Intelligent analysis method for physiological signals in cardiac intervention based on multi-source data fusion

Abstract

The present application relates to the technical field of task management, and more particularly to a cardiac intervention physiological signal intelligent analysis method based on multi-source data fusion. The method comprises the following steps: acquiring multi-source cardiac data, and marking the multi-source cardiac data with intervention operation events to obtain operation marked data; constructing multi-source features according to the operation marked data to obtain multi-source feature data; acquiring operation process data; constructing a stage flowchart according to the operation process data and the multi-source feature data to obtain stage flowchart data; and calculating load budget according to the stage flowchart data to obtain stage load data for cardiac intervention physiological signal reminding work. The present application effectively improves the flow management accuracy and system response capability of the operation process by structurally associating multi-source physiological signals with intraoperative operation processes, constructing a quantifiable stage load budget model, and combining a graph structure propagation mechanism.

Description

Technical Field

[0001] This invention relates to the field of task management technology, and in particular to an intelligent analysis method for cardiac interventional physiological signals based on multi-source data fusion. Background Technology

[0002] With the increasing sophistication and standardization of interventional cardiac procedures, intraoperative physiological load management has become a crucial aspect of surgical safety control. Traditional interventional cardiac procedures rely on the surgeon's experience for operational control, and operative actions (such as guidewire advancement and balloon dilation) are often not actively recorded by the equipment, resulting in a lack of precise mapping between intraoperative data and operative events. Although some studies have attempted to monitor intraoperative status using physiological signals such as electrocardiogram, pressure wave, and blood oxygenation, the lack of key technologies such as multimodal data fusion, automated operative event labeling, and process-level budget modeling makes it difficult to meet the needs for real-time intraoperative load perception and alerts. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes an intelligent analysis method for cardiac interventional physiological signals based on multi-source data fusion, thereby resolving at least one of the aforementioned technical issues.

[0004] This application provides an intelligent analysis method for cardiac interventional physiological signals based on multi-source data fusion, comprising the following steps:

[0005] S1: Acquire multi-source cardiac data and mark interventional operation events on the multi-source cardiac data to obtain operation mark data;

[0006] S2: Construct multi-source features based on the operation label data to obtain multi-source feature data;

[0007] S3: Obtain operation process data; construct stage flowcharts based on operation process data and multi-source feature data to obtain stage flowchart data;

[0008] S4: Calculate the load budget based on the phase flowchart data to obtain phase load data for cardiac intervention physiological signal alerts.

[0009] This invention constructs a graph-based interventional procedure model, achieving high-precision identification and dynamic annotation of weak perturbation features in multi-source cardiac physiological signals, thus solving the problem of inability to directly collect or vaguely label intraoperative events. By fusing multimodal data such as ECG, blood pressure, blood oxygenation, and catheter movement, and combining it with operation labeling data to construct cross-modal features, the accuracy of node physiological state representation is effectively improved. Based on the stage flowchart execution budget propagation and load integral analysis, high-load accumulation areas can be dynamically identified, outputting load reminder information that combines temporality and structure, which helps to improve the real-time performance and stage granularity of intelligent perception of intraoperative instrument use.

[0010] Preferably, acquiring multi-source cardiac data specifically involves:

[0011] ECG signal data is obtained by acquiring ECG signals through a preset ECG signal acquisition module.

[0012] Arterial pressure data is obtained by collecting arterial pressure through a preset arterial pressure sensor module;

[0013] Blood oxygen data is obtained by monitoring blood oxygen using a pre-set blood oxygen monitoring model;

[0014] The catheter position is acquired using a catheter position signal acquisition module to obtain catheter position data;

[0015] By integrating electrocardiogram signal data, arterial pressure data, blood oxygen data, and catheter position data, multi-source cardiac data is obtained.

[0016] This invention achieves parallel acquisition and standardized integration of multi-source cardiac physiological data during surgery by setting up dedicated modules for acquiring ECG, arterial pressure, blood oxygen, and catheter position signals. This effectively avoids the limitations of single-modal signals in terms of temporal integrity and response specificity. After aligning the various modal signals along the time axis, a unified data input structure is formed, providing a stable data foundation for weak perturbation identification, cross-modal coupling feature extraction, and stage state modeling. This enhances the system's sensitivity to operational behaviors and its real-time perception of physiological changes at the data level.

[0017] Preferably, the intervention operation event marker is specifically as follows:

[0018] Weak variation feature extraction is performed on multi-source cardiac data to obtain weak variation feature data;

[0019] Graph attention is used to identify weakly changing feature data to obtain feature-weighted data;

[0020] Operation event segment identification is performed based on feature-weighted data to obtain operation event segment data;

[0021] The operation event types are classified based on the operation event segment data to obtain operation tag data.

[0022] This invention extracts weak perturbation features from multi-source cardiac data and combines this with a graph attention mechanism to dynamically weight the importance of features across modalities, thereby improving the ability to identify interventional events under low-amplitude signal changes. Compared to traditional methods that rely on fixed thresholds or single-channel mutation detection, this invention can identify the actual occurrence interval of an interventional event segment based on multi-channel collaborative responses and achieve accurate event type classification through intra-segment feature aggregation. This process effectively reduces reliance on manual annotation, improves the automation level of event recognition, and enhances the ability to finely analyze physiological responses.

[0023] Preferably, the extraction of weak change features specifically includes:

[0024] The interval rate of change and QRS wave fluctuation time were calculated from the electrocardiogram signal data in the multi-source cardiac data to obtain the interval rate of change and QRS wave fluctuation time, respectively.

[0025] Waveform micro-variation rise time and peak change are extracted from arterial pressure data in multi-source cardiac data to obtain waveform micro-variation data and peak change data, respectively.

[0026] The blood oxygen fluctuation intensity data is obtained by calculating the mean instantaneous slope of blood oxygen data from multi-source cardiac data.

[0027] The catheter jitter peak value was calculated from the catheter position data in the multi-source cardiac data to obtain the catheter jitter peak value data;

[0028] By integrating the interval change rate, QRS wave fluctuation time, waveform micro-change data, peak change data, blood oxygen fluctuation intensity data, and catheter jitter peak data, weak change characteristic data are obtained.

[0029] This invention constructs a highly sensitive feature set capable of characterizing the perturbation effects of interventional procedures by calculating and extracting fine-grained parameters from multi-source cardiac signals. Specifically, the interval change rate and QRS oscillation time of the electrocardiogram (ECG) signal reflect instantaneous rhythm fluctuations; the waveform variations and peak values ​​of arterial pressure are used to capture hemodynamic responses; the mean slope of blood oxygen reflects the rate of change in oxygenation levels; and the peak value of catheter jitter reflects mechanical disturbances caused by device manipulation. Integrating these modal features helps to form a multi-faceted characterization of minor physiological disturbances caused by procedural actions, improving the detectability and robustness of procedural event identification in weak signal backgrounds.

[0030] Preferably, graph attention recognition specifically includes:

[0031] Feature maps are constructed based on weakly changing feature data to obtain feature map data;

[0032] Feature encoding is performed on the feature map data to obtain feature-encoded data;

[0033] Graph attention weights are calculated based on the feature encoding data to obtain feature graph attention data;

[0034] We obtain weighted feature data by weighting and aggregating the attention data from the feature maps.

[0035] This invention constructs a graph structure from weakly changing feature data, connecting time windows or modal nodes with edges to form a perturbation correlation network. A graph attention mechanism is then used to model the perturbation coupling relationships between different nodes. After strengthening node representation through feature encoding, the graph attention weights are calculated to adaptively enhance the saliency of key perturbation nodes, effectively highlighting spatiotemporal feature regions highly correlated with operational events. The system utilizes attention-weighted aggregation to generate feature-weighted data, which not only enhances the response to weak perturbations but also improves the efficiency of capturing non-local cooperative changes, providing a structurally enhanced input representation for event segment identification and improving the accuracy and anti-interference capability of event identification.

[0036] Preferably, the operation event segment identification specifically includes:

[0037] Threshold-based continuous detection is performed on the feature-weighted data to obtain preliminary event segment data;

[0038] The time period boundaries are refined based on the initial time period data to obtain secondary time period data;

[0039] The data for the operation event segment is obtained by deduplicating the data from the secondary time period.

