Deep learning-based ventricle implantation auxiliary device operation monitoring method

Abstract

The invention discloses a deep learning-based operation monitoring method for an auxiliary device implanted into a ventricle. The method comprises the following steps of collecting multi-source time sequence monitoring data related to operation of the auxiliary device of the ventricle; constructing an operation state prediction model comprising a basic monitoring module and an accurate monitoring module; the basic monitoring module comprises a local feature extraction module, a feature summarization module and a first classification module which are cascaded; the fine monitoring module comprises a high-frequency time sequence data feature extraction module, a low-frequency asynchronous data condition coding module, a fusion module and a second classification module. And analyzing the real-time monitoring data by using the running state prediction model, and outputting the current running state information of the ventricular assist device. The method provided by the invention has the advantages of timely fault early warning, high identification precision, strong model generalization and the like, and can realize dynamic monitoring and intelligent diagnosis of the operation state of the ventricular assist device.

Description

Technical Field

[0001] The present invention belongs to the field of medical artificial intelligence technology, and in particular relates to a method for monitoring the operation of an implanted ventricular assist device based on deep learning. Background Art

[0002] A ventricular assist device (VAD) is a circulatory support device implanted in patients with severe heart failure. It assists or replaces ventricular function through a mechanical pump and is currently an important treatment for end-stage heart failure. With the increasing clinical adoption of VADs, particularly left ventricular assist devices (LVADs), their long-term operation has a critical impact on patients' quality of life and physiological safety. Therefore, accurately assessing the operation of implantable devices and providing fault warnings have become crucial issues in post-implant remote management and precision maintenance.

[0003] Currently, many VAD devices still rely primarily on pre-set programs or manual follow-up to assess device status, lacking intelligent operational monitoring mechanisms. Over long-term operation, VADs can experience mechanical fatigue, flow anomalies, sensor drift, intra-pump thrombosis, motor overheating, and other issues, leading to device performance degradation or even failure. Failure to promptly detect these anomalies can severely impact VAD performance and even endanger the patient's life.

[0004] Current technologies for monitoring the operation of ventricular assist devices still have the following problems:

[0005] 1. The monitoring indicators are single and mainly rely on threshold analysis of parameters such as pump speed and flow rate, which cannot capture the correlation between deep time series characteristics and multi-source data;

[0006] 2. Low intelligence level, lack of deep learning or data-driven fault identification mechanisms, making it difficult to achieve early trend prediction and high-level feature extraction;

[0007] 3. Poor generalization ability: diagnostic performance degrades when faced with different device models, individual differences, or differences in usage stages. Summary of the Invention

[0008] To solve the above problems, the present invention provides a method for monitoring the operation of an implanted ventricular assist device based on deep learning.

[0009] The technical solution for achieving the purpose of the present invention is:

[0010] A method for monitoring the operation of an implanted ventricular assist device based on deep learning, comprising:

[0011] By building a deep monitoring model that integrates convolutional neural networks, recurrent neural networks, and an attention mechanism, we can achieve intelligent monitoring of the operating status of implanted ventricular assist devices. The deep monitoring model includes a basic monitoring module, which includes a cascaded local feature extraction module, a feature aggregation module, and a first classification module:

[0012] The local feature extraction module includes a cascaded first one-dimensional convolution layer and a second one-dimensional convolution layer to extract local features of the model input;

[0013] The feature aggregation module includes a first global average pooling layer to compress local features to form a global statistical feature representation;

[0014] The first classification module includes a cascaded first fully connected layer, a second fully connected layer and a Sigmoid activation function, which is used to output the operating status information of the implanted ventricular assist device.

[0015] As a further optimization solution of the present invention, the steps of forming a training data set for the basic monitoring module include:

[0016] Collect high-frequency time series data, including VAD device operating parameters and patient vital signs recorded at a fixed sampling rate;

[0017] Preprocess the collected high-frequency time series data, including missing value filling, outlier processing, and time synchronization;

[0018] The preprocessed high-frequency time series data is labeled with basic status labels, where the basic status labels include normal operation and abnormal operation.

