Methods and systems for processing monitoring data of implantable medical pressure monitors

Abstract

The application provides a monitoring data processing method and system for an implantable medical pressure monitor, and relates to the technical field of data processing. By acquiring a pressure monitoring reference signal, a pressure monitoring target signal and a pressure monitoring target signal category, data conversion is performed to obtain a converted data stream, the converted data stream is input into a target neural network to obtain a target recognition result, and the target recognition result includes a signal correlation recognition result of the pressure monitoring target signal compared with the pressure monitoring reference signal, a signal category recognition result and a signal segment of interest recognition result. The application can perform multiple recognition branches of the same data source based on a single neural network, has high accuracy and processing efficiency. Efficient data processing technology is adopted to improve the flexibility of data use, facilitate long-term and continuous neural signal monitoring data processing and use, and achieve the technical effects of improved accuracy, reduced risk and lower power consumption.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and system for processing monitoring data for implantable medical pressure monitors. Background Technology

[0002] In modern medicine, especially in pressure monitoring related to implantable medical devices, accurate analysis and processing of pressure monitoring data is crucial for patient diagnosis, treatment, and health monitoring. Traditional pressure monitoring data processing methods often have several limitations. In the data acquisition and initial processing stages, there is a lack of a systematic and comprehensive approach to consider various key elements in the pressure monitoring process. For example, some existing methods may only focus on the target pressure signal itself, neglecting the importance of the control signal and the category of the target signal. The control signal serves as a reference standard, providing a benchmark for judging the normality of the target signal, while the target signal category helps in classifying and managing pressure signals. A lack of comprehensive consideration of these elements leads to incomplete data information, thus affecting the accuracy of subsequent analysis. Furthermore, traditional neural networks, when processing pressure monitoring data, often construct multiple independent networks or adopt a step-by-step processing approach for different processing tasks (such as signal correlation identification, signal category identification, and identification of signal segments of interest). This approach has significant drawbacks. First, there is the problem of error propagation; in the step-by-step processing, errors generated in previous steps may accumulate in subsequent steps, seriously affecting the accuracy of the final result. Secondly, independent networks or step-by-step processing methods cannot fully utilize the inherent correlations between data, resulting in slower processing speeds. When each task is processed individually, multiple features of the data cannot be comprehensively considered simultaneously, failing to maximize the utilization of data information and thus reducing overall processing efficiency. Summary of the Invention

[0003] In view of this, this application provides a method and system for processing monitoring data of implantable medical pressure monitors. The technical solution of this application is implemented as follows:

[0004] On one hand, this application provides a monitoring data processing method for implantable medical pressure monitors, comprising: acquiring a pressure monitoring control signal, a pressure monitoring target signal, and a pressure monitoring target signal category; performing data conversion on the pressure monitoring control signal, the pressure monitoring target signal, and the pressure monitoring target signal category to obtain a converted data stream; inputting the converted data stream into a target neural network to obtain a target recognition result, wherein the target recognition result includes a signal correlation recognition result, a signal category recognition result, and a signal segment of interest recognition result of the pressure monitoring target signal compared to the pressure monitoring control signal; wherein the target neural network is obtained by supervised learning of a base neural network based on the output result of the training converted data stream and training prior labels, the training converted data stream is obtained by data conversion of the training pressure monitoring control signal, the training pressure monitoring target signal, and the training pressure monitoring target signal category, and the training prior labels include a training correlation prior label, a training signal category prior label, and a training signal segment of interest prior label of the training pressure monitoring target signal compared to the training pressure monitoring control signal.

[0005] On the other hand, this application provides a computer system including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the program to implement the steps in the method described above.

[0006] The beneficial effects of this application include at least the following: this application can acquire pressure monitoring control signals, pressure monitoring target signals, and pressure monitoring target signal categories; convert the pressure monitoring control signals, pressure monitoring target signals, and pressure monitoring target signal categories into a converted data stream; input the converted data stream into a target neural network to obtain target recognition results, which include signal correlation recognition results, signal category recognition results, and signal segment of interest recognition results for the pressure monitoring target signal compared to the pressure monitoring control signals; wherein, the target neural network is obtained by supervised learning of the basic neural network based on the output results of the training converted data stream and training prior labels, the training converted data stream is obtained by converting the training pressure monitoring control signals, training pressure monitoring target signals, and training pressure monitoring target signal categories, and the training prior labels include training correlation prior labels, training signal category prior labels, and training signal segment of interest prior labels for the training pressure monitoring target signals compared to the training pressure monitoring control signals. Based on this, after acquiring the pressure monitoring control signal, the pressure monitoring target signal, and the pressure monitoring target signal category, this application uses data conversion to obtain a converted data stream as input to the target neural network. Furthermore, by combining the various network components in the target neural network, the data items in the converted data stream of the input data are processed simultaneously, enabling the network to comprehensively measure the characteristics of all data items in the converted data stream and strengthening the signal correlation between multiple processing branches. In addition, after processing the converted data stream, identification labels for multiple processing branches are obtained, such as the identification results of the signal segment of interest and signal category of the pressure monitoring target signal, as well as the signal correlation identification results between the pressure monitoring target signal and the pressure monitoring control signal. Based on this, executing multiple processing branches with the same signal correlation simultaneously using a single neural network can prevent error propagation and improve the accuracy and speed of processing. Attached Figure Description

[0007] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with this application and, together with the specification, serve to explain the technical solutions of this application.

[0008] Figure 1 This is a schematic diagram illustrating the implementation process of a monitoring data processing method for an implantable medical pressure monitor, as provided in an embodiment of this application.

[0009] Figure 2 This is a schematic diagram of the hardware entity of a computer system provided in an embodiment of this application. Detailed Implementation

[0010] This application provides a data processing method for implantable medical pressure monitors, which can be executed by a processor of a computer system. The computer system can refer to a device with data processing capabilities, such as a server, laptop, tablet, desktop computer, or mobile device.

[0011] Figure 1 This is a schematic diagram illustrating the implementation flow of a monitoring data processing method for implantable medical pressure monitors provided in an embodiment of this application, as shown below. Figure 1 As shown, the method includes:

[0012] Step S100: Obtain the pressure monitoring control signal, the pressure monitoring target signal, and the pressure monitoring target signal category.

[0013] A pressure monitoring control signal is a signal acquired by a computer system from a specific data source. In the context of implantable medical pressure monitoring, it can be considered a reference standard signal. For example, when monitoring intracardiac pressure, this control signal might be based on a known model of normal cardiac physiological pressure. The computer system can acquire this signal, for instance, through a communication interface with the implantable medical device. Assuming the implantable medical device uses a specific communication protocol, such as Bluetooth Low Energy (BLE), the computer system will receive the pressure monitoring control signal according to the specifications of that protocol. This signal may exist in the form of a digital signal, such as a series of discrete voltage values ​​or encoded numerical values.

[0014] The target pressure signal is also crucial data acquired by the computer system. It is the actual pressure signal directly collected from the implanted medical pressure monitor. Continuing with the example of cardiac pressure monitoring, this target signal is the electrical signal corresponding to the real-time pressure changes within the heart, monitored by a pressure sensor implanted near the heart. The principle by which the sensor converts pressure into an electrical signal is likely based on the piezoresistive effect; that is, when pressure acts on the sensitive element in the sensor, its resistance changes, causing a change in current or voltage. The computer system acquires this signal in a similar way to acquiring the control signal, through a communication link with the implanted medical device. However, because this target signal is an actually monitored signal, it may be affected by various factors, such as the patient's body movement and the influence of surrounding electromagnetic fields, so its value may fluctuate and contain noise.

[0015] A pressure monitoring target signal category is an identifier used to classify pressure monitoring target signals. In the medical field, different physiological processes or pathological states may correspond to different categories of pressure signals. For example, in the context of cardiac pressure monitoring, there may be pressure signal categories for a normal heartbeat, pressure signal categories for arrhythmias, and pressure signal categories for cardiac congestion. Computer systems can acquire this category information in several ways. One approach is to pre-set classification tags in implantable medical devices; when the device acquires a pressure signal, it simultaneously sends the corresponding category information to the computer system. Another approach is for the computer system to perform preliminary analysis of the acquired pressure monitoring target signal based on pre-stored algorithms and models to determine its possible category.

[0016] For example, suppose a classification algorithm determines the category by analyzing the frequency and amplitude characteristics of a pressure signal. Let the pressure signal be P(t), with frequency f and amplitude A. If f is within the normal heart rate range (e.g., the frequency range corresponding to 60-100 beats / minute), and A is within the normal pressure amplitude range (let's assume it's A0), then... min To A max If the pressure monitoring target signal is within the normal heartbeat range, then the computer system can initially classify this pressure monitoring target signal as a pressure signal category. The formula can be expressed as: if 60 ≤ f ≤ 100 (unit: frequency units converted from times / minute) and A min ≤A≤A max If so, the category is normal cardiac stress signal.

[0017] By acquiring the pressure monitoring control signal, the pressure monitoring target signal, and the category of the pressure monitoring target signal, the computer system lays the foundation for subsequent data processing steps. This data will be transformed and analyzed in subsequent steps to comprehensively and accurately evaluate the pressure data collected by the implantable medical pressure monitor. This includes determining the correlation between the target signal and the control signal, further accurately identifying the signal category, and identifying signal segments of interest, thereby providing valuable information for medical diagnosis and patient health monitoring.

[0018] Step S200: Convert the pressure monitoring control signal, the pressure monitoring target signal, and the pressure monitoring target signal category to obtain a converted data stream.

