for generating the quantization feedback label includes: The system provides healthcare professionals with an effectiveness evaluation option that includes at least three levels: improvement, no change, and deterioration.
[0024] The evaluation options selected by medical staff are mapped to preset numerical codes to form the quantitative feedback labels. The mapping relationship is as follows: ;in, For the first The quantitative feedback label value corresponding to the step adjustment operation; +1 indicates that the patient's respiratory parameters improved in a favorable direction after the step adjustment operation, 0 indicates that the patient's respiratory parameters did not change after the step adjustment operation, and -1 indicates that the patient's respiratory parameters deteriorated in an unfavorable direction after the step adjustment operation.
[0025] Furthermore, the quantized feedback label is also used as an immediate reward signal in reinforcement learning training, and the quantized feedback label value is normalized. The normalization expression is as follows: ;in, The normalized reward signal value, with a range of [-1, 1], is used to eliminate differences in different evaluation scales and adapt to the input requirements of reinforcement learning algorithms; For the first The original quantization feedback label value of the step operation; The maximum absolute value of the quantized feedback label value is used as a scaling factor for normalization.
[0026] Furthermore, the construction step of the state transition data pair includes: […]. The preprocessed data collected before the step adjustment operation is defined as a state vector: ;in, For the first The system state vector before the operation comprehensively describes the overall state of the ventilator and the patient at that moment. It is a waveform data matrix containing time-series waveform data of airway pressure, airflow velocity, and tidal volume; The monitoring parameter vector includes real-time monitoring values of blood oxygen saturation and end-tidal carbon dioxide partial pressure; The parameter vector is set to include configuration parameters for the current breathing mode of the ventilator and the set tidal volume; This is a vector of patient physiological indicators, containing physiological parameters completed by the respiratory mechanics model.
[0027] The first The operation content of step adjustment is encoded as an action vector: ;in, For the first The step's motion vector describes the specific adjustment operations performed by medical personnel; This serves as a type identifier for the adjusted parameter, indicating which ventilator parameter was adjusted in this operation; The adjustment range indicates the amount of change in the value of this parameter during this operation.
[0028] Will Combined into a state transition data pair, where To execute the first The new state vector is acquired after each step; multiple consecutive state transition data pairs constitute a complete ventilator adjustment trajectory. ;in, This is a complete ventilator adjustment trajectory, recording the entire adjustment process from the initial state to the final state; This represents the total number of operation steps in this adjustment process.
[0029] Furthermore, prior to step S6, a data verification and annotation correction step is also included: The system displays respiratory waveform data, monitoring parameter data, setting parameter data, and corresponding quantitative feedback labels at each time point using a visual interface. It supports medical staff in correcting the quantitative feedback labels at any time point, as well as adding, deleting, or reordering time points in the data acquisition sequence. After correction, the system re-executes the state transition data pair construction step S5 based on the correction results.
[0030] Compared with the prior art, the beneficial effects of the present invention are: This invention overcomes the limitations of traditional ventilator data recording methods, which lack temporal sequence and operational correspondence, by organizing the entire process data of ventilator adjustment into a state transition data pair structure. It realizes the transformation from single-time-point parameter snapshots to a full-process structured dataset. Through a four-level data preprocessing pipeline, it achieves transmission error correction, data fragment reconstruction, missing parameter completion, and continuous waveform reconstruction, significantly improving the integrity and usability of the data. The structured training dataset output by this invention can be directly used for training various artificial intelligence algorithms such as reinforcement learning, imitation learning, and temporal reasoning, promoting the development of intelligent decision-making systems for ventilators. Attached Figure Description
[0031] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only a part of the embodiments of this invention, not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.
[0033] Example like Figure 1 The diagram shown is a flowchart of a ventilator data processing method according to the present invention. The method includes the following steps: Step S1: Collect multimodal raw data output by the ventilator in real time through the ventilator's data transmission interface. The multimodal raw data includes respiratory waveform data, monitoring parameter data, and setting parameter data.
[0034] In the multimodal raw data: The respiratory waveform data includes airway pressure-time waveform data, airflow velocity-time waveform data, and tidal volume-time waveform data; wherein, the airway pressure-time waveform data is a waveform signal reflecting the continuous change of airway pressure in the patient's breathing circuit over time, the airflow velocity-time waveform data is a waveform signal reflecting the change of gas flow velocity over time during inspiration and expiration, and the tidal volume-time waveform data is a waveform signal reflecting the cumulative change of the volume of gas entering and leaving the lungs over time in a single respiratory cycle.