[0019] As a further optimization solution of the present invention, the step of forming a training data set for the basic monitoring module further includes:

[0020] Statistical features are extracted from the preprocessed high-frequency time series data, and the statistical features are copied to form a statistical feature tensor with the same length as the high-frequency time series data;

[0021] The preprocessed high-frequency time series data and the corresponding statistical feature tensor are used as input to the basic monitoring module.

[0022] As a further optimization solution of the present invention, the deep monitoring model also includes a fine monitoring module. When the classification result of the basic monitoring module is abnormal operation, the fine monitoring module is activated; the fine monitoring module includes a high-frequency time series data feature extraction module, a low-frequency asynchronous data conditional encoding module, a fusion module, and a second classification module:

[0023] The high-frequency time series data feature extraction module includes a cascaded third one-dimensional convolution layer, a fourth one-dimensional convolution layer, a fifth one-dimensional convolution layer, a first LSTM layer, a second LSTM layer, a sixth one-dimensional convolution layer, a first upsampling layer, and a one-dimensional refinement convolution layer to obtain high-dimensional and high-resolution time series features.

[0024] The low-frequency asynchronous data conditional encoding module includes a cascaded third fully connected layer and a fourth fully connected layer to obtain a high-dimensional feature vector; the high-dimensional feature vector is respectively used as the input of two parallel parameter branches, one of which includes a cascaded channel scaling unit and a first dimension expansion unit, and the other parameter branch includes a cascaded channel bias unit and a second dimension expansion unit. The two parameter branches generate channel scaling parameters and channel bias parameters, respectively.

[0025] The fusion module includes a Film adjustment unit, which uses the FiLM method to apply channel scaling parameters and channel bias parameters to high-dimensional and high-resolution time series features, and achieves conditional modulation through channel-by-channel affine transformation to obtain the fused time series features;

[0026] The second classification module includes a cascaded temporal attention pooling layer, a fifth fully connected layer, a sixth fully connected layer, and a Softmax activation function to output the classification results.

[0027] As a further optimization solution of the present invention, the one-dimensional refinement convolution layer includes a cascaded seventh one-dimensional convolution layer, a ReLU function, and layer normalization.

[0028] As a further optimization solution of the present invention, the fused temporal features are input into the time attention pooling layer, and the attention weight is calculated for each time step. As a time point index, where is the attention weight at time step t.

[0029] As a further optimization solution of the present invention, the steps of forming a training data set for the fine monitoring module include:

[0030] Multi-source data collection, including high-frequency time-series data and low-frequency asynchronous data. High-frequency time-series data includes VAD device operating parameters and patient vital signs recorded at a fixed sampling rate, while low-frequency asynchronous data includes structured information obtained from electronic medical record systems and postoperative follow-up.

[0031] Preprocess the collected high-frequency time series data, including missing value filling, outlier processing, and time synchronization;

[0032] Preprocess the collected low-frequency asynchronous data, including encoding, missing value filling, and generating missing value masks;

[0033] The preprocessed high-frequency time series data and low-frequency asynchronous data are annotated with fine state labels, where the fine state labels include normal state, warning state and abnormal state.

[0034] As a further optimization solution of the present invention, the step of forming a training data set for the fine monitoring module further includes:

[0035] Statistical features are extracted from the preprocessed high-frequency time series data, and the statistical features are copied to form a statistical feature tensor with the same length as the high-frequency time series data;

[0036] The preprocessed high-frequency time series data and the corresponding statistical feature tensor, the preprocessed low-frequency asynchronous data and the corresponding mask tensor are used as the input of the fine monitoring module.

[0037] As a further optimization solution of the present invention, the step of forming a training data set for the fine monitoring module further includes:

[0038] The time interval feature is added to the preprocessed low-frequency asynchronous data according to the following formula, and normalized by log compression to generate a time encoding tensor:

[0039]

[0040] in, is the time interval between two measurements, To set the time limit cutoff, is the time scale parameter;

[0041] The preprocessed high-frequency time series data and the corresponding statistical feature tensor, the preprocessed low-frequency asynchronous data and the corresponding mask tensor and time coding tensor are used as the input of the fine monitoring module.