[0019] The pressure monitoring control signal, as a reference signal, has a specific format and numerical range. It is a pressure signal representation determined based on a certain standard or model. For example, in a cardiovascular implantable pressure monitoring system, the control signal might be an ideal pressure waveform signal constructed based on a large amount of normal cardiovascular pressure data, existing in a specific digital encoding form, such as representing the value of each pressure sampling point with a 16-bit binary number. The pressure monitoring target signal, on the other hand, is the actual pressure signal directly acquired from the implantable pressure monitor. Due to the influence of various factors such as the human physiological environment and the characteristics of the monitor itself, its data format and numerical distribution may differ from the control signal. It may contain more noise, fluctuations, and special characteristics caused by individual physiological differences. For example, the target signal may experience temporary abnormal fluctuations in pressure values ​​during acquisition due to slight movements of the patient, and its quantization accuracy may differ from the control signal. The pressure monitoring target signal category is a classification identifier for the target signal, which helps to differentiate between different types of pressure signals in subsequent processing. Different categories may correspond to different physiological or pathological conditions. For example, in pulmonary pressure monitoring, it can be divided into normal breathing pressure category, dyspnea pressure category, etc. This category information may exist in the form of digital codes or character identifiers.

[0020] When performing data conversion, computer systems need to comprehensively consider the characteristics of these three types of data. One technical approach is to use standardized data conversion algorithms. For the pressure monitoring control signal and the pressure monitoring target signal, since they are both pressure signal-related data, the computer system can first perform normalization processing on them. Assume the pressure monitoring control signal is... ,in This represents the value of the i-th sample point in the control signal, where n is the total number of sample points. The purpose of normalization is to map these values ​​to a specific interval, such as [0, 1]. The normalization formula can use the min-max normalization method: Where x is the original value (here it is...) X is the original data set (here, C), and x' is the normalized value. For pressure monitoring target signals... (where m is the total number of its sampling points), and this formula can also be used for normalization.

[0021] For pressure monitoring target signal categories, since their data format differs from that of pressure signals, the computer system needs to convert them into a format compatible with the converted pressure signal data. For example, if the category information exists in the form of character identifiers, such as "Normal" and "Abnormal", a unique numerical code can be assigned to each category, such as 0 for "Normal" and 1 for "Abnormal".

[0022] After processing these three types of data individually, the computer system will integrate and transform the data. This integration and transformation needs to consider the logical relationships between the different data and the requirements of subsequent processing. For example, to facilitate subsequent neural network processing, the computer system can arrange and combine these three types of data into a new data structure in a certain order. Assume the normalized pressure monitoring control signal is... The pressure monitoring target signal is The target signal for pressure monitoring is coded as L after category conversion. Computer systems can construct new data structures, such as two-dimensional arrays or vectors, and combine them together. For example, a vector D=[C', T', L] can be created, where D is the integrated data representation.

[0023] In practical applications, more complex transformations can be performed based on the specific neural network structure and data processing requirements. For example, if the neural network requires the input data to have specific dimensions and data types, the computer system can further adjust the integrated data. Suppose the neural network requires the input data to be a fixed-length vector, and the length of the currently integrated vector D does not meet the requirement. The computer system can use padding or data sampling techniques to adjust the data length. If the length of D is less than the required length, padding can be used by adding specific values ​​(such as 0) to the end of the vector to achieve the required length; if the length of D is greater than the required length, data sampling techniques, such as uniform sampling, can be used, selecting an element every certain number of elements to shorten the vector length.

[0024] Another technical approach is to employ data mapping technology. Based on predefined mapping rules, the pressure monitoring control signal, the pressure monitoring target signal, and the pressure monitoring target signal category are mapped to a completely new numerical space. For example, for the amplitude and frequency characteristics of a pressure signal, a mapping function can be established based on the known range and frequency range of physiological pressure signals, mapping the original amplitude and frequency values ​​to a new numerical range. This new numerical range may be more suitable for the input requirements of a neural network or more conducive to feature extraction from the data.

[0025] Through the series of data transformation operations described above, the computer system converts the pressure monitoring control signal, the pressure monitoring target signal, and the pressure monitoring target signal category into a unified, transformed data stream suitable for subsequent processing. This transformed data stream is then passed as input to the subsequent target neural network, providing fundamental data support for further signal analysis, recognition, and other operations. This enables effective processing of data from implantable medical pressure monitors, such as accurately determining the relationship between the target signal and the control signal, precisely identifying signal categories, and discovering signal segments of interest.

[0026] Step S300: Input the transformed data stream into the target neural network to obtain the target recognition result. The target recognition result includes the signal correlation recognition result of the pressure monitoring target signal compared with the pressure monitoring control signal, the signal category recognition result, and the signal segment of interest recognition result. The target neural network is obtained by supervised learning of the basic neural network based on the output result of the training transformed data stream and the training prior labels. The training transformed data stream is obtained by transforming the training pressure monitoring control signal, the training pressure monitoring target signal, and the training pressure monitoring target signal category. The training prior labels include the training correlation prior label of the training pressure monitoring target signal compared with the training pressure monitoring control signal, the training signal category prior label, and the training signal segment of interest prior label.

[0027] In this embodiment of the application, the target identification results include the signal correlation identification results of the pressure monitoring target signal compared with the pressure monitoring control signal, the signal category identification results, and the signal segment of interest identification results.

[0028] A target neural network is a specially trained artificial neural network model capable of performing complex analysis and processing of input transformed data streams. In medical stress monitoring scenarios, the structure and parameters of the neural network are designed and adjusted according to the specific needs of processing stress monitoring data. For example, it may contain multiple hidden layers, each consisting of multiple neurons connected by specific weights to extract and analyze features from the input data. When a computer system inputs a transformed data stream into the target neural network, this transformed data stream serves as the input vector or matrix and is processed according to the neural network's forward propagation algorithm. For example, if the transformed data stream is a vector containing multiple values... The first layer of neurons in a neural network will be based on a pre-set weight matrix. and bias vector The input data undergoes a linear transformation and activation function processing. The linear transformation formula is as follows: Then, the output of the first layer is obtained through an activation function (such as the ReLU function f(x) = max(0, x)). This process occurs sequentially in each layer of the neural network, with the output of each layer serving as the input for the next layer, until the final layer produces the output.

[0029] Signal correlation identification refers to the degree of correlation between a target pressure monitoring signal and a control pressure monitoring signal, determined by a computer system using a target neural network. For example, in cardiac pressure monitoring, if the control signal represents a normal pattern of systolic and diastolic pressure, and the target signal is the patient's actual cardiac pressure signal, the neural network might determine the correlation between the two by analyzing features such as waveform, amplitude, and frequency. If the waveform, amplitude, and frequency of the target signal are highly similar to the control signal, the correlation is high; otherwise, it is low. From a technical implementation perspective, the neural network might calculate correlation metrics such as the Pearson correlation coefficient. For two signal sequences... and Pearson correlation coefficient formula ,in These are the means of x and y, respectively. Neural networks may implicitly calculate similar metrics internally or determine correlations by learning feature patterns in the data.

[0030] Signal category recognition is the determination of the target signal category by the target neural network. Taking lung pressure monitoring as an example, if there are categories such as normal breathing pressure, expiratory pressure, and inspiratory pressure, the neural network classifies the target signal based on the feature information in the transformed data stream. During the training process, the neural network has learned the characteristic patterns of different categories of pressure signals. When a new transformed data stream is input, the neural network matches and judges based on these learned patterns. For example, if certain features in the transformed data stream (such as the rate of pressure change, the magnitude of peak pressure, etc.) match the characteristic patterns of the expiratory pressure category in the training data, then the neural network will classify the target signal as the expiratory pressure category.

[0031] The identification of segments of interest (SIOs) is the process by which a computer system uses a target neural network to identify specific segments within the target signal for pressure monitoring. For example, when monitoring intracranial pressure, researchers or physicians may be interested in signal segments related to specific physiological events (such as changes in intracranial pressure during abnormal brainwave discharges). The neural network identifies these significant signal segments by analyzing the transformed data stream. This is based, for example, on certain abrupt changes in the signal or its correlation with other physiological signals (such as EEG signals). For instance, if a sudden increase or decrease in intracranial pressure occurs within a certain timeframe and is temporally correlated with abnormal fluctuations in the EEG signal, then the pressure signal segment during that timeframe might be identified as an SIO.

[0032] The training process of the target neural network is crucial for accurately obtaining these recognition results. During training, a large amount of training transformed data streams and corresponding training prior labels are used. This training data enables the neural network to learn the mapping relationship between input data (similar to transformed data streams) and desired outputs (such as signal correlation, signal category, and signal segments of interest). When faced with new transformed data streams, the neural network can accurately identify and judge based on the previously learned mapping relationship.

[0033] As one implementation method, step S300 involves inputting the converted data stream into the target neural network to obtain the target recognition result, including:

[0034] Step S310: Based on the target neural network, perform information enhancement representation on each data item in the transformed data stream to obtain a set of implicit representations of data items. The set of implicit representations of data items includes the implicit representation of each data item.

[0035] Step S320: Integrate and map the implicit representation of each data item with the implicit representation set of data items in the forward and backward transmissions to obtain the set of recognition tag support corresponding to the implicit representation of each data item;

[0036] Step S330: Based on the multiple label types in each recognition label support set and the label support corresponding to each label type, determine the target label trajectory. The target label trajectory is used to instruct the target neural network to output the target recognition result.

[0037] In step S310, each data item in the transformed data stream undergoes special processing to obtain its implicit representation set. This transformed data stream is the data format suitable for input to the target neural network after processing in the previous steps; it contains various information related to stress monitoring. For each data item in the transformed data stream, the target neural network employs a specific mechanism for information reinforcement representation.

