A respiratory state induced aerosolization scheduling method based on causal reasoning

Abstract

The application discloses a kind of based on causal inference's respiratory state evoked atomization scheduling method, comprising the following steps: S1, establish causal atlas;S2, execute pattern recognition, generate state marker sequence;S3, input to improved Autoformer model, determine to execute atomization evoked scheduling operation or terminate this evoked procedure;S4, generate atomization trigger instruction, record atomization trigger time point, atomization dose and execution parameter, and generate evoked event entry;S5, collect the respiratory state feedback signal after atomization trigger, determine evoked success event or evoked failure event;S6, by causal inference stack structure executes causal structure update and attribution type classification operation, generates strategy failure label or physiological non-response label;S7, construct feedback memory set, form self-adapting evolution's atomization scheduling mechanism.The application improves the real-time responsiveness of atomization evoked, decision accuracy and individual adaptability.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical scheduling and control technology, and in particular to a respiratory state-induced nebulization scheduling method based on causal reasoning. Background Technology

[0002] With the continuous increase in the number of patients with chronic respiratory diseases and the widespread use of nebulizer therapy in clinical and home care, the demand for intelligent nebulizer management based on the patient's real-time physiological status is becoming increasingly prominent. Existing nebulizer scheduling methods mainly rely on manually setting time or simple preset threshold rules to control the nebulizer's start and stop, which generally suffers from the following problems in practical applications: The nebulization scheduling process cannot accurately perceive the evolution of a patient's respiratory status, especially in patients with irregular breathing or pathological features such as asthma or COPD. Fixed-time nebulization or periodic spraying not only fails to match the optimal timing for inhaled medication efficacy but may also exacerbate symptoms or lead to resource waste. Current methods lack causal analysis based on real-time respiratory status, making it difficult to determine whether respiratory changes truly trigger clinically necessary nebulization events, resulting in oversensitivity or delayed response in nebulization behavior. Some threshold-based models are easily affected by inter-individual differences, respiratory rhythm fluctuations, and environmental disturbances, making it difficult to achieve dynamic adaptation to triggering mechanisms. Furthermore, existing solutions generally do not consider the role of feedback information in the scheduling mechanism; the post-nebulization state is not effectively monitored and utilized, and the nebulization strategy lacks closed-loop optimization capabilities, resulting in low overall system control capacity and delayed response, thus hindering the development and promotion of intelligent nebulization scheduling.

[0003] Therefore, how to provide a respiratory state-induced nebulization scheduling method based on causal reasoning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a respiratory state-induced nebulization scheduling method based on causal reasoning. This invention fully utilizes state feature vector construction, feedback feature extraction, and causal reasoning stack structure. It describes in detail how, based on the collection of inspiratory flow rate signals, tidal volume signals, and rhythm change signals, a feedback feature sequence group is constructed to determine successful and failed induction events. Furthermore, based on pre-induction state features, induction behavior parameters, and feedback response features, an attribution sample data sequence is constructed. Attribution labels are generated using causal input tensors and attribution candidate sets. This method has the advantages of reasonable induction timing scheduling, accurate induction response discrimination, and clear identification of failure causes.

[0005] A respiratory state-induced nebulization scheduling method based on causal reasoning according to an embodiment of the present invention includes the following steps: S1. Collect inspiratory flow rate, tidal volume, thoracic expansion volume, lung sound spectrum and rhythm change signals, construct respiratory state data in time series format, and establish a causal map by combining historical evoked records; S2. Perform pattern recognition on the respiratory state data, label the respiratory phase, rhythm stability and airway resistance category, and generate a state label sequence; S3. Input the state marker sequence into the improved Autoformer model to generate prediction results within multiple preset time intervals; set the atomization trigger threshold condition to determine whether to execute the atomization induction scheduling operation or terminate the current induction process; S4. Based on the node states in the causal graph and the prediction results of the improved Autoformer model, generate atomization trigger instructions, record the atomization trigger time, atomization dose and execution parameters, and generate induced event entries. S5. Collect respiratory status feedback signals after nebulization triggering, construct induced response data sequence, determine tidal volume change value, peak inspiratory volume change amplitude and rhythm fluctuation range; set response threshold conditions to determine induced successful event or induced failed event; S6. Based on the induced failure event, construct an attribution sample data sequence, perform causal structure update and attribution type classification operations through the causal inference stack structure, and generate strategy failure label or physiological non-response label. S7. Construct a feedback memory set and dynamically update the atomization trigger threshold parameters, the induced scheduling strategy and the causal graph path weights to form an adaptive atomization scheduling mechanism.

[0006] Preferably, the step of collecting inspiratory flow rate, tidal volume, thoracic cavity expansion, lung sound spectrum and rhythm variation signals to construct time-series respiratory state data specifically involves: Set sampling period and data synchronization rules to collect inspiratory flow rate signals, tidal volume signals, chest expansion volume signals, lung sound spectrum signals and rhythm change signals; Multi-scale smoothing processing is performed on the inhalation velocity signal to extract the instantaneous rate of change and the periodic average value, and an inhalation velocity feature sequence is constructed. Perform respiratory cycle segmentation and peak-valley detection on the tidal volume signal, extract the extreme values ​​and variation amplitude of tidal volume in each cycle, and construct a tidal volume feature sequence; The morphology normalization and tension estimation operations were performed on the thoracic expansion signal to extract the expansion peak, expansion duration and tension fluctuation index, and to construct the thoracic expansion feature sequence. Short-time Fourier transform processing was performed on the lung sound spectrum signal to extract the power distribution and main frequency jump points of different frequency bands and construct the lung sound spectrum feature sequence; Perform periodic stability analysis and phase calibration on the rhythmic variation signal, extract the rhythm continuity value and the difference value between adjacent periods, and construct the rhythmic feature sequence; By aligning the inspiratory flow rate feature sequence, tidal volume feature sequence, chest expansion feature sequence, lung sound spectrum feature sequence, and rhythm feature sequence along a unified time axis, respiratory state data in time series format is constructed.