[0035] The monitoring parameters include blood oxygen saturation, end-tidal carbon dioxide partial pressure, measured respiratory rate, measured tidal volume, measured minute ventilation, peak airway pressure, and plateau airway pressure.
[0036] The setting parameters include breathing mode, set respiratory rate, set tidal volume, positive end-expiratory pressure, inhaled oxygen concentration, inspiratory pressure support level, and inspiratory-to-expiratory ratio.
[0037] For example, in the intensive care unit, medical staff adjust the ventilator parameters for a patient with acute respiratory distress syndrome (ARDS) and use the method described in this invention to collect and process data. The specific process is as follows: Data Acquisition: Medical staff activate the data acquisition software and connect to the ventilator's data transmission port via an RJ45 interface. The system automatically identifies the ventilator model and loads the corresponding communication protocol. The data acquisition software begins real-time acquisition of multimodal raw data output by the ventilator, including: respiratory waveform data (airway pressure-time waveform, airflow velocity-time waveform, and tidal volume-time waveform, sampling frequency of 200Hz); monitoring parameter data (blood oxygen saturation SpO2, end-tidal carbon dioxide partial pressure EtCO2, measured respiratory rate, measured tidal volume, peak airway pressure, airway plateau pressure, etc.); and setting parameter data (breathing mode, PEEP, inspiratory oxygen concentration FiO2, inspiratory pressure support level, inspiratory-expiratory ratio, etc.).
[0038] Meanwhile, medical staff manually enter the patient's basic information, including age, gender, height, weight, and diagnosis, as the basic data for the patient's condition. In this example, the patient is a 56-year-old male, 172cm tall, weighing 75kg, and diagnosed with ARDS.
[0039] Step S2: In response to the step-by-step adjustment operation of the ventilator by medical staff, data acquisition is triggered after each adjustment operation is performed, the temporal correspondence between each adjustment operation and the acquired data is established, and a temporal operation data sequence with operation timestamp and operation step identifier is generated.
[0040] Step-by-step operation steps: After observing the patient's current respiratory status, medical staff develop a step-by-step adjustment plan. In this embodiment, the medical staff plans to break down the ventilator adjustment process into 5 steps: Step 1: Increase PEEP from 8 cmH2O to 10 cmH2O; Step 2: Reduce FiO2 from 60% to 55%; Step 3: Adjust the tidal volume from 420mL to 400mL; Step 4: Adjust the inhalation pressure support from 12 cmH2O to 14 cmH2O; Step 5: After observing for 30 seconds, decide whether to fine-tune FiO2 based on the changes in SpO2.
[0041] After each adjustment step is completed, medical staff click the acquisition button in the data acquisition software to trigger the system to collect all ventilator data at that moment. The system automatically records the precise timestamp of the acquisition time and the operation step identifier (such as Step-1, Step-2, etc.), establishing a time-series correspondence between the acquired data and the corresponding operation steps. After each acquisition trigger, the system continuously records data for 30 to 60 seconds to fully capture the dynamic changes in respiratory parameters after the adjustment operation.
[0042] Step S3: Receive the effect evaluation information of the respiratory parameter changes after each adjustment operation from medical staff, convert the effect evaluation information into a quantitative feedback label, and associate the quantitative feedback label with the time sequence operation data of the corresponding step.
[0043] Feedback label generation steps: After each adjustment operation is completed and changes in respiratory parameters are observed, medical staff evaluate the adjustment effect of the current step on the same interface of the data acquisition software. The system provides three evaluation levels: improvement, no change, and deterioration.
[0044] In this embodiment, the effects of each step were evaluated as follows: Step 1 (increasing PEEP): SpO2 increased from 89% to 92%, evaluated as "improvement"; Step 2 (decreasing FiO2): SpO2 remained at 91%, evaluated as "no change"; Step 3 (adjusting tidal volume): Airway plateau pressure decreased from 29 cmH2O to 27 cmH2O, evaluated as "improvement"; Step 4 (increasing inspiratory pressure support): Tidal volume recovered to 415 mL and plateau pressure did not increase significantly, evaluated as "improvement"; Step 5 (fine-tuning FiO2 to 53% after observation): SpO2 stabilized at 92%, evaluated as "no change".