[0042] As a further optimization solution of the present invention, the loss function consisting of three-classification loss and time index positioning loss is used in the training process of the fine monitoring module. :

[0043]

[0044] in is the three-category loss, is the time index positioning loss, is the weight, , , is the number of training samples, For the three classification serial numbers, is the training sample number, is the probability of three classifications, is the real fine state label, The weight assigned to the model at the correct time point, is the set of training samples that need time positioning, is the sequence number of the training sample that needs time positioning, and When the set is empty .

[0045] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the above-mentioned method for monitoring the operation of an implanted ventricular assist device based on deep learning are implemented.

[0046] The present invention further provides an electronic device, characterized by comprising:

[0047] memory for storing computer programs;

[0048] A processor is used to implement the steps of the above-mentioned deep learning-based implanted ventricular assist device operation monitoring method when executing the computer program.

[0049] Compared with the prior art, the present invention has the following significant advantages:

[0050] The present invention introduces deep learning technology into the field of ventricular assist device health monitoring. The model can automatically learn complex features and correlations from multi-source data, far exceeding the method of manually setting thresholds; by modeling time dependence and attention mechanism to focus on key signs, it can discover in advance the performance degradation and fault signs that are difficult to detect in time with traditional methods, and realize early prediction and warning of anomalies; through the data window division method driven by abnormal events as the core, it efficiently captures the key signs in the early stage of abnormal signals, greatly improving the sensitivity of abnormal monitoring and warning; for the first time, a two-layer monitoring model combining basic monitoring and fine monitoring is proposed, combining the rapid warning of the basic model with the precise judgment of the fine model, significantly improving the efficiency and accuracy of the system, and overcoming the traditional problems of large computational complexity and response delay of a single complex model; it has the advantages of timely fault warning, high recognition accuracy, and strong model generalization, and can realize dynamic monitoring and intelligent diagnosis of the operating status of ventricular assist devices, and is suitable for remote evaluation and precise maintenance management of postoperative implanted equipment.

[0051] The present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is an overall flow chart of the method of the present invention.

[0053] Figure 2 It is the structural diagram of the basic monitoring model.

[0054] Figure 3 It is a structural diagram of the fine monitoring model.

[0055] Figure 4 It is a real-time monitoring and early warning output flow chart.

[0056] Figure 5 It is a one-dimensional convolution structure diagram.

[0057] Figure 6 It is a one-dimensional refined convolution structure diagram. DETAILED DESCRIPTION

[0058] The present invention provides a method for monitoring the operation of an implanted ventricular assist device based on deep learning, such as Figure 1 As shown, the following steps are included:

[0059] S1. Multi-source data acquisition and preprocessing: Collect high-frequency time-series data such as VAD device operating parameters and patient physiological indicators, and low-frequency asynchronous data such as postoperative follow-up information, and perform preprocessing, including:

[0060] S11. Multi-source data acquisition: High-frequency time-series data acquisition includes device operating parameters (pump power, pump speed, and estimated flow) output in real time by the VAD controller, as well as patient vital signs (heart rate, arterial pressure, blood oxygen saturation, electrocardiogram, etc.) collected in real time by medical monitoring devices such as bedside monitors. This data is continuously recorded at a fixed sampling rate (100-500Hz). Low-frequency asynchronous data acquisition includes structured information collected by electronic medical records and postoperative follow-up visits, such as patient symptoms and signs (fatigue, edema, dizziness, etc.), lactate, NYHA class, medication adherence, and days after surgery. This data is updated less frequently and has an uncertain time step.

[0061] S12. Data preprocessing:

[0062] S121. Preprocessing of high-frequency time series data: Missing value filling, outlier processing, and time synchronization are performed on seven types of high-frequency time series data, including pump power, pump speed, expected flow, heart rate, arterial pressure, blood oxygen saturation, and electrocardiogram. Linear interpolation is used to fill in missing values ​​for short-term data. For noise-type anomalies such as transient jumps and single-point noise, numerical clipping or elimination is performed directly. Finally, all channel data are resampled according to a unified time axis, with T = 128 sampling points per cardiac cycle to ensure that multi-source signals correspond to and match at the same time step.

