Deep learning-based implantable ventricular assist device operational monitoring method

Abstract

The application discloses a kind of based on deep learning's implantation ventricular assist device operation monitoring method, comprising the following steps: collection ventricular assist device operation related multi-source time series monitoring data;Build including basic monitoring module and accurate monitoring module operation state prediction model;Basic monitoring module includes cascaded local feature extraction module, feature summary module and first classification module;Fine monitoring module includes high-frequency time series data feature extraction module, low-frequency asynchronous data condition coding module, fusion module and second classification module.Analysis is carried out to real-time monitoring data using operation state prediction model, and the current operation state information of ventricular assist device is output.The method of the present application has the advantages of timely fault warning, high recognition accuracy, strong model generalization, etc., and can realize dynamic monitoring and intelligent diagnosis of the operation state of ventricular assist device.

Description

Technical Field

[0001] This invention belongs to the field of medical artificial intelligence technology, and specifically relates to a method for monitoring the operation of an implantable ventricular assist device based on deep learning. Background Technology

[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 one of the important means of treating end-stage heart failure. With the increasing clinical application of VADs, especially left ventricular assist devices (LVADs), their long-term operational status has a crucial impact on patients' quality of life and physiological safety. Therefore, how to accurately assess the operational status of implanted devices and provide early warning of malfunctions has become an important issue for remote management and precise maintenance after implantation.

[0003] Currently, many ventricular assist devices (VADs) still rely primarily on preset programs or manual follow-up to assess device status, lacking intelligent operational monitoring mechanisms. During long-term operation, VADs may experience problems such as mechanical fatigue, abnormal flow, sensor misalignment, intracavitary thrombosis, and motor overheating, leading to device performance degradation or even malfunction. Failure to detect these abnormalities in a timely manner will severely impact the effectiveness of the VAD and may 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 singular, mainly relying on threshold analysis of parameters such as pump speed and flow rate, which cannot capture deep-seated time-series characteristics and the correlation between multi-source data;

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

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

[0008] To address the aforementioned problems, this invention provides a deep learning-based method for monitoring the operation of implanted ventricular assist devices.

[0009] The technical solution for achieving the objective of this invention is as follows:

[0010] A deep learning-based method for monitoring the operation of implanted ventricular assist devices, comprising:

[0011] By constructing a deep monitoring model that integrates convolutional neural networks, recurrent neural networks, and attention mechanisms, intelligent monitoring of the operational status of implanted ventricular assist devices is achieved. The deep monitoring model includes a basic monitoring module, which comprises a cascaded local feature extraction module, a feature summarization module, and a first classification module.

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

[0013] The feature aggregation module includes a first global average pooling layer, which is used 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, used to output the operational status information of the implanted ventricular assist device.

[0015] As a further optimization of the present invention, the steps for constructing the training dataset of 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] The high-frequency time-series data collected is preprocessed, including missing value imputation, outlier handling, and time synchronization.

[0018] Basic status labels are applied to the preprocessed high-frequency time series data, which include two categories: normal operation and abnormal operation.

[0019] As a further optimization of the present invention, the step of constructing the training dataset 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 inputs to the basic monitoring module.

[0022] As a further optimization of the present invention, the deep monitoring model also includes a fine monitoring module. When the classification result of the basic monitoring module indicates 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 coding 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 convolutional layer, a fourth one-dimensional convolutional layer, a fifth one-dimensional convolutional layer, a first LSTM layer, a second LSTM layer, a sixth one-dimensional convolutional layer, a first upsampling layer, and a one-dimensional thinning convolutional layer, in order to obtain high-dimensional and high-resolution time series features.

[0024] The low-frequency asynchronous data conditional coding module includes a cascaded third fully connected layer and a fourth fully connected layer to obtain high-dimensional feature vectors. The high-dimensional feature vectors are used as inputs to two parallel parameter branches. One parameter branch 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 offset parameters to high-dimensional and high-resolution time-series features. Conditional modulation is achieved 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, which are used to output the classification results.

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

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

[0029] As a further optimization of the present invention, the steps for constructing the training dataset for the fine-grained monitoring module include:

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

[0031] The high-frequency time-series data collected is preprocessed, including missing value imputation, outlier handling, and time synchronization.

[0032] The collected low-frequency asynchronous data is preprocessed, including encoding, missing value imputation, and generation of a missing value mask.

[0033] Fine-grained status labels are applied to the preprocessed high-frequency time-series data and low-frequency asynchronous data. These fine-grained status labels include two categories: normal status, warning status, and abnormal status.