[0045] Step S4: Preprocess the multimodal raw data. The preprocessing includes: transmission error detection and correction, data segment time sequence reconstruction, missing parameter calculation and completion, and waveform reconstruction.
[0046] The time-series reassembly of the data segments includes: The discrete data segments output by the ventilator at first preset time intervals are sequentially concatenated along the time axis to form a continuous data segment covering a second preset time window. The concatenation relationship is represented as follows: ;in, The number of data segments required for splicing represents the number of discrete data segments contained within a complete time window; The first preset time interval is the time interval between two consecutive data outputs from the ventilator, ranging from 3 to 10 seconds. The second preset time window is the total time length covered by the continuous data segments after splicing, and the value ranges from 30 to 120 seconds.
[0047] During the splicing process, boundary consistency is checked at the junctions of adjacent data segments. When a value jump at the junction is detected to exceed a preset threshold, interpolation smoothing is used to eliminate splicing artifacts.
[0048] The missing parameter calculation and completion includes: determining the target parameter that the ventilator does not directly output in the current working mode of the ventilator; and calculating the value of the target parameter using a respiratory mechanics relationship model based on the associated parameters collected in the multimodal raw data.
[0049] The respiratory mechanics model includes: calculating respiratory rate based on the respiratory cycle. ;in, The calculated respiratory rate is expressed in breaths per minute, representing the number of times the patient breathes per minute. The duration of a single respiratory cycle, expressed in seconds, represents the time elapsed from the start of one inhalation to the start of the next.
[0050] Dynamic lung compliance is calculated based on tidal volume and pressure difference: ;in, Dynamic lung compliance reflects the elastic expansion capacity of lung tissue under dynamic respiratory conditions; Tidal volume is the volume of gas that enters the lungs during a single respiratory cycle. Peak airway pressure, which is the highest pressure value in the airway during the inspiratory phase; Positive end-expiratory pressure (PEEP) is the positive pressure maintained in the airway at the end of expiration.
[0051] Calculate airway resistance based on inspiratory flow rate and pressure difference: ;in, Airway resistance reflects the magnitude of the frictional resistance encountered by gas as it passes through the respiratory tract; The plateau pressure is the static pressure in the airway at the end of inspiration when the patient holds their breath, reflecting the pressure within the alveoli. This is the peak inlet velocity, which is the maximum inlet velocity of the gas in the inlet phase.
[0052] The continuous waveform reconstruction of discrete numerical sequences includes: extracting airway pressure, airflow velocity, and tidal volume sample values stored in numerical list form from the multimodal raw data; determining the sampling time corresponding to each sample value according to the transmission protocol of the data transmission interface; and reconstructing the discrete sampling points into a continuous waveform curve using a piecewise linear interpolation method with the sampling time as the horizontal axis and the sample value as the vertical axis. The interpolation expression is as follows: ;in, For any time The waveform signal value obtained by interpolation at that location; and For two adjacent sampling times, represent the th sampling time, respectively. The and the first The time coordinates corresponding to each sampling point; and At the sampling time respectively and The actual signal value collected at the location.
[0053] The acquired multimodal raw data undergoes a four-stage preprocessing pipeline: Level 1: Transmission Error Detection and Correction; The system performs integrity verification on the raw data received through the RJ45 interface. A statistical anomaly detection method is used; when the value of a sampling point deviates from the local mean by more than three times the standard deviation, an error is detected. The system identifies this point as a transmission error and corrects it using linear interpolation of adjacent sampling points. Simultaneously, redundant characters and frame header / tail markers generated during protocol parsing are removed from the data stream. In this embodiment, the system detects and automatically corrects approximately 0.3% of the abnormal data points in all collected data.
[0054] Level 2: Data Segment Time Sequence Reassembly; The ventilator outputs data segments at 5-second intervals. The system sequentially splices these discrete segments along the time axis into a continuous 60-second data segment. The splicing relationship is as follows: Where N=12, =5 seconds, =60 seconds. During the splicing process, boundary consistency is checked at the junctions of adjacent segments. When a jump in airway pressure exceeding 2 cmH2O is detected at the junction, cubic spline interpolation is used for smoothing to eliminate splicing artifacts caused by transmission delay.
