AP-AKI Prediction Method and Bedside Tool Based on Liquid Neural Network

Abstract

This invention discloses an AP-AKI prediction method and bedside tool based on a liquid neural network, specifically relating to the field of machine learning. It dynamically collects latent clinical features of AP patients as input features, constructs a temporal feature sequence, and preprocesses this sequence as the basic data for the LNN's input layer. A dynamic time constant containing the input features is introduced into the LNN's liquid layer to achieve adaptive response to rapidly / slowly changing clinical signals in AP patients. Weight pruning and sparse neuron activation are introduced into the LNN's liquid layer to construct a sparse topology. The numerical solution of the neural ordinary differential equations in the LNN's liquid layer is replaced with a closed-form analytical solution, and the output model's predicted probability is linearly interpreted after feature dimensionality reduction. The LNN uses the KDIGO AKI diagnostic criteria as the gold standard for training labels, linking the LNN's predicted probability to the KDIGO grading to form a clinical closed-loop strategy.

Description

Technical Field

[0001] This invention relates to the field of machine learning, specifically to an AP-AKI prediction method and bedside tool based on liquid neural networks. Background Technology

[0002] Early identification of the risk of acute pancreatitis (AP) patients developing acute kidney injury (AKI) is crucial for clinical intervention and improved prognosis. Currently, there are no specific drugs that can effectively reverse AKI. While dialysis can partially alleviate symptoms, it is costly and cannot fundamentally improve prognosis. Therefore, early identification of high-risk patients and timely intervention are key to reducing mortality and improving outcomes. Traditional clinical assessment tools such as the Ranson score and the Bedside Index of Acute Pancreatitis Severity (BISAP) score have been used to predict disease severity, but they still have limitations in data collection, computational complexity, and clinical interactivity. Summary of the Invention

[0003] To address these issues, the present invention provides an AP-AKI prediction method and bedside tool based on liquid neural networks.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an AP-AKI prediction method based on a liquid neural network, comprising the following steps:

[0005] Using the admission time as the baseline, clinical latent characteristic variables of AP patients were collected as input features to construct a time-series feature sequence and preprocess the time-series feature sequence as the basic data for the input layer of the liquid neural network (LNN).

[0006] A dynamic time constant containing the temporal feature sequence of the input features is introduced into the liquid layer of the liquid neural network (LNN) to achieve adaptive response to fast / slow-changing clinical signals of AP patients. Weight pruning and sparse activation of neurons are introduced into the liquid layer of the liquid neural network (LNN) to construct a sparse topology. The numerical solution of the neural ordinary differential equations of the liquid layer of the liquid neural network (LNN) is changed to closed-form analytical solution to solve the problems of traditional LNNs, such as inability to distinguish between fast / slow-changing signals, large volume, and slow speed. After feature dimensionality reduction, the output model predicts probabilities linearly.

[0007] The Liquid Neural Network (LNN) uses the KDIGO AKI diagnostic criteria as the gold standard for training labels, classifying patients into AKI positive / negative and subclinical stages. By determining the risk stratification threshold through the Youden index, the predicted probability of the Liquid Neural Network (LNN) is linked to the KDIGO classification, forming a clinical closed-loop strategy of prediction-classification-intervention-feedback, ensuring that the output of the Liquid Neural Network (LNN) is consistent with the clinical diagnosis and treatment pathway.

[0008] Preferably, the construction of the time-series feature sequence specifically includes: uniformly collecting clinical indicators from AP patients at five time points after admission: 0h (baseline), 6h, 12h, 24h, and 48h, incorporating six key features: albumin, blood urea nitrogen, alkaline phosphatase, blood potassium, serum creatinine, and NLR, forming a time-series data sequence of 6 features × 5 time points, denoted as... ,in .

[0009] Preferably, the time-series feature sequences are preprocessed, specifically including:

[0010] Missing value handling: Time series interpolation (i.e., linear interpolation and weighted completion of adjacent time points) is used. For missing values ​​at individual time points, they are filled in according to the data trend of the two adjacent time points, preserving the time series change characteristics.

