Neural network model for predicting neonatal acute renal failure based on urine volume time sequence data
By using a neural network model based on urine volume time series data, combined with multiple time windows and parallel network structure, and dynamically adjusting sample weights, the timeliness and accuracy issues of predicting acute renal failure in neonates were solved, achieving early warning and personalized medical treatment.
Patent Information
- Application Number
- CN202510755267.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-06
- Publication Date
- 2025-09-19
AI Technical Summary
Existing technologies lack timeliness, accuracy, and generalization capabilities in predicting acute renal failure in neonates. They rely on the delayed and invasive detection of biomarkers such as serum creatinine. Traditional machine learning methods have strong reliance on feature engineering, poor information fusion capabilities, and weak adaptability to small samples, making them unable to achieve early warning and personalized medicine.
A neural network model based on urine volume time series data is adopted. Features are extracted through a multi-time window mechanism, basic patient information is integrated, and a parallel convolutional neural network is combined with a fully connected network branch. The loss function and sample scheduling strategy are used to dynamically adjust the sample weights. The dynamic gating mechanism regulates feature fusion, and transfer learning and knowledge distillation technology are used to improve model performance.
It achieves multi-level capture of early signals of acute renal failure in neonates, reduces the missed detection rate, ensures the timeliness, accuracy and generalization ability of prediction, and supports personalized medicine.
Smart Images
Figure CN120674067A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of medical artificial intelligence technology, and specifically to a neural network model for predicting acute renal failure in neonates based on urine volume time series data. Background Art
[0002] Neonatal acute kidney injury (AKI) is a significant clinical issue in neonatal intensive care, and early, accurate prediction is crucial for improving outcomes. Existing technologies primarily rely on clinical laboratory tests (such as serum creatinine) and expert judgment. However, biomarkers like serum creatinine require several hours to days after renal impairment to detect lesions, making early warning difficult. These tests rely on invasive testing methods (such as renal biopsy), increasing the risk of infection and medical burden on neonates. Furthermore, expert judgment involves infrequent manual monitoring, which fails to capture the dynamic fluctuations of physiological indicators such as urine output.
[0003] In addition, although traditional machine learning methods attempt to use urine volume data for prediction, they face the following technical bottlenecks: strong dependence on feature engineering: manual design of time window statistics is required; poor information fusion capabilities: inability to effectively integrate heterogeneous data (such as collaborative analysis of patient vital signs and time series); weak adaptability to small samples: the performance of existing methods significantly decreases when the data volume is insufficient; insufficient false alarm suppression: limited ability to identify false-negative samples.
[0004] Therefore, existing technologies have significant deficiencies in the timeliness, accuracy, and generalization ability of neonatal AKI prediction, and an intelligent solution is urgently needed to achieve early warning and personalized medical treatment. Summary of the Invention
[0005] The purpose of this application is to provide a neural network model for predicting neonatal acute renal failure based on urine volume time series data, so as to solve the technical problems raised in the above background technology.
[0006] To achieve the above objectives, the present application discloses the following technical solutions: a neural network model for predicting neonatal acute renal failure based on urine volume time series data, comprising: A hybrid feature extraction module is configured to extract urine volume temporal features through a multi-time window mechanism and fuse basic patient information to form a multidimensional feature matrix, wherein the basic information includes gestational age, birth weight, and Apgar score; The patient feature fusion module is configured to: standardize the patient's basic information and then splice it with the urine volume time series feature matrix; The multi-scale classification network is configured as follows: based on parallel convolutional neural network branches and fully connected network branches, cross-branch feature weighted fusion is performed through the attention mechanism; The loss function optimization module is configured to dynamically adjust the sample weights using the loss function and sample scheduling strategy.
[0007] Preferably, the hybrid feature extraction module includes: The short time window difference calculation unit is configured to: extract the difference between the short-term and long-term urine volume change rates; The shock information weighting unit is configured to: model sudden urine volume fluctuations through a nonlinear activation function; The dynamic average calculation unit is configured to use a sliding window method and age group Z-Score standardization to process the mean urine volume.