[0040] This invention achieves preliminary identification of potential operational event time periods by performing threshold-based continuous detection on feature-weighted data, effectively avoiding misjudgments caused by point-like abrupt changes. Combined with time period boundary refinement, the start and end positions of events can be dynamically corrected based on changes in confidence gradients, thereby improving the accuracy of event segment boundaries and time sequence alignment. By cleaning and merging overlapping or redundant segments, the system constructs a clearly structured set of non-conflicting operational event segments, enhancing the automated identification of event intervals in continuous operation scenarios and providing a stable and boundary-accurate data foundation for event type classification and stage flow mapping.

[0041] Preferably, the operation event type classification is as follows:

[0042] Event segment feature data is obtained by aggregating event segment features based on the operation event segment data.

[0043] The event segment feature data is input into the preset operation event type classification model to obtain type confidence data;

[0044] The label confidence is evaluated based on the type confidence data to obtain the operation label data.

[0045] This invention aggregates multimodal features from identified operational event segments to extract a unified representation of the physiological response characteristics of that segment, ensuring that the model input possesses a comprehensive expressive capability across modalities, physiological responses, and perturbation morphologies. Inputting this feature into a pre-defined classification model allows for event type discrimination based on historical samples, avoiding subjective bias and inconsistencies caused by manual annotation. The system evaluates and filters the type confidence of the classification results, outputting stable and reliable event labels, thus achieving automatic annotation of intraoperative operational events.

[0046] Preferably, S2 specifically comprises:

[0047] Modal feature data is obtained by extracting intra-event modal features from operation marker data;

[0048] Modal collaborative feature data is obtained by extracting intermodal collaborative change features from modal feature data.

[0049] Multi-source feature data is obtained by constructing feature vectors from modal feature data and modal co-modal feature data.

[0050] This invention, based on operational event labeling, extracts features from each modal signal within the event segment, preserving key response features of physiological signals such as ECG, pressure, blood oxygenation, and catheter displacement under local perturbations, thereby ensuring the independent feature representation capability of each modality. Subsequently, by analyzing the cooperative change patterns between different modalities, the synchronicity, similarity, and coupling structure between modalities are extracted, capturing the composite response features of multi-source signals to the same operational event. Through unified feature vector construction, single-modal expressions and cooperative features are fused to generate structurally consistent multi-source feature data, effectively improving the accuracy and discriminative power of stage mapping and load calculation in representing event states.

[0051] Preferably, S3 specifically comprises:

[0052] The operation stage data is obtained by performing stage analysis based on the operation process data.

[0053] Multi-source feature data is matched to operational stage data to obtain operational stage matched data;

[0054] Based on the matching data of the operation stages, inter-stage dependencies are identified to obtain stage edge data;

[0055] Based on the stage edge data, a graph is constructed from the matching data of the operation stages to obtain the stage flowchart data.

[0056] This invention clarifies the time intervals and sequence of various key operations during surgery by performing stage-by-stage analysis on the operational process data, forming a structured representation of the operational stages. By combining multi-source feature data to match the physiological response state of each stage, stage nodes not only possess time and type information but also specific physiological characteristics, thereby enhancing the semantic expressive power of the nodes. The system extracts stage dependency edges by analyzing multi-dimensional relationships such as the temporal sequence, device calls, and signal mutations between adjacent stages, constructing a flowchart structure with causal chain characteristics. The final stage flowchart not only possesses a topological structure but also integrates multi-source sensory information.

[0057] Preferably, S4 specifically comprises:

[0058] Based on the phase flowchart data, the initial budget base value is set using preset operation budget data to obtain the flowchart budget data;

[0059] Based on the budget data in the flowchart, stage load estimation is performed to obtain preliminary stage load data;

[0060] Budget offset propagation is performed based on the preliminary stage load data to obtain stage impact chain data;

[0061] Graph aggregation processing is performed on the stage impact chain data to obtain stage load data for cardiac interventional physiological signal alerts.

[0062] This invention, based on a staged flowchart structure, introduces pre-set operational budget data to establish initial budget base values ​​for each stage node, thus achieving pre-constraint modeling of operational resources and physiological load. Combining the multi-source characteristics of each stage, the system estimates the actual load, calculates the deviation from the budget value, and propagates this deviation along the graph structure, constructing load influence chains between stages to reflect the diffusion trend of accumulated operational risk among nodes in the graph. Graph aggregation processing identifies high-density load areas and key abnormal nodes, effectively improving the ability to detect high-risk stages in advance.

[0063] The beneficial effects of this invention are as follows: By constructing an interventional procedure event labeling mechanism based on multi-source cardiac data, it achieves automatic time-segment-level identification and type classification of intraoperative procedures without external synchronization signals, improving the automation and accuracy of event extraction. Based on the labeling results, the system extracts modal features within event segments and intermodal collaborative change features, constructing a high-dimensional feature vector reflecting the physiological response of the procedure, laying the data foundation for stage-level modeling. A stage flowchart incorporating temporal sequence, equipment coupling, and feature association is established by combining the procedure flow data, realizing interventional process modeling under a graph structure. Based on the graph structure, budget baseline allocation, load estimation, and budget offset diffusion are performed, forming graph aggregation analysis results that can be used for high-load node detection and alerts. This invention enhances the dynamic perception capability of intraoperative procedure device parameter changes and the stage response tracking capability of physiological parameter changes. Attached Figure Description

[0064] Other features, objects, and advantages of this application will become more apparent from the following detailed description of the non-limiting embodiments, taken with reference to the accompanying drawings:

[0065] Figure 1 A flowchart illustrating the steps of an intelligent analysis method for cardiac interventional physiological signals based on multi-source data fusion is shown in one embodiment.

[0066] Figure 2 A flowchart illustrating the steps of a multi-source cardiac data acquisition method according to an embodiment is shown.

[0067] Figure 3 A flowchart illustrating the steps of a multi-source feature construction method according to an embodiment is shown.

[0068] Figure 4 A flowchart illustrating the steps of a stage flowchart construction method according to an embodiment is shown.

[0069] Figure 5 A flowchart illustrating the steps of a load budget calculation method according to an embodiment is shown. Detailed Implementation

[0070] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0071] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. Functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0072] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0073] Please see Figures 1 to 5 A method for intelligent analysis of cardiac interventional physiological signals based on multi-source data fusion includes the following steps:

[0074] S1: Acquire multi-source cardiac data and mark interventional operation events on the multi-source cardiac data to obtain operation mark data;

[0075] Specifically, the system acquires multi-source cardiac data and performs corresponding interventional operation event labeling processing to form structured operation labeling data. The system acquires ECG signals, arterial pressure signals, blood oxygen saturation signals, and catheter position or acceleration signals through a pre-set multi-channel data acquisition device. The system performs unified time synchronization on all data channels, aligning them to the same global time axis to obtain a standardized multi-source cardiac data stream. The system extracts multimodal perturbation features related to interventional actions from the multi-source cardiac data stream. On the ECG signal side, it extracts subtle perturbations in the trend of heart rate interval changes and the duration of ECG waveforms; on the pressure signal side, it identifies the amplitude of pressure peak changes and the time characteristics of the rapid rise phase; on the blood oxygen data side, it calculates the average slope of the blood oxygen curve within a sliding time window to reflect the slow changes in oxygenation levels; and on the catheter position signal side, it identifies device operation actions by analyzing the amplitude of positional jitter per unit time. Based on the above multimodal perturbation features, the system constructs a graph structure representation within the sliding time window, treating each time window as a graph node, and establishing connection relationships and weights between nodes based on the consistency of perturbations between different modes. The system employs graph attention computation to evaluate the perturbation intensity and correlation of each time window, obtaining the corresponding perturbation significance index and filtering out candidate event segments for corresponding interventional actions. For these candidate event segments, the system performs continuous interval detection, refining the boundaries of regions with less interference but blurred boundaries to obtain refined event segments. The system extracts feature vectors from each refined event segment and inputs them into a pre-defined event type classification model (e.g., a lightweight neural network classifier) ​​to identify the operation type to which the event belongs. The system outputs structured operation label data, including event type, start and end times, corresponding confidence indices, and corresponding multi-source cardiac data.