[0011] Standardization: Z-score standardization is applied to the time series sequences of each feature to eliminate the influence of dimensions. The formula is as follows: ,in , For the first The feature is the mean and standard deviation of the data at all time points; These are the original feature values, i.e., the latent clinical feature variables of AP patients; It is the first The standardized eigenvalues ​​of the original eigenvalues ​​at time t;

[0012] Outlier handling: Outliers were removed using the 3σ principle combined with clinical common sense. For clinical extreme values ​​such as serum potassium > 6.5 mmol / L and serum creatinine > 353.6 μmol / L, the original data were retained and marked as rapid change signal feature points.

[0013] Preferably, a classification standard for time-series signals in the time-series feature sequence is set to clarify the threshold for determining fast / slow-changing signals. The determination standard is as follows:

[0014] A rapid change signal is defined as an increase of ≥50% in NLR within 6 hours, an increase of ≥26.5 μmol / L in serum creatinine within 24 hours, or a decrease of ≥20 mmHg in systolic blood pressure within 1 hour.

[0015] Albumin drift ≤10% within 48 hours and baseline heart rate fluctuation ≤10 beats / min within 24 hours are considered slow-changing signals.

[0016] The preferred and improved liquid neural network (LNN) is constructed as follows:

[0017] Among them, dynamic time constant The calculation formula is as follows:

[0018] ;

[0019] in, The Sigmoid activation function maps the output to the (0,1) interval; and The weights and biases for the first fully connected layer; and The weights and biases of the second fully connected layer together constitute the generation. Subnetwork parameters ={ , , , }; , These are the upper and lower limits of the time constant, set according to the rate of change of the clinical signal; fast-changing signals are selected as... Slow-changing signal acquisition ; , ; These are the latent clinical characteristic variables corresponding to time point t;

[0020] Dynamic time constant The ordinary differential equation of an embedded LNN, and the LNN state update formula:

[0021] ;

[0022] in, This represents the state of neurons in the liquid layer of an LNN. It is a non-linear activation function; These are the weights of the LNN main network;

[0023] During training, Subnetwork parameters With LNN main network weights As a whole parameter set, it is jointly optimized through backpropagation (BP), and the loss function adopts binary classification cross-entropy (BCE):

[0024] ;

[0025] in This indicates the sample size, i.e., the number of samples currently used to calculate the loss. This is the actual label for AKI. Predict probabilities for the model; It is the sample index, from 1 to... Iterate through all samples in sequence; The average loss of all samples is obtained by averaging the loss values ​​of the entire sample, thus defining the subnetwork. Training parameters: The sub-network optimizer is Adam, the learning rate is 1e-4, the sample size is 32, the number of training epochs is 100, and the early stopping patience is 20.

[0026] Preferably, the liquid neural network (LNN) feature dimensionality reduction layer uses principal component analysis (PCA) to reduce the 256-dimensional high-dimensional temporal features of the LNN liquid layer to 10 dimensions, retaining more than 95% of the information and reducing redundancy. The 10-dimensional features after dimensionality reduction are input into the LR output layer to output the AKI prediction probability. Based on the LR layer coefficients, the contribution of the 10-dimensional features after dimensionality reduction to AKI prediction is quantified and back-mapped to the original 6 temporal features, giving a ranking of the clinical importance of the original features at each time point. Based on SHAP interpretability analysis and combined with the importance of LR features, a two-dimensional explanatory report of temporal feature change trend and single feature contribution is generated.

[0027] Preferably, the steps for constructing a sparse topology are as follows:

[0028] After training, L1 regularization is applied to the weights of the LNN liquid layer and fully connected layer, selecting weights whose absolute values ​​are less than a threshold. The weights are reset to 0, and 20% of the key weights are retained after pruning;

[0029] A Dropout layer is introduced into the LNN liquid layer to ensure that 30% of neurons are activated during training, while maintaining a fixed sparse topology; the dropout value is 0.7.