[0008] Preferably, the multi-scale classification network is specifically configured as follows: The convolutional neural network branch uses an expanded convolution structure to capture multi-scale local features; The fully connected network branch enhances deep feature expression through the residual module; A channel-spatial dual attention mechanism is used to achieve feature fusion, where the attention weights are dynamically generated by the input features.
[0009] Preferably, the multi-scale classification network is further configured as follows: the convolutional neural network branch and the bidirectional LSTM network form a hybrid architecture, and the hybrid architecture is specifically: After the dilated convolutional layer, a bidirectional LSTM layer with a time step of 6 hours is connected, and the shallow features of the convolutional neural network branch are spliced with the LSTM time series features through jump connections; The gating mechanism is used to adjust the feature fusion weight, where the gating value , is the feature matrix extracted by the convolutional neural network branch, The temporal feature matrix extracted by the bidirectional LSTM network, is the gating mechanism weight matrix, is the gating mechanism bias vector, is the Sigmoid activation function.
[0010] Preferably, the loss function optimization module specifically includes: The loss function unit is configured as follows: The loss function is specifically: , through the dynamic sample weight coefficient Balance the positive and negative sample losses, where is the regulating factor, is the predicted probability that the sample belongs to its true category; The sample ratio adjustment unit is configured to: dynamically adjust the sample ratio during training, and gradually increase the positive sample ratio to 1:5 in the late training period; The transfer learning strategy unit is configured to: use multi-center data for pre-training and fine-tune to adapt to the local data of the target hospital.
[0011] Preferably, the model, when deployed, has: The urine volume data standardization interface is configured to: support automatic interpolation of sampling frequencies of 1-4 hours / time; The model interpretation module is configured to: output feature importance heatmap; The safety threshold mechanism is configured so that abnormal data triggers the switching of the baseline model.
[0012] As a preferred approach, the model adopts the following optimization strategies for small sample scenarios: The feature space mapping module is configured to: project the input features into the latent space of the pre-trained model; Knowledge distillation technology is configured to: use large sample models to guide small sample model training; Adversarial training augmentation,is configured to add Gaussian noise and random masks.
[0013] Preferably, the input preprocessing of the model includes: Logarithmic transformation of urine volume data: ,in, , is the urine volume value; Difference processing of time series: calculate the second-order difference of the urine volume change rate at adjacent time points; Outlier correction: The cubic spline interpolation method was used to correct urine volume data that exceeded the physiological range.
[0014] Preferably, the training of the model includes the following steps: Staged frozen training: the first stage fixes the feature extraction layer to train the classifier, and the second stage uses elastic weight update; The course learning strategy prioritizes training fuzzy samples with a risk score of 0.3-0.7 in the initial stage, and dynamically adjusts the sample priority in the middle and late stages.
[0015] Preferably, the model supports dynamic expansion, which specifically includes: Widen the network by adding parallel convolution branches or deepen the network by adding residual modules; Set up an expandable channel after the pooling layer, when the feature matrix rank deficiency Dimension expansion is triggered when , and the expansion scale does not exceed 50% of the original number of channels, where is the feature matrix dimension.
[0016] Beneficial effects: The neural network model of this application for predicting neonatal acute renal failure based on urine volume time series data extracts urine volume time series features through a multi-time window mechanism, integrates the patient's basic information to form a multi-dimensional feature matrix, and combines the cross-scale feature extraction capabilities of parallel convolutional neural networks and fully connected network branches to achieve multi-level capture of early signals of neonatal acute renal failure. It also uses loss functions and sample scheduling strategies to dynamically adjust sample weights, reduce the missed detection rate of minority samples, ensure high prediction sensitivity, and thus ensure the reliability of the timeliness, accuracy and generalization ability of neonatal AKI prediction, thereby achieving early warning and personalized medicine. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without paying any creative work.