[0076] S2: Construct multi-source features based on the operation label data to obtain multi-source feature data;

[0077] Specifically, the system constructs multi-source features based on the aforementioned operation marker data to obtain multi-source feature data. For each marked operation event segment, the system extracts the electrocardiogram (ECG), arterial pressure, blood oxygenation, and catheter position signals within that time period according to its corresponding start and end times. The system extracts representative intervention response indicators from each modality. For example, it extracts ST segment displacement features from the ECG signal to represent changes in myocardial electrical activity; it extracts the average rate of ascent from the pressure signal to represent hemodynamic changes; it extracts the lowest blood oxygenation level from the blood oxygenation signal to represent oxygenation abnormalities; and it extracts the main peak features in the acceleration spectrum of the catheter signal to represent catheter micromovement or impact behavior. Through the above processing, the system obtains a separate feature vector for each modality. After completing intramodal feature extraction, the system constructs intermodal collaborative features and evaluates the consistency of intervention responses among different modalities. For example, it calculates the degree of time delay matching between peak variations and arterial pressure changes in electrocardiogram signals, representing an index of modal response synchronicity (i.e., using feature-based change processing to obtain feature change data, and using the time delay matching corresponding to the feature change data as an index of modal response synchronicity). Alternatively, the system can also generate collaborative feature vectors of intermodal dependencies through correlation coefficient analysis, canonical correlation analysis, or multimodal feature concatenation and self-attention mechanisms. The system concatenates intramodal features and intermodal collaborative features, unifying the length and numerical scale of each dimension to construct a feature representation with a unified structure.

[0078] S3: Obtain operation process data; construct stage flowcharts based on operation process data and multi-source feature data to obtain stage flowchart data;

[0079] Specifically, the system acquires operational process data and, combined with the aforementioned multi-source feature data, constructs a stage flowchart representing the evolutionary relationships of surgical or interventional procedures. The system extracts a series of chronologically ordered operational stages by parsing pre-set operational logs or tagged operational events. Each operational stage includes stage type, start and end times, corresponding operator identity, and equipment information, and is bound to multi-source feature vectors within the corresponding time period to form structured stage node information. After extracting the stage nodes, the system establishes basic directed connections based on the chronological order of each stage to represent the natural flow of operational behavior. Simultaneously, if the system detects continuous equipment use, significant feature changes, or a continuous abnormal trend in physiological signals between adjacent stages, it will construct reinforced connections between the corresponding stage nodes. It can assign edge weights to these connections based on the feature change rate or propagation weights calculated by the model to express the strength of the correlation between operations. The system supports combining and encapsulating multiple closely related operational units (such as guidewire advancement and injection operations) to construct subgraph structures and expresses hierarchical dependencies between operational stages through nested relationships, improving the modeling visualization capabilities and structural clarity of the process. The system combines the nodes, connections, and edge weights of each stage to generate a structured stage flowchart.

[0080] S4: Calculate the load budget based on the phase flowchart data to obtain phase load data for cardiac intervention physiological signal alerts.

[0081] Specifically, the system performs load budget calculations based on the constructed stage flowchart data, generating stage load data for cardiac interventional procedures to support real-time alerts and postoperative analysis. The system sets corresponding expected load baseline values ​​for different types of operation nodes. For example, for specific operations such as balloon dilation, the system can set its standard load reference value. When setting the budget baseline value, the system can match pre-operative textual descriptions of the patient's condition (such as past medical history, cardiac function level, pre-operative electrocardiogram assessment, etc.) with a pre-set budget baseline parameter mapping library to determine a budget baseline value consistent with the patient's condition. This budget baseline parameter library is constructed from clinical expert experience and intraoperative data statistics, covering the expected load for different operation types in multiple patient conditions, and outputs a set of budget reference data for each stage node. Based on the feature vector bound to each stage node, the system estimates its actual operation load value using a pre-set model (such as a prediction model or rule system trained based on historical data, where the prediction model is generated based on empirical or historical data and corresponding pre-set actual operation load values ​​through neural network algorithms or decision generation tree algorithms). The system compares the budgeted reference value with the actual operating load value, calculating the load offset for each node, representing the degree of difference between the actual operation and the expectation. Based on the load offset calculation, the system performs graph propagation processing of the budget offset according to the stage flowchart structure. That is, the system propagates the offset of each node to downstream nodes along the directed connections between nodes to construct a load influence chain across the entire graph. If a node simultaneously receives offset influences from multiple predecessor nodes, the system will perform weighted accumulation processing based on the weights of the propagation path (the weight coefficient is an inverse proportional function based on the time interval, e.g., ...). , The propagation weight coefficients from the predecessor node to the current node. This is the node time interval, which is the time difference between the end time of the predecessor node and the start time of the current node. This is a stability constant or zero-prevention factor used to avoid the abnormal situation where the denominator is zero when the time interval is zero. If the predecessor node is closer to the current node in time, it indicates that its load offset has a more direct impact on the current stage, and its weight should be higher. This represents the cumulative load effect on the node. The system performs a high-load area identification operation on the nodes in the entire flowchart. That is, the system combines (e.g., adds) the load value of each node with the offset of its neighboring nodes to calculate the load aggregation index of the node. If the load aggregation index of a node exceeds a preset threshold, it is marked as a load alert candidate, and its current stage load value is attached as a reference for triggering the alert. The system outputs structured stage load data, including the budget reference value, actual estimated value, offset propagation value, and alert marking status for each stage.

[0082] Preferably, acquiring multi-source cardiac data specifically involves:

[0083] S11: Acquire electrocardiogram (ECG) signals through a preset ECG signal acquisition module to obtain ECG signal data;

[0084] Specifically, the system performs ECG data acquisition through a pre-set ECG signal acquisition module. The system activates the configured ECG acquisition device, which can use a five-lead or twelve-lead ECG acquisition system, with a sampling frequency set to no less than 500Hz. During acquisition, the system uses bipolar or multipolar lead wiring and standardizes the position of the grounding electrode. The acquired signals from each lead are organized according to a time-series structure and cached in a database. The structure includes the voltage amplitude and acquisition time corresponding to each time point. The system sets the timestamp accuracy to the millisecond level. The output ECG signal data is organized in a multi-lead signal stream structure, including complete time-series data for each lead and unified time index information.

[0085] S12: Arterial pressure is collected through a preset arterial pressure sensor module to obtain arterial pressure data;

[0086] Specifically, the system acquires arterial pressure data by using a pre-set arterial pressure sensor module to perform arterial pressure signal acquisition. The system can use an insertable pressure catheter or sheath pressure sensor as the acquisition device, with the sampling frequency set to no less than 125Hz. The system organizes the acquired arterial pressure data according to a time series, recording the corresponding voltage value and timestamp for each valid sampling point, forming a continuous and structured data stream. If the system is configured with multiple pressure acquisition channels (e.g., simultaneous monitoring of the left and right coronary arteries), the acquisition paths for different channels must be clearly marked during data acquisition. The arterial pressure data output by the system is organized in a structure of channel identification and time series, with each channel corresponding to a set of continuous data consisting of time points and pressure values.

[0087] S13: Blood oxygen monitoring is performed using a preset blood oxygen monitoring model to obtain blood oxygen data;

[0088] Specifically, the system performs blood oxygen signal acquisition through a preset blood oxygen monitoring module to obtain blood oxygen data. The system uses a finger clip or ear clip pulse oximeter sensor, with a sampling frequency set to at least once per second (≥1Hz) to continuously acquire blood oxygen saturation. The system synchronously extracts pulse wave signals. Each valid sampling point includes the acquisition time and blood oxygen saturation value, which are stored and managed in a structured format. If the system needs to be used in conjunction with other medical devices (such as anesthesia systems), it also provides standardized data interfaces to achieve data exchange and unified processing between different devices. The output blood oxygen data is presented with timestamps, The data is presented in a time series format.

[0089] S14: The catheter position is acquired through the catheter position signal acquisition module to obtain catheter position data;

[0090] Specifically, the system continuously acquires the spatial motion of the catheter during surgery using a catheter position signal acquisition module, obtaining catheter position data. The system can utilize positioning devices such as magnetic positioning systems, impedance positioning modules, or inertial measurement units (IMUs) to acquire the catheter's three-dimensional coordinate position information (X, Y, Z) and attitude parameters (such as Euler angles or quaternions) in real time during the surgical procedure. The system sets the sampling rate of the catheter position signal to the range of 50 to 200 Hz. In addition to position data acquisition, the system also simultaneously records the catheter's acceleration or angular velocity data. Each sampling result is structured and stored as a multi-dimensional data record including a timestamp, three-dimensional position information, and three-axis acceleration information. The exported catheter position signal data is presented in a time-series structure.