[0055] Level 3: Missing parameter calculation and completion; Based on the current operating mode of the ventilator (SIMV mode in this example), determine the parameters that the ventilator does not directly output in this mode. Based on the collected associated parameters, calculate the target parameters using a respiratory mechanics model: The respiratory rate was calculated based on the respiratory cycle. In this embodiment, the respiratory cycle was determined by detecting the zero-crossing point of the airflow velocity waveform, and the actual respiratory rate was calculated to be 18 breaths / minute. Dynamic lung compliance was calculated based on tidal volume and pressure difference. In this embodiment, the tidal volume was 400 mL, the peak pressure was 32 cmH2O, and the PEEP was 10 cmH2O, resulting in a dynamic lung compliance of 18.2 mL / cmH2O. Airway resistance was calculated based on inspiratory flow rate and pressure difference. In this embodiment, the peak pressure was 32 cmH2O, the plateau pressure was 27 cmH2O, and the peak flow rate was 0.5 L / s, resulting in an airway resistance of 10 cmH2O / (L / s).
[0056] Level 4: Reconstruction of continuous waveforms from discrete numerical sequences; The sampled values of airway pressure, airflow velocity, and tidal volume, stored in the form of a numerical list, are used to determine the sampling time based on a sampling frequency of 200Hz, and piecewise linear interpolation is employed to reconstruct the continuous waveform. The reconstructed waveform is processed by bandpass filtering, with a passband frequency of 0.1Hz to 40Hz, to remove high-frequency electromagnetic interference noise and baseline drift, and output standard respiratory waveform data; each 60-second waveform contains 12,000 sampling points, which can clearly show the complete morphology of the inspiratory phase, plateau phase and expiratory phase in the respiratory cycle.
[0057] After the piecewise linear interpolation reconstruction, the method further includes: performing bandpass filtering on the reconstructed waveform curve. The bandpass filtering is used to remove high-frequency noise interference and low-frequency baseline drift components in the signal, retain the effective signal components that match the frequency range of respiratory physiological activities, and output the filtered standard respiratory waveform data for use in subsequent data processing steps.
[0058] Step S5: Pair the preprocessed data according to adjacent operation steps in the time-series operation data sequence to construct state transition data pairs. Each state transition data pair is associated with corresponding operation action information and quantization feedback label.
[0059] The method for generating the quantization feedback label includes: The system provides healthcare professionals with an effectiveness evaluation option that includes at least three levels: improvement, no change, and deterioration.
[0060] The evaluation options selected by medical staff are mapped to preset numerical codes to form the quantitative feedback labels. The mapping relationship is as follows: ;in, For the first The quantitative feedback label value corresponding to the step adjustment operation; +1 indicates that the patient's respiratory parameters improved in a favorable direction after the step adjustment operation, 0 indicates that the patient's respiratory parameters did not change after the step adjustment operation, and -1 indicates that the patient's respiratory parameters deteriorated in an unfavorable direction after the step adjustment operation.
[0061] The system maps effect evaluation to quantitative feedback labels. In this embodiment, the feedback labels for each step are: +1, 0, +1, +1, 0. Each quantitative feedback label is automatically associated with the time-series operation data of the corresponding step.
[0062] The quantized feedback label is also used as an immediate reward signal in reinforcement learning training, and the quantized feedback label value is normalized. The normalization expression is as follows: ;in, The normalized reward signal value, with a range of [-1, 1], is used to eliminate differences in different evaluation scales and adapt to the input requirements of reinforcement learning algorithms; For the first The original quantization feedback label value of the step operation; The maximum absolute value of the quantized feedback label value is used as a scaling factor for normalization.
[0063] Step S6: The state transition data pairs and their associated operation action information and quantization feedback labels are encapsulated in a structured manner according to a preset data pattern, and output as a structured training dataset for training artificial intelligence models.
[0064] The steps for constructing the state transition data pair include: ... The preprocessed data collected before the step adjustment operation is defined as a state vector: ;in, For the first The system state vector before the operation comprehensively describes the overall state of the ventilator and the patient at that moment. It is a waveform data matrix containing time-series waveform data of airway pressure, airflow velocity, and tidal volume; The monitoring parameter vector includes real-time monitoring values of blood oxygen saturation and end-tidal carbon dioxide partial pressure; The parameter vector is set to include configuration parameters for the current breathing mode of the ventilator and the set tidal volume; This is a vector of patient physiological indicators, containing physiological parameters completed by the respiratory mechanics model.