[0038] Taking cardiac pressure data from medical stress monitoring as an example, the data items in the transformed data stream include pressure values ​​collected at different time points and related auxiliary information (such as pre-processed feature values ​​related to the pressure signal). The computer system processes these data items using the structure and algorithm of the target neural network. One technique involves using neuron weights and activation functions within the neural network. Assuming a data item in the transformed data stream is x, and a neuron in a certain layer of the target neural network processes it with weight w and bias b, then after linear combination, we obtain z = wx + b. Then, an activation function (such as the sigmoid function) is used. The result y=f(z) after enhanced representation is obtained. This result y is a representation of the data item after enhanced representation, that is, the implicit representation of the data item. For each data item in the transformed data stream, the computer system performs this operation to obtain a set of implicit representations of the data item. Each element in this set corresponds to the implicit representation of a data item in the transformed data stream. It contains the feature information of the original data item after processing by the neural network. This feature information is more suitable for subsequent analysis of the pressure monitoring target signal, such as judging the correlation with the control signal, identifying the signal category, and finding the signal segment of interest.

[0039] In step S320, after obtaining the set of implicit representations of data items, the computer system performs this step to further explore the relationships between the implicit representations of data items and perform integration transformation. Here, bidirectional feature propagation is an information transmission mechanism in the target neural network that can comprehensively consider the information of the implicit representations of data items in the forward and backward directions. For example, when analyzing continuous cardiac pressure monitoring data, the implicit representation of a data item at a certain moment is not only related to the implicit representations of other data items at the current moment, but also to the implicit representations of data items at previous and subsequent moments. The computer system implements this integration and mapping transformation through a specific algorithm. Assume the set of implicit representations of data items is... For one of the data items, implicitly represented The computer system will integrate this with the implicit representations associated with it in the preceding and following processes. One possible integration method is weighted summation, for example, by assigning an implicit representation to the current data item. Assign a weight 'a' to the relevant implicit representations in the forward and backward propagation. If we assign weights b and c respectively, then the integrated result is: .

[0040] Then, the computer system performs a mapping transformation on the integrated result, for example, using a linear mapping function f(z) = kz + d (where k and d are pre-defined parameters), to obtain a numerical value that represents the implicit representation of the data item. For each identification marker, the support level is calculated. This process is repeated for different identification markers (e.g., signal correlation markers, signal category markers, signal segment of interest markers, etc.) to obtain the implicit support level set for each data item. Each element in this set corresponds to the support level of an identification marker, reflecting the implicit support level of the data item for each marker, considering the information passed before and after. This helps determine the target identification result in subsequent steps.

[0041] In step S330, the computer system determines the target marker trajectory based on the set of recognition marker support obtained in the previous steps. The set of recognition marker support includes multiple marker types (such as signal correlation recognition markers, signal category recognition markers, and signal segment of interest recognition markers) and the marker support corresponding to each marker type.

[0042] Taking lung pressure monitoring as an example, suppose there are three label types in the label support set: normal breathing pressure category label, expiratory obstruction pressure category label, and inspiratory obstruction pressure category label, and their corresponding label support are respectively The computer system needs to comprehensively consider these tag types and tag support to determine the target tag trajectory.

[0043] One technical approach is to construct a labeled trajectory evaluation function. For example, for a labeled trajectory... (in For the tag type), the evaluation function can be defined as (in Is with tag type Related weights, It is a tag type (Label support in the set of identified label support). The computer system determines the target label trajectory by calculating the evaluation function values ​​of different label trajectories.

[0044] In practice, multiple candidate labeled trajectories may be available. For example, when analyzing cardiac stress data, one candidate labeled trajectory might first determine that the signal correlation is high (corresponding to one label type and label support), then identify the signal category as normal cardiac stress (corresponding to another label type and label support), and finally determine the signal segment of interest as a specific time period (corresponding to a third label type and label support). Another candidate labeled trajectory might first determine that the signal correlation is low, then identify the signal category as abnormal cardiac stress, and then determine the signal segment of interest as another time period. The computer system calculates the evaluation function value of each candidate labeled trajectory and selects the labeled trajectory with the optimal evaluation function value as the target labeled trajectory. This target labeled trajectory instructs the target neural network to output target recognition results, such as accurate signal correlation recognition results, signal category recognition results, and signal segment of interest recognition results, thereby providing accurate and valuable information for implantable medical stress monitoring.

[0045] In this process, the computer system must handle a large amount of complex data relationships and calculations. It must consider not only the support values ​​of each identification marker set, but also the logical relationships between marker types and their significance within the overall medical stress monitoring scenario. For example, the determination of signal correlation may affect the subsequent identification of signal categories, which in turn may be related to the identification of the signal segment of interest. Through carefully designed algorithms and models, the computer system finds the target marker trajectory that best matches the characteristics and requirements of stress monitoring data from numerous possible marker trajectories, which is crucial for accurate medical diagnosis and monitoring.

[0046] Through steps S310-S330, the computer system processes the transformed data stream step by step within the framework of the target neural network. This process involves strengthening the representation of information for each data item, integrating information from the preceding and following transmissions to obtain a set of recognition label support, and then determining the target label trajectory to output the target recognition result. This enables in-depth analysis and accurate identification of monitoring data from implantable medical pressure monitors, providing effective technical support for pressure monitoring in the medical field.

[0047] As one implementation method, the target neural network includes a feature extraction component. In step S310, each data item in the transformed data stream is subjected to information enhancement representation to obtain an implicit representation set of the data items, including:

[0048] Step S311: Perform implicit representation transformation on each data item in the transformed data stream based on the feature extraction component to obtain an implicit representation set, which includes an implicit representation array corresponding to each data item;

[0049] Step S312: Based on the forward and backward features of each implicit representation array in the implicit representation set, perform feature extraction on each implicit representation array to obtain the implicit representation set of data items after information enhancement representation.

[0050] When the computer system executes step S311, it performs implicit representation transformation on each data item in the transformed data stream based on the feature extraction component in the target neural network, resulting in an implicit representation set. The transformed data stream is the previously processed data format, where each data item contains information related to implantable medical stress monitoring. Taking cardiovascular stress monitoring as an example, the data items in the transformed data stream may include stress values ​​collected at different time points, stress feature values ​​after preprocessing, etc. The feature extraction component is a part of the target neural network specifically used to extract data features; it can contain multiple neurons and corresponding connection weights.

[0051] When a computer system processes a data item, let's say this data item is x, the neuron weight matrix in the feature extraction component is W, and the bias vector is b. The computer system calculates z = Wx + b through linear combination. This step essentially performs a preliminary transformation of the original data item x within the space of the feature extraction component. This z is the result of the linear transformation of the data item x; it's a vector containing multiple elements, each corresponding to a neuron in the feature extraction component. These linearly transformed results z then form the elements in the implicit representation set. For each data item in the transformed data stream, the computer system repeats this process, resulting in an implicit representation set. Each element in this set is an implicit representation array, a representation of the corresponding data item after the preliminary processing by the feature extraction component. This representation, compared to the original data item, begins to exhibit features relevant to subsequent processing by the target neural network. For example, when processing cardiac pressure monitoring data, if a data item in the transformed data stream is the cardiac pressure value at a certain moment and related auxiliary feature values, after the above processing by the feature extraction component, the resulting implicit representation array may contain feature information related to pressure change trends, pressure fluctuation frequency, etc., and this information exists in the implicit representation array in the form of a vector.

[0052] Next, the computer system executes step S312, which extracts features (e.g., encodes) for each implicit representation array based on the forward and backward features of each implicit representation array in the implicit representation set, to obtain the implicit representation set of data items after information enhancement representation.

[0053] In an implicit representation set, each implicit representation array contains certain feature information. This feature information not only has its own numerical characteristics but also forward and backward correlation characteristics. Taking cardiac pressure monitoring as an example mentioned earlier, an implicit representation array may represent the changes in cardiac pressure within a certain time period, and its forward and backward characteristics may reflect the relationship with the changes in pressure data in adjacent time periods.

[0054] To further mine this feature information, computer systems employ feature extraction (encoding) techniques. A common technique is to use the convolutional layer principle of Convolutional Neural Networks (CNNs) for encoding operations. Assume the implicit representation array is A, the convolutional kernel is K, and the stride is s. The computer system calculates this through convolutional operations. Convolutional operations can effectively extract local features from the implicit representation array, and by using multiple convolutional kernels, different types of features can be obtained.

[0055] Besides convolution operations, the principles of recurrent neural networks (RNNs) can also be used to consider forward and backward features. For an implicit representation of the sequence of elements in an array, let's assume it's a1, a2, ..., a... n The hidden state update formula in RNN is h t =f(Ua t +Wh t-1 ), where U and W are weight matrices, h t It is the hidden state at the current moment, h t-1 is the hidden state from the previous time step, and f is the activation function (such as the tanh function). In this way, forward and backward features can be incorporated into the feature extraction process.

[0056] Through these feature extraction (such as encoding) operations, the computer system processes each implicit representation array in the implicit representation set, resulting in a set of implicit representations of data items after information enhancement representation. Each implicit representation of a data item in this set, compared to the implicit representation array obtained in step S311, delves deeper into the feature information within the data. This feature information is more aligned with the subsequent processing needs of the target neural network, such as integration with other implicit representations of data items in later steps and determination of recognition label support. This lays the foundation for accurately analyzing and identifying the monitoring data from implantable medical pressure monitors.