[0063] S122. Asynchronous data preprocessing: For low-frequency or structured information, such as patient symptoms and signs, lactate, NYHA cardiac function class, medication compliance, and postoperative days, perform the following steps:

[0064] S1221. Feature coding: Continuous variables such as lactate and postoperative days were standardized using Z-score; hierarchical variables (NYHA) were coded using one-hot encoding (1 to 4); whether medication was taken on time (i.e., medication compliance), whether fatigue symptoms occurred, whether edema symptoms occurred, and whether dizziness symptoms occurred were binarized using Boolean variables (Yes=1, No=0).

[0065] S1222. Missing Value Handling and Masking: For continuous variables such as lactate and postoperative days, missing values ​​were imputed using mean-filling. For data with clear hierarchical trends, such as the NYHA, a trend-guided segmented interpolation strategy was used, with intermediate levels interpolated proportionally to maintain trend consistency within the hierarchical sequence. For binary data, forward imputation was performed using the most recent valid follow-up record. A missing mask was also generated (if the mask is 0, the data is missing; if the mask is 1, the data is not missing). This mask tensor was input into the subsequent model along with the asynchronous data encoding features to enhance the model's ability to identify and utilize missing patterns and data integrity.

[0066] S2. Data enhancement and dataset design

[0067] S21. Data Enhancement

[0068] S211. High-Frequency Time Series Data Enhancement: Statistical feature enhancement is performed on high-frequency time series data. Statistical features are extracted from T data points within each cardiac cycle, including but not limited to mean, variance, and rate of change, to further enhance the ability to distinguish event features. The rate of change is the first-order derivative approximation of the difference between adjacent time points, reflecting the intensity of instantaneous fluctuations. The three statistical quantities are replicated T times, ultimately forming a tensor of the same length as the time series feature.

[0069] S212, low-frequency asynchronous data enhancement: add the time interval feature of "from the current measurement time" and pass The compression is normalized and can be calculated as follows:

[0070]

[0071] in, is the time interval from the last measurement, The upper limit is cut off, and any value exceeding this limit is considered equally "obsolete". The generated time encoding tensor is input into the subsequent model together with the asynchronous data encoding feature to enhance the model's perception of data freshness.

[0072] S22. Data window division: The enhanced high-frequency and low-frequency continuous data streams are divided into windows in units of 5 cardiac cycles, with a data number of 5T, and each window is a sample unit. The data is manually or semi-automatically annotated for abnormalities. The abnormal event determination mainly consists of the following two steps: (1) Automatic detection based on pathological abnormalities: Using the statistical characteristics of the calculated pump power, pump speed, expected flow, heart rate, arterial pressure and other data, when any indicator continuously deviates from the physiological reasonable range and exceeds the preset amplitude and duration threshold, it is automatically marked as an abnormal candidate interval. (2) Expert annotation or recorded abnormal event log: Combining historical monitoring curves, clinical intervention records and physician experience, clinical experts mark the key state change time points including suction events, organ hypoperfusion, aortic valve closure time abnormalities, etc. in the abnormal candidate area. The system uses the abnormal samples obtained above as event node n and marks the abnormal label; at the same time, the n-1th window is marked as an abnormal precursor, the n-2th window is marked as a possible abnormal precursor, and the rest are marked as normal. This approach implements an event-driven data annotation framework, significantly improving the targeting and scarcity utilization efficiency of precursor identification samples.

[0073] S23. Category ratio control: Based on the annotated device operating status labels, a sample set is constructed in the ratio of 6:1:1:1, corresponding to the four categories of normal, abnormal, abnormal precursor, and possible abnormal precursor, to alleviate the category imbalance problem.

[0074] S3. Operation Monitoring Model Construction: Build a layered dual-channel deep learning diagnostic framework, including a basic monitoring module and a fine-grained monitoring module, to achieve the synergy between low-latency initial warning and high-precision deep diagnosis. The model construction steps include:

[0075] S31. Basic monitoring module: Build a lightweight convolutional neural network (CNN) that uses only basic high-frequency time series data, namely VAD device pump speed, pump power, expected flow rate and heart rate for rapid initial warning.