[0007] Preferably, the step of establishing a causal graph by combining historical induced records specifically includes: Based on the inspiratory flow rate, tidal volume, thoracic expansion volume, lung sound spectrum and rhythm changes in historical evoked records, a pre-evoked state vector sequence is constructed. The pre-induction state vector sequence is associated and labeled with the corresponding nebulization trigger command, nebulization dose, execution parameters and induced response results to generate an induced event control dataset. Based on the induced event control dataset, state change patterns are extracted, and node sets and causal connections are constructed. Define the directionality and weight generation method of causal connections, and construct a causal graph containing node sets and causal connections.

[0008] Preferably, S2 specifically includes: Phase recognition is performed on the inspiratory flow rate feature sequence to extract the positions of the inspiratory rise, inspiratory peak, and inspiratory fall, and to generate a respiratory phase label sequence; Perform window stationarity calculation and period difference evaluation on the rhythm feature sequence, extract rhythm continuity index and mutation index, and generate rhythm stability label sequence; Joint fluctuation analysis was performed on the tidal volume characteristic sequence and the thoracic expansion characteristic sequence. Combined with the high-frequency energy distribution in the lung audio spectrum, the airway impedance change characteristics were extracted to generate an airway resistance category label sequence. The respiratory phase label sequence, rhythm stability label sequence, and airway resistance category label sequence are aligned by time index to construct a state label sequence.

[0009] Preferably, the improved Autoformer model is as follows: The improved Autoformer model includes an input encoding module, a trend decomposition module, a residual attention module, and a prediction output module. The input encoding module receives the state label sequence, performs time index embedding on the respiratory phase label, rhythm stability label and airway resistance category label, and generates a time-series embedding vector. The trend decomposition module includes a dynamic boundary-aware sliding window structure, which identifies sequence boundary points based on the rate of change of inhalation velocity characteristics and the location of rhythmic continuity changes, constructs a variable-length sliding window group, performs local trend fitting operations within each variable-length sliding window, and generates trend sequences and residual sequences. The residual attention module performs sliding window attention computation and multi-head self-attention computation on the residual sequence to generate multi-scale residual representations; The prediction output module performs a fusion calculation on the trend sequence and the multi-scale residual representation, and outputs prediction results within multiple preset time intervals; the prediction results include the predicted peak value of inhalation velocity, the predicted value of rhythm change trend, and the induction feasibility score.

[0010] Preferably, the step of setting the atomization trigger threshold condition to determine whether to execute the atomization-induced scheduling operation or terminate the current induction process specifically includes: Set the inspiratory amplitude threshold, rhythm stability threshold, and evoked confidence threshold; The atomization triggering threshold conditions include: the predicted peak inspiratory flow rate is greater than the inspiratory amplitude threshold; the predicted rhythm change trend is within the interval corresponding to the rhythm stability threshold; and the induction feasibility score is greater than the induction confidence threshold. If the predicted peak inspiratory flow rate, the predicted rhythm change trend, and the induction feasibility score threshold all meet the corresponding threshold parameter conditions, then the atomization induction scheduling operation will be executed; if any prediction result does not meet the corresponding threshold parameter conditions, then the induction process will be terminated.

[0011] Preferably, S4 specifically comprises: Extract the path weight distribution and node activation value of the state nodes in the causal graph within the corresponding time interval, and construct the induced context state vector. The induced context state vector is jointly modeled with the predicted peak inhalation velocity, the predicted rhythm change trend, and the induced feasibility score output by the improved Autoformer model to generate an atomization trigger command vector. Extract the predicted peak inspiratory flow rate from the nebulization trigger command vector and set it as the nebulization trigger time point; combine the rhythm stability label and airway resistance category label in the induced context state vector to obtain the corresponding dose parameter value from the table and set it as the nebulization dose; combine the nebulization device model identifier and feedback channel status to generate execution parameters; encode the nebulization trigger time point, nebulization dose, and execution parameters to construct the nebulization trigger command structure; The atomization trigger command structure, prediction results, and context state vector are integrated to generate induced event entries.

[0012] Preferably, S5 specifically includes: Collect inspiratory flow rate, tidal volume, and rhythmic variation signals after atomization triggering to construct a set of feedback feature sequences; Local extremum detection and trend fitting are performed on the inspiratory flow rate signal to extract the feedback inspiratory peak value and fluctuation amplitude, and to construct the inspiratory peak value variation characteristics. The tidal volume signal is divided into periods and the change ratio is calculated. The difference and change ratio of tidal volume before and after induction are extracted to construct the tidal volume change characteristics. Short-time stability assessment and frequency domain differential analysis are performed on the rhythmic variation signal to extract the rhythmic fluctuation range and the degree of continuity degradation, and to construct rhythmic fluctuation characteristics; Set thresholds for peak inspiratory variation, tidal volume change rate, and rhythm fluctuation amplitude to construct response threshold conditions. Match the peak inspiratory variation feature, tidal volume change feature, and rhythm fluctuation feature with the response threshold conditions respectively. If the response threshold conditions are met, record it as a successful event. If any feature does not meet the corresponding response threshold conditions, record it as a failed event.

[0013] Preferably, S6 specifically includes: Collect pre-induction state characteristics, induction behavior parameters and feedback response characteristics corresponding to induction failure events, and construct attribution sample data sequences; The attribution sample data sequence is input into the causal inference stack structure according to the time index, and a causal input tensor is generated based on the node state, path weight and triggering event entries. Based on the causal input tensor, the structure traversal and dependency path retrieval are performed to identify the offset path between the pre-induction state features and the feedback response features, and to update the path weights and node activation values ​​in the causal inference stack structure. The joint bias pattern of induced behavior parameters and pre-induced state features is extracted, an attribution candidate set is constructed, and an attribution type classification operation is performed to generate classification results. Bias related to the induced scheduling strategy in the classification results are marked as strategy failure labels, and biases related to abnormal physiological performance are marked as physiological non-response labels. Attribution labels include strategy failure labels and physiological non-response labels.