[0091] S15: Integrate ECG signal data, arterial pressure data, blood oxygen data, and catheter position data to obtain multi-source cardiac data.

[0092] Specifically, the system integrates the aforementioned acquired ECG signal data, arterial pressure data, blood oxygen saturation data, and catheter position data to generate multi-source cardiac data in a unified format. The system constructs a unified timeline, preferably using the ECG signal channel as the primary clock reference, and aligns all modal data to a time scale measured in milliseconds. The system employs temporal resampling methods such as linear interpolation to map the original sampled data from different modalities to a unified time point. The system uses standardized timestamps to uniformly calibrate all signals, including absolute timestamps based on Unix time or relative time scales with the intraoperative start point as zero. The aligned data is organized in structured blocks, with each time point corresponding to a multi-source data block containing the corresponding ECG signal value, arterial pressure value, blood oxygen saturation value, and catheter spatial status. During integration, if a modality lacks valid data at a specific time point, the system marks that modality as missing. The system outputs a complete multi-source cardiac data sequence, with each time point corresponding to a multi-modal data block, possessing a unified time index and structured content.

[0093] Preferably, the intervention operation event marker is specifically as follows:

[0094] Weak variation feature extraction is performed on multi-source cardiac data to obtain weak variation feature data;

[0095] Specifically, the system performs weak change feature extraction on synchronized multi-source cardiac data to generate weak change feature data. The multi-source data stream is divided into time slices according to a preset sliding window strategy. The window length can be set to 2 seconds, and the sliding step size is 0.5 seconds. The system divides the data stream into a series of time segments, each corresponding to an independent data window. Within each time window, the system performs feature extraction for different modalities. For ECG signal segments, the system calculates the rate of change between adjacent heartbeat cycles (i.e., the change in RR interval) and calculates the standard deviation of the QRS complex duration. For arterial pressure segments, the system extracts the time features of the signal rising edge, specifically the rising time interval before and after the maximum slope point, while also assessing the offset amplitude of the pressure wave peak. For blood oxygenation signals, the system calculates the average slope per second, representing the short-term fluctuation of the blood oxygenation curve and identifying low-amplitude but continuous oxygenation trends. For catheter acceleration signals, the system extracts the maximum acceleration peak value within the current window and calculates the peak value variation frequency per unit time. After extracting features from each modality, the system concatenates these multimodal perturbation features to construct a unified feature vector, representing the weakly changing features within the current time window. The system repeats this process for all time windows, generating a sequence of weakly changing features composed of multiple feature vectors. Weakly changing features refer to data that, within a continuous time scale, do not reach anomaly thresholds or abrupt change criteria, but exhibit persistence, low amplitude, consistent direction, or slight rhythmic shifts.

[0096] This invention selects sensitive and physiologically significant weak variation features from multi-source cardiac data to construct a time-window-level multimodal perturbation representation. These features are chosen with specific functional orientations: the RR interval variation rate and the standard deviation of the QRS duration in electrocardiogram (ECG) signals reflect subtle fluctuations in heart rhythm, helping to capture the autonomic nervous system's response to minor manipulations; in arterial pressure signals, the slope change of the rising edge and the peak offset reflect subtle hemodynamic adjustments, suitable for identifying physiological changes induced by non-drastic postural changes or low-amplitude stimuli; the average slope change of blood oxygenation signals reveals slow and continuous fluctuations in oxygenation status, suitable for identifying blood oxygenation responses under low-intensity interventions; and the peak frequency and peak magnitude in catheter acceleration signals serve as direct indicators of physical contact or catheter fine-tuning, exhibiting high operational relevance. By performing temporal sliding extraction and feature concatenation of the above modal features, the system can suppress transient noise interference while maintaining sensing sensitivity, thereby providing accurate and reliable feature input for operational event recognition.

[0097] Graph attention is used to identify weakly changing feature data to obtain feature-weighted data;

[0098] Specifically, the system, based on a graph neural network mechanism, performs graph attention recognition processing on the extracted weakly changing feature data to generate feature-weighted data. The system constructs a graph structure to represent the relationships between different time windows. The system uses the weakly changing feature vector corresponding to each time window as a node in the graph, forming an initial node set. By default, adjacent time windows are connected by directed edges, representing the temporal continuity of operational or physiological signals. The system compares the degree of modal feature change between different time windows. If the difference is less than a preset similarity threshold (e.g., changes in ECG and pressure features within two windows are both within a reasonable range), a bidirectional connection is added between the corresponding nodes, constructing edges representing weak approximation relationships, thus forming a feature graph. After the feature graph is constructed, the system uses a graph attention mechanism to model the intensity of feature interactions between nodes. The system performs a linear embedding transformation (i.e., a linear transformation based on a preset trainable parameter matrix) on the original feature vector of each node to generate an internal representation vector. For each pair of connected nodes, the system calculates attention weights based on the importance of the feature combinations between nodes, representing the degree of influence between nodes during propagation. These weights are generated through a trainable attention mechanism and can be adaptively adjusted based on the correlation between node features. The system uses these attention weights to weight and aggregate the feature / internal representation vectors of each node's neighboring nodes, completing the node's feature update operation. The aggregated feature representation is the weighted feature of the current node, used to replace the original weakly changing features. The system outputs a complete sequence of weighted feature data, with each item corresponding to the perception result of a time window.

[0099] Operation event segment identification is performed based on feature-weighted data to obtain operation event segment data;

[0100] Specifically, based on the aforementioned feature-weighted data, the system performs operation event segment identification processing and outputs operation event segment data. The system performs confidence assessment on the feature-weighted data for each time window. Using a preset operation event segment identification model, such as a sigmoid mapping or a lightweight multilayer perceptron model, the system generates a confidence score for each time window, with values ​​ranging from 0 to 1, representing the probability of the operation occurring at that time point. The system uses continuous judgment, searching for a series of continuous window segments in the time series with confidence scores higher than a set threshold (e.g., 0.6), considering them as preliminary operation event candidate segments, and recording their start and end times. Each candidate segment is a possible intervention action period and is marked as a draft event segment. The system then refines and adjusts the start and end boundaries of the candidate segments. The system sets fixed-length buffer regions (e.g., 0.5 seconds before and after) at both ends of the candidate segment and extends them along the time axis using a sliding window until a significant decrease in intensity is detected. Based on this, the system determines the actual boundary position of the current segment, thereby correcting the candidate segment range and obtaining a more accurate event segment division. The criteria for determining a significant decrease in intensity include setting an intensity decrease threshold (e.g., a decrease exceeding 30% of the previous value); or using a moving average window (e.g., 3 windows) to calculate the local confidence average. When the current window value is lower than a certain percentage (e.g., 70%) of the moving average, a significant decrease is considered to have occurred; alternatively, the first-order difference signal of the confidence can be combined. When two consecutive windows show negative changes and the decrease exceeds a set difference threshold (e.g., -0.2), it is determined to be an inflection point of intensity decrease. The system executes a deduplication strategy to eliminate duplicate or redundant identified segments. On the one hand, for adjacent event segments with a time interval less than a set threshold (e.g., 1 second), the system merges them into a single complete event; on the other hand, for highly overlapping candidate segments (e.g., an overlap rate exceeding 70%), the system retains only the segment with the highest total confidence. The system outputs a structured dataset of operation event segments, with each record containing the start time, end time, and corresponding perturbation confidence level of the operation event.

[0101] The operation event types are classified based on the operation event segment data to obtain operation tag data.

[0102] Specifically, based on the identified operation event segment data, the system performs operation event type classification processing and outputs operation label data. For each operation event segment, the system calls back the previously generated weighted feature sequence within its corresponding time range, extracting feature representations for all time points within that segment. The system aggregates the weighted feature vectors within the segment, using average pooling or max pooling to generate representative segment-level feature vectors. The system inputs these segment-level feature vectors into a pre-trained operation event classification model (trained based on historical data and corresponding historical operation label data). This model can be constructed using a shallow neural network structure (such as a three-layer multilayer perceptron) and trained based on existing labeled data. The model outputs a category probability distribution, representing the likelihood that the event segment belongs to each operation type. The system selects the category with the highest confidence as the initial classification type for the event segment; if the confidence of all categories is below a preset threshold (e.g., 0.6), the segment is marked as "uncertain" to prompt manual confirmation or semi-automatic correction in subsequent processes. The system also performs confidence evaluation processing on the classification results of each segment. This processing includes retaining the maximum class confidence value output by the model and metrics used for internal consistency judgment, such as the reasonableness assessment of the classification result within the temporal logic. If the system detects a significant conflict between the current classification result and the type or duration of adjacent event segments, such as an unreasonable jump in operation type, the system will invoke preset prior sequence rules or context constraint logic to fine-tune and correct the label results. The system outputs an operation label dataset, where each record contains the start and end times of the event segment, the operation type label, and its corresponding classification confidence score.