[0018] Figure 1 This is a structural diagram of a neural network model for predicting neonatal acute renal failure based on urine volume time series data; Figure 2 Schematic diagram of dynamic model training. DETAILED DESCRIPTION
[0019] The following is a clear and complete description of the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments of this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.
[0020] In this document, the term "comprising" is intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising..." does not preclude the presence of additional identical elements in the process, method, article, or apparatus that includes the element.
[0021] See also Figure 1-2 This embodiment provides a Figure 1 The neural network model shown for predicting neonatal acute renal failure based on urine volume time series data includes: A hybrid feature extraction module is configured to extract urine volume temporal features through a multi-time window mechanism and fuse basic patient information to form a multidimensional feature matrix, wherein the basic information includes gestational age, birth weight, and Apgar score; The patient feature fusion module is configured to: standardize the patient's basic information and then splice it with the urine volume time series feature matrix; The multi-scale classification network is configured as follows: based on parallel convolutional neural network branches and fully connected network branches, cross-branch feature weighted fusion is performed through the attention mechanism; The loss function optimization module is configured to dynamically adjust the sample weights using the loss function and sample scheduling strategy.
[0022] The above model extracts urine volume time series features through a multi-time window mechanism, integrates basic patient information to form a multi-dimensional feature matrix, and combines the cross-scale feature extraction capabilities of parallel convolutional neural networks and fully connected network branches to achieve multi-level capture of early signals of neonatal acute renal failure. It also uses loss functions and sample scheduling strategies to dynamically adjust sample weights, reduce the missed detection rate of minority samples, ensure high prediction sensitivity, and thus ensure the timeliness, accuracy and generalization ability of neonatal AKI prediction, thereby achieving early warning and personalized medical care.
[0023] Specifically, in this embodiment, the hybrid feature extraction module includes: The short time window difference calculation unit is configured to: extract the difference between the urine volume change rate in the short term (6-12 hours) and the long term (24-48 hours); The shock information weighting unit is configured to: model sudden urine volume fluctuations through a nonlinear activation function; The dynamic average calculation unit is configured to use a sliding window method and age group Z-Score standardization to process the mean urine volume.
[0024] Through a multi-time window mechanism, the temporal characteristics of urine volume (such as the difference in change rate between short-term 8 hours and long-term 36 hours) are extracted, and the patient's basic information (gestational age, birth weight, etc.) is integrated to form a multi-dimensional feature matrix. Combined with the cross-scale feature extraction capabilities of parallel convolutional neural networks and fully connected network branches, multi-level capture of early signals of acute renal failure in neonates can be achieved.
[0025] Furthermore, in the impact information weighting unit, the method for extracting the impact signal is specifically as follows: Step 1: Define the local difference signal as in, Represents the urine volume data at the current time point t (unit: ml / kg / h). It is the original urine volume monitoring value, which comes from the neonatal urine volume time series monitoring sequence and is the core time series feature of the model input; Indicates the current time point (unit: hour), which is used as the reference time for calculating the local difference signal. The time interval is usually 1 hour (it can support unified granularity after interpolation); Represents the urine volume data at the i-th time point in the time window, where the value range of i is , the window range is to extend k hours to the left and right from the current time t to form a symmetrical time window, for example: When the window contains There are 7 time points in total (i.e., i traverses these 7 time points); It represents the number of sample points in the time window, that is, the total number of time points contained in the window, which is determined by the window radius k. The formula is ,when hour, , that is, the window contains urine volume data at 7 time points; is the window radius, =3; Step 2: Capture the pattern association of adjacent difference signals through the convolution kernel of CNN (size 5×1); Step 3: Use the LSTM gating unit to identify persistent abnormal fluctuation patterns.
[0026] The above-mentioned method of extracting the impact signal calculates the urine volume at the current time point The difference between the average urine volume in the window is used to detect local abnormal fluctuations in urine volume. The absolute value of is large, indicating that the current urine volume deviates significantly from the recent average level, which may indicate abnormal renal function. In this way, the long-short time window difference calculation unit can identify the trend difference of urine volume change, and the impact information weighting unit uses the hyperbolic tangent activation function to detect sudden fluctuations (such as Nonlinear modeling is performed on the CNN convolution kernel (size 5×1) and the LSTM gating unit described later to achieve early warning of latent renal damage.