[0065] The first The operation content of step adjustment is encoded as an action vector: ;in, For the first The step's motion vector describes the specific adjustment operations performed by medical personnel; This serves as a type identifier for the adjusted parameter, indicating which ventilator parameter was adjusted in this operation; The adjustment range indicates the amount of change in the value of this parameter during this operation.
[0066] Will Combined into a state transition data pair, where To execute the first The new state vector is acquired after each step; multiple consecutive state transition data pairs constitute a complete ventilator adjustment trajectory. ;in, This is a complete ventilator adjustment trajectory, recording the entire adjustment process from the initial state to the final state; This represents the total number of operation steps in this adjustment process.
[0067] The preprocessed data was paired according to adjacent operation steps in the time-series operation data sequence, including three channels: airway pressure, flow rate, and tidal volume, with 12,000 sampling points in each channel. To monitor parameter vectors (including SpO2, EtCO2, measured respiratory rate, etc.). To set the parameter vector (including PEEP, FiO2, tidal volume settings, etc.). This is a vector of patient physiological indicators (including dynamic compliance, airway resistance, etc., completed in step S4). Parameter type codes include PEEP (coded as 1), FiO2 (coded as 2), tidal volume (coded as 3), and inspiratory pressure support (coded as 4). For example, the action code for step 1 in this embodiment is... This indicates that PEEP will be increased by 2 cmH2O.
[0068] This embodiment generates a total of 5 state transition data pairs, forming a complete ventilator adjustment trajectory, where n=5. This is the initial state before adjustment. This is the final state after all adjustments are completed.
[0069] Before executing the structured output, medical staff review all collected data through the data verification interface. The interface displays respiratory waveform data, monitoring parameters, setting parameters, and feedback labels at each time point in a timeline format. The medical staff finds that the evaluation in step 2 may need modification (because although SpO2 did not change significantly, the decrease in FiO2 itself is positive in ARDS protective ventilation strategies), and changes its feedback label from 0 to +1. The system automatically regenerates the corresponding state transition data pairs based on the correction result.
[0070] Finally, the system structurally encapsulates all state transition data pairs, their associated action information, and quantized feedback labels according to a preset JSON data pattern. Each state transition data pair is encapsulated as an independent training sample, containing a complete state vector, action vector, reward value, and next state vector. All training samples form a structured training dataset, which is stored in a database for easy access during subsequent AI model training.
[0071] Before step S6, a data verification and annotation correction step is also included: the respiratory waveform data, monitoring parameter data, setting parameter data and corresponding quantitative feedback labels at each time point are displayed in a visual interface; medical staff can correct the quantitative feedback labels at any time point, and add, delete or reorder the time points in the data acquisition sequence; after the correction is completed, the state transition data pair construction step of step S5 is re-executed according to the correction result.
[0072] We continuously collected ventilator adjustment data for 6 months in the intensive care unit of a tertiary-level hospital, obtaining a total of 1,200 complete ventilator adjustment trajectory data, covering 8 common clinical scenarios such as ARDS, acute exacerbation of COPD, and postoperative respiratory support.
[0073] The obtained structured training dataset was used to train a ventilator parameter recommendation model based on a deep Q-network (DQN). In the training data, the state vector served as the network input, the action vector served as the output action space, and the quantized feedback label served as the reward signal. After 500 rounds of training, the model achieved a parameter recommendation accuracy of 82.3% on the test set, meaning that the proportion of adjustment directions recommended by the model that were consistent with the actual adjustment directions of experienced medical personnel was 82.3%.
[0074] Furthermore, the same dataset was used to train a temporal decision model based on the Transformer architecture. This model takes state transition trajectories as input sequences and can automatically generate adjustment action suggestions based on the current state and expected reward level. Test results show that the model achieves an average cumulative reward of 4.2 (out of 1, based on the number of adjustment steps) in a simulated environment, which is superior to the baseline DQN model's 3.8.
[0075] The above verification results show that the structured training dataset collected by the method described in this invention can effectively support the training needs of various artificial intelligence algorithms, verifying the practical value of the data processing method.