[0103] Preferably, the extraction of weak change features specifically includes:

[0104] The interval rate of change and QRS wave fluctuation time were calculated from the electrocardiogram signal data in the multi-source cardiac data to obtain the interval rate of change and QRS wave fluctuation time, respectively.

[0105] Specifically, the system extracts temporal features from electrocardiogram (ECG) signal data in multi-source cardiac data, including two sub-processing steps: interval variability calculation and QRS wave fluctuation time calculation. For interval variability calculation, the system uses a QRS wave detection algorithm (such as the Pan-Tompkins algorithm) to analyze the complete ECG signal beat-by-beat, extracting the R-wave positions corresponding to all heartbeats. Based on this, the system calculates the interval between two adjacent R waves, forming a continuous heartbeat cycle sequence. The system performs first-order difference processing on this cycle sequence to represent the amplitude of change between adjacent cycles, and calculates the average variability within this interval using a sliding time window (e.g., containing five heartbeat cycles). The above are then integrated to obtain the interval variability. For QRS wave fluctuation time calculation, for each identified QRS segment, the system extracts an ECG signal segment within a time range of approximately 40 milliseconds before and 60 milliseconds after the R wave. The system calculates the second derivative of this signal segment to identify key peaks and troughs, and locates the time period where the signal energy is mainly concentrated. The signal energy is the integral or cumulative value of the square of the ECG signal amplitude. The system calculates the duration of this energy concentration region (e.g., the time interval covering 95% of the total energy) as the QRS wave fluctuation time. The system outputs interval rate of change data and QRS fluctuation time data, where the former reflects the characteristics of heart rhythm fluctuations, and the latter reflects the fine structural changes of the electrocardiogram waveform.

[0106] Waveform micro-variation rise time and peak change are extracted from arterial pressure data in multi-source cardiac data to obtain waveform micro-variation data and peak change data, respectively.

[0107] Specifically, the system extracts features from arterial pressure signals in multi-source cardiac data, comprising two sub-modules: waveform micro-variation rise time extraction and pressure peak change calculation. For waveform micro-variation rise time extraction, the system identifies the starting point (i.e., the position where pressure begins to rise rapidly) and the peak point (i.e., the position of maximum pressure within this cycle) in the corresponding pressure waveform for each cardiac cycle. The system calculates the time interval from the starting point to the peak point, representing the duration of the hemodynamic rise phase in each cycle. The system calculates the mean and standard deviation of this rise time over multiple consecutive cycles (e.g., 3-5 cycles) to obtain waveform micro-variation data. For peak change calculation, the system extracts the maximum pressure value occurring in each cycle and calculates the variation amplitude of the maximum value between adjacent cycles, representing the periodic blood pressure peak fluctuation trend. The system smooths these differences using a moving average to output continuous peak change data. The system outputs two types of feature data sequences: one is the pressure rise phase time series / waveform micro-variation data, representing the characteristics of the rise phase in the cardiac cycle; the other is the pressure peak difference sequence / peak change data, representing the dynamic changes in blood pressure intensity during the cycle.

[0108] The blood oxygen fluctuation intensity data is obtained by calculating the mean instantaneous slope of blood oxygen data from multi-source cardiac data.

[0109] Specifically, the system extracts short-term variation features from multi-source cardiac data. The system processes consecutive sampling points in the blood oxygen signal pairwise, calculating the rate of change between two adjacent time points, i.e., the blood oxygen slope value within each time period. The system performs statistical processing on the continuous slope sequence at preset analysis times (e.g., every 10 seconds). Within each analysis time, the system extracts the absolute values ​​of all slope values ​​and calculates their average, using this as the blood oxygen fluctuation intensity feature for the current time period. The system outputs blood oxygen fluctuation intensity data.

[0110] The catheter jitter peak value was calculated from the catheter position data in the multi-source cardiac data to obtain the catheter jitter peak value data;

[0111] Specifically, the system extracts the physical perturbation amplitude from catheter position data in multi-source cardiac data, calculates the peak vibration generated by the catheter during the procedure, and identifies potential operative actions or local mechanical stimulation events. The system extracts triaxial acceleration data at corresponding time points from the catheter signals, including acceleration values ​​in the X, Y, and Z directions. Based on the data in these three directions, the system calculates the resultant acceleration value at each time point. The system analyzes the changes in resultant acceleration within a time window (e.g., one second per window) and extracts the difference between the maximum and minimum acceleration values ​​within that window, i.e., the peak catheter vibration value for that time period. The system outputs the peak catheter vibration data in time series format.

[0112] By integrating the interval change rate, QRS wave fluctuation time, waveform micro-change data, peak change data, blood oxygen fluctuation intensity data, and catheter jitter peak data, weak change characteristic data are obtained.

[0113] Specifically, the system fuses / combines / semblages the extracted multi-modal cardiac features, aligning and integrating various temporal physiological features according to temporal consistency requirements to generate weakly variable feature data. Using a sliding time window as a unit, the system performs temporal alignment on six features extracted within each time segment: interval rate of change, QRS fluctuation time, pressure waveform rise time / waveform micro-variation data, pressure peak change amplitude / peak change data, blood oxygen fluctuation intensity data, and catheter jitter peak value. The system then concatenates these data from different modalities to form a feature vector. The weakly variable feature data output by the system is presented as a sequence of temporal feature vectors, with each time window corresponding to a standardized fused feature vector.

[0114] Preferably, graph attention recognition specifically includes:

[0115] Feature maps are constructed based on weakly changing feature data to obtain feature map data;

[0116] Specifically, the system constructs a feature graph structure based on the aforementioned weakly changing feature data. The system uses a sliding time window as a unit, representing the weakly changing feature vector within each time window as a node in the graph. The attributes carried by each node represent the set of multimodal physiological signal feature values ​​for that time period. Regarding edge construction in the graph structure, the system generates connections between nodes according to two types of rules: first, the temporal adjacency rule, where the system automatically connects two temporally adjacent nodes; second, the feature similarity rule, where the system measures the difference in feature vectors between any two nodes, and if the difference is less than a preset similarity threshold, it is considered to have potential commonalities or pattern recurrence trends, and the system adds additional connecting edges to express weak correlations in features. The system also sets a sliding window connection range limit, allowing each node to only connect to nodes within a few time windows before and after it (e.g., a maximum of three). The system outputs graph structure data containing a set of nodes and a set of edges. Each node in the graph represents the physiological perturbation state of a time segment, and the edges represent temporal dependencies or feature commonalities (temporal proximity or similarity of feature changes).

[0117] Feature encoding is performed on the feature map data to obtain feature-encoded data;

[0118] Specifically, the system performs feature encoding on the constructed feature map data, mapping the original node feature vectors to a unified-dimensional embedding representation. The system performs a linear transformation on the original weakly variable feature vectors carried by each node in the graph, mapping them from the original feature space to a preset embedding space dimension. This linear transformation is constructed using preset weight matrices and preset bias parameters, both of which can be learned and optimized during model training. After feature mapping, the system sets a non-linear activation function, preferably LeakyReLU. After this mapping and activation process, the system generates a unified-dimensional latent vector representation for each node in the graph. This latent vector retains the core information of the original node features and serves as the basic input unit in the graph attention mechanism. All latent vectors are cached uniformly.