[0057] As one implementation, the target neural network further includes an LSTM component connected to the feature extraction component. Step S320 involves integrating and mapping the implicit representation of each data item with the implicit representation set of data items in the preceding and following passes to obtain a set of recognition label support corresponding to each data item's implicit representation, including:

[0058] Step S321: Based on the LSTM component, perform a preorder pass on the implicit representations of each data item in the implicit representation set to obtain the forward implicit representation corresponding to each data item implicit representation;

[0059] Step S322: Pass the implicit representations of each data item in the implicit representation set in the next order to obtain the backward implicit representation corresponding to each data item implicit representation;

[0060] Step S323: Integrate the implicit representation of each data item with the corresponding forward implicit representation and the corresponding backward implicit representation to obtain the integrated implicit representation corresponding to the implicit representation of each data item;

[0061] Step S324: Map the integrated implicit representation to obtain the set of recognition tag support corresponding to the implicit representation of each data item.

[0062] When the computer system executes step S321, it performs a preorder pass on the implicit representations of each data item in the implicit representation set based on the LSTM component in the target neural network, obtaining the forward implicit representation corresponding to each data item implicit representation. The LSTM (Long Short-Term Memory) component is a special neural network structure, particularly suitable for processing data with time-series characteristics. When processing implantable medical stress monitoring data, it can effectively capture the sequential relationship between the implicit representations of data items.

[0063] Taking cardiac pressure monitoring data as an example, assume that the data items in the set are implicitly represented and arranged in chronological order, reflecting the characteristic information related to cardiac pressure at different times. For one of the data items, the implicit representation h... i LSTM components perform preorder pass operations through their internal gating mechanism. LSTM has input gates i t Forgotten Gate t and output gate o t The calculation process is as follows:

[0064] First, the forget gate determines which information to discard from the cell state, using the following formula: W f It is the weight matrix of the forget gate, h i-1 It is the implicit representation of the preceding data item (the implicit representation of the preceding data item in the time series), x t It is the input at the current moment (here, the implicit representation of the data item h). i ), b f It is a bias vector. It is the Sigmoid activation function.

[0065] Next, the input gate determines which values ​​to update, calculated as follows: W i W C This is the corresponding weight matrix, b i b C It is the bias vector.

[0066] Then, update the cell state. .

[0067] Finally, the output gate determines which values ​​to output, using the following formula: The implicit representation h of the current data item is obtained. i The corresponding forward implicit representation h i,f =o t *tanh(C t ).

[0068] In this way, for each implicit representation of a data item in the set of implicit representations, the computer system can obtain its corresponding forward implicit representation. These forward implicit representations contain the influence of the preceding information of the data item's implicit representation in the time series and are an important component of subsequent integration operations.

[0069] After completing step S321, the computer system executes step S322, which performs a backward pass on the implicit representations of each data item in the implicit representation set to obtain the backward implicit representation corresponding to each data item implicit representation. This step is similar to the forward pass, but in the opposite direction, and also utilizes the characteristics of the LSTM component to capture the reverse order relationship between the implicit representations of data items. For example, for the same implicit representation h... i Now, we perform similar calculations from the end to the beginning. Assuming the data is in reverse chronological order, the next data item is implicitly represented as h_{i+1} (the implicit representation of the subsequent data item in the time series). Using an LSTM gating mechanism similar to step S321, we obtain h. i The corresponding backward implicit representation h i,b This backward implicit representation includes the impact of the implicit representation of data items on subsequent information in the time series, providing another piece of information for comprehensively considering the sequential relationship of the implicit representation of data items.

[0070] When executing step S323, the computer system integrates the implicit representation of each data item with its corresponding forward implicit representation and its corresponding backward implicit representation to obtain the integrated implicit representation corresponding to each data item's implicit representation. This integration operation is to comprehensively consider all the information passed between the preceding and following sequences of the data item's implicit representation.

[0071] Continuing with cardiac stress monitoring data as an example, for the data item implicitly represented as h... i Its corresponding forward implicit representation is h i,f The backward implicit representation is h i,b Computer systems can be integrated using a weighted summation method. Let the integrated result be implicitly represented as h. i,integrated Then there can be h i,integrated =αh i +βh i,f +γh i,bHere, α, β, and γ are pre-defined weighting coefficients that can be adjusted based on the characteristics of the data and processing requirements. For example, if forward information is considered more important, the value of β can be increased; if backward information is more critical, the value of γ can be increased; if the implicit representation of the original data item itself has a high weight, the value of α can be increased. Through this integration method, each data item's implicit representation incorporates its own original information as well as relevant information from preceding and following sequences, forming a more comprehensive integrated implicit representation that reflects the data characteristics.

[0072] Finally, the computer system executes step S324, mapping the integrated implicit representation (e.g., using a sigmoid function) to obtain the set of support values ​​for the recognition markers corresponding to the implicit representation of each data item. The purpose of the mapping operation is to convert the integrated implicit representation into support values ​​associated with the recognition markers, so as to subsequently determine the trajectory of the target marker.

[0073] Assume the integrated implicit representation is h i,integrated The Sigmoid function is used for mapping. For different identification markers (e.g., signal correlation markers, signal category markers, markers for segments of interest, etc.), h is... i,integrated The values ​​are input into the corresponding mapping functions to obtain the support scores for different recognition labels. For example, for signal correlation recognition labels, the support score s is obtained. i,correlation =f(h i,integrated For signal category identification labels, the support s is obtained. i,category =f(h i,integrated For the signal segment of interest, identify the label and obtain the support s. i,segment =f(h i,integrated These support values ​​constitute the set of support values ​​for the recognition markers corresponding to the implicit representation of each data item. Each element in this set represents the degree of support that the implicit representation of the data item provides for different recognition markers after considering the preceding and following information and integrating and mapping it. This provides an important basis for determining the target marker trajectory based on these support values ​​in step S330.

[0074] Throughout the entire implantable medical pressure monitoring data processing process, steps S321-S324 use the LSTM component to perform sequential transmission, integration, and mapping operations on the implicit representation of data items, deeply mining the sequential relationships and feature information in the data. This allows the obtained recognition label support set to more accurately reflect the relationship between the data and different recognition labels, thereby helping to improve the accuracy of the target neural network in analyzing and recognizing pressure monitoring data, and providing a more reliable basis for medical diagnosis and monitoring.

[0075] As one implementation, the target neural network further includes a label mapping component connected to the LSTM component. Step S330, determining the target label trajectory based on multiple label types in each recognition label support set and the label support corresponding to each label type, includes:

[0076] Step S331: Based on the data item position of the transformed data stream, the tag mapping component combines any tag type selected from each identification tag support set to obtain multiple candidate tag trajectories. Each candidate tag trajectory includes the tag type of each data item.

[0077] Step S332: Based on the multiple label types in each identification label support set and the label support corresponding to each label type, determine the label support of each label type represented in each candidate label trajectory;

[0078] Step S333: Based on the label support of each label type represented in each candidate label trajectory and the label dependency between adjacent label types, determine the trajectory support of the corresponding candidate label trajectory;

[0079] Step S334: Based on the trajectory support value, determine the target marker trajectory from multiple candidate marker trajectories.

[0080] When the computer system executes step S331, based on the label mapping component, it combines any label type selected from each identification label support set according to the data item position in the transformed data stream, resulting in multiple candidate label trajectories. The label mapping component (such as a classifier) ​​plays a crucial role in the target neural network by mapping data features to different label types. The data item position in the transformed data stream contains sequence information, which is critical for constructing label trajectories because different data item positions may correspond to different stress monitoring states or stages. Taking cardiac stress monitoring as an example, suppose the identification label support set includes signal correlation identification labels (label type 1), signal category identification labels (label type 2, such as normal heartbeat category, abnormal heartbeat category, etc.), and signal segment of interest identification labels (label type 3). The data items in the transformed data stream correspond to the results of cardiac stress data collected at different times after processing in the previous steps. For each data item position, the computer system selects a label type from the identification label support set for combination. For example, for the first data item location, the computer system can choose a signal correlation identification marker; for the second data item location, it can choose a signal category identification marker such as "normal heartbeat category"; and for the third data item location, it can choose a marker for a specific signal segment of interest. This forms a candidate marker trajectory. By selecting different combinations of marker types for each data item location, the computer system can obtain multiple candidate marker trajectories. One technique is to use an exhaustive method, calculating all possible combinations based on the number of marker types in the identification marker support set and the number of data items in the transformed data stream, thus obtaining all candidate marker trajectories.

[0081] Next, the computer system executes step S332, determining the label support for each label type represented in each candidate label trajectory based on multiple label types in each identification label support set and the label support corresponding to each label type. The label support in each identification label support set reflects the degree of support that each data item implicitly represents for the corresponding label type. For each candidate label trajectory, the computer system needs to determine the label support corresponding to each label type.

[0082] For example, in the cardiac pressure monitoring example above, a candidate label trajectory might be [signal correlation identification label (support s1), normal heartbeat category (support s2), specific signal segment of interest (support s3)], where s1, s2, and s3 are obtained from the corresponding identification label support sets. The computer system can determine the label support of each label type in the candidate label trajectory by looking up the identification label support set for the corresponding data item location. This step is crucial for evaluating the rationality and accuracy of each candidate label trajectory, as label support reflects the strength of the association between the data and the label type. Then, the computer system executes step S333, determining the trajectory support of the corresponding candidate label trajectory based on the label support of each label type represented in each candidate label trajectory and the label dependency relationship between adjacent label types (referring to the transformation rules or constraints between adjacent labels in the label trajectory). The label dependency relationship reflects the logical connection between different label types.