[0076] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific 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 ventilator data processing method, characterized in that, The method includes the following steps: Step S1: Collect multimodal raw data output by the ventilator in real time through the data transmission interface of the ventilator. The multimodal raw data includes respiratory waveform data, monitoring parameter data, and setting parameter data. Step S2: In response to the step-by-step adjustment operation of the ventilator by medical staff, data acquisition is triggered after each adjustment operation is performed, the temporal correspondence between each adjustment operation and the acquired data is established, and a temporal operation data sequence with operation timestamp and operation step identifier is generated. Step S3: Receive the effect evaluation information of the respiratory parameter changes after each adjustment operation from medical staff, convert the effect evaluation information into quantitative feedback labels, and associate the quantitative feedback labels with the time sequence operation data of the corresponding step. Step S4: Preprocess the multimodal raw data. The preprocessing includes: transmission error detection and correction, data segment time sequence reconstruction, missing parameter calculation and completion, and waveform reconstruction. Step S5: Pair the preprocessed data according to adjacent operation steps in the time-series operation data sequence to construct state transition data pairs. Each state transition data pair is associated with corresponding operation action information and quantization feedback label. Step S6: The state transition data pairs and their associated operation action information and quantization feedback labels are encapsulated in a structured manner according to a preset data pattern, and output as a structured training dataset for training artificial intelligence models.
2. The ventilator data processing method according to claim 1, characterized in that, In the multimodal raw data: The respiratory waveform data includes airway pressure-time waveform data, airflow velocity-time waveform data, and tidal volume-time waveform data; wherein, the airway pressure-time waveform data is a waveform signal reflecting the continuous change of airway pressure in the patient's breathing circuit over time, the airflow velocity-time waveform data is a waveform signal reflecting the change of gas flow velocity over time during inspiration and expiration, and the tidal volume-time waveform data is a waveform signal reflecting the cumulative change of the volume of gas entering and leaving the lungs over time in a single respiratory cycle; The monitoring parameters include blood oxygen saturation, end-tidal carbon dioxide partial pressure, measured respiratory rate, measured tidal volume, measured minute ventilation, peak airway pressure, and airway plateau pressure. The setting parameters include breathing mode, set respiratory rate, set tidal volume, positive end-expiratory pressure, inhaled oxygen concentration, inspiratory pressure support level, and inspiratory-to-expiratory ratio.
3. The ventilator data processing method according to claim 2, characterized in that, The time-series reassembly of the data segments includes: The discrete data segments output by the ventilator at first preset time intervals are sequentially concatenated along the time axis to form a continuous data segment covering a second preset time window. The concatenation relationship is represented as follows: ; in, The number of data segments required for splicing represents the number of discrete data segments contained within a complete time window; The first preset time interval is the time interval between two consecutive data outputs from the ventilator, ranging from 3 to 10 seconds. The second preset time window is the total time length covered by the continuous data segments after splicing, and the value ranges from 30 to 120 seconds. During the splicing process, boundary consistency is checked at the junctions of adjacent data segments. When a value jump at the junction is detected to exceed a preset threshold, interpolation smoothing is used to eliminate splicing artifacts.
4. The ventilator data processing method according to claim 3, characterized in that, The missing parameter calculation and completion includes: Determine the target parameters that the ventilator does not directly output in the current operating mode of the ventilator. Based on the associated parameters collected in the original multimodal data, the value of the target parameter is calculated using a respiratory mechanics relationship model; The respiratory mechanics model includes: Calculate respiratory rate based on respiratory cycle: ;in, The calculated respiratory rate is expressed in breaths per minute, representing the number of times the patient breathes per minute. The duration of a single respiratory cycle, expressed in seconds, represents the time elapsed from the start of one inhalation to the start of the next inhalation. Dynamic lung compliance is calculated based on tidal volume and pressure difference: ;in, Dynamic lung compliance reflects the elastic expansion capacity of lung tissue under dynamic respiratory conditions; Tidal volume is the volume of gas that enters the lungs during a single respiratory cycle. Peak airway pressure, which is the highest pressure value in the airway during the inspiratory phase; Positive end-expiratory pressure (PEEP) is the positive pressure maintained in the airway at the end of expiration. Calculate airway resistance based on inspiratory flow rate and pressure difference: ;in, Airway resistance reflects the magnitude of the frictional resistance encountered by gas as it passes through the respiratory tract; The plateau pressure is the static pressure in the airway at the end of inspiration when the patient holds their breath, reflecting the pressure within the alveoli. This is the peak inlet velocity, which is the maximum inlet velocity of the gas in the inlet phase.