[0030] The number of parameters in the original LNN model was compressed from approximately 100,000 to less than 10,000 to adapt to the low memory requirements of bedside tools.

[0031] Preferably, the numerical solution of the original LNN's Neural ODE is bypassed, and a closed-form analytical solution is used to solve for the neuron state changes, eliminating the computational time of numerical solutions and improving the inference speed. The process is as follows:

[0032] For any time interval The changes in neuron state are as follows:

[0033] ;

[0034] By using numerical integration to approximate the integral term, the original iterative numerical solution is replaced, reducing the single-sample inference time from milliseconds to microseconds.

[0035] The preferred formula for calculating the Youden index is:

[0036] ;

[0037] This refers to the Youden exponent, which has a value range of [−1, 1].

[0038] YI=1: Indicates that the model perfectly distinguishes between positive and negative samples;

[0039] YI=0: This indicates that the model's predictive ability is no different from random guessing;

[0040] YI<0: This indicates that the model's prediction is worse than random guessing;

[0041] Sensitivity: True positive rate, which is the proportion of patients who actually have AKI but are correctly predicted as positive by the model.

[0042] Specificity: True negative rate, which is the proportion of patients who are actually not AKI but are correctly predicted as negative by the model.

[0043] This invention also discloses a bedside tool for AP-AKI prediction based on a liquid neural network, using the Python-Streamlit framework. This bedside tool performs the aforementioned AP-AKI prediction method based on a liquid neural network, including:

[0044] The temporal feature input module exports the liquid neural network (LNN) and loads it in the Streamlit environment. Clinical medical staff can input the patient's key clinical features through numerical input boxes or drop-down options in the interactive interface. This bedside tool automatically performs format verification and standardization on the input data to ensure that it is consistent with the data structure in the training phase.

[0045] The intervention strategy recommendation module uses the KDIGO AKI diagnostic criteria as the gold standard for training labels to classify patients into AKI positive / negative and subclinical stages. The risk stratification threshold is determined by the Youden index, and the prediction probability of the liquid neural network LNN is linked to the KDIGO classification to obtain a clinical closed-loop strategy of prediction-classification-intervention-feedback.

[0046] The risk level display module presents the calculated risk probability on the interface and simultaneously displays the corresponding risk level. This bedside tool utilizes SHAP interpretability analysis to visualize the contribution of each feature to the prediction results, enabling clinicians to quickly understand the decision-making basis of the model and assist in personalized diagnosis and treatment.

[0047] The feedback data upload module adopts an incremental learning approach, incorporating clinical feedback data to iteratively update the parameters of the liquid neural network (LNN).

[0048] This bedside tool converts liquid neural networks (LNNs) into ONNX format and supports dual Python / embedded C++ runtime environments.

[0049] The incremental learning optimizer is AdamW, with a learning rate of 5e-5, a batch size of 16, and the weights of the original model's underlying layers are frozen. Only the parameters of the top-level classifier and the τ sub-network are fine-tuned, and the number of training rounds is 50.

[0050] Preferably, the bedside tool reserves a multimodal data fusion interface, which supports data sources such as pancreatic CT classification, real-time vital signs (heart rate, blood oxygen, blood pressure), medication history, and urine output monitoring data. The interface data format is HL7 / FHIR, supporting seamless integration with medical information systems and reserving space for subsequent feature dimension expansion. The interface transmission rate is ≥10MB / s, the data latency is ≤1s, it supports ≥100 concurrent accesses, and it is compatible with Python 3.8+ and C++11+ development environments.

[0051] This invention abandons static single-time-point baseline features and adopts multi-time-scale dynamic temporal features, giving full play to the advantages of the liquid time constant of the liquid neural network (LNN) to achieve adaptive modeling of clinical signals with different rates of change in AP patients, thereby improving the sensitivity and specificity of identifying subclinical phase AP-AKI.