[0083] For example, in cardiac stress monitoring, if the signal correlation identification markers show low correlation, then the abnormal heartbeat category is more likely to appear in the signal category identification markers; this is a type of label dependency. For a candidate label trajectory, the computer system calculates its trajectory support. Suppose a candidate label trajectory is [low correlation (support s1), abnormal heartbeat category (support s2), specific signal segment of interest (support s3)], and let the weight matrix of the label dependency be W, and the dependency function between label types be f. The computer system can first calculate the dependency score d = f(s1, s2, s3, W) based on the label dependency. Here, the function f can be designed according to the specific label dependency definition. For example, if s1 is low correlation and s2 is the abnormal heartbeat category, d may get a higher score because this matches the expected label dependency. Then, the computer system can combine the label support for each label type with the dependency score to obtain the trajectory support S=g(s1, s2, s3, d), where g is a comprehensive function, which can be in the form of weighted summation, such as S=αs1+βs2+γs3+δd, where α, β, γ and δ are pre-set weight coefficients.

[0084] Finally, the computer system executes step S334, determining the target marker trajectory from multiple candidate marker trajectories based on the trajectory support value. The target marker trajectory is the one that best reflects the characteristics of the pressure monitoring data and conforms to logical relationships.

[0085] For example, the computer system calculates the trajectory support of multiple candidate labeled trajectories, namely S1, S2, S3, etc. If the value of S1 is the largest among all trajectory support values, then the corresponding candidate labeled trajectory is the target labeled trajectory. This target labeled trajectory will be used to instruct the target neural network to output target recognition results, such as accurate signal correlation recognition results, signal category recognition results, and signal segment of interest recognition results.

[0086] Throughout the medical stress monitoring data processing, steps S331-S334, by constructing candidate labeled trajectories, determining the support of each label type within the trajectory, calculating trajectory support considering label dependencies, and selecting the target labeled trajectory, enable the computer system to find the result that best matches the characteristics and logical relationships of the stress monitoring data from numerous possibilities. This helps improve the accuracy of the target neural network in analyzing implantable medical stress monitoring data, providing more reliable and valuable information for medical diagnosis and treatment. For example, in brain stress monitoring, accurate target labeled trajectories can help doctors determine the presence of brain lesions, the type of lesion, and critical time periods related to stress changes, which plays an irreplaceable role in developing appropriate treatment plans.

[0087] As one implementation, step S333, based on the label support of each label type represented in each candidate label trajectory and the label dependency between adjacent label types, calculates the trajectory support of each candidate label trajectory, including:

[0088] Step S3331: Determine the label state support corresponding to each candidate label trajectory based on the label support in each recognition label support set;

[0089] Step S3332: Determine the label dependency support based on the label dependency relationships among multiple label types in each candidate label trajectory;

[0090] Step S3333: Determine the trajectory support based on the label state support and label dependency support corresponding to each candidate label trajectory.

[0091] When executing step S3331, the computer system determines the label state support corresponding to each candidate label trajectory based on the label support in each label support set. The label support in the label support set reflects the degree of support of the implicit representation of the data item for each label type, while the label state support is a comprehensive consideration of the label support from the perspective of the entire candidate label trajectory.

[0092] Taking cardiac stress monitoring as an example, suppose a candidate marker trajectory contains three marker types: signal correlation identification markers, signal category identification markers (e.g., normal or abnormal heartbeat categories), and signal segment of interest identification markers. For each marker type in this candidate marker trajectory, there is a corresponding marker support, denoted as s1, s2, and s3, respectively. The computer system can use a weighted summation method to determine the marker state support. Let the weight coefficients be α, β, and γ (these weight coefficients can be preset based on experience or data analysis, for example, determined according to the importance of different marker types in the overall judgment), then the marker state support S state It can be done through formula S state =αs1+βs2+γs3 is calculated. This calculation method integrates the support of each label type in the candidate label trajectory, so that the label state support can reflect the rationality of the entire candidate label trajectory based solely on label support.

[0093] For example, if the support s1 of the signal correlation identification label is high, it indicates strong evidence supporting the signal correlation; if the support s2 of the signal category identification label is also high, it indicates a good match between the data and the identified signal category; and if the support s3 of the signal segment of interest identification label is also high, it means that the data also has some basis for identifying the signal segment of interest. The support S of the label state is obtained through weighted summation. state This can comprehensively reflect the overall support level of the alternative marker trajectory in these aspects.

[0094] Next, the computer system executes step S3332, determining the label dependency support based on the label dependencies between multiple label types in each candidate label trajectory. Label dependencies reflect the transformation rules or constraints between adjacent labels in a label trajectory, while label dependency support is a quantitative assessment of whether the candidate label trajectories conform to these logical relationships.

[0095] Taking cardiac stress monitoring as an example, assuming that under normal physiological conditions, if the signal correlation identification marker shows a high correlation, then the signal category identification marker is more likely to be a normal heartbeat category; this is a marker dependency. For a candidate marker trajectory, the computer system needs to evaluate whether the marker types conform to this dependency. Suppose there is a function f to quantify this marker dependency; the input of this function is the marker type and its corresponding marker support, and the output is a numerical value representing the degree to which the marker dependency conforms.

[0096] For example, given a candidate labeled trajectory [high correlation (support s1), normal heartbeat category (support s2)], function f might be calculated according to predefined rules. If s1 is highly correlated and s2 is the normal heartbeat category, function f might output a higher value, indicating that the candidate labeled trajectory has a good fit in terms of label dependency; conversely, if s1 is highly correlated but s2 is the abnormal heartbeat category, function f might output a lower value. By applying function f to all adjacent label type pairs in the candidate labeled trajectory and combining the results (e.g., summing or weighted summation), the computer system can obtain the label dependency support S. dependency .

[0097] Finally, the computer system executes step S3333, determining the trajectory support based on the label state support and label dependency support corresponding to each candidate label trajectory. Trajectory support is a comprehensive evaluation index for candidate label trajectories, which comprehensively considers label state support (a comprehensive consideration based on label support) and label dependency support (a consideration based on the logical relationship between label types).

[0098] For example, let the support of the labeled state be S. state The tag dependency support is S. dependency Computer systems can use a linear combination approach to determine the trajectory support S, such as S = λS. state +μS dependency Here, λ and μ are pre-defined weighting coefficients. These two weighting coefficients can be adjusted according to the specific application scenario and data characteristics. If a medical stress monitoring scenario places more emphasis on direct evidence based on label support, then the value of λ can be appropriately increased; if the logical relationship between label types is emphasized, then the value of μ can be increased. Taking lung stress monitoring as an example, suppose there is a candidate labeled trajectory with label state support S. state This indicates that the trajectory has some reasonableness in terms of individual support for each label type, and the label-dependent support S dependency This indicates that the trajectory's logical transformation between marker types is also relatively reasonable. The trajectory support S, calculated using the above formula, comprehensively reflects the overall rationality and accuracy of the candidate marker trajectory, providing an important basis for the subsequent computer system to determine the target marker trajectory from multiple candidate marker trajectories.

[0099] Throughout the implantable medical stress monitoring data processing, steps S3331-S3333 determine the labeled state support and labeled dependency support, ultimately obtaining the trajectory support. This allows the computer system to more comprehensively and deeply evaluate the quality of each candidate labeled trajectory. This evaluation method considers both the labeled support itself and the logical relationship between labeled types, helping to improve the accuracy of the target recognition results output by the target neural network, thereby providing more reliable and valuable information for medical diagnosis and treatment.

[0100] In one implementation, step S200 involves converting the pressure monitoring control signal, the pressure monitoring target signal, and the pressure monitoring target signal category to obtain a converted data stream, including:

[0101] Step S210: Merge the pressure monitoring control signal, the pressure monitoring target signal, and the pressure monitoring target signal category to obtain merged data;

[0102] Step S220: Split the merged data into data items to obtain the transformed data stream.

[0103] In the specific implementation of step S200, which converts the pressure monitoring control signal, the pressure monitoring target signal, and the pressure monitoring target signal category to obtain a converted data stream, steps S210 and S220 are included. This is an important process for the computer system to integrate and convert different types of pressure monitoring related data.

[0104] When executing step S210, the computer system merges the pressure monitoring control signal, the pressure monitoring target signal, and the pressure monitoring target signal category to obtain merged data. The pressure monitoring control signal is pressure signal-related data used as a reference standard. For example, in a cardiac pressure monitoring scenario, it might be an ideal pressure waveform signal constructed based on a large amount of normal cardiac pressure data. The pressure monitoring target signal is the pressure signal data actually collected from the implanted pressure monitor, which is affected by various factors and will differ from the control signal. The pressure monitoring target signal category is an identifier for classifying the target signal. For example, in pulmonary pressure monitoring, it can be divided into normal breathing pressure category, dyspnea pressure category, etc.

[0105] Taking cardiac stress monitoring as an example, suppose the stress monitoring control signal is a vector C={c1, c2, …, c} containing n sampling points. n Each sampling point represents a pressure value or pressure-related characteristic value at a specific moment; the pressure monitoring target signal is a vector T = {t1, t2, …, t} containing m sampling points. mThe target signal category for pressure monitoring is represented by a numerical code. Assuming normal cardiac pressure is category 0, abnormal cardiac pressure is category 1, and the currently monitored target signal category is L=0, the computer system merges these three types of data into a new data structure. One merging method is to arrange them sequentially into a new vector, for example, M={C, T, L}, where M represents the merged data. Of course, in practical applications, the merging method may be adjusted based on subsequent processing needs and data characteristics. For example, if the data formats of the pressure monitoring control signal and the target signal differ (e.g., different sampling frequencies), the computer system can first interpolate or decimate one signal to match it with the other signal in the time dimension before merging.

[0106] Next, the computer system executes step S220, which splits the merged data into data items to obtain a transformed data stream. Data item splitting involves breaking down the merged data into smaller data units according to specific rules or requirements. The data stream composed of these data units is the transformed data stream, which is more suitable for subsequent processing, such as inputting it into a target neural network.