[0014] Preferably, S7 specifically includes: Collect the nebulization trigger time, nebulization dose, execution parameters, context state vector and prediction results from the induced event entries, extract the attribution tags, construct the feedback unit structure, write it into the feedback memory set, and establish a multi-dimensional association mapping according to the time index and event identifier; Based on the distribution of attribution labels recorded in the feedback memory set, the correlation pattern between the recurrence frequency of statistical strategy failure labels and the triggering parameters is analyzed, and a numerical adjustment operation is performed on the atomization trigger threshold. The distribution of associated state features of physiological non-response tags is statistically analyzed to identify state pattern clusters. Update operations are performed on the state adaptation rules in the induced scheduling strategy to adjust the scheduling activation range and time window settings. Analyze the state offset paths in the feedback memory set, and combine the path weight evolution trend output by the causal inference stack structure to update the weight values ​​of the corresponding paths in the causal graph. The update results of the atomization trigger threshold parameter, the adjustment results of the induced scheduling strategy, and the update records of the causal graph path weight are synchronously written into the feedback memory set to complete one feedback loop and form an adaptive evolutionary atomization scheduling mechanism.

[0015] The beneficial effects of this invention are: This invention addresses the issues of inaccurate triggering timing, ambiguous feedback on induction failure, and rigid strategy scheduling in respiratory-induced nebulization by constructing time-series respiratory state data and combining causal mapping with an improved Autoformer model. It generates a state-labeled sequence by collecting inspiratory flow rate, tidal volume, thoracic expansion volume, lung sound spectrum, and rhythm change signals. A multi-scale prediction path is constructed using a dynamic boundary-aware sliding window structure and residual attention mechanism, outputting predicted peak inspiratory flow rate, predicted rhythm change trend, and induction feasibility score. Furthermore, it sets inspiratory amplitude thresholds and rhythm stability thresholds in the nebulization scheduling decision. By jointly determining the induced confidence threshold, a nebulization trigger command is constructed. In the induced response determination, inspiratory peak change features, tidal volume change features, and rhythmic fluctuation features are extracted, and response threshold conditions are constructed to clearly distinguish between successful and failed induced events. Based on failed induced events, an attribution sample data sequence is constructed and input into a causal inference stack structure to complete causal structure updates and attribution type classification operations, generating strategy failure labels or physiological non-response labels. Finally, a feedback memory set is constructed to dynamically update the nebulization trigger threshold parameters, induced scheduling strategy, and causal graph path weights, forming an adaptive evolutionary nebulization scheduling mechanism. This method achieves accurate prediction of nebulization induction, clear attribution of failures, and dynamic strategy optimization, improving the intelligence level and intervention reliability of respiratory state-induced nebulization scheduling. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a respiratory state-induced nebulization scheduling method proposed in this invention; Figure 2 This is a schematic diagram of the structure of the improved Autoformer model proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figure 1-2 A respiratory state-induced nebulization scheduling method based on causal reasoning includes the following steps: S1. Collect inspiratory flow rate, tidal volume, thoracic expansion volume, lung sound spectrum and rhythm change signals, construct respiratory state data in time series format, and establish a causal map by combining historical evoked records; S2. Perform pattern recognition on the respiratory state data, label the respiratory phase, rhythm stability and airway resistance category, and generate a state label sequence; S3. Input the state marker sequence into the improved Autoformer model to generate prediction results within multiple preset time intervals; set the atomization trigger threshold condition to determine whether to execute the atomization induction scheduling operation or terminate the current induction process; S4. Based on the node states in the causal graph and the prediction results of the improved Autoformer model, generate atomization trigger instructions, record the atomization trigger time, atomization dose and execution parameters, and generate induced event entries. S5. Collect respiratory status feedback signals after nebulization triggering, construct induced response data sequence, determine tidal volume change value, peak inspiratory volume change amplitude and rhythm fluctuation range; set response threshold conditions to determine induced successful event or induced failed event; S6. Based on the induced failure event, construct an attribution sample data sequence, perform causal structure update and attribution type classification operations through the causal inference stack structure, and generate strategy failure label or physiological non-response label. S7. Construct a feedback memory set and dynamically update the atomization trigger threshold parameters, the induced scheduling strategy and the causal graph path weights to form an adaptive atomization scheduling mechanism.

[0019] This implementation method achieves a respiratory state-induced nebulization scheduling method with causal reasoning capabilities by sequentially executing respiratory state acquisition, pattern recognition, prediction generation, trigger command issuance, feedback judgment, causal attribution, and strategy update operations. First, it collects inspiratory flow rate, tidal volume, thoracic expansion volume, lung sound spectrum, and rhythm change signals to construct time-series respiratory state data. Combined with historical induction records, a causal graph is established, enabling unified modeling of multi-dimensional physiological parameters and structured expression of historical causal relationships, improving the interpretability of the scheduling process. Further, pattern recognition is performed on the respiratory state data to label respiratory phase, rhythm stability, and airway resistance category, accurately capturing current respiratory state characteristics and supporting subsequent state classification and response modeling. Further, the state-labeled sequence is input into an improved Autoformer model to generate prediction results within multiple time intervals, and nebulization trigger threshold conditions are set, enabling early identification of potential intervention windows and improving the foresight and response efficiency of nebulization scheduling. Simultaneously, combined with the causal graph... The node status and model prediction results generate nebulization trigger commands, record the nebulization trigger time, dosage, and execution parameters, and establish induced event entries to ensure standardized execution and process traceability of nebulization behavior. Furthermore, it collects respiratory status feedback signals after nebulization triggering, constructs an induced response data sequence, and judges tidal volume changes, peak inspiratory volume changes, and rhythm fluctuation ranges. Based on set response threshold conditions, it determines whether the induction was successful, thus establishing a complete closed-loop feedback chain. Based on induced failure events, it constructs an attribution sample data sequence, calls the causal inference stack structure to perform causal structure updates and attribution type classification operations, and generates strategy failure labels or physiological non-response labels. This system can identify the reasons for ineffective intervention and improve the model's causal correction capability. Finally, it constructs a feedback memory set to dynamically update the nebulization trigger threshold, induced scheduling strategy, and causal graph path weights, forming an adaptive evolutionary nebulization scheduling mechanism. This effectively enhances the adaptability and individualized response capability of the nebulization scheduling system under various respiratory states, achieving the goals of precise, intelligent, and real-time nebulization intervention.