[0027] Specifically, the multi-scale classification network includes a convolutional neural network branch that adopts a three-layer dilated convolution structure (with dilation coefficients of 1, 3, and 5 respectively); the multi-scale classification network is specifically configured as follows: The convolutional neural network branch uses an expanded convolution structure to capture multi-scale local features; The fully connected network branch enhances deep feature expression through the residual module; A channel-spatial dual attention mechanism is used to achieve feature fusion, where the attention weights are dynamically generated by the input features.
[0028] Based on the above, the three-layer expanded convolution structure (expansion coefficient 1 / 3 / 5) of the convolutional neural network branch captures multi-scale local features, the fully connected network branch enhances deep semantic expression through the residual module, and the channel-space dual attention mechanism dynamically weights and fuses the two types of features, improving the model's ability to analyze the association of heterogeneous data (time series features + patient vital signs).
[0029] Further, if Figure 2 As shown, the multi-scale classification network is further configured as follows: the convolutional neural network branch and the bidirectional LSTM network form a hybrid architecture, and the hybrid architecture is specifically: After the dilated convolutional layer, a bidirectional LSTM layer with a time step of 6 hours is connected, and the shallow features of the convolutional neural network branch are spliced with the LSTM time series features through jump connections; The gating mechanism is used to adjust the feature fusion weight, where the gating value , It is the feature matrix extracted by the convolutional neural network branch, which is the local urine volume mutation feature output by the dilated convolution layer. The dimension is usually [batch, time step, feature channel]; It is the time series feature matrix extracted by the bidirectional LSTM network, and the long-term trend feature output by the bidirectional LSTM layer has the same dimension as Consistent, easy for channel splicing; is the gating mechanism weight matrix, which is used to perform linear transformation on the concatenated feature matrix, with a size of , is the number of feature channels; is the gating mechanism bias vector, which is used to adjust the output offset after linear transformation, and its size is [ ]; Sigmoid activation function is used to map the linear transformation result to a gated weight value between 0 and 1.
[0030] Based on this, the gate value calculation process is as follows: (1) Feature concatenation: concatenate the feature matrix With the characteristic matrix Splicing along the channel dimension to obtain the joint feature matrix:
[0031] (2) Linear transformation: weight matrix through gating mechanism and the gating mechanism bias vector Perform linear mapping on the joint features:
[0032] The output dimension is , which is consistent with the number of single-branch feature channels; (3) Activation function generates gate value: The linear output is compressed to the (0,1) interval through the Sigmoid activation function to obtain the gate weight matrix :
[0033] matrix Each element of represents the fusion weight of the CNN feature and LSTM feature at the corresponding position (0 means only LSTM features are retained, 1 means only CNN features are retained).
[0034] Secondly, the parameter learning method is as follows: (1) Training objectives: and It is a learnable parameter that is optimized through the back-propagation algorithm. The goal is to minimize the loss function (such as FocalLoss used in this embodiment) to maximize the prediction accuracy of the model for neonatal acute renal failure.
[0035] (2) Biological significance: The gating mechanism simulates the physiological regulation process of the kidney, and high gating value ( ): indicates that urine volume mutation characteristics (such as acute oliguria) dominate, which may indicate early renal damage; low gate value ( ): It suggests that the trend characteristics of urine volume (such as continuous low-level fluctuations) are more critical and reflect chronic changes in renal function.
[0036] Based on the above, this embodiment implements adaptive adjustment of feature fusion weights through a dynamic gating mechanism, solving the following problems: Feature redundancy: Avoiding interference of irrelevant features (such as CNN details of normal urine volume fluctuations) on LSTM trend analysis; Time series dependency: Dynamically assign feature weights based on the time scale of urine volume changes (e.g., 6-hour steps) to improve prediction timeliness; Interpretability: Gating values can be mapped to feature importance indicators to assist clinicians in understanding the model’s decision logic (e.g., heatmap output as described in the manual).