[0052] The beneficial effects of this invention are as follows:

[0053] This invention utilizes a hybrid architecture combining liquid neural networks (LNN) and logistic regression (LR), sparse topology, and closed-loop analytical solution for triple optimization. This approach balances model prediction performance, clinical interpretability, parameter count, and inference speed, making it suitable for edge computing and deployment needs of emergency bedside tools.

[0054] This invention breaks through the limitations of traditional models that only focus on risk prediction. It deeply integrates the prediction model with the clinical stratified intervention process, and constructs a closed-loop system of risk prediction, stratified decision-making, intervention execution, result feedback and model iteration, so as to realize the implementation from risk prediction to clinical application. Attached Figure Description

[0055] Figure 1 The diagram shows the structural framework of the improved liquid neural network (LNN) model provided by this invention.

[0056] Figure 2 The coefficient path diagram of the risk predictor of AKI in AP patients in the LASSO regression analysis provided by this invention.

[0057] Figure 3 A comparison of AUC values ​​of the five models provided in this invention before and after applying SMOTE oversampling.

[0058] Figure 4 The calibration curves, DCA curves, and predicted probability distribution histograms of the five models provided in this invention in the validation queue;

[0059] Figure 5 SHAP feature analysis of the improved liquid neural network (LNN) model provided by this invention on the test set.

[0060] Figure 6 Force graph analysis diagram of the improved liquid neural network (LNN) model in the test set provided by this invention.

[0061] Figure 7 A visual illustration of the bedside tool for predicting the risk of AKI in AP patients, developed based on the improved liquid neural network (LNN) model provided by this invention.

[0062] Figure 8 Performance evaluation of the bedside tools provided by this invention on a test set. Detailed Implementation

[0063] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] This invention proposes an AP-AKI prediction method based on a liquid neural network, comprising the following steps:

[0065] Using the admission time as the baseline, clinical latent characteristic variables of AP patients were collected as input features to construct a time-series feature sequence and preprocess the time-series feature sequence as the basic data for the input layer of the liquid neural network (LNN).

[0066] A dynamic time constant containing the temporal feature sequence of the input features is introduced into the liquid layer of the liquid neural network (LNN) to achieve adaptive response to fast / slow-changing clinical signals of AP patients. Weight pruning and sparse activation of neurons are introduced into the liquid layer of the LNN to construct a sparse topology. The numerical solution of the neural ordinary differential equations of the liquid layer of the LNN is changed to closed-form analytical solution to solve the problems of traditional LNNs, such as inability to distinguish between fast / slow-changing signals, large volume, and slow speed. After feature dimensionality reduction, the output model predicts probabilities linearly.

[0067] The Liquid Neural Network (LNN) uses the KDIGO AKI diagnostic criteria as the gold standard for training labels, classifying patients into AKI positive / negative and subclinical stages. By determining the risk stratification threshold through the Youden index, the predicted probability of the Liquid Neural Network (LNN) is linked to the KDIGO classification, forming a clinical closed-loop strategy of prediction-classification-intervention-feedback, ensuring that the output of the Liquid Neural Network (LNN) is consistent with the clinical diagnosis and treatment pathway.

[0068] As an example of this invention, the construction of the time-series feature sequence specifically includes: uniformly collecting clinical indicators from AP patients at five time points after admission: 0h (baseline), 6h, 12h, 24h, and 48h, incorporating six key features: albumin, blood urea nitrogen, alkaline phosphatase, blood potassium, serum creatinine, and NLR, forming a time-series data sequence of 6 features × 5 time points, denoted as... ,in .

[0069] As an example of the present invention, preprocessing of the time-series feature sequence specifically includes:

[0070] Missing value handling: Time series interpolation (i.e., linear interpolation and weighted completion of adjacent time points) is used. For missing values ​​at individual time points, they are filled in according to the data trend of the two adjacent time points, preserving the time series change characteristics.