[0107] Continuing with the cardiac stress monitoring example above, for the merged data M={C, T, L}, assume the computer system needs to split it into data items for each sampling point. For the stress monitoring control signal portion C={c1,c2,…,c… n After splitting, n data items are obtained; for the pressure monitoring target signal part T={t1, t2,…, t…}, the result is n data items. m After splitting, we get m data items; while for the pressure monitoring target signal category L, since it is a single encoded value, it can be regarded as a data item. Thus, the transformed data stream after splitting contains n+m+1 data items.

[0108] In practice, data item splitting may involve more complex technical methods. For example, if the merged data contains different types of data (such as numerical stress values ​​and categorical codes), the computer system can use tokenization techniques. For numerical data, it can be split according to a certain numerical range or precision. Assuming the stress value range is [a, b], the computer system can divide this range into k intervals, each interval corresponding to a data item. For discrete data such as categorical codes, it can be directly treated as an independent data item. In addition, to make the transformed data stream more suitable for the input requirements of the target neural network, the computer system may also need to standardize or normalize the split data items. For example, for numerical data item x, min-max normalization can be used to map the value of the data item to the interval [0, 1]. This results in a more uniform numerical range for the data items in the transformed data stream, which is beneficial for the target neural network to perform subsequent processing, such as feature extraction and classification.

[0109] Through steps S210 and S220, the computer system converts the pressure monitoring control signal, the pressure monitoring target signal, and the pressure monitoring target signal category from their original independent data forms into a transformed data stream suitable for the input of the target neural network. This transformation process comprehensively considers various factors such as the characteristics of different data, the needs of subsequent processing, and the input requirements of the neural network, laying the foundation for accurate analysis of the pressure monitoring data in subsequent steps. For example, in step S300, the transformed data stream is processed by the target neural network to obtain important information such as the signal correlation identification result, signal category identification result, and signal segment of interest identification result of the pressure monitoring target signal compared with the pressure monitoring control signal.

[0110] As one implementation method, the method also includes a network training step, specifically including:

[0111] Step S10: Obtain the training stress monitoring control signal, the training stress monitoring target signal, and the training stress monitoring target signal category;

[0112] Step S20: Convert the training stress monitoring control signal, the training stress monitoring target signal, and the training stress monitoring target signal category data to obtain a training conversion data stream;

[0113] Step S30: Input the training transformation data stream into the basic neural network and obtain multiple alternative label trajectories obtained by the basic neural network when processing the training transformation data stream;

[0114] Step S40: Obtain training prior labels, and instruct the basic neural network based on the training prior labels to determine the target candidate label trajectory among multiple candidate label trajectories, so as to output the target inference and recognition result;

[0115] Step S50: Generate the target training cost based on the proportion of trajectory support of the target reasoning and recognition results, and repeatedly debug the basic neural network based on the target training cost. When the preset evaluation conditions are met, the debugged target neural network is obtained.

[0116] The computer system first executes step S10 to acquire training stress monitoring control signals, training stress monitoring target signals, and training stress monitoring target signal categories. In medical stress monitoring scenarios, this training data is crucial for the neural network to learn the characteristics and patterns of stress signals. For example, in a cardiac stress monitoring training dataset, the training stress monitoring control signals are, for instance, standard stress signal patterns constructed by screening cardiac stress data from a large number of healthy individuals. These patterns reflect the ideal pressure changes of a normal heart under different physiological states. These control signals can be preprocessed digital signals stored in a specific data structure, such as a vector, where each element represents the pressure value at a specific moment or a stress-related feature value. The training stress monitoring target signals are cardiac stress signals actually collected from different individuals (including healthy individuals and those with specific heart diseases). These signals contain rich information on individual differences and pressure changes under various physiological and pathological states. The training stress monitoring target signal category is used to identify the category to which the target signal belongs, such as the normal cardiac stress category or the stress category corresponding to different types of heart disease (such as coronary heart disease, cardiomyopathy, etc.). This category information helps the neural network learn to distinguish different types of stress signals.

[0117] Next, the computer system executes step S20, converting the training stress monitoring control signal, the training stress monitoring target signal, and the training stress monitoring target signal category data into a training transformation data stream. This step is similar to the operation in step S200, and is to convert different types of training data into a uniform data format suitable for neural network processing. For example, the training stress monitoring control signal and the target signal can be normalized to ensure that the data are within the same numerical range. Assume the training stress monitoring control signal is C={c1, c2,…, c…} n The data can be normalized to the [0, 1] interval using min-max normalization. For the target signal category of the training stress monitoring, if the category name is represented in character form (such as "Normal", "Disease1", etc.), it can be converted into a numerical code (such as 0 for normal, 1 for disease 1, etc.). Then, these processed control signals, target signals, and target signal categories are merged and split according to certain rules to obtain the training transformation data stream.

[0118] In step S30, the computer system inputs the training-transformed data stream into the basic neural network and obtains multiple candidate labeled trajectories obtained by the basic neural network when processing the training-transformed data stream. The basic neural network is an initial neural network structure with a certain number of neuron layers, connection weights, and activation functions. When the training-transformed data stream is input, the basic neural network performs calculations according to its internal forward propagation algorithm. Taking the aforementioned cardiac pressure monitoring as an example, assume that the input layer of the basic neural network receives the training-transformed data stream, and then performs feature extraction and information transmission through the hidden layers. Each neuron obtains its output based on the weighted sum of its inputs, and then passes it through an activation function (such as the ReLU function f(x)=max(0, x)), and passes the output to the next layer. In this process, the basic neural network generates multiple candidate labeled trajectories based on the features of the input data and its internal calculation mechanism. These candidate labeled trajectories are the different interpretations and classification results of the training-transformed data stream by the basic neural network. Each trajectory contains a series of label types, corresponding to different judgments on the pressure signal, such as signal correlation, signal category, and signal segment of interest.

[0119] In step S40, the computer system acquires training prior labels and, based on these labels, instructs the basic neural network to determine the target candidate label trajectory from multiple alternative label trajectories, thereby outputting the target inference and recognition result. The training prior labels are known correct results, obtained based on professional medical knowledge or accurate annotations. For example, in cardiac stress monitoring training data, if a target signal for a particular training stress monitoring signal comes from a patient with coronary artery disease, its corresponding training prior label will accurately indicate the signal category as coronary artery disease stress, as well as the accurate signal correlation (difference from the control signal) and the signal segment of interest (such as the time period of pressure change associated with myocardial ischemia). The computer system compares these training prior labels with multiple alternative label trajectories obtained by the basic neural network. For each label type in the alternative label trajectory, it matches it with the corresponding part in the training prior label. For example, if the signal category in an alternative label trajectory is labeled as normal cardiac stress, while the training prior label indicates it is coronary artery disease stress, then this alternative label trajectory does not meet the requirements. Through this comparison and selection, the computer system instructs the basic neural network to determine the target candidate marker trajectory that is closest to the training prior label, thereby obtaining the target inference and recognition result. This result represents the best judgment made by the basic neural network on the pressure signal based on training data and prior knowledge.

[0120] Finally, the computer system executes step S50, generating a target training cost based on the proportion of trajectory support of the target inference recognition result, and repeatedly debugging the basic neural network based on the target training cost. When the preset evaluation conditions are met, the debugged target neural network is obtained. The trajectory support of the target inference recognition result reflects the credibility or reasonableness of this result in the current network state. For example, if the trajectory support corresponding to the target inference recognition result is high, it indicates that the basic neural network is relatively certain about this result; if the trajectory support is low, it indicates that the credibility of the result is low. The computer system generates the target training cost based on the proportion of trajectory support. One possible approach is to let the trajectory support of the target inference recognition result be S, and the sum of the support of all possible trajectories be T (including the trajectory support of the target inference recognition result and other alternative labeled trajectories). The target training cost C can be defined as C = 1 - S / T. This formula indicates that when the trajectory support of the target inference recognition result is higher, the target training cost is lower, indicating better network performance; conversely, the target training cost is higher, indicating that the network needs more adjustments.

[0121] The computer system iteratively debugs the basic neural network based on the target training cost. This debugging process involves adjusting the weights and parameters of the neural network, for example, using the backpropagation algorithm. The backpropagation algorithm calculates the impact of each weight on the cost based on the target training cost (by taking partial derivatives), and then applies a certain learning rate (e.g., ...). Update the weights. Assume a weight w in the neural network, and its update formula is: By continuously adjusting the weights based on the target training cost, the performance of the basic neural network gradually improves. Preset evaluation conditions can take various forms, such as setting a maximum number of training epochs (e.g., 1000 epochs), stopping training when this number is reached; or setting a threshold for the target training cost (e.g., 0.01), considering the network to have reached good performance and stopping training when the target training cost is less than this threshold. Through such repeated adjustments, the computer system finally obtains a well-tuned target neural network. This network can accurately process monitoring data from implantable medical stress monitors. For example, in actual cardiac stress monitoring, it can accurately determine the type, correlation, and signal segment of interest of the stress signal, providing reliable support for medical diagnosis and patient health monitoring.

[0122] Throughout the entire network training process, steps S10-S50 form a complete system, with each step closely related. From acquiring training data to converting it into a form suitable for a neural network, to obtaining preliminary results through the basic neural network, determining the target result based on prior labels, and finally debugging the network based on the evaluation of the target result, this series of operations enables the neural network to learn the complex patterns and relationships in stress monitoring data, thereby possessing the ability to accurately process actual monitoring data.

[0123] As one implementation method, step S50, generating the target training cost based on the proportion of trajectory support of the target reasoning and recognition results, includes:

[0124] Step S51: Determine the inference label state support and inference label dependency support for each candidate label trajectory, and determine the candidate trajectory support for the corresponding candidate label trajectory based on the inference label state support and inference label dependency support;

[0125] Step S52: Determine the total support of candidate trajectories corresponding to the support of multiple candidate trajectories, and search for the target trajectory support corresponding to the target inference recognition result among the multiple candidate trajectory supports;

[0126] Step S53: Generate the target training cost based on the weight ratio between the target trajectory support and the total support of the inference candidate trajectories.