[0037] The loss function of this embodiment (adjustment factor γ = 2, dynamic sample weight coefficient α t ) and a dynamic sample ratio adjustment strategy (increasing the positive sample ratio to 1:5 in the middle and late stages of training) significantly reduce the missed detection rate of minority samples and maintain predictive sensitivity even in extremely unbalanced scenarios with a sample ratio of 1:33.
[0038] Secondly, the hybrid architecture of convolutional neural network branches and bidirectional LSTM networks simultaneously captures local mutation patterns (such as 5×1 convolution kernels detecting adjacent difference signals) and long-term evolution trends (LSTM gated units identifying sustained abnormal fluctuations) in urine volume data, shortening prediction delay by nearly 12 hours compared to traditional LSTM models.
[0039] Specifically, the loss function optimization module specifically includes: The loss function unit is configured as follows: The loss function is specifically: , through the dynamic sample weight coefficient Balance the positive and negative sample losses, where is the regulating factor and , is the predicted probability that the sample belongs to its true category; for positive samples (neonatal acute renal failure cases), ,in The probability that the model predicts that the sample is a positive class (the value range is ; For negative samples (non-neonatal acute renal failure cases), ,at this time Indicates the probability of predicting the negative class. In the formula, the positive sample loss term is: ,when When it is close to 1 (the prediction is correct and the confidence is high), Approaching 0, reducing the loss weight of the sample (weight decay of easy-to-classify samples); when When it is close to 0 (prediction error), the weight remains high, forcing the model to focus on difficult-to-classify samples. Negative sample loss term: , the logic is symmetrical with the positive sample, through Suppress the loss contribution of easily classified negative samples. Dynamic sample weight coefficient By formula Calculate, where is the proportion of positive samples in the current batch; It is a learnable parameter used to balance the imbalance of positive and negative samples. , which controls the weight decay rate of difficult and easy samples and is the core hyperparameter of FocalLoss; The sample ratio adjustment unit is configured to dynamically adjust the sample ratio during training, maintaining the original sample ratio (1:33) at the beginning of training and gradually increasing the positive sample ratio to 1:5 in the middle and late stages of training; The transfer learning strategy unit is configured to: use multi-center data for pre-training and fine-tune to adapt to the local data of the target hospital.
[0040] Furthermore, during model training, the five model parameters with the highest AUC of the validation set during the training process are saved and the average is taken.
[0041] In a feasible implementation, the neural network model for predicting neonatal acute renal failure based on urine volume time series data of this embodiment, when deployed, has: The urine volume data standardization interface is configured to: support automatic interpolation of 1-4 hours / time sampling frequency; The model interpretation module is configured to: output feature importance heatmap; The safety threshold mechanism is configured so that abnormal data triggers the switching of the baseline model.
[0042] In this way, clinical credibility is enhanced and the reliability of prediction results is ensured.
[0043] In a feasible implementation, the neural network model for predicting neonatal acute renal failure based on urine volume time series data in this embodiment adopts the following optimization strategy for small sample scenarios: The feature space mapping module is configured to: project the input features into the latent space of the pre-trained model; Knowledge distillation technology is configured to: use large sample models to guide small sample model training; Adversarial training enhancement, configured to: add Gaussian noise ( ) and random mask (ratio ≤ 15%).
[0044] In this way, the feature space mapping module projects the input features into the latent space of the pre-trained model. Combining knowledge distillation technology (large sample model guided training) with adversarial training enhancement, it can maintain prediction accuracy when the sample size is insufficient, solving the problem of scarce neonatal AKI data.
[0045] In a feasible implementation, the neural network model for predicting neonatal acute renal failure based on urine volume time series data in this embodiment has its input preprocessing including: Logarithmic transformation of urine volume data: ,in, , is the urine volume value; Difference processing of time series: calculate the second-order difference of the change rate of urine volume at adjacent time points; Outlier correction: Cubic spline interpolation was used to correct urine volume data that exceeded the physiological range.