[0071] Standardization: Z-score standardization is applied to the time series sequences of each feature to eliminate the influence of dimensions. The formula is as follows: ,in , For the first The feature is the mean and standard deviation of the data at all time points; These are the original feature values, i.e., the latent clinical feature variables of AP patients; It is the first The standardized eigenvalues ​​of the original eigenvalues ​​at time t;

[0072] Outlier handling was performed before standardization: outliers were removed using the 3σ principle combined with clinical common sense. For clinical extreme values ​​such as serum potassium > 6.5 mmol / L and serum creatinine > 353.6 μmol / L, the original data were retained and labeled as rapid change signal feature points.

[0073] After cleaning the data, check if it is balanced. If it is imbalanced, use SMOTE oversampling to oversample the minority class. Based on the balanced training set, apply the Least Absolute Shrinkage and Selection Operator (LASSO) algorithm for feature selection, and determine the optimal penalty parameter ( The feature subset was determined through 5-fold cross-validation and then filtered, such as... Figure 2 As shown.

[0074] As an example of the present invention, a classification standard for time-series signals in a time-series feature sequence is set to clarify the threshold for determining fast / slow-changing signals. The determination standard is as follows:

[0075] A rapid change signal is defined as an increase of ≥50% in NLR within 6 hours, an increase of ≥26.5 μmol / L in serum creatinine within 24 hours, or a decrease of ≥20 mmHg in systolic blood pressure within 1 hour.

[0076] Albumin drift ≤10% within 48 hours and baseline heart rate fluctuation ≤10 beats / min within 24 hours are considered slow-changing signals.

[0077] As an example of the present invention, the input layer of the liquid neural network (LNN) is a 6-feature × 5-time-point temporal feature matrix (30-dimensional).

[0078] like Figure 1 As shown, the liquid layer of the liquid neural network (LNN) contains 32 excitatory neurons and 16 inhibitory neurons. It extracts temporal dynamic features through a dynamic time constant τ, outputting 256-dimensional high-dimensional temporal features. The improved liquid neural network (LNN) is constructed as follows:

[0079] Among them, dynamic time constant The calculation formula is as follows:

[0080] ;

[0081] in, The Sigmoid activation function maps the output to the (0,1) interval; and The weights and biases for the first fully connected layer; and The weights and biases of the second fully connected layer together constitute the generation. Subnetwork parameters ={ , , , }; , These are the upper and lower limits of the time constant, set according to the rate of change of the clinical signal; fast-changing signals are selected as... Slow-changing signal acquisition ; , ; These are the latent clinical characteristic variables corresponding to time point t;

[0082] Dynamic time constant The adaptive control logic is as follows:

[0083] Rapid change signals: When rapid change signals such as a sudden drop in blood pressure (systolic blood pressure decrease of ≥20 mmHg within 1 hour), an increase in NLR of ≥50% within 6 hours, or an increase in serum creatinine of ≥26.5 μmol / L within 24 hours are detected, the subnetwork generates... (0.1~2), enabling LNN neurons to respond quickly to signal changes and capture abnormal fluctuations in the subclinical phase;

[0084] Slowly varying signals: For slowly varying signals such as baseline heart rate and albumin drift ≤10% within 48 hours, the subnetwork generates... (5~10) enables LNN neurons to retain long-term memory and capture slow, trend-like changes in features.

[0085] Dynamic time constant The ordinary differential equation of an embedded LNN, and the LNN state update formula:

[0086] ;

[0087] in, This represents the state of neurons in the liquid layer of an LNN. It is a non-linear activation function; These are the weights of the LNN main network;

[0088] During training, Subnetwork parameters With LNN main network weights As a whole parameter set, it is jointly optimized through backpropagation (BP), and the loss function adopts binary classification cross-entropy (BCE):

[0089] ;

[0090] in This indicates the sample size, i.e., the number of