[0034] As a further optimization of the present invention, the step of constructing the training dataset 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, as well as the preprocessed low-frequency asynchronous data and the corresponding mask tensor, are used as inputs to the fine monitoring module.

[0037] As a further optimization of the present invention, the step of constructing the training dataset for the fine monitoring module further includes:

[0038] The preprocessed low-frequency asynchronous data is augmented with time interval features using the following formula, and then normalized using log compression to generate a time-coded tensor:

[0039]

[0040] in, The time interval between the two measurements, To set a time limit for truncation, For time scale parameters;

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

[0042] As a further optimization of the present invention, the training process of the fine monitoring module adopts a loss function consisting of two parts: a three-class classification loss and a time index localization loss. :

[0043]

[0044] in For three-class loss, Locating the loss for time indexing, As weight, , , The number of training samples, It is a three-category serial number. The training sample number. For three-class classification probabilities, For realistic and detailed status labels, The weights assigned to the model at the correct time points. This is the set of training samples that require time-based localization. This refers to the sequence number of the training sample that requires time-based localization, and When it is an empty set .

[0045] The present invention also provides a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements the steps of the deep learning-based implantable ventricular assist device operation monitoring method described above.

[0046] The present invention also provides an electronic device, characterized in that it comprises:

[0047] Memory, used to store computer programs;

[0048] A processor is used to execute the computer program to implement the steps of the deep learning-based implantable ventricular assist device operation monitoring method described above.

[0049] Compared with the prior art, the significant advantages of this invention are:

[0050] This invention introduces deep learning technology into the field of ventricular assist device (VAD) health monitoring. The model can automatically learn complex features and correlations from multi-source data, far exceeding the ability to manually set thresholds. By modeling time dependence and focusing on key signs through attention mechanisms, it can detect performance degradation and fault signs that are difficult to detect in time using traditional methods, achieving early prediction and warning of anomalies. Through a data window partitioning method driven by abnormal events, it efficiently captures key signs in the early stages of abnormal signals, significantly improving the sensitivity of anomaly monitoring and warning. For the first time, a dual-layer monitoring model combining basic and fine-grained monitoring is proposed. The combination of rapid early warning from the basic model and accurate judgment from the fine-grained model significantly improves system efficiency and accuracy, overcoming the traditional problems of large computational load and response delay of single complex models. It has the advantages of timely fault warning, high recognition accuracy, and strong model generalization, enabling dynamic monitoring and intelligent diagnosis of the operating status of VADs. It is suitable for remote evaluation and precise maintenance management of implanted devices after surgery.

[0051] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0052] Figure 1 This is an overall flowchart of the method of the present invention.

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

[0054] Figure 3 This is a structural diagram of the detailed monitoring model.

[0055] Figure 4 This is a flowchart of real-time monitoring and early warning output.

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

[0057] Figure 6 It is a one-dimensional thinned convolution structure diagram. Detailed Implementation

[0058] This invention provides a deep learning-based method for monitoring the operation of implanted ventricular assist devices, such as... Figure 1 As shown, it includes the following steps:

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

[0060] S11. Multi-source data acquisition: High-frequency time-series data acquisition includes real-time output of equipment operating parameters (pump power, pump speed, estimated flow rate) from the VAD controller and real-time acquisition of patient vital signs (heart rate, arterial pressure, blood oxygen saturation, electrocardiogram, etc.) from 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 from electronic medical record systems and postoperative follow-up, such as patient symptoms and signs (fatigue, edema, dizziness, etc.), lactate, NYHA functional classification, medication adherence, postoperative days, etc. This data is updated at a lower frequency and the time step is uncertain.

[0061] S12, Data Preprocessing:

[0062] S121. High-frequency time-series data preprocessing: For seven types of high-frequency time-series data—pump power, pump speed, estimated flow rate, heart rate, arterial pressure, blood oxygen saturation, and electrocardiogram—missing value imputation, outlier processing, and time synchronization are performed sequentially. For short-term missing data, linear interpolation is used for completion. For noise-type anomalies such as transient jumps and single-point noise, numerical values ​​are directly pruned or removed. Finally, all channel data are resampled along a unified time axis, with 128 sampling points per cardiac cycle (T=128), ensuring that multi-source signals correspond 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 functional class, medication adherence, and postoperative days, perform the following steps:

[0064] S1221. Feature Coding: For continuous variables such as lactate and postoperative days, Z-score standardization is used; for ordinal variables (NYHA), one-hot coding (1 to 4) is used; for whether medication is taken on time (i.e., medication adherence), whether fatigue symptoms occur, whether edema symptoms occur, and whether dizziness symptoms occur, Boolean variable binarization (Yes=1, No=0) is used respectively.