[0046] In this way, the dynamic range of urine volume data is compressed through logarithmic transformation, the change rate characteristics are enhanced by second-order difference calculation, and outliers are corrected by cubic spline interpolation. After preprocessing, the signal-to-noise ratio of the data is greatly improved, and the noise interference of the model input is significantly reduced.
[0047] In a feasible implementation, the training process of the neural network model for predicting neonatal acute renal failure based on urine volume time series data in this embodiment includes the following steps: The training is frozen in stages. In the first stage, the classifier is trained with a fixed feature extraction layer (50 rounds). In the second stage, elastic weight update is used. The update formula is: in, Represents model parameters During the model training process, the model parameters need to be continuously adjusted to minimize the loss function. That is, the parameters are updated every time The change value of The learning rate controls the step size of the model's parameter updates during training. If the learning rate is set too high, the model may skip the optimal solution during training, resulting in failure to converge. If the learning rate is set too low, the model's training speed will be very slow. In actual training, the learning rate is usually adjusted based on the model's performance and the characteristics of the dataset. is a parameter The gradient represents the rate of change of the loss function with respect to the parameters, which is calculated in the back propagation algorithm. It indicates the direction in which the parameters are updated to make the loss function drop fastest, and sets the gradient cutoff threshold. ; The parameter update variance records the changes in the parameter update process and is used to adaptively adjust the step size of the parameter update. By considering the parameter update variance, the model can be made more stable during the training process and avoid excessive or unstable parameter updates. is a very small constant (usually very small, such as , its role is to prevent the denominator from being zero and ensure that the calculation of the formula is numerically stable. In actual calculations, due to the limitations of the computer's floating-point precision, when Very small hours, A division by zero error may occur, add This can be avoided; The course learning strategy prioritizes training fuzzy samples with a risk score of 0.3-0.7 in the initial stage, and dynamically adjusts the sample priority in the middle and late stages.
[0048] Furthermore, the sample input order optimization method corresponding to the course learning strategy includes the following steps: Step 1: Implement a curriculum learning strategy, prioritizing risk score input at the beginning of training. Blurred samples of Step 2: Dynamic priority sorting is used in the middle and late stages, and the sample weights are: in Predict probabilities for the current model; Step 3: Set up a sample buffer to keep the latest 200 misclassified samples for cyclic retraining.
[0049] Furthermore, the risk score of the course learning strategy is calculated as follows: The pre-trained risk scoring model uses the XGBoost algorithm, and the input includes: urine volume variation coefficient, negative fluctuation duration ratio, and the area under the urine volume curve (AUC) in the past 24 hours. , Scoring output is calibrated: , is the cumulative distribution function (CDF) of the standard normal distribution.
[0050] In this way, the course learning strategy optimizes the sample input order in stages (initial training prioritizes fuzzy samples with risk scores of 0.3-0.7), combined with elastic weight update Combined with the gradient truncation technology, it effectively suppresses overfitting and reduces the fluctuation of the AUC of the validation set by more than 30% during training.
[0051] In a feasible implementation, the neural network model for predicting neonatal acute renal failure based on urine volume time series data in this embodiment includes the following steps for verification: A time-series cross-validation strategy was used to group patients by ID to ensure that the training set and validation set patients did not overlap; Evaluation metrics include time-sensitive AUC-ROC and AUC-PR; A delayed verification strategy was implemented, and the predicted results were compared with the clinical diagnosis results 48 hours later.
[0052] In a feasible implementation, the neural network model for predicting neonatal acute renal failure based on urine volume time series data in this embodiment supports dynamic expansion, and the dynamic expansion specifically includes: Widen the network by adding parallel convolution branches or deepen the network by adding residual modules; Set up an expandable channel after the pooling layer, when the feature matrix rank deficiency Dimension expansion is triggered when , and the expansion scale does not exceed 50% of the original number of channels, where is the feature matrix and adjust the feature normalization method by replacing the pooling layer type (max pooling, average pooling, mixed pooling).