[0020] In this embodiment, the acquisition of inspiratory flow rate, tidal volume, thoracic cavity expansion, lung sound spectrum and rhythm variation signals to construct time-series respiratory state data specifically involves: The sampling period and data synchronization rules were set to collect inspiratory flow rate signals, tidal volume signals, thoracic expansion volume signals, lung sound spectrum signals, and rhythm change signals. Multi-scale smoothing was performed on the inspiratory flow rate signals to extract instantaneous rate of change and periodic average values, constructing an inspiratory flow rate feature sequence. Respiratory cycle segmentation and peak-valley detection were performed on the tidal volume signals to extract the extreme values ​​and amplitudes of tidal volume within each cycle, constructing a tidal volume feature sequence. Morphological normalization and tension estimation were performed on the thoracic expansion volume signals to extract expansion peak values, expansion duration, and tension fluctuation indices, constructing a thoracic expansion feature sequence. Short-time Fourier transform processing was performed on the lung sound spectrum signals to extract power distribution and dominant frequency jump points in different frequency bands, constructing a lung sound spectrum feature sequence. Periodic stability analysis and phase calibration were performed on the rhythm change signals to extract rhythm continuity values ​​and differences between adjacent cycles, constructing a rhythm feature sequence. The inspiratory flow rate feature sequence, tidal volume feature sequence, thoracic expansion feature sequence, lung sound spectrum feature sequence, and rhythm feature sequence were aligned along a unified time axis to construct time-series formatted respiratory state data.

[0021] In this embodiment, the step of establishing a causal graph by combining historical evoked records specifically includes: Based on inspiratory flow rate, tidal volume, thoracic expansion volume, lung sound spectrum and rhythm changes in historical induced events, a pre-induced state vector sequence is constructed. The pre-induced state vector sequence is associated and labeled with the corresponding nebulization trigger command, nebulization dose, execution parameters and induced response results to generate an induced event control dataset. Based on the induced event control dataset, state change patterns are extracted, and node sets and causal connections are constructed. The directionality and weight value generation method of causal connections are set to construct a causal graph containing node sets and causal connections.

[0022] In this embodiment, S2 specifically refers to: Phase recognition was performed on the inspiratory flow rate feature sequence. By extracting the trends of the first and second derivatives of the flow rate, the starting point of the inspiratory rise, the peak point of inspiration, and the ending point of the inspiratory fall were identified, constructing the inspiratory phase time interval and generating corresponding respiratory phase label sequences. The rhythm feature sequence was divided into fixed time windows, and the standard deviation of fluctuation amplitude and cycle length within the window was calculated. The rhythm stability was judged based on a set threshold. Furthermore, the absolute value of the cycle difference between adjacent cycles was normalized, and rhythm continuity indicators and mutation indices were extracted to generate rhythm stability label sequences. A joint normalization operation was performed on the tidal volume feature sequence and the thoracic expansion feature sequence to construct a joint fluctuation ratio matrix and analyze its performance during the inspiratory rise. The relative rate of change between the peak and decline phases is used to extract the fluctuation asymmetry factor. Combined with the high-frequency energy distribution density in the lung audio spectrum features, the time offset of the energy center position within the high-frequency band is calculated. The airway impedance change trend is fitted using a multi-factor weighting method to generate an airway resistance category label sequence. The respiratory phase label sequence, rhythm stability label sequence, and airway resistance category label sequence are aligned by time index, and missing markers are filled in and timestamps are updated synchronously to construct a state label sequence. This state label sequence provides structured label support for subsequent prediction model inputs, possessing temporal consistency, feature completeness, and semantic clarity, thereby improving the nebulization scheduling model's ability to accurately characterize and respond to the current respiratory state.

[0023] In this embodiment, the improved Autoformer model specifically refers to: The improved Autoformer model includes an input encoding module, a trend decomposition module, a residual attention module, and a prediction output module; The input encoding module receives a sequence of state labels and performs time index embedding operations on respiratory phase labels, rhythm stability labels, and airway resistance category labels. The embedding operation includes constructing a corresponding category vector embedding table for each label, constructing a position vector for the label's time index in the sequence using a position encoding function, and concatenating the category vector and position vector to generate a temporal embedding vector. The temporal embedding vector serves as a unified input basis for trend modeling and residual analysis. This structure enhances the ability to understand the temporal changes of multiple types of physiological states. The trend decomposition module includes a dynamic boundary-aware sliding window structure. This structure uses variable-length sliding windows instead of fixed windows in the time series, dynamically adjusting the window position and length through a boundary-aware mechanism. The boundary-aware mechanism is built based on the rate of change of inspiratory velocity characteristics and the location of rhythmic continuity changes. The rate of change uses the sliding derivative method to calculate the local slope, and the location of rhythmic continuity changes is identified using a mutation point detection function based on the rhythm stability label. A group of variable-length sliding windows is constructed, and trend separation operations are performed within each window. The trend separation operation calls a quadratic polynomial fitting function to fit the data within the local window, outputting a trend sequence and a residual sequence. The residual sequence is the time series residual obtained by subtracting the fitted trend sequence from the original sequence, preserving abnormal fluctuations and short-term mutation features. This mechanism improves the model's responsiveness to physiological rhythm mutation points and the accuracy of trend modeling by dynamically adjusting the window length and fitting boundary. The residual attention module performs sliding window attention computation and multi-head self-attention computation on the residual sequence. Sliding window attention computation constructs a local weight distribution based on the residual data within a local window, enhancing the ability to focus on short-term abrupt changes. Multi-head self-attention computation involves inputting the residual sequence into multiple parallel attention heads, calculating the attention distribution across the entire sequence in each head, and then concatenating and transforming the outputs of all attention heads to generate a global residual representation. The local residual representation is then fused with the global residual representation to form a multi-scale residual representation. The multi-scale residual representation takes into account both short-term abrupt changes and long-term fluctuation trends, improving the modeling ability for non-stationary respiratory states. The prediction output module receives the trend sequence and multi-scale residual representation, constructs a joint representation vector, and inputs it into a fully connected mapping network. The output results include the predicted peak inspiratory flow rate, the predicted rhythm change trend, and the induction feasibility score. The predicted rhythm change trend is generated by performing point prediction and sliding smoothing on the rhythm change curve within a future time period. The induction feasibility score is implemented by constructing a feasibility scoring function. The scoring function takes the smoothness of rhythm change, respiratory phase stability, and the length of the time interval in which the predicted inspiratory peak is greater than a preset threshold as inputs, and calls a normalized weighted scoring function for scoring. The weights of each parameter in the weighted scoring function are obtained by minimizing the scoring error objective function through the training process of historical respiratory induction samples. This score represents the physiological feasibility of inducing external intervention in the current respiratory state, and the higher the value, the more suitable the induction timing.