[0065] S1222 Missing Value Handling and Mask Labeling: For continuous variables such as lactate and postoperative days, missing values ​​are filled using the mean imputation method. For data with a clear rank trend, such as NYHA, a trend-guided segmented interpolation strategy is used to proportionally interpolate and fill in intermediate ranks, maintaining the trend consistency of the rank sequence. For binarized data, forward imputation is performed using the most recent valid follow-up record. A missing value mask is generated (if the mask is 0, the data is missing; if the mask is 1, the data is not missing). The generated mask tensor is input into the subsequent model along with asynchronous data encoding features to enhance the model's ability to identify and utilize missing patterns and data integrity.

[0066] S2, Data Augmentation and Dataset Design

[0067] S21, Data Augmentation

[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 discriminative power of event features. The rate of change is approximated by the first derivative of the difference between adjacent time points, reflecting the intensity of instantaneous fluctuations. The three statistics are copied T times respectively, ultimately forming a tensor of the same length as the time series features.

[0069] S212, Low-Frequency Asynchronous Data Enhancement: Adds a time interval feature of "time since current measurement", and through... Compression normalization can be calculated using the following formula:

[0070]

[0071] in, The time interval since the last measurement. The upper limit is used as a cutoff point; anything exceeding it is considered equally "outdated." The time scale is used. The generated time-encoded tensor is input into subsequent models along with asynchronous data-encoded features 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 volume of 5T. Each window is a sample unit. Manual or semi-automatic anomaly labeling is performed on the data. The judgment of abnormal events mainly consists of the following two steps: (1) Automatic detection based on pathological abnormalities: Utilizing the statistical characteristics of calculated pump power, pump speed, expected flow, heart rate, arterial pressure, etc., when any indicator continuously deviates from the physiological reasonable range beyond the preset amplitude and duration threshold, it is automatically marked as an abnormal candidate interval. (2) Expert-annotated or recorded abnormal event logs: Combining historical monitoring curves, clinical intervention records, and physician experience, clinical experts mark key state change time points in the abnormal candidate area, including aspiration events, organ hypoperfusion, aortic valve closure time abnormalities, etc. The system uses the abnormal samples obtained above as event nodes n and labels them with abnormal tags; at the same time, it labels the n-1th window as an abnormal precursor, the n-2th window as a possible abnormal precursor, and the rest as normal. This approach enables 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 labeled device operating status tags, construct a sample set in a ratio of 6:1:1:1, corresponding to four categories: normal, abnormal, abnormal precursor, and possible abnormal precursor, in order to alleviate the category imbalance problem.

[0074] S3. Operational Monitoring Model Construction: A hierarchical, dual-channel deep learning diagnostic framework is constructed, including a basic monitoring module and a fine-grained monitoring module, to achieve synergy between low-latency preliminary early warning and high-precision deep diagnosis. The model construction steps include:

[0075] S31, Basic Monitoring Module: Construct 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, to provide rapid preliminary early warning.

[0076] S311. Input: The input to this network consists of four types of high-frequency time-series data: pump speed, pump power, projected flow rate, and heart rate (these four types of high-frequency data that are easy to monitor are preferred as input; in specific applications, the selection can be made according to needs, or all high-frequency data can be selected as input), and 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 (4 channels, 5T in length time series); therefore, the overall input structure after merging 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 (combining ReLU activation function and max pooling) to capture short-term trend changes in the signal; then, a global average pooling layer is used to compress the time dimension, forming a compact global statistical feature representation; finally, two fully connected layers output the diagnostic results, and a sigmoid activation function is used to estimate the anomaly probability. When an anomaly is detected, an anomaly warning is triggered, and subsequent fine-tuning monitoring modules are 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, along with their statistical characteristics and corresponding labels. In the label definition, "abnormal," "abnormal precursor," and "potential abnormal precursor" are uniformly labeled as the abnormal category, thus forming a binary classification task. During training, the binary cross-entropy loss function is used as the objective function, and the optimizer chosen is Adam (learning rate 1×10⁻⁻⁶). 4 The model incorporates an early stopping mechanism to prevent overfitting; model performance is comprehensively evaluated using accuracy, recall, and F1-score.