[0053] In a feasible implementation, the neural network model for predicting neonatal acute renal failure based on urine volume time series data of this embodiment has a dynamic information injection mechanism, specifically: Auxiliary input nodes are set in the middle layer of the network to allow clinical indicators to be injected during the feature abstraction stage.
[0054] The injection method uses adaptive scaling: , is the scaling factor; Clinical information includes but is not limited to: frequency of antibiotic use and vasoactive inhaler score (VIS).
[0055] In this way, by adding parallel convolution branches to widen the network or by adding residual modules to deepen the network, an expandable channel is set after the pooling layer, so that the model can adapt to different data complexities, and the intermediate layer clinical indicator injection interface (such as the frequency of antibiotic use and vasoactive drug scores) allows the real-time integration of the latest diagnosis and treatment information, thereby improving the generalization ability of the model.
[0056] In a feasible implementation, the neural network model for predicting neonatal acute renal failure based on urine volume time series data of this embodiment has an information aggregation method, specifically: Set distribution alignment loss between multiple layers of MLP: .
[0057] Implement layer-by-layer information distillation: the output of the lth layer contains the attention-weighted sum of the features of the previous l-1 layers.
[0058] A manifold learning constraint is adopted: the cosine distance of clinically similar patients is kept ≤ 0.2 in the latent space.
[0059] In summary, the neural network model of this embodiment for predicting neonatal acute renal failure based on urine volume time series data integrates the CNN-LSTM network to capture urine volume mutation patterns and time series trends, addresses the limitations of traditional feature engineering, balances sample distribution through optimized loss functions and curriculum learning strategies, improves small sample generalization capabilities, supports dynamic injection of clinical indicators and adaptive expansion of the architecture, enhances cross-scenario adaptability and interpretability, and significantly improves the timeliness, accuracy, and generalization capabilities of neonatal AKI prediction, enabling early warning and personalized medicine.
[0060] In the embodiments provided herein, it should be understood that the embodiments described herein can be implemented using hardware, software, firmware, middleware, code, or any appropriate combination thereof. For hardware implementation, the processor may be implemented in one or more of the following: an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field-programmable gate array (FPGA), a processor, a controller, a microcontroller, a microprocessor, or other electronic units designed to implement the functionality described herein, or any combination thereof. For software implementation, some or all of the processes of the embodiments may be performed by a computer program instructing the relevant hardware. During implementation, the program may be stored in a computer-readable storage medium or transmitted as one or more instructions or codes on a computer-readable storage medium. Computer-readable storage media include computer storage media and communication media, wherein communication media includes any medium that facilitates the transmission of a computer program from one location to another. The storage medium may be any available medium that can be accessed by a computer. Computer-readable storage media may include, but are not limited to, RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing the desired program code in the form of instructions or data structures and accessible by a computer.
[0061] Finally, it should be noted that the above is only a preferred embodiment of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or make equivalent replacements for some of the technical features therein. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.
Claims
1. A neural network model for predicting neonatal acute renal failure based on urine volume time series data, characterized in that: include: A hybrid feature extraction module is configured to extract urine volume temporal features through a multi-time window mechanism and fuse basic patient information to form a multidimensional feature matrix, wherein the basic information includes gestational age, birth weight, and Apgar score; The patient feature fusion module is configured to: standardize the patient's basic information and then splice it with the urine volume time series feature matrix; The multi-scale classification network is configured as follows: based on parallel convolutional neural network branches and fully connected network branches, cross-branch feature weighted fusion is performed through the attention mechanism; The loss function optimization module is configured to dynamically adjust the sample weights using the loss function and sample scheduling strategy.
2. The neural network model for predicting neonatal acute renal failure based on urine volume time series data according to claim 1, characterized in that: The hybrid feature extraction module includes: The short time window difference calculation unit is configured to: extract the difference between the short-term and long-term urine volume change rates; The shock information weighting unit is configured to: model sudden urine volume fluctuations through a nonlinear activation function; The dynamic average calculation unit is configured to use a sliding window method and age group Z-Score standardization to process the mean urine volume.