[0076] S311. Input: The network input consists of four types of high-frequency time series data: pump speed, pump power, estimated flow rate, and heart rate (the four types of high-frequency data that are easy to monitor are preferred here; in specific applications, you can select other types as needed, or even all high-frequency data as input), along with their corresponding three types of statistical feature data. The structure of a single data channel (such as pump speed) and its statistical features is 4×5T (a 4-channel, 5T-long time series); therefore, the combined input structure of the four types of data is 16×5T.

[0077] S312, Structure: The model first extracts local feature patterns through two layers of one-dimensional convolution (combined with ReLU activation function and max pooling) to capture short-term signal trends. A global average pooling layer then compresses the time dimension to form a compact global statistical feature representation. Finally, two fully connected layers output the diagnosis results, using the Sigmoid activation function to estimate the probability of anomalies. If an anomaly is identified, an anomaly warning is triggered and the subsequent fine-grained monitoring module is activated.

[0078] S313. Training: The training data for the basic monitoring model includes four types of high-frequency time series data: pump speed, pump power, expected flow rate, and heart rate; their statistical characteristics; and their corresponding labels. In the label definition, "abnormal," "abnormal precursor," and "possible abnormal precursor" are uniformly labeled as abnormal categories, forming a binary classification task. During training, a binary cross-entropy loss function is used as the objective function, and the optimizer is Adam (learning rate 1×10⁻). 4 ), and introduced an early stopping mechanism to prevent overfitting; the model performance was comprehensively evaluated through accuracy, recall, and F1-score.

[0079] S32, fine monitoring module (activated only when the basic monitoring module triggers an abnormal warning): When the basic monitoring module detects an abnormality, the fine monitoring module is immediately activated to perform in-depth global diagnostic analysis on a wider range of high-frequency time series data and low-frequency asynchronous data.

[0080] S321, Input

[0081] S3211, high-frequency time series channel input: This channel takes seven types of high-frequency time series data, including pump speed, pump power, expected flow, heart rate, arterial pressure, blood oxygen saturation, and electrocardiogram, and their corresponding statistical features as input. The corresponding data structure is 28×5T.

[0082] S3212, Low-Frequency Asynchronous Channel Input: This channel receives seven types of low-frequency asynchronous data and their corresponding time codes and masks. These three types of information constitute nine feature channels, with the corresponding initial data structure being 9×1, meaning each sample corresponds to a set of static asynchronous features.

[0083] S322, Structure: Divided into high-frequency time series data feature extraction channel and low-frequency asynchronous data conditional encoding channel.

[0084] S3221, high-frequency time series data feature extraction channel: high-frequency time series data first passes through three layers of one-dimensional convolution (such as Figure 5As shown in Figure 1 ), each layer uses ReLU activation and max pooling to extract local temporal pattern features layer by layer. The feature sequence is then fed into a two-layer bidirectional LSTM network to model long-term trends and dynamic changes. The LSTM output undergoes channel projection via a 1×1 one-dimensional convolution, and then upsampling to restore the time length to the original resolution. Finally, it undergoes a layer of refinement convolution (Conv1d→ReLU→LayerNorm, as shown in Figure 1 ). Figure 6 As shown in Figure 2, high-dimensional and high-resolution temporal features are obtained.

[0085] S3222, low-frequency asynchronous data conditional encoding channel: low-frequency asynchronous data is first encoded into a compact high-dimensional feature vector through a two-layer fully connected network. On this basis, two parallel parameter branches are set to generate channel scaling respectively. With bias , where the scaling parameter is further adopted To ensure numerical stability. 、 It is extended in the time dimension to align with high-frequency features, providing modulation parameters for subsequent FiLM fusion.

[0086] S3223, dual-channel fusion: The fusion module uses the FiLM method to scale the channel generated by the low frequency and channel offset parameters Acting on high-frequency features, conditional modulation is achieved through channel-by-channel affine transformation to obtain fused time series features.