5. The ventilator data processing method according to claim 4, characterized in that, The continuous waveform reconstruction of discrete numerical sequences includes: extracting airway pressure, airflow velocity, and tidal volume sample values stored in numerical list form from the multimodal raw data; determining the sampling time corresponding to each sample value according to the transmission protocol of the data transmission interface; and reconstructing the discrete sampling points into a continuous waveform curve using a piecewise linear interpolation method with the sampling time as the horizontal axis and the sample value as the vertical axis. The interpolation expression is as follows: ;in, For any time The waveform signal value obtained by interpolation at that location; and For two adjacent sampling times, represent the th sampling time, respectively. The and the first The time coordinates corresponding to each sampling point; and At the sampling time respectively and The actual signal value collected at the location.
6. The ventilator data processing method according to claim 5, characterized in that, After piecewise linear interpolation reconstruction, the process further includes: bandpass filtering of the reconstructed waveform curve. The bandpass filter is used to remove high-frequency noise interference and low-frequency baseline drift components from the signal, retain the effective signal components that match the frequency range of respiratory physiological activities, and output the filtered standard respiratory waveform data for use in subsequent data processing steps.
7. The ventilator data processing method according to claim 6, characterized in that, The method for generating the quantization feedback label includes: Provide medical personnel with an effectiveness evaluation option that includes at least three levels, namely, improvement, no change, and deterioration; The evaluation options selected by medical staff are mapped to preset numerical codes to form the quantitative feedback labels. The mapping relationship is as follows: ;in, For the first The quantitative feedback label value corresponding to the step adjustment operation; +1 indicates that the patient's respiratory parameters improved in a favorable direction after the step adjustment operation, 0 indicates that the patient's respiratory parameters did not change after the step adjustment operation, and -1 indicates that the patient's respiratory parameters deteriorated in an unfavorable direction after the step adjustment operation.
8. The ventilator data processing method according to claim 7, characterized in that, The quantized feedback label is also used as an immediate reward signal in reinforcement learning training, and the quantized feedback label value is normalized. The normalization expression is as follows: ;in, The normalized reward signal value, with a range of [-1, 1], is used to eliminate differences in different evaluation scales and adapt to the input requirements of reinforcement learning algorithms; For the first The original quantization feedback label value of the step operation; The maximum absolute value of the quantized feedback label value is used as a scaling factor for normalization.
9. The ventilator data processing method according to claim 8, characterized in that, The steps for constructing the state transition data pair include: ... The preprocessed data collected before the step adjustment operation is defined as a state vector: ;in, For the first The system state vector before the operation comprehensively describes the overall state of the ventilator and the patient at that moment. It is a waveform data matrix containing time-series waveform data of airway pressure, airflow velocity, and tidal volume; The monitoring parameter vector includes real-time monitoring values of blood oxygen saturation and end-tidal carbon dioxide partial pressure; The parameter vector is set to include configuration parameters for the current breathing mode of the ventilator and the set tidal volume; This is a vector of patient physiological indicators, containing physiological parameters completed by the respiratory mechanics model. The first The operation content of step adjustment is encoded as an action vector: ;in, For the first The step's motion vector describes the specific adjustment operations performed by medical personnel; This serves as a type identifier for the adjusted parameter, indicating which ventilator parameter was adjusted in this operation; The adjustment range indicates the amount of change in the value of this parameter during this operation; Will Combined into a state transition data pair, where To execute the first The new state vector is acquired after each step; multiple consecutive state transition data pairs constitute a complete ventilator adjustment trajectory. ;in, This is a complete ventilator adjustment trajectory, recording the entire adjustment process from the initial state to the final state; This represents the total number of operation steps in this adjustment process.
10. The ventilator data processing method according to claim 9, characterized in that, Before step S6, there are also data verification and annotation correction steps: The system displays respiratory waveform data, monitoring parameter data, setting parameter data, and corresponding quantitative feedback labels at various time points using a visual interface. It supports medical staff in modifying quantitative feedback labels at any time point, as well as adding, deleting, or reordering time points in the data collection sequence; After the correction is completed, the state transition data pair construction step of step S5 is re-executed based on the correction result.