[0079] S32. Fine Monitoring Module (Activated only when the Basic Monitoring Module triggers an anomaly warning): When the Basic Monitoring Module detects an anomaly, 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 timing channel input: This channel takes seven high-frequency timing data and their corresponding statistical characteristics as input: pump speed, pump power, expected flow rate, heart rate, arterial pressure, blood oxygen saturation, and electrocardiogram. 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-coded and masked data. These three types of information constitute nine feature channels, with an initial data structure of 9×1, meaning each sample corresponds to a set of static asynchronous features.

[0083] S322, Structure: It is divided into a high-frequency time-series data feature extraction channel and a low-frequency asynchronous data conditional coding channel.

[0084] S3221, High-frequency time series data feature extraction channel: High-frequency time series data first undergoes three layers of one-dimensional convolution (e.g., Figure 5As shown in the diagram, each layer employs ReLU activation and max pooling to extract local temporal pattern features layer by layer. Subsequently, the feature sequence is input 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, then upsampling to restore the time length to the original resolution, and finally passes through a thinning convolution (Conv1d→ReLU→LayerNorm, as shown in the diagram). Figure 6 As shown, 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. Based on this, two parallel parameter branches are set to generate channel scaling respectively. With bias The scaling parameter is further adopted To ensure numerical stability. , Extending the time dimension to align with high-frequency features provides modulation parameters for subsequent FiLM fusion.

[0086] S3223, Dual-channel fusion: The fusion module uses the FiLM method to scale the channel parameters generated by the low frequency. and channel offset parameters By applying the high-frequency features and performing conditional modulation through channel-by-channel affine transformation, the fused time-series features are obtained.

[0087] S3224, Output Module: Inputs the fused feature sequence into the temporal attention pooling layer and calculates the attention weights for each time step. (After softmax normalization), and then weighted summation to obtain the global feature vector. The vector is sequentially processed through a fully connected layer, ReLU activation, Dropout, another fully connected layer, and Softmax, outputting a three-class classification result: Normal, Anomaly Warning (including possible early signs and early signs of anomalies), and Abnormal. When the classification result is "Anomaly Warning" or "Abnormal," based on... Time distribution for key time indexing: taking the location of the main peak It serves as a time point index and can be converted to physical time based on the sampling rate; if the category is "normal", the time index is not returned.

[0088] S323. Training and Optimization: A multi-task learning framework is adopted, and the loss function consists of two parts: a three-class classification loss and a temporal index localization loss.

[0089]

[0090] in For three-class loss, Locating the loss for time indexing, As weight.

[0091]

[0092]

[0093] in For the sample size, It is a three-category serial number. For three-class classification probabilities, This is a real label. The weights assigned to the model at the correct time points. For the sample set that requires time-based location (the actual category is "warning / anomaly"), if the current batch ,definition .

[0094] The optimization process uses the Adam optimizer combined with a cosine annealing learning rate scheduling strategy, while introducing weight decay to reduce the risk of overfitting. Performance evaluation employs three types of metrics for comprehensive verification: First, classification performance metrics, including overall classification accuracy, F1-score, and AUC curve, measure the model's ability to distinguish different health states, with a focus on the sensitivity to anomaly detection; second, temporal localization performance metrics, using mean absolute error (MAE) to measure the deviation between the predicted temporal index and the true index; and third, runtime efficiency metrics, statistically analyzing the inference latency of the fine-grained diagnostic module after triggering, ensuring that the model meets accuracy requirements while possessing real-time diagnostic capabilities.

[0095] S4. Real-time operation monitoring and early warning output: such as Figure 4 As shown, the basic monitoring module and the fine monitoring module, which have been trained, are used to analyze 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 of the patient's implanted ventricular assist device, and acquisition of the most recent low-frequency asynchronous data. All acquired data are normalized, time-aligned, and augmented according to the same preprocessing procedures as in the training phase to ensure that the format and distribution of the 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 invoked, inputting complete high-frequency time-series data and low-frequency asynchronous data together. After processing, a three-category health status (normal, abnormal warning, abnormal) is output, and a 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, and key time points or time intervals are located through peak values, providing a reference for clinical tracing and decision-making.

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

[0099] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations 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, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0100] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0101] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The steps of the function specified in one or more boxes.