[0087] S3224, output module: input the fused feature sequence into the temporal attention pooling layer and calculate the attention weight for each time step (normalized by softmax), and the weighted sum is used to obtain the global feature vector The vector is sequentially connected, ReLU activated, Dropout, fully connected, and Softmax, and outputs three classification results: normal, abnormal warning (including possible abnormal precursors and abnormal precursors), and abnormal. When the classification result is "abnormal warning" or "abnormal", based on Time distribution positioning key time index: take the main peak position As a time point index, it can be converted to physical time according to the sampling rate. If the classification is "normal", no time index is returned.

[0088] S323, training and optimization: Using a multi-task learning framework, the loss function consists of two parts: three-classification loss and time index positioning loss:

[0089]

[0090] in is the three-category loss, is the time index positioning loss, is the weight.

[0091]

[0092]

[0093] in is the number of samples, For the three classification serial numbers, is the probability of three classifications, is the true label. The weight assigned to the model at the correct time point, For a sample set that needs time positioning (the actual category is "warning / abnormal"), if the current batch ,definition .

[0094] The optimization process uses the Adam optimizer, combined with a cosine annealing learning rate scheduling strategy, and introduces weight decay to reduce the risk of overfitting. Performance evaluation uses three metrics for comprehensive verification: first, classification performance metrics, including overall classification accuracy, F1-score, and AUC curve, which measure the model's ability to distinguish different health states, focusing on the sensitivity of abnormal state detection; second, temporal localization performance metrics, using mean absolute error (MAE) to measure the deviation between the predicted time index and the actual index; and third, operational efficiency metrics, which measure the inference delay after the fine diagnosis module is triggered, ensuring that the model meets accuracy requirements while also providing real-time diagnostic capabilities.

[0095] S4. Real-time operation monitoring and early warning output: Figure 4 As shown, the trained basic monitoring module and fine monitoring module are used to analyze the high-frequency and low-frequency data and output the real-time operation monitoring results of the ventricular assist device.

[0096] S41. Data Acquisition and Preprocessing: Real-time acquisition of high-frequency time series data from patients with implanted ventricular assist devices, along with the most recent low-frequency asynchronous data. All acquired data undergoes normalization, time alignment, and data augmentation according to the same preprocessing process used during the training phase, ensuring that the format and distribution of input features are consistent with the training data.

[0097] S42. Result Generation and Time Index Output: First, real-time, high-frequency basic data is input into the basic monitoring module to obtain a binary classification result (normal or abnormal). If it is determined to be abnormal, the fine monitoring module is called to input the complete high-frequency time series data and low-frequency asynchronous data. After processing, the three-category health status (normal, abnormal warning, abnormal) is output and the corresponding abnormal or abnormal warning time index is generated. The time index is based on the time weight distribution calculated in the attention pooling stage. The peak locates the key time point or time interval, providing a reference for clinical traceability and decision-making.

[0098] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0099] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0100] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0101] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1The steps for the function specified in one or more boxes.

[0102] The above embodiments are only for illustrating the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for monitoring the operation of an implanted ventricular assist device based on deep learning, characterized in that: include: By building a deep monitoring model, intelligent monitoring of the operating status of implanted ventricular assist devices is achieved. The deep monitoring model includes a basic monitoring module, which includes a cascaded local feature extraction module, a feature aggregation module, and a first classification module: The local feature extraction module includes a cascaded first one-dimensional convolution layer and a second one-dimensional convolution layer to extract local features of the model input; The feature aggregation module includes a first global average pooling layer to compress local features to form a global statistical feature representation; The first classification module includes a cascaded first fully connected layer, a second fully connected layer and a Sigmoid activation function, which is used to output the operating status information of the implanted ventricular assist device.

2. The method according to claim 1, characterized in that The steps for constructing the training data set for the basic monitoring module include: Collect high-frequency time series data, including VAD device operating parameters and patient vital signs recorded at a fixed sampling rate; Preprocess the collected high-frequency time series data, including missing value filling, outlier processing, and time synchronization; The preprocessed high-frequency time series data is labeled with basic status labels, where the basic status labels include normal operation and abnormal operation.

3. The method according to claim 2, characterized in that The steps for constructing the training data set for the basic monitoring module also include: Statistical features are extracted from the preprocessed high-frequency time series data, and the statistical features are copied to form a statistical feature tensor with the same length as the high-frequency time series data; The preprocessed high-frequency time series data and the corresponding statistical feature tensor are used as input to the basic monitoring module.