[0127] When executing step S51, the computer system determines the inference label state support and inference label dependency support for each candidate label trajectory, and determines the candidate trajectory support for the corresponding candidate label trajectory based on the inference label state support and inference label dependency support. The inference label state support is a result of comprehensively considering the label type within each candidate label trajectory and its corresponding inference recognition label support. The inference recognition label support reflects the degree of support the underlying neural network provides for the inferred label type of each data item when processing the training transformation data stream.

[0128] Taking cardiac stress monitoring as an example, suppose a candidate marker trajectory contains three marker types: signal correlation identification markers, signal category identification markers (e.g., normal or abnormal heartbeat categories), and signal segment of interest identification markers. For each marker type, the basic neural network generates corresponding inference marker support during inference, denoted as s1, s2, and s3, respectively. The computer system can use a weighted summation method to determine the inference marker state support. Let the weight coefficients be α, β, and γ (these weight coefficients can be preset based on experience or data analysis, for example, determined according to the importance of different marker types in the overall judgment), then the inference marker state support S state It can be done through formula S state=αs1+βs2+γs3 is calculated. This calculation method integrates the inference recognition tag support of each tag type in the candidate tag trajectory, so that the inference tag state support can reflect the rationality of the entire candidate tag trajectory based solely on the inference tag support level.

[0129] Inference label dependency support is the support determined by considering the label dependencies between multiple label types in each candidate label trajectory. Label dependencies embody the transition rules or constraints between adjacent labels in a label trajectory. For example, in cardiac stress monitoring, if a signal correlation identification label shows high correlation, then a signal category identification label would more likely be a normal heartbeat category; this is a label dependency. For a candidate label trajectory, the computer system needs to evaluate whether its label types conform to this dependency. Suppose there is a function f to quantify this label dependency; the input of this function is the label type and its corresponding inference identification label support, and the output is a numerical value representing the degree to which the label dependency conforms.

[0130] For example, given a candidate labeled trajectory [high relevance (inference label support s1), normal heartbeat category (inference label support s2)], function f might be calculated according to predefined rules. If s1 is highly relevant and s2 is a normal heartbeat category, function f might output a higher value, indicating that the candidate labeled trajectory has a good fit in terms of label dependency; conversely, if s1 is highly relevant but s2 is an abnormal heartbeat category, function f might output a lower value. By applying function f to all adjacent label type pairs in the candidate labeled trajectory and combining the results (e.g., summing or weighted summation), the computer system can obtain the inference label dependency support S. dependency .

[0131] After obtaining the support S of the inference labeled state state and inference label dependency support S dependency Then, the computer system can determine the candidate trajectory support S of the corresponding alternative labeled trajectories through a linear combination method. candidate For example, S candidate =λS state +μS dependency λ and μ are pre-defined weighting coefficients. These two weighting coefficients can be adjusted according to the specific application scenario and data characteristics. If a medical stress monitoring scenario places more emphasis on direct evidence based on inference label support, then the value of λ can be appropriately increased; if the logical relationship between label types is more important, then the value of μ can be increased.

[0132] Next, the computer system executes step S52 to determine the total support of the candidate trajectories corresponding to the support of multiple candidate trajectories, and to search for the target trajectory support corresponding to the target inference and recognition result among the multiple candidate trajectory supports. For all candidate labeled trajectories, the computer system has already calculated the respective candidate trajectory support S. candidate1 S candidate2 S candidate3 Wait, the total support of the candidate trajectory is the sum of the support of these candidate trajectories.

[0133] Then, the computer system finds the target trajectory support corresponding to the target inference recognition result among these candidate trajectory supports. The target inference recognition result is the best result determined from multiple candidate labeled trajectories based on training prior labels in step S40. The computer system determines the target trajectory support S by finding the candidate trajectory support of the candidate labeled trajectories corresponding to the target inference recognition result. target .

[0134] Finally, the computer system executes step S53, generating the target training cost based on the weight ratio between the target trajectory support and the total support of the inference candidate trajectories. The target training cost is an indicator used to measure the performance of the underlying neural network; it reflects the relative reasonableness of the target inference recognition result among all possible outcomes. Let the target training cost be C, which the computer system can obtain using the formula C=1-S. target / S total Generate the target training cost. The significance of this formula lies in the fact that when the target trajectory support S... target The higher the value of S, the more plausible the target reasoning and recognition result is among all candidate marker trajectories. target / S total The larger the value of S, the smaller the target training cost C, indicating better performance of the basic neural network; conversely, when S is smaller... target When the value is small, the training cost C of the target is large, indicating that the basic neural network needs more adjustments and optimizations.

[0135] For example, in a network training scenario for lung pressure monitoring, if the target trajectory support S targetA relatively high target training cost C indicates that the target inference and recognition results (such as the judgment of stress signal category, correlation, and signal segment of interest) obtained by the basic neural network when processing the training transformation data stream are more in line with expectations. Therefore, the target training cost C will be relatively low. The computer system will then adjust the basic neural network based on this target training cost, such as adjusting parameters like the network weights, to reduce the target training cost and improve the network's accuracy. This process is also applicable in cardiac stress monitoring or other implantable medical stress monitoring scenarios. By continuously adjusting the basic neural network according to the target training cost, it can process monitoring data more accurately, providing more reliable results for medical diagnosis and patient health monitoring.

[0136] Throughout the process, steps S51-S53 meticulously analyze the support-related indicators of each candidate labeled trajectory and calculate the target training cost, providing a basis for the effective debugging of the basic neural network. This target training cost calculation method based on labeled trajectory support fully considers the inference label state, inference label dependencies, and the relative rationality of the target inference recognition result among all possible outcomes. This allows the basic neural network to be adjusted towards more accurate processing of implantable medical stress monitoring data, thereby improving the performance and reliability of the entire data processing system.

[0137] As one implementation, step S51, determining the inference tag state support and inference tag dependency support for each candidate tag trajectory, includes:

[0138] Step S511: Obtain the set of inference recognition label support when the basic neural network processes each data item in the training transformation data stream. The set of inference recognition label support includes multiple inference label types and the inference recognition label support corresponding to each inference label type.

[0139] Step S512: Determine the target inference tag type corresponding to each data item in each candidate tag trajectory, and determine the inference tag state support of the corresponding candidate tag trajectory based on the inference recognition tag support corresponding to each target inference tag type;

[0140] Step S513: Determine the type dependency support between any two adjacent target inference tag types in each candidate tag trajectory, and determine the inference tag dependency support of the corresponding candidate tag trajectory based on multiple type dependency supports.

[0141] When executing step S511, the computer system acquires a set of inference recognition label support scores for each data item in the training-transformed data stream processed by the basic neural network. The inference recognition label support score set contains multiple inference label types and the corresponding inference recognition label support score for each inference label type. In the context of implantable medical stress monitoring, such as cardiac stress monitoring, each data item in the training-transformed data stream is processed by the basic neural network, and for each data item, the network outputs inference recognition label support scores for different label types.

[0142] Suppose a data item in the training transformation data stream is related to cardiac pressure monitoring data at a certain moment. The basic neural network might output inference recognition label support for this data item, including signal correlation identification label, signal category identification label (e.g., normal heartbeat or abnormal heartbeat category), and signal segment of interest identification label. These support values ​​reflect the network's acceptance of different label types based on the current data item. For example, for signal correlation identification labels, the inference recognition label support might be a value calculated based on the similarity between the current data item and a reference control signal. If the data item and the control signal are very similar in waveform, amplitude, frequency, etc., then the inference recognition label support for signal correlation identification labels might be high, such as 0.8; for signal category identification labels, if the features of the current data item are more inclined towards a normal heartbeat pattern, then the inference recognition label support for the normal heartbeat category might be 0.7; for signal segment of interest identification labels, if the current data item is in a potentially critical time period (e.g., a specific phase of diastole), then its corresponding inference recognition label support might be 0.6. These different label types and their corresponding inference recognition label support constitute the inference recognition label support set for that data item. The computer system performs a similar operation for each data item in the training transformed data stream to obtain the inference recognition label support set for each data item.

[0143] Next, the computer system executes step S512 to determine the target inference label type corresponding to each data item in each candidate label trajectory, and to determine the inference label state support of the corresponding candidate label trajectory based on the inference recognition label support corresponding to each target inference label type. Each candidate label trajectory is a label sequence hypothesis of stress monitoring data generated by the basic neural network when processing the training transformation data stream.

[0144] Taking cardiac stress monitoring as an example, a candidate label trajectory might be: [signal correlation identification label (high correlation at a certain data item, corresponding inference label support s1), normal heartbeat category (at another data item, corresponding inference label support s2), specific signal segment of interest (at yet another data item, corresponding inference label support s3)]. For each data item in this candidate label trajectory, there is a corresponding target inference label type (such as the aforementioned signal correlation identification label, signal category identification label, and signal segment of interest identification label).

[0145] One possible technique for computer systems to determine the inference label state support is to use a weighted summation method. Assume that for different label types, weight coefficients are pre-defined based on their importance in the overall judgment. Let α be the weight of the signal correlation identification label, β be the weight of the signal category identification label, and γ be the weight of the signal segment of interest identification label. For the above candidate label trajectories, its inference label state support S state It can be done through formula S state =αs1 + βs2 + γs3 is calculated. For example, if α = 0.3, β = 0.5, γ = 0.2, and s1 = 0.8, s2 = 0.7, s3 = 0.6, then S state =0.3×0.8+0.5×0.7+0.2×0.6=0.71. The support S of this inference labeled state is... state The inference recognition tag support of the target inference tag type corresponding to each data item in the candidate tag trajectory is comprehensively reflected, which shows the rationality of the candidate tag trajectory in terms of inference recognition tag support.