3. The neural network model for predicting neonatal acute renal failure based on urine volume time series data according to claim 2, characterized in that: The multi-scale classification network is specifically configured as follows: The convolutional neural network branch uses an expanded convolution structure to capture multi-scale local features; The fully connected network branch enhances deep feature expression through the residual module; A channel-spatial dual attention mechanism is used to achieve feature fusion, where the attention weights are dynamically generated by the input features.
4. The neural network model for predicting neonatal acute renal failure based on urine volume time series data according to claim 3, characterized in that: The multi-scale classification network is further configured as follows: the convolutional neural network branch and the bidirectional LSTM network form a hybrid architecture, and the hybrid architecture is specifically: After the dilated convolutional layer, a bidirectional LSTM layer with a time step of 6 hours is connected, and the shallow features of the convolutional neural network branch are spliced with the LSTM time series features through jump connections; The gating mechanism is used to adjust the feature fusion weight, where the gating value , is the feature matrix extracted by the convolutional neural network branch, The temporal feature matrix extracted by the bidirectional LSTM network, is the gating mechanism weight matrix, is the gating mechanism bias vector, is the Sigmoid activation function.
5. The neural network model for predicting neonatal acute renal failure based on urine volume time series data according to claim 1, characterized in that: The loss function optimization module specifically includes: The loss function unit is configured as follows: The loss function is specifically: , through the dynamic sample weight coefficient Balance the positive and negative sample losses, where is the regulating factor, is the predicted probability that the sample belongs to its true category; The sample ratio adjustment unit is configured to: dynamically adjust the sample ratio during training, and gradually increase the positive sample ratio to 1:5 in the late training period; The transfer learning strategy unit is configured to: use multi-center data for pre-training and fine-tune to adapt to the local data of the target hospital.
6. The neural network model for predicting neonatal acute renal failure based on urine volume time series data according to claim 1, characterized in that: When deployed, the model has: The urine volume data standardization interface is configured to: support automatic interpolation of sampling frequencies of 1-4 hours / time; The model interpretation module is configured to: output feature importance heatmap; The safety threshold mechanism is configured so that abnormal data triggers the switching of the baseline model.
7. The neural network model for predicting neonatal acute renal failure based on urine volume time series data according to claim 1, characterized in that: The model adopts the following optimization strategies for small sample scenarios: The feature space mapping module is configured to: project the input features into the latent space of the pre-trained model; Knowledge distillation technology is configured to: use large sample models to guide small sample model training; Adversarial training augmentation,is configured to add Gaussian noise and random masks.
8. The neural network model for predicting neonatal acute renal failure based on urine volume time series data according to claim 1, characterized in that: The input preprocessing of the model includes: Logarithmic transformation of urine volume data: ,in, , is the urine volume value; Difference processing of time series: calculate the second-order difference of the urine volume change rate at adjacent time points; Outlier correction: The cubic spline interpolation method was used to correct urine volume data that exceeded the physiological range.
9. The neural network model for predicting neonatal acute renal failure based on urine volume time series data according to claim 1, characterized in that: The training of the model includes the following steps: Staged frozen training: the first stage fixes the feature extraction layer to train the classifier, and the second stage uses elastic weight update; The course learning strategy prioritizes training fuzzy samples with a risk score of 0.3-0.7 in the initial stage, and dynamically adjusts the sample priority in the middle and late stages.
10. The neural network model for predicting neonatal acute renal failure based on urine volume time series data according to claim 1, characterized in that: This model supports dynamic expansion, which specifically includes: Widen the network by adding parallel convolution branches or deepen the network by adding residual modules; Set up an expandable channel after the pooling layer, when the feature matrix rank deficiency Dimension expansion is triggered when , and the expansion scale does not exceed 50% of the original number of channels, where is the feature matrix dimension.