[0024] This implementation improves the ability to accurately express trends and abrupt changes in complex respiratory states through an improved Autoformer model, enabling multi-objective joint prediction of inspiratory flow rate, rhythm stability, and induction feasibility, and significantly improving the timeliness, stability, and individual adaptability of induced behaviors in the intelligent scheduling system.

[0025] In this embodiment, the step of setting the atomization trigger threshold condition to determine whether to execute the atomization-induced scheduling operation or terminate the current induction process specifically includes: Set the inspiratory amplitude threshold, rhythm stability threshold, and evoked confidence threshold; The atomization triggering threshold conditions include: the predicted peak inspiratory flow rate is greater than the inspiratory amplitude threshold; the predicted rhythm change trend is within the interval corresponding to the rhythm stability threshold; and the induction feasibility score is greater than the induction confidence threshold. If the predicted peak inspiratory flow rate, the predicted rhythm change trend, and the induction feasibility score threshold all meet the corresponding threshold parameter conditions, then the atomization induction scheduling operation will be executed; if any prediction result does not meet the corresponding threshold parameter conditions, then the induction process will be terminated.

[0026] In this embodiment, S4 specifically refers to: The path weight distribution and node activation values ​​within the time interval corresponding to the state nodes in the causal graph are extracted. The path weight distribution is constructed based on the causal strength value output by the structural attention structure, and the node activation value is the maximum response value of each node in the state vector within the time interval. The two are concatenated after being aligned by node number and time index to generate a context state tensor. The context state tensor is vectorized, and a positional encoding function is used to encode the location of path weight changes and the frequency of node activation changes. The context state vector is constructed by combining node attributes. The positional encoding function uses sine and cosine functions to alternately construct the encoding components, and forms an embedding structure through multi-dimensional periodic basis combination to enhance the model's ability to perceive time changes. The context state vector is concatenated with the predicted peak inspiratory flow rate, the predicted rhythm change trend, and the evoked feasibility score from the improved Autoformer model using a unified time index to generate the trigger input vector. The predicted peak inspiratory flow rate is the interpolated peak point obtained by curve fitting the model to the maximum value in the future inspiratory flow rate sequence. The predicted rhythm change trend is calculated by the slope fitting function of the trend sequence, and the slope value reflects the acceleration or deceleration trend of the respiratory rhythm. The evoked feasibility score is constructed based on the relative variability index between the trend sequence and the residual sequence, and is obtained by weighted summation of the state label stability coefficient and the airway resistance index score matrix, with the output range normalized to [0,1]. A fully connected mapping operation is performed on the trigger input vector to generate a nebulization trigger command vector. The nebulization trigger command vector structure includes a trigger time point index, a dose parameter label, a device channel code, and a synchronization control code. All parameter structures correspond one-to-one with the time index according to the position index mapping method. The peak inspiratory flow rate prediction time point is extracted from the nebulization trigger command vector and set as the nebulization trigger time point. Combining the rhythm stability label and airway resistance category label in the context state vector, the corresponding dose value in the preset dose parameter matrix is ​​obtained by looking up a table and set as the nebulization dose. The dose parameter matrix is ​​a three-dimensional tensor with dimensions of rhythm label type, resistance label type, and device channel type. Discrete label combinations are used as indexes to directly access the set values. The current nebulization device model identifier and feedback channel status code are obtained to construct an execution parameter field structure. The execution parameter fields include execution duration, nebulization rate level, and feedback synchronization parameter value. The nebulization triggering time, nebulization dose, and execution parameters are encoded to form a nebulization triggering instruction structure in a structure format. Each field in the nebulization triggering instruction structure is represented by a 32-bit fixed-length encoding method to ensure consistency with the communication protocol of the execution interface. By integrating the atomization trigger command structure, prediction results, and context state vector, induced event entries are constructed. These entries are indexed by timestamps and include the event trigger basis, target state label, prediction confidence level, and execution confirmation identifier. Based on these induced event entries, precise response control and traceable status recording of the atomization device can be achieved.

[0027] This implementation achieves an intelligent atomization control mechanism based on multi-dimensional state fusion, which improves the accuracy and response efficiency of atomization induction timing selection, reduces the risk of excessive induction or delayed response, and significantly enhances the system's adaptability to dynamic changes in respiratory state and induction effectiveness.

[0028] In this embodiment, S5 specifically refers to: Collect inspiratory flow rate, tidal volume, and rhythm change signals after nebulization triggering. Establish a time window based on the nebulization triggering time point in the nebulization triggering instruction structure. Extract the inspiratory cycle data segment, tidal volume data segment, and rhythm change data segment within the corresponding time window to construct a feedback feature sequence group. Local extremum detection is performed on the inspiratory flow rate signal to extract the peak inspiratory flow rate within the inspiratory cycle after triggering. The fluctuation amplitude is calculated by combining the peak inspiratory flow rate with the average flow rate within the inspiratory cycle. The fluctuation amplitude is defined as the difference between the maximum flow rate and the average flow rate. The peak inspiratory flow rate and the fluctuation amplitude are combined to form a feedback inspiratory flow feature vector. The inspiratory flow rate sequence is trend-fitted using cubic spline interpolation. The inspiratory flow trend offset is calculated by the mean square error index between the fitted curve and the original signal. The inspiratory flow trend offset and the feedback inspiratory flow feature vector are combined to form the inspiratory peak change feature. The tidal volume signal is divided into cycles based on the respiratory cycle markers in the rhythm change signal. Tidal volume sequences of equal length before and after triggering are divided. The average tidal volume before triggering and the average tidal volume after triggering are calculated respectively. The tidal volume difference is defined as the difference between the two. The tidal volume change ratio is defined as the ratio of the tidal volume difference to the average tidal volume before triggering. The tidal volume difference and the change ratio are concatenated to form the tidal volume change feature. Short-time stability assessment was performed on the rhythm change signal. The sliding window method was used to calculate the respiratory cycle variation coefficient index and frequency domain difference analysis was performed. The fast Fourier transform was used to obtain the rhythm main frequency distribution before and after triggering. The main frequency change amplitude and main frequency drift rate were calculated as rhythm fluctuation descriptors. The above rhythm fluctuation indexes were constructed as rhythm fluctuation features. A set of response threshold conditions is constructed by setting thresholds for peak inspiratory velocity variation, tidal volume change rate, and rhythm fluctuation amplitude. The peak inspiratory velocity variation characteristics are matched with the peak inspiratory velocity variation thresholds. If the increase in peak inspiratory velocity is greater than the threshold and the trend deviation is within the set range, the inspiratory characteristics are deemed to meet the requirements. The tidal volume variation characteristics are matched with the tidal volume change rate thresholds. If the change rate is greater than the set threshold, the tidal volume characteristics are deemed to meet the requirements. The rhythm fluctuation characteristics are matched with the rhythm fluctuation amplitude thresholds. If both the dominant frequency change amplitude and the drift rate are within the set range, the rhythm characteristics are deemed to meet the requirements. If all three types of features meet the corresponding response threshold conditions, it is recorded as a successful induction event; if any type of feature does not meet the response threshold conditions, it is recorded as a failed induction event. The classification of induction results is achieved by matching and judging the multi-dimensional feedback features with the response threshold conditions, thereby improving the feedback evaluability and strategy adaptability of the atomization induction strategy.