[0119] Graph attention weights are calculated based on the feature encoding data to obtain feature graph attention data;

[0120] Specifically, based on the aforementioned generated feature encoding data, the system executes a graph attention mechanism on the constructed feature graph structure to calculate the attention weights between nodes. The system sequentially traverses each defined edge in the graph, and for each pair of connected nodes, it calculates the importance score of that neighboring node to the center node, i.e., the attention score, based on its feature encoding vector. ,in For nodes For nodes The original attention score, For activation function, The attention scoring vector is preset for the system, representing the degree of influence of the perceived features of neighboring nodes on the central node (trainable parameter). For nodes eigenvectors, This is the feature transformation matrix (weight matrix). For nodes The system calculates attention scores by combining a pre-defined trainable parameter vector with concatenated node features using a non-linear activation function. This process senses the degree of feature difference between adjacent nodes and generates an initial attention score representing the strength of the perturbation association. After obtaining the initial attention scores of all neighboring nodes, the system normalizes the neighboring attention values ​​of each central node using the Softmax function, ensuring that the sum of the attention weights of all adjacent edges is 1. The system implements sparsity, pruning low-weight attention edges with a threshold. If an attention weight is below a set threshold (e.g., between 0.05 and 0.1), the system can reset the weight to zero. The system outputs attention weight data in a sparse matrix structure, retaining valid weight values ​​only between node pairs connected in the graph.

[0121] We obtain weighted feature data by weighting and aggregating the attention data from the feature maps.

[0122] Specifically, based on the calculated feature map attention data, the system performs a weighted aggregation operation in the graph neural network to generate a weighted feature representation for each node within the graph structure. The system traverses all its neighboring nodes sequentially, using the corresponding attention weights to weight the feature representations of these neighboring nodes. By aggregating the feature information of neighboring nodes, the system achieves the fusion and propagation of perturbation features between nodes and enhances the context of the current node's pattern. For feature aggregation, the system can employ various strategies, including weighted summation, average pooling of neighboring features, max pooling, or concatenating features from multiple neighboring nodes before processing. The aggregation result, after being processed by a non-linear activation function, serves as the weighted feature representation of the current node. If the system employs a multi-head attention mechanism, multiple independent attention subspaces are computed in parallel at each node, feature aggregation is performed separately, and the results of multiple sub-attentions are combined through concatenation or averaging to improve the model's ability to represent diverse patterns. The system also performs standardization processing on the obtained weighted feature representations, such as using L2 normalization or batch normalization, to unify the output features of different nodes to a similar scale. The system outputs feature-weighted data, representing the features of each graph node within its context.

[0123] Preferably, the operation event segment identification specifically includes:

[0124] Threshold-based continuous detection is performed on the feature-weighted data to obtain preliminary event segment data;

[0125] Specifically, based on the aforementioned generated feature-weighted data sequence, the system performs a threshold continuous detection operation to identify a set of preliminary event segment data. The system performs confidence scoring on the weighted feature vector corresponding to each time window, using a set of preset linear transformation weights and activation functions to calculate the intensity of each feature vector. The output range is normalized to the interval between 0 and 1, such as... , For the confidence score of the k-th time window, Here, is the activation function, and is the sigmoid function. The feature intensity transformation matrix is ​​a preset value. For the feature weighted data of the k-th time window, This is the bias term, and this is a preset value. After obtaining the confidence score sequence for consecutive time windows, the system sets a confidence threshold (e.g., 0.6) and performs sliding detection on the entire time series based on this threshold. For time window segments with confidence scores continuously higher than this threshold, the system aggregates them into consecutive time segments and determines whether their duration exceeds a preset minimum duration threshold (e.g., 1 second or 3 consecutive windows). Only when consecutive high-confidence segments meet the duration requirement is the system identified as preliminary candidate event segments and considered as preliminary event segment data. Each candidate event segment is defined by its start and end times and includes the average confidence score for that time period. The system outputs a set of preliminary event segment data, where each item includes the time range and average confidence information.

[0126] The time period boundaries are refined based on the initial time period data to obtain secondary time period data;

[0127] Specifically, after obtaining the initial event segment data, the system performs boundary refinement processing to improve the accuracy of the start and end times of the event segments, avoiding premature or delayed boundary truncation, thereby outputting secondary event segment data. The system performs forward and backward boundary expansion operations on each initial event segment. During the expansion process, the system sequentially slides towards both ends of the event segment to judge the confidence values ​​of adjacent time windows. If the confidence value of an adjacent frame is still higher than the set secondary judgment threshold (e.g., 0.4), the time window is included in the current event segment range. This process continues until the confidence value of several consecutive frames (e.g., 2 to 3 frames) is lower than the secondary threshold, at which point the system considers the boundary to have reached the natural termination point of the perturbation region. If only a single frame with a jump higher than the threshold appears at the beginning or end of a segment without continuity, the system will determine it as a false positive and backtrack the boundary position. The validity of candidate segment boundaries can be filtered by combining the set minimum consecutive rising frame number parameter. After completing the above processing, the system updates the start and end times of each event segment to the refined time points, and records the difference information before and after the boundary changes. The system outputs a set of event segments with corrected boundaries, each segment being finely adjusted based on the original candidates through dynamic expansion and contraction strategies.

[0128] The data for the operation event segment is obtained by deduplicating the data from the secondary time period.

[0129] Specifically, after refining the event segment boundaries, the system performs deduplication and fusion processing on the secondary time segment data to generate operational event segment data. The system employs an overlap merging strategy, evaluating the overlap between any two event segments. If two time segments have a significant overlap, and the proportion of their time overlap to the total duration exceeds a set threshold (e.g., 0.6), the system considers them different detection results of the same event and merges the two segments into a new event segment. The start time is the minimum of the two, and the end time is the maximum, thus forming a continuous and complete event interval. The system applies time proximity fusion to identify adjacent event segments that are close in time but do not overlap. If the interval between two event segments is less than a set threshold (e.g., 1 second or 2 frames), the system determines them as the same event operating continuously and performs segment merging processing. Short segments with a duration below the minimum time threshold (e.g., 0.5 seconds) are discarded directly. Low-confidence segments with an average confidence level below a preset threshold (e.g., 0.5) are also excluded. The operational event segment data set output by the system includes the start time, end time, and average confidence level of each valid event.

[0130] Preferably, the operation event type classification is as follows:

[0131] Event segment feature data is obtained by aggregating event segment features based on the operation event segment data.

[0132] Specifically, based on the identified operational event segment data, the system performs multimodal feature aggregation processing at the event segment level to generate event segment feature data. The system extracts multi-source physiological signal fragments within the corresponding time range from each operational event segment. The modalities involved may include signal types such as electrocardiogram (ECG), arterial pressure, blood oxygen saturation, and catheter displacement. For each type of signal, the system performs multi-dimensional feature extraction operations within the event segment, including basic statistical features (such as mean, standard deviation, maximum, and minimum values), time-domain dynamic features (such as the rate of change of ECG intervals, the slope of arterial pressure rise, and the intensity of blood oxygen fluctuations), and frequency-domain features (such as energy distribution indicators extracted through wavelet transform). After completing the extraction of each modal feature, the system concatenates and integrates various features into a unified vector representation according to a preset structural template. The event segment feature data output by the system consists of multiple standardized vectors, each vector corresponding to one operational event segment.

[0133] The event segment feature data is input into the preset operation event type classification model to obtain type confidence data;

[0134] Specifically, the system inputs the constructed event segment feature data into a preset operation event type classification model, performs corresponding classification inference processing, and outputs the confidence distribution results representing each candidate operation type. The classification model can employ a lightweight neural network structure, such as a multilayer perceptron (MLP), or use a tree-based ensemble method, such as XGBoost or a random forest model. The training process of the classification model can be based on real intraoperative operation records or historical datasets annotated by clinicians, enabling the model to identify physiological perturbation patterns corresponding to common intervention behaviors. In the actual classification process, the system uses the feature vector corresponding to each event segment as the model input and uniformly performs standardization or normalization processing. The model output is a set of confidence distributions, representing the degree of matching of the event segment under various preset operation types. For example, the output results include "angiography injection: 0.82", "balloon dilation: 0.10", "guidewire advancement: 0.05", etc., representing the model's confidence level in judging various operations. The type confidence data output by the system is presented in a structured form, providing multi-category confidence assessment results for each operation event segment.

[0135] The label confidence is evaluated based on the type confidence data to obtain the operation label data.