4. The method according to claim 1, wherein The deep monitoring model also includes a fine monitoring module. When the basic monitoring module classifies the module as abnormal operation, the fine monitoring module is activated. The fine monitoring module includes a high-frequency time series data feature extraction module, a low-frequency asynchronous data conditional encoding module, a fusion module, and a second classification module: The high-frequency time series data feature extraction module includes a cascaded third one-dimensional convolution layer, a fourth one-dimensional convolution layer, a fifth one-dimensional convolution layer, a first LSTM layer, a second LSTM layer, a sixth one-dimensional convolution layer, a first upsampling layer, and a one-dimensional refinement convolution layer to obtain high-dimensional and high-resolution time series features. The low-frequency asynchronous data conditional encoding module includes a cascaded third fully connected layer and a fourth fully connected layer to obtain a high-dimensional feature vector; The high-dimensional feature vectors are respectively used as inputs of two parallel parameter branches, one of which includes a cascaded channel scaling unit and a first dimension expansion unit, and the other parameter branch includes a cascaded channel bias unit and a second dimension expansion unit. The two parameter branches generate channel scaling parameters and channel bias parameters, respectively. The fusion module includes a Film adjustment unit, which uses the FiLM method to apply channel scaling parameters and channel bias parameters to high-dimensional and high-resolution time series features, and achieves conditional modulation through channel-by-channel affine transformation to obtain the fused time series features; The second classification module includes a cascaded temporal attention pooling layer, a fifth fully connected layer, a sixth fully connected layer, and a Softmax activation function to output the classification results.

5. The method according to claim 4, characterized in that The one-dimensional refinement convolution layer includes a cascaded seventh one-dimensional convolution layer, a ReLU function, and layer normalization.

6. The method according to claim 4, characterized in that The fused temporal features are input into the time attention pooling layer, and the attention weight is calculated for each time step. As a time point index, where is the attention weight at time step t.

7. The method according to claim 4, characterized in that The steps for constructing the training dataset for the fine monitoring module include: Multi-source data collection, including high-frequency time-series data and low-frequency asynchronous data. High-frequency time-series data includes VAD device operating parameters and patient vital signs recorded at a fixed sampling rate, while low-frequency asynchronous data includes structured information obtained from electronic medical record systems and postoperative follow-up. Preprocess the collected high-frequency time series data, including missing value filling, outlier processing, and time synchronization; Preprocess the collected low-frequency asynchronous data, including encoding, missing value filling, and generating missing value masks; The preprocessed high-frequency time series data and low-frequency asynchronous data are annotated with fine state labels, where the fine state labels include normal state, warning state and abnormal state.

8. The method according to claim 7, characterized in that The steps for constructing the training data set for the fine monitoring module also include: Statistical features are extracted from the preprocessed high-frequency time series data, and the statistical features are copied to form a statistical feature tensor with the same length as the high-frequency time series data; The preprocessed high-frequency time series data and the corresponding statistical feature tensor, the preprocessed low-frequency asynchronous data and the corresponding mask tensor are used as the input of the fine monitoring module.

9. The method according to claim 7, characterized in that The steps for constructing the training data set for the fine monitoring module also include: The time interval feature is added to the preprocessed low-frequency asynchronous data according to the following formula, and normalized by log compression to generate a time encoding tensor: ， in, is the time interval between two measurements, To set the time limit cutoff, is the time scale parameter; The preprocessed high-frequency time series data and the corresponding statistical feature tensor, the preprocessed low-frequency asynchronous data and the corresponding mask tensor and time coding tensor are used as the input of the fine monitoring module.

10. The method according to claim 4, characterized in that The training process of the fine monitoring module adopts a loss function consisting of three-classification loss and time index positioning loss. : ， in is the three-category loss, is the time index positioning loss, is the weight, , , is the number of training samples, For the three classification serial numbers, is the training sample number, is the probability of three classifications, is the real fine state label, The weight assigned to the model at the correct time point, is the set of training samples that need time positioning, is the sequence number of the training sample that needs time positioning, and When the set is empty .

Images