[0146] Finally, the computer system executes step S513 to determine the type dependency support (also called transition support) between any two adjacent target inference tag types in each candidate tag trajectory, and to determine the inference tag dependency support of the corresponding candidate tag trajectory based on multiple type dependency supports. Type dependency support reflects the degree of logical relationship between adjacent tag types.

[0147] For example, in a candidate marker trajectory for cardiac stress monitoring, if the signal correlation identification marker is highly correlated, according to normal physiological logic, the next signal category identification marker is more likely to be the normal heartbeat category. The computer system quantifies this type dependency using a predefined function. Suppose there is a function f to calculate type dependency support for two adjacent marker types (signal correlation identification marker and signal category identification marker). If the inferred identification marker support for the current signal correlation identification marker is s1, and the inferred identification marker support for the signal category identification marker is s2, function f might calculate type dependency support d1 based on the values ​​of s1 and s2 and the expected logical relationship between them. For example, if s1 is highly correlated (assumed value 0.8) and s2 is the normal heartbeat category (assumed value 0.7), function f calculates d1 = 0.8 based on internally defined rules (possibly based on extensive prior knowledge and statistical data), indicating a high type dependency support between the two adjacent marker types, consistent with the expected logical relationship.

[0148] The computer system performs a similar operation on each pair of adjacent target inference label types in the candidate label trajectories, obtaining multiple type dependency support values ​​d1, d2, d3, etc. Then, the inference label dependency support is determined through a comprehensive approach. For example, summation or weighted summation can be used. Assuming a weighted summation approach is used, for each type dependency support value d1, d2, d3, etc., the inference label dependency support is determined. i Each has a pre-set weight w i Then the inference label dependency support S dependency It can be done through formula S dependency = The calculation yielded the inference label, which depends on the support S. dependency The rationality of the alternative tag trajectories is reflected from the perspective of the logical dependencies between tag types.

[0149] Figure 2 A hardware entity diagram of a computer system provided in an embodiment of this application is shown below. Figure 2 As shown, the hardware entity of the computer system 1000 includes a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can run on the processor 1001, and the processor 1001 executes the program to implement the steps in the method of any of the above embodiments.

[0150] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method of processing monitoring data for an implantable medical pressure monitor, characterized by, The method comprises: obtaining a pressure monitoring reference signal, a pressure monitoring target signal and a pressure monitoring target signal category; performing data conversion on the pressure monitoring reference signal, the pressure monitoring target signal and the pressure monitoring target signal category to obtain a converted data stream; inputting the converted data stream into a target neural network to obtain a target recognition result, specifically comprising: performing information enhancement representation on each data item in the converted data stream according to the target neural network to obtain a data item implicit representation set, the data item implicit representation set comprising a data item implicit representation corresponding to each data item; integrating and mapping transforming each data item implicit representation and the implicit representation in the front and back transmission of the data item implicit representation set to obtain a recognition mark support degree set corresponding to each data item implicit representation; determining a target mark trajectory according to a plurality of mark types in each recognition mark support degree set and a mark support degree corresponding to each mark type, the target mark trajectory being used to indicate that the target neural network outputs a target recognition result; the target recognition result comprising a signal correlation recognition result, a signal category recognition result and a signal segment of interest recognition result of the pressure monitoring target signal compared with the pressure monitoring reference signal; wherein the target neural network is obtained by supervised learning of a basic neural network according to an output result of a training converted data stream and a training prior mark, the training converted data stream being obtained by data conversion on a training pressure monitoring reference signal, a training pressure monitoring target signal and a training pressure monitoring target signal category, and the training prior mark comprising a training correlation prior mark, a training signal category prior mark and a training signal segment of interest prior mark of the training pressure monitoring target signal compared with the training pressure monitoring reference signal.

2. The method of claim 1, wherein, The target neural network comprises a feature extraction component, and the information enhancement representation on each data item in the converted data stream to obtain the data item implicit representation set comprises: performing implicit representation conversion on each data item in the converted data stream based on the feature extraction component to obtain an implicit representation set, the implicit representation set comprising an implicit representation array corresponding to each data item; performing feature extraction on each implicit representation array according to the forward and backward features of each implicit representation array in the implicit representation set to obtain the data item implicit representation set after information enhancement representation.

3. The method of claim 2, wherein, The target neural network further comprises an LSTM component connected with the feature extraction component, and the integration and mapping transformation of each data item implicit representation and the implicit representation in the front and back transmission of the data item implicit representation set to obtain the recognition mark support degree set corresponding to each data item implicit representation comprises: performing front sequence transmission on each data item implicit representation in the data item implicit representation set based on the LSTM component to obtain a forward implicit representation corresponding to each data item implicit representation; performing back sequence transmission on each data item implicit representation in the data item implicit representation set to obtain a backward implicit representation corresponding to each data item implicit representation; integrating each data item implicit representation with the corresponding forward implicit representation and the corresponding backward implicit representation to obtain a corresponding integrated implicit representation of each data item implicit representation; mapping the integrated implicit representation to obtain a corresponding set of recognition mark support degrees of each data item implicit representation.

4. The method of claim 3, wherein, The target neural network further comprises a mark mapping component connected with the LSTM component, and the target mark track is determined according to a plurality of mark types in each set of recognition mark support degrees and a mark support degree corresponding to each mark type, comprising: combining any one of the filtered mark types in each set of recognition mark support degrees according to the data item position of the converted data stream based on the mark mapping component to obtain a plurality of alternative mark tracks, each of which includes the mark type of each data item; determining the mark support degree of each mark type represented in each alternative mark track according to a plurality of mark types in each set of recognition mark support degrees and a mark support degree corresponding to each mark type; determining the track support degree of the corresponding alternative mark track based on the mark support degree of each mark type represented in each alternative mark track and the mark dependency relationship between adjacent mark types; determining the target mark track among the plurality of alternative mark tracks according to the numerical value of the track support degree.

5. The method of claim 4, wherein, The mark support degree of each mark type represented in each alternative mark track and the mark dependency relationship between adjacent mark types are used to calculate the track support degree of each alternative mark track, comprising: determining the mark state support degree corresponding to each alternative mark track according to the mark support degree in each set of recognition mark support degrees; determining the mark dependency support degree according to the mark dependency relationship between a plurality of mark types in each alternative mark track; determining the track support degree according to the mark state support degree and the mark dependency support degree corresponding to each alternative mark track.

6. The method of claim 1, wherein, The pressure monitoring reference signal, the pressure monitoring target signal and the pressure monitoring target signal category are data converted to obtain a converted data stream, comprising: merging the pressure monitoring reference signal, the pressure monitoring target signal and the pressure monitoring target signal category to obtain merged data; performing data item splitting on the merged data to obtain a converted data stream.

7. The method according to any one of claims 1 to 6, characterized in that, The method further comprises: obtaining a training pressure monitoring reference signal, a training pressure monitoring target signal and a training pressure monitoring target signal category; data converting the training pressure monitoring reference signal, the training pressure monitoring target signal and the training pressure monitoring target signal category to obtain a training converted data stream; inputting the training converted data stream into the basic neural network and obtaining a plurality of alternative mark tracks obtained by the basic neural network when processing the training converted data stream; obtaining a training prior mark and instructing the basic neural network to determine a target alternative mark track in the plurality of alternative mark tracks based on the training prior mark to output a target reasoning recognition result; The target training cost is generated based on a proportion of the trajectory support degree of the target inference recognition result, and the base neural network is repeatedly debugged based on the target training cost, and a debugged target neural network is obtained when a preset evaluation condition is reached.

8. The method of claim 7, wherein, The target training cost is generated based on a proportion of the trajectory support degree of the target inference recognition result, and the base neural network is repeatedly debugged based on the target training cost, and a debugged target neural network is obtained when a preset evaluation condition is reached. The inference label state support degree and the inference label dependency support degree of each candidate label trajectory are determined, and the candidate trajectory support degree of the corresponding candidate label trajectory is determined based on the inference label state support degree and the inference label dependency support degree; The candidate trajectory total support degree corresponding to a plurality of candidate trajectory support degrees is determined, and the target trajectory support degree corresponding to the target inference recognition result is searched in the plurality of candidate trajectory support degrees; The target training cost is generated based on a proportion of the trajectory support degree of the target inference recognition result, and the base neural network is repeatedly debugged based on the target training cost, and a debugged target neural network is obtained when a preset evaluation condition is reached. The inference label state support degree and the inference label dependency support degree of each candidate label trajectory are determined, and the candidate trajectory support degree of the corresponding candidate label trajectory is determined based on the inference label state support degree and the inference label dependency support degree; The inference recognition label support degree set of the base neural network when processing each data item in the training conversion data stream is obtained, and the inference recognition label support degree set contains a plurality of inference label types and the inference recognition label support degree corresponding to each inference label type; The target inference label type corresponding to each data item in each candidate label trajectory is determined, and the inference label state support degree of the corresponding candidate label trajectory is determined based on the inference recognition label support degree corresponding to each target inference label type; The type dependency support degree between any two adjacent target inference label types in each candidate label trajectory is determined, and the inference label dependency support degree of the corresponding candidate label trajectory is determined based on a plurality of type dependency support degrees.

9. A computer system comprising a memory and a processor, said memory storing a computer program operable on the processor, characterised in that, The processor implements the steps in the method of any one of claims 1 to 8 when executing the program.

Images