[0136] Specifically, the system judges the model output based on a set confidence threshold. If the confidence level corresponding to a certain operation type is the highest among all categories and its value is higher than a preset threshold (e.g., 0.75), the system uses that category as the label for that event segment and marks it as "confirmed." If the highest confidence level does not reach the threshold, the system marks the event segment as "uncertain" or "candidate event," indicating that it needs to be manually confirmed or included in the subsequent review process. If the system detects that the confidence level distribution of an event segment across all categories is extremely even (i.e., the distribution entropy value is high), the system can identify it as a case of poor signal quality or feature representation failure, and record relevant abnormal information as an important trigger for manual intervention or model rollback. The operation label data output by the system is presented in the form of structured data, with each record containing an event segment identifier, operation type label, confidence score, and label status.

[0137] Preferably, S2 specifically comprises:

[0138] S21: Extract intra-event modal features based on operation marker data to obtain modal feature data;

[0139] Specifically, the system receives input operation tag data, where each record contains the start and end times of the event segment and the corresponding operation type label. The system retrieves raw physiological signal data corresponding to the time range of the event segment, including multiple modalities such as electrocardiogram (ECG), arterial pressure, blood oxygen saturation, and catheter position / acceleration. Within each operation event segment, the system extracts data segments from each modal signal according to the corresponding time axis. For each modal signal, the system performs feature extraction operations in the following dimensions: ECG modality: The system calculates the mean and standard deviation of the intracardiac repetition interval (RR interval) within the event segment, calculates the average width of the QRS wave, and extracts heart rate variability indicators such as root mean square deviation (RMSSD) and percentage of rhythm change (pNN50). Arterial pressure modality: The system calculates the mean arterial pressure (MAP) within the event segment, extracts the maximum amplitude between systolic and diastolic pressure, and calculates the rise time of the pressure waveform and the rate of change of pulse pressure. Blood oxygenation modality: The system extracts the lowest and average blood oxygenation values ​​from the blood oxygenation signal, and simultaneously calculates the average slope of the blood oxygenation decline phase. Catheter position modality: The system calculates the average change in catheter position coordinates within this segment, counts the maximum value among the three-axis accelerations, and extracts the frequency and peak amplitude of catheter jitter during the operation. These features are organized according to modal dimensions, generating a multimodal feature set for each event segment, including ECG modal feature vectors, pressure modal feature vectors, blood oxygenation modal feature vectors, and catheter modal feature vectors. Each type of modal feature vector has a fixed dimension, representing the dynamic response pattern of that operation event segment in different physiological channels. The system outputs modal feature data.

[0140] S22: Extract intermodal collaborative change features based on modal feature data to obtain modal collaborative feature data;

[0141] Specifically, the system inputs a set of multimodal feature vectors corresponding to each operational event segment, covering channels such as ECG, arterial pressure, blood oxygenation, and catheter position. The system calculates the synergistic change indices between each mode in the form of mode pairs. These synergistic relationships include, but are not limited to: ECG and arterial pressure mode: calculating the correlation coefficient between heart rate variability and blood pressure fluctuation parameters to represent the linkage between rhythm and hemodynamics; arterial pressure and blood oxygenation mode: calculating the partial derivative trend of blood oxygenation slope with respect to systolic blood pressure changes to represent the degree of response of oxygen supply changes to blood flow changes; ECG and catheter mode: measuring the difference between QRS band features and catheter acceleration change curves (e.g., dynamic time warping distance) to indicate the coupling relationship between electrical activity disturbances and physical operations; blood oxygenation and catheter mode: assessing the synchronicity of the rapid drop in blood oxygenation and the peak of catheter operation on the time axis, using indicators such as time intersection-combination ratio to quantify their overlap. In terms of collaborative feature aggregation, the system employs two approaches: one is to directly construct collaborative feature vectors using manually defined key indicators (such as correlation coefficient, temporal intersection-over-union ratio, DTW distance, etc.) and then concatenate them to form a structured output; the other approach is to concatenate multiple modal features and input them into an autoencoder structure, extracting potential modal fusion representations through nonlinear dimensionality reduction. The system's output modal collaborative feature data is organized in units of event segments, with each segment containing a set of feature representations describing the collaborative relationships between modalities.

[0142] S23: Construct feature vectors from modal feature data and modal co-modal feature data to obtain multi-source feature data.

[0143] Specifically, the system receives various modal feature vectors corresponding to each operational event segment, including ECG feature vectors, arterial pressure feature vectors, blood oxygenation feature vectors, and catheter position feature vectors, while also inputting the modal co-modality feature vectors generated in the preceding steps. The system concatenates all the above features in a uniform order to form a multi-source fusion feature vector. The multi-source feature data output by the system is presented in a standard vector format, with each operational event segment corresponding to a feature vector structure of uniform dimension.

[0144] Preferably, S3 specifically comprises:

[0145] S31: Perform stage analysis based on the operation process data to obtain the operation stage data;

[0146] Specifically, the operational process data can be a pre-defined standard operational step template or an actual operational log recorded during the procedure, including various operation names and their corresponding timing information. For well-structured operational procedures (such as coronary intervention), the system prioritizes parsing based on a preset standard operational template, dividing the entire operation into several logical stages by matching the operational steps with their expected execution order. Under non-template conditions, the system can perform timing reconstruction and stage division based on the operational log. The system sorts operational events by timestamp, identifies operational segments with boundary characteristics through text recognition / a preset operational process text parameter library, and automatically identifies and marks stage boundaries by combining the operation type and duration of each segment. Each parsed operational stage is assigned a unique identifier (e.g., stage_01, stage_02, etc.) and includes corresponding operation type information, time range, or expected timing position and other attribute data. The system outputs a set of operational stage data containing stage identifiers, corresponding operation types, and timing parameters.

[0147] S32: Match the multi-source feature data to the operation stage data to obtain the operation stage matching data;

[0148] Specifically, the system receives two types of input data: first, structured operation stage information, including the name, start and end times, and operation type of each stage; second, multi-source fused feature data corresponding to each operation event segment, with each event data also including its time range. During the matching process, the system traverses each event segment and determines its corresponding operation stage based on its start and end times and the time boundaries of each stage. Matching operations include: complete inclusion, meaning that if the time interval of an event segment falls entirely within a certain stage, it is directly assigned to that stage; or using time overlap rate judgment, if the time interval of an event segment significantly overlaps with that of a stage (e.g., more than half), then the event segment is considered to belong to that stage. Through this matching process, the system can accurately aggregate scattered event feature data into each operation stage, constructing a fused feature set at the stage dimension.

[0149] S33: Identify inter-stage dependencies based on the operation stage matching data to obtain stage edge data;

[0150] Specifically, the system determines the sequential dependencies between stages through time sequence analysis. For any two stages, if all their related event segments exhibit a non-overlapping, continuous temporal order—that is, the completion time of all operations in one stage is earlier than the start time of the other—the system identifies the former as the precursor stage of the latter, thus establishing a temporal dependency edge between stages. Based on business rules in the procedural flow (such as certain operations requiring the completion of other operations) or on frequently co-occurring event sequence patterns in historical data, the system determines the causal or triggering relationship between stages. When an abnormal physiological response in a stage is detected to cause the premature start of the next stage, the system marks such edges as soft dependency edges, distinguishing them from strictly sequential temporal edges. The system attaches weight information to stage edges, such as calculating the average time interval between operations represented by the edge, success rate, or anomaly frequency, as edge attribute information.

[0151] S34: Construct a graph from the phase edge data to the operation phase matching data to obtain the phase flowchart data.

[0152] Specifically, the system treats each identified operation stage as a node in the graph. Each node contains the stage name, time sequence number, and the aggregated result of multi-source feature vectors corresponding to all event segments within that stage. For example, average pooling or max pooling can be used to reduce the features, serving as the node's attribute information. The graph edges in the stage flowchart originate from the identified inter-stage dependencies. These edges are directional, indicating the execution order or triggering logic between successive stages. They can also be labeled with edge types (e.g., time sequence dependency, soft dependency) and weight information (e.g., average time delay or stage switching success rate). The graph structure is presented as a directed acyclic graph, but it can also support stage rollback in special scenarios, constructing a directed graph structure with partial back edges.

[0153] Preferably, S4 specifically comprises:

[0154] S41: Based on the stage flowchart data, set the initial budget base value using the preset operation budget data to obtain the flowchart budget data;

[0155] Specifically, the system loads graph data containing stage nodes and their dependencies, and imports an externally defined operation budget template. This template sets corresponding budget caps for different operation types, such as the maximum allowed time for each type of operation, consumable usage restrictions, or the duration / range of parameter change restrictions. During budget allocation, the system iterates through each stage node in the graph, finds the corresponding budget value in the budget template based on its identified operation type, and assigns it to the node. The system updates the structure of each node, adding the corresponding budget base value as an attribute field to the graph node, forming flowchart data with initial budget configuration.