[0029] In this embodiment, S6 specifically refers to: Collect the pre-induction state features, induction behavior parameters, and feedback response features corresponding to the induction failure event, and construct an attribution sample data sequence according to the induction execution time order. The attribution sample data sequence includes an induction time index, a pre-induction state feature vector, a combination of induction behavior parameters, and a feedback response feature vector. The attribution sample data sequence is input into the causal inference stack structure according to the time index. The causal inference stack structure includes a node set, a path set, a state activation vector, and a path weight matrix. Based on the pre-induction state feature vector and feedback response feature vector in the attribution sample data, a causal node state encoding vector is constructed. The encoding vector is generated by normalization and difference fitting function. The difference fitting function takes the mean square error between the pre-induction state feature and the feedback response feature as the offset function form. A causal input tensor is constructed based on the combination of causal node state encoding vectors and induced behavior parameters; the causal input tensor includes state node indexes, path connection relationships, node initial activation values, and parameter deviation indices. The causal input tensor is input into the causal inference stack structure, and the path structure traversal operation and dependency path retrieval operation are performed to identify the set of paths where there is a significant deviation between the pre-induction state features and the feedback response features. The path offset value is calculated using the path offset intensity function. The path offset value is obtained by weighting the change in activation value of each node in the causal path. The change in activation value is taken as the difference between the pre-induced state and the feedback response in the attribution sample. The path weights in the causal inference stack structure are updated based on the path offset values. The update rules include: if the offset value is greater than a set offset threshold, the path weight is increased and the path is marked as an abnormal path; if the offset value is less than the offset threshold, the path weight remains unchanged. Extract the joint bias pattern between induced behavioral parameters and pre-induced state features from the attribution samples. The joint bias pattern is used to calculate the offset contribution rate of behavioral parameters to induced failure through a difference mapping function. A candidate attribution set is constructed based on the induced behavior parameters and the offset contribution rate. The candidate attribution set is then input into the classification operation module. The classification operation module uses a rule mapping method based on the discrimination threshold to perform label judgment on each group of biases in the candidate attribution set and generate an attribution type classification result. Deviations identified in the classification results as having unreasonable scheduling rules or mismatched response timing are recorded as strategy failure labels; deviations identified in the classification results as failing to activate physiological responses such as abnormal inhalation, sudden drop in tidal volume, or rhythm disorder are recorded as physiological non-response labels. The attribution label set includes strategy failure labels and physiological non-response labels; the attribution results serve as the basis for feedback analysis of induced failure events, enabling multi-dimensional classification of the causes of failure and generation of scheduling optimization suggestions.

[0030] In this embodiment, S7 specifically refers to: Collect the nebulization trigger time, nebulization dose, execution parameters, context state vector and prediction results from the induced event entries, extract the attribution tags, construct the feedback unit structure, write it into the feedback memory set, and establish a multi-dimensional association mapping according to the time index and event identifier; Based on the distribution of attribution labels recorded in the feedback memory set, the correlation pattern between the recurrence frequency of statistical strategy failure labels and the triggering parameters is analyzed, and numerical adjustment operations are performed on the atomization trigger threshold. The distribution of associated state features of physiological non-response tags is statistically analyzed to identify state pattern clusters. Update operations are performed on the state adaptation rules in the induced scheduling strategy to adjust the scheduling activation range and time window settings. Analyze the state offset paths in the feedback memory set, and combine the path weight evolution trend output by the causal inference stack structure to update the weight values ​​of the corresponding paths in the causal graph. The update results of the atomization trigger threshold parameter, the adjustment results of the induced scheduling strategy, and the update records of the causal graph path weight are synchronously written into the feedback memory set to complete one feedback loop and form an adaptive evolutionary atomization scheduling mechanism.

[0031] Example 1: To verify the feasibility of this invention in practice, it was applied to the nebulization therapy management system of the respiratory department of a large tertiary hospital. The system covers 32 beds and receives an average of more than 130 patients per day. The patients are mainly chronic respiratory diseases such as COPD, asthma, and bronchiectasis. Conventional nebulization therapy has problems such as inaccurate timing of induction, delayed treatment feedback, and low resource allocation efficiency, which seriously restricts clinical efficiency and patient experience.

[0032] In practical deployment, the system integrates with the hospital's existing respiratory monitoring platform, using respiratory sensors to collect key physiological indicators such as respiratory rate, tidal volume, oxygenation index, and forced expiratory parameters in real time. It also combines medication history and individual response records to construct a causal chain structure for nebulization-induced responses. The system employs an improved GCIN model to identify abrupt changes in the patient's respiratory state, calculates causal strength values, generates nebulization-induced candidate sequences, and determines the final induction node by combining the confidence level of the target node and the causal strength threshold.

[0033] To evaluate the improvement of the present invention in terms of the accuracy of nebulization timing and resource allocation efficiency, three consecutive months of clinical operation data were selected. Comparative data on key indicators such as response delay, resource utilization, average patient treatment time, and symptom relief time after intervention were statistically analyzed between the traditional approach and the present invention. The results are summarized in Table 1. Table 1. Comparison of key operating indicators between the traditional method and the method of this invention.

[0034] As shown in Table 1, the method of this invention exhibits significant advantages in several key performance indicators. Response latency was reduced from 52.3 seconds to 18.9 seconds, a decrease of 63.9%, significantly improving the immediacy of nebulization intervention after sudden changes in respiratory status; resource utilization increased from 66.2% to 87.5%, optimizing the scheduling mechanism of nebulization equipment and human resources; the average treatment time was shortened by approximately 9 minutes, reducing patient waiting time and operation time; symptom relief time was also significantly shortened by 35.5%, indicating that the intelligent selection of induction timing significantly improved the treatment effect.

[0035] Furthermore, to further evaluate the suitability of the method of the present invention in different types of patients, a sample of 60 patients who received nebulization treatment was randomly selected. After being divided into disease categories, the treatment efficacy rates of the traditional regimen and the regimen of the present invention were compared and summarized as shown in Table 2: Table 2 Comparison of treatment effectiveness rates for different disease types

[0036] As can be seen from the results in Table 2, the method of the present invention showed higher treatment efficacy in different types of respiratory diseases, with an efficacy increase of 20% in patients with bronchiectasis, 16.8% in patients with asthma, and 18.8% in patients with COPD. This demonstrates the robustness and adaptability of the causal chain modeling mechanism and confidence assessment strategy of the present invention across a wide range of diseases.

[0037] This embodiment constructs a respiratory state-induced nebulization scheduling method based on causal reasoning, which enables rapid identification of the optimal intervention node after a sudden change in the patient's respiratory state and intelligent activation of the nebulization scheduling process. This not only significantly improves response speed and resource utilization efficiency, but also enhances clinical treatment effectiveness and patient experience, and has good promotional value and clinical application prospects.

[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A respiratory state-induced nebulization scheduling method based on causal reasoning, characterized in that, Includes the following steps: S1. Collect inspiratory flow rate, tidal volume, thoracic expansion volume, lung sound spectrum and rhythm change signals, construct respiratory state data in time series format, and establish a causal map by combining historical evoked records; S2. Perform pattern recognition on the respiratory state data, label the respiratory phase, rhythm stability and airway resistance category, and generate a state label sequence; S3. Input the state marker sequence into the improved Autoformer model to generate prediction results within multiple preset time intervals; set the atomization trigger threshold condition to determine whether to execute the atomization induction scheduling operation or terminate the current induction process; S4. Based on the node states in the causal graph and the prediction results of the improved Autoformer model, generate atomization trigger instructions, record the atomization trigger time, atomization dose and execution parameters, and generate induced event entries. S5. Collect respiratory status feedback signals after nebulization triggering, construct induced response data sequence, and determine tidal volume change value, peak inspiratory volume change amplitude and rhythm fluctuation range. Set response threshold conditions to determine whether to trigger a successful event or a failed event; S6. Based on the induced failure event, construct an attribution sample data sequence, perform causal structure update and attribution type classification operations through the causal inference stack structure, and generate strategy failure label or physiological non-response label. S7. Construct a feedback memory set and dynamically update the atomization trigger threshold parameters, the induced scheduling strategy and the causal graph path weights to form an adaptive atomization scheduling mechanism.

2. The respiratory state-induced nebulization scheduling method based on causal reasoning according to claim 1, characterized in that, The process involves collecting inspiratory flow rate, tidal volume, thoracic expansion, and lung sound spectrum and rhythm changes to construct time-series respiratory status data, specifically: Set sampling period and data synchronization rules to collect inspiratory flow rate signals, tidal volume signals, chest expansion volume signals, lung sound spectrum signals and rhythm change signals; Multi-scale smoothing processing is performed on the inhalation velocity signal to extract the instantaneous rate of change and the periodic average value, and an inhalation velocity feature sequence is constructed. Perform respiratory cycle segmentation and peak-valley detection on the tidal volume signal, extract the extreme values ​​and variation amplitude of tidal volume in each cycle, and construct a tidal volume feature sequence; The morphology normalization and tension estimation operations were performed on the thoracic expansion signal to extract the expansion peak, expansion duration and tension fluctuation index, and to construct the thoracic expansion feature sequence. Short-time Fourier transform processing was performed on the lung sound spectrum signal to extract the power distribution and main frequency jump points of different frequency bands and construct the lung sound spectrum feature sequence; Perform periodic stability analysis and phase calibration on the rhythmic variation signal, extract the rhythm continuity value and the difference value between adjacent periods, and construct the rhythmic feature sequence; By aligning the inspiratory flow rate feature sequence, tidal volume feature sequence, chest expansion feature sequence, lung sound spectrum feature sequence, and rhythm feature sequence along a unified time axis, respiratory state data in time series format is constructed.

3. The respiratory state-induced nebulization scheduling method based on causal reasoning according to claim 1, characterized in that, The establishment of a causal graph by combining historical triggering records specifically includes: Based on the inspiratory flow rate, tidal volume, thoracic expansion volume, lung sound spectrum and rhythm changes in historical evoked records, a pre-evoked state vector sequence is constructed. The pre-induction state vector sequence is associated and labeled with the corresponding nebulization trigger command, nebulization dose, execution parameters and induced response results to generate an induced event control dataset. Based on the induced event control dataset, state change patterns are extracted, and node sets and causal connections are constructed. Define the directionality and weight generation method of causal connections, and construct a causal graph containing node sets and causal connections.

4. The respiratory state-induced nebulization scheduling method based on causal reasoning according to claim 1, characterized in that, Specifically, S2 is: Phase recognition is performed on the inspiratory flow rate feature sequence to extract the positions of the inspiratory rise, inspiratory peak, and inspiratory fall, and to generate a respiratory phase label sequence; Perform window stationarity calculation and period difference evaluation on the rhythm feature sequence, extract rhythm continuity index and mutation index, and generate rhythm stability label sequence; Joint fluctuation analysis was performed on the tidal volume characteristic sequence and the thoracic expansion characteristic sequence. Combined with the high-frequency energy distribution in the lung audio spectrum, the airway impedance change characteristics were extracted to generate an airway resistance category label sequence. The respiratory phase label sequence, rhythm stability label sequence, and airway resistance category label sequence are aligned by time index to construct a state label sequence.

5. The respiratory state-induced nebulization scheduling method based on causal reasoning according to claim 1, characterized in that, The improved Autoformer model is specifically as follows: The improved Autoformer model includes an input encoding module, a trend decomposition module, a residual attention module, and a prediction output module. The input encoding module receives the state label sequence, performs time index embedding on the respiratory phase label, rhythm stability label and airway resistance category label, and generates a time-series embedding vector. The trend decomposition module includes a dynamic boundary-aware sliding window structure, which identifies sequence boundary points based on the rate of change of inhalation velocity characteristics and the location of rhythmic continuity changes, constructs a variable-length sliding window group, performs local trend fitting operations within each variable-length sliding window, and generates trend sequences and residual sequences. The residual attention module performs sliding window attention computation and multi-head self-attention computation on the residual sequence to generate multi-scale residual representations; The prediction output module performs a fusion calculation on the trend sequence and the multi-scale residual representation, and outputs prediction results within multiple preset time intervals; the prediction results include the predicted peak value of inhalation velocity, the predicted value of rhythm change trend, and the induction feasibility score.

6. The respiratory state-induced nebulization scheduling method based on causal reasoning according to claim 1, characterized in that, The setting of the atomization trigger threshold condition determines whether to execute the atomization induction scheduling operation or terminate the current induction process, specifically as follows: Set the inspiratory amplitude threshold, rhythm stability threshold, and evoked confidence threshold; The atomization triggering threshold conditions include: the predicted peak inspiratory flow rate is greater than the inspiratory amplitude threshold; the predicted rhythm change trend is within the interval corresponding to the rhythm stability threshold; and the induction feasibility score is greater than the induction confidence threshold. If the predicted peak inspiratory flow rate, the predicted rhythm change trend, and the induction feasibility score threshold all meet the corresponding threshold parameter conditions, then the atomization induction scheduling operation will be executed; if any prediction result does not meet the corresponding threshold parameter conditions, then the induction process will be terminated.

7. The respiratory state-induced nebulization scheduling method based on causal reasoning according to claim 1, characterized in that, Specifically, S4 is: Extract the path weight distribution and node activation value of the state nodes in the causal graph within the corresponding time interval, and construct the induced context state vector. The induced context state vector is jointly modeled with the predicted peak inhalation velocity, the predicted rhythm change trend, and the induced feasibility score output by the improved Autoformer model to generate an atomization trigger command vector. Extract the predicted peak inspiratory flow rate from the nebulization trigger command vector and set it as the nebulization trigger time point; combine the rhythm stability label and airway resistance category label in the induced context state vector to obtain the corresponding dose parameter value from the table and set it as the nebulization dose; combine the nebulization device model identifier and feedback channel status to generate execution parameters; encode the nebulization trigger time point, nebulization dose, and execution parameters to construct the nebulization trigger command structure; The atomization trigger command structure, prediction results, and context state vector are integrated to generate induced event entries.

8. The respiratory state-induced nebulization scheduling method based on causal reasoning according to claim 1, characterized in that, Specifically, S5 is: Collect inspiratory flow rate, tidal volume, and rhythmic variation signals after atomization triggering to construct a set of feedback feature sequences; Local extremum detection and trend fitting are performed on the inspiratory flow rate signal to extract the feedback inspiratory peak value and fluctuation amplitude, and to construct the inspiratory peak value variation characteristics. The tidal volume signal is divided into periods and the change ratio is calculated. The difference and change ratio of tidal volume before and after induction are extracted to construct the tidal volume change characteristics. Short-time stability assessment and frequency domain differential analysis are performed on the rhythmic variation signal to extract the rhythmic fluctuation range and the degree of continuity degradation, and to construct rhythmic fluctuation characteristics; Set thresholds for peak inspiratory variation, tidal volume change rate, and rhythm fluctuation amplitude to construct response threshold conditions; match the peak inspiratory variation characteristics, tidal volume change characteristics, and rhythm fluctuation characteristics with the response threshold conditions respectively; if the response threshold conditions are met, record it as a successful event. If any feature does not meet the corresponding response threshold condition, it is recorded as an event that triggers failure.

9. The respiratory state-induced nebulization scheduling method based on causal reasoning according to claim 1, characterized in that, Specifically, S6 is: Collect pre-induction state characteristics, induction behavior parameters and feedback response characteristics corresponding to induction failure events, and construct attribution sample data sequences; The attribution sample data sequence is input into the causal inference stack structure according to the time index, and a causal input tensor is generated based on the node state, path weight and triggering event entries. Based on the causal input tensor, the structure traversal and dependency path retrieval are performed to identify the offset path between the pre-induction state features and the feedback response features, and to update the path weights and node activation values ​​in the causal inference stack structure. The joint bias pattern of induced behavior parameters and pre-induced state features is extracted, an attribution candidate set is constructed, and an attribution type classification operation is performed to generate classification results. Bias related to the induced scheduling strategy in the classification results are marked as strategy failure labels, and biases related to abnormal physiological performance are marked as physiological non-response labels. Attribution labels include strategy failure labels and physiological non-response labels.

10. The respiratory state-induced nebulization scheduling method based on causal reasoning according to claim 1, characterized in that, Specifically, S7 is: Collect the nebulization trigger time, nebulization dose, execution parameters, context state vector and prediction results from the induced event entries, extract the attribution tags, construct the feedback unit structure, write it into the feedback memory set, and establish a multi-dimensional association mapping according to the time index and event identifier; Based on the distribution of attribution labels recorded in the feedback memory set, the correlation pattern between the recurrence frequency of statistical strategy failure labels and the triggering parameters is analyzed, and numerical adjustment operations are performed on the atomization trigger threshold. The distribution of associated state features of physiological non-response tags is statistically analyzed to identify state pattern clusters. Update operations are performed on the state adaptation rules in the induced scheduling strategy to adjust the scheduling activation range and time window settings. Analyze the state offset paths in the feedback memory set, and combine the path weight evolution trend output by the causal inference stack structure to update the weight values ​​of the corresponding paths in the causal graph. The update results of the atomization trigger threshold parameter, the adjustment results of the induced scheduling strategy, and the update records of the causal graph path weight are synchronously written into the feedback memory set to complete one feedback loop and form an adaptive evolutionary atomization scheduling mechanism.

Images