[0102] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A deep learning based implanted ventricular assist device operational monitoring method, characterized in that, The application relates to a deep monitoring model for monitoring the running state of an implanted ventricular assist device. The deep monitoring model comprises a basic monitoring module, which comprises a cascaded local feature extraction module, a feature summary module and a first classification module. The local feature extraction module comprises a cascaded first one-dimensional convolution layer and a second one-dimensional convolution layer, which are used to extract local features of the model input. The feature summary module comprises a first global average pooling layer, which is used to compress the local features to form a global statistical feature representation. The first classification module comprises a cascaded first full connection layer, a second full connection layer and a Sigmoid activation function, which are used to output the running state information of the implanted ventricular assist device. The deep monitoring model further comprises a fine monitoring module, which is started when the classification result of the basic monitoring module is abnormal running. The fine monitoring module comprises a high-frequency time series data feature extraction module, a low-frequency asynchronous data condition coding module, a fusion module and a second classification module. The high-frequency time series data feature extraction module comprises 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 up-sampling layer and a one-dimensional refining convolution layer, which are used to obtain high-dimensional and high-resolution time series features. The low-frequency asynchronous data condition coding module comprises a cascaded third full connection layer and a fourth full connection layer, which are used to obtain a high-dimensional feature vector. The high-dimensional feature vector is input into two parallel parameter branches, one of which comprises a channel scaling unit and a first dimension expansion unit, and the other of which comprises a channel bias unit and a second dimension expansion unit.

2. The method of claim 1, wherein, The fusion module comprises a FilM adjustment unit, which adopts a FiLM method to apply the channel scaling parameter and the channel bias parameter to the high-dimensional and high-resolution time series features, realizes condition modulation through a channel-by-channel affine transformation, and obtains fused time series features. The second classification module comprises a cascaded time attention pooling layer, a fifth full connection layer, a sixth full connection layer and a Softmax activation function, which are used to output a classification result. The training data set construction steps of the basic monitoring module comprise: Collecting high-frequency time series data, including recording VAD device running parameters and patient vital signs at a fixed sampling rate; 3. The method of claim 2, wherein, Pretreating the collected high-frequency time series data, including missing value filling, abnormal value processing and time synchronization; Labeling the pretreated high-frequency time series data with a basic state label, wherein the basic state label comprises two categories of normal running and abnormal running. The training data set construction steps of the basic monitoring module further comprise:

4. The method of claim 1, wherein, Extracting statistical features from the pretreated high-frequency time series data, and copying the statistical features to form a statistical feature tensor with the same length as the high-frequency time series data; 5. The method of claim 1, wherein, The fused time sequence feature inputs a time attention pooling layer, calculates attention weights for each time step, and takes as a time point index, wherein is the attention weight of the time step t.

6. The method of claim 1, wherein, Taking the pretreated high-frequency time series data and the corresponding statistical feature tensor as the input of the basic monitoring module. The one-dimensional refining convolution layer comprises a cascaded seventh one-dimensional convolution layer, a ReLU function and a layer normalization. The training data set construction steps of the fine monitoring module comprise: Multi-source data collection, including high-frequency time series data and low-frequency asynchronous data, wherein the high-frequency time series data includes VAD device operating parameters and patient vital signs recorded at a fixed sampling rate, and the low-frequency asynchronous data includes structured information obtained from electronic medical record systems and postoperative follow-up; Preprocessing the collected high-frequency time series data, including missing value filling, outlier processing, and time synchronization; Preprocessing the collected low-frequency asynchronous data, including encoding, missing value filling, and generating a missing mask; Labeling the preprocessed high-frequency time series data and low-frequency asynchronous data with fine state labels, including normal state, early warning state, and abnormal state.

7. The method of claim 6, wherein, The training data set construction step of the fine monitoring module further includes: Statistical feature extraction is performed on 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 corresponding statistical feature tensor, preprocessed low-frequency asynchronous data and corresponding mask tensor are used as inputs of the fine monitoring module.

8. The method of claim 6, wherein, The training data set construction step of the fine monitoring module further includes: According to the following formula, time interval features are added to the preprocessed low-frequency asynchronous data, and normalized by log compression to generate a time encoding tensor: ， wherein, is the time interval between two measurements, is the set upper time cut-off, is the time scale parameter; The preprocessed high-frequency time series data and corresponding statistical feature tensor, preprocessed low-frequency asynchronous data and corresponding mask tensor and time encoding tensor are used as inputs of the fine monitoring module.

9. The method of claim 1, wherein, The loss function used in the training process of the fine monitoring module is composed of a three-classification loss and a time index positioning loss : ， wherein is a tri-class loss, is a time index localization loss, is a weight, , , is the number of training samples, is a tri-class number, is a training sample number, is a tri-class probability, is a true fine state label, is a weight assigned by the model on the correct time point, is a set of training samples that need time localization, is a training sample number that needs time localization, and is empty .

Images