[0156] S42: Perform phase load estimation based on the flowchart budget data to obtain preliminary phase load data;

[0157] Specifically, the system extracts fusion feature information from each node, including standardized indicators such as heart rate variability, blood pressure fluctuation amplitude, blood oxygen saturation level, and catheter manipulation intensity (catheter acceleration, manipulation frequency, or catheter curve change frequency, the latter requiring acquisition via image reconstruction or displacement sensors). It then calls the budget ceiling value set for each node in the previous step. Based on a set of weighted rules or a built-in lightweight estimation model (e.g., support vector regression or multilayer perceptron) or parameter subtraction, the system estimates the load intensity of each node and outputs its actual load value. The system calculates the difference between the actual load value and the budget ceiling to obtain the budget offset for that node. If the offset of a node is greater than zero, meaning its load exceeds the budget value, the system marks it as a primary risk node for budget overrun. The system outputs the estimated load value, offset result, and preliminary risk label for each stage node, constituting the preliminary stage load data.

[0158] S43: Based on the preliminary stage load data, perform budget offset propagation to obtain stage impact chain data;

[0159] Specifically, the system is based on the topology of a phased flowchart, preserving the original phase sequence and node connection relationships. In this graph structure, if a node is identified as a primary risk node for budget overruns (meaning its actual load has exceeded the budget limit), the system initiates the budget offset propagation process starting from that node. During propagation, the system weighted accumulates and diffuses the risk impact based on the connection strength of the edges in the graph and a preset propagation coefficient. The propagation process employs local diffusion, similar to the distributed propagation mechanism of information in a network, gradually transmitting the offset impact to adjacent downstream phase nodes. The connection strength of the edges can be set with reference to factors such as the time interval between nodes and the degree of operational coupling to reflect the differences in the strength of the impact between different paths. The system continuously records the propagation path during the propagation process, forming a complete impact chain. For example, if the risk offset propagates from phase A to phase B, and then further to phase C, the corresponding impact path is recorded as [A→B→C]. The system outputs the impact chain data corresponding to each risk node, with a structure including the source node identifier, a list of affected nodes, and the corresponding offset propagation trajectory, constituting the phase impact chain data.

[0160] S44: Perform graph aggregation processing based on the stage impact chain data to obtain stage load data for cardiac interventional physiological signal alerting.

[0161] Specifically, based on the propagated budget offset information, the system aggregates and calculates nodes across the entire stage graph to identify systemic risk areas and key nodes. The system assesses the overall load overrun ratio, representing the severity of the overall budget offset. Based on the node's structural attributes and offset degree, the system marks two categories of important nodes: first, "core risk nodes," which are nodes with high in-degree and large budget offsets, serving as core control points in the impact chain; and second, "chain endpoint risk nodes," which, although their own budgets are not overrun, possess high early warning value due to risk transmission from upstream stages. For each stage node, the system outputs an information structure containing its identification information, budget value, estimated load, offset, risk level, and the propagation chain it belongs to. This structure can be directly used as input to the alert module to support intraoperative risk alerts. High-risk nodes and their corresponding impact chains are packaged and sent to the alert module, which can trigger real-time voice alerts or graphical interface alarms to assist users in identifying current signal anomalies and stage execution risks.

Claims

1. A method for intelligent analysis of cardiac interventional physiological signals based on multi-source data fusion, characterized in that, Includes the following steps: S1: Acquire multi-source cardiac data, extract weak variation features from the multi-source cardiac data to obtain weak variation feature data. The weak variation feature data is obtained by dividing the multi-source cardiac data into time slices according to a preset sliding window strategy. Within each time window, the system performs feature extraction for different modalities of data. After completing the feature extraction of each modality, the system concatenates the above multi-modal features to construct a feature vector with a unified structure, representing the weak variation features within the current time window. Graph attention recognition is performed on the weak variation feature data to obtain feature weighted data. Operation event segment identification is performed based on feature-weighted data to obtain operation event segment data; operation event type classification is performed based on operation event segment data to obtain operation tag data. S2: Construct multi-source features based on the operation label data to obtain multi-source feature data; S3: Obtain operation process data; construct stage flowcharts based on operation process data and multi-source feature data to obtain stage flowchart data; S4: Calculate the load budget based on the phase flowchart data to obtain phase load data for cardiac intervention physiological signal alerts.

2. The method according to claim 1, characterized in that, The specific steps for obtaining multi-source cardiac data are as follows: ECG signal data is obtained by acquiring ECG signals through a preset ECG signal acquisition module. Arterial pressure data is obtained by collecting arterial pressure through a preset arterial pressure sensor module; Blood oxygen data is obtained by monitoring blood oxygen using a pre-set blood oxygen monitoring model; The catheter position is acquired using a catheter position signal acquisition module to obtain catheter position data; By integrating electrocardiogram signal data, arterial pressure data, blood oxygen data, and catheter position data, multi-source cardiac data is obtained.

3. The method according to claim 1, characterized in that, The extraction of weakly changing features is specifically as follows: The interval rate of change and QRS wave fluctuation time were calculated from the electrocardiogram signal data in the multi-source cardiac data to obtain the interval rate of change and QRS wave fluctuation time, respectively. Waveform micro-variation rise time and peak change are extracted from arterial pressure data in multi-source cardiac data to obtain waveform micro-variation data and peak change data, respectively. The blood oxygen fluctuation intensity data is obtained by calculating the mean instantaneous slope of blood oxygen data from multi-source cardiac data. The catheter jitter peak value was calculated from the catheter position data in the multi-source cardiac data to obtain the catheter jitter peak value data; By integrating the interval change rate, QRS wave fluctuation time, waveform micro-change data, peak change data, blood oxygen fluctuation intensity data, and catheter jitter peak data, weak change characteristic data are obtained.

4. The method according to claim 1, characterized in that, Graph attention recognition specifically involves: Feature maps are constructed based on weakly changing feature data to obtain feature map data; Feature encoding is performed on the feature map data to obtain feature-encoded data; Graph attention weights are calculated based on the feature encoding data to obtain feature graph attention data; We obtain weighted feature data by weighting and aggregating the attention data from the feature maps.

5. The method according to claim 1, characterized in that, The operation event segment identification is specifically as follows: Threshold-based continuous detection is performed on the feature-weighted data to obtain preliminary event segment data; The time period boundaries are refined based on the initial time period data to obtain secondary time period data; The data for the operation event segment is obtained by deduplicating the data from the secondary time period.

6. The method according to claim 1, characterized in that, The operation event types are specifically categorized as follows: Event segment feature data is obtained by aggregating event segment features based on the operation event segment data. The event segment feature data is input into the preset operation event type classification model to obtain type confidence data; The label confidence is evaluated based on the type confidence data to obtain the operation label data.

7. The method according to claim 1, characterized in that, S2 specifically refers to: Modal feature data is obtained by extracting intra-event modal features from operation marker data; Modal collaborative feature data is obtained by extracting intermodal collaborative change features from modal feature data. Multi-source feature data is obtained by constructing feature vectors from modal feature data and modal co-modal feature data.

8. The method according to claim 1, characterized in that, S3 specifically refers to: The operation stage data is obtained by performing stage analysis based on the operation process data. Multi-source feature data is matched to operational stage data to obtain operational stage matched data; Based on the matching data of the operation stages, inter-stage dependencies are identified to obtain stage edge data; Based on the stage edge data, a graph is constructed from the matching data of the operation stages to obtain the stage flowchart data.

9. The method according to claim 1, characterized in that, S4 specifically refers to: Based on the stage flowchart data, the initial budget base value is set through the preset operation budget data to obtain the flowchart budget data. The operation budget data includes the maximum allowed time for each type of operation, the consumable usage restrictions, or the duration / range of parameter change restrictions. Based on the budget data in the flowchart, stage load estimation is performed to obtain preliminary stage load data; Budget offset propagation is performed based on the preliminary stage load data to obtain stage impact chain data; Graph aggregation processing is performed on the stage impact chain data to obtain stage load data for cardiac interventional physiological signal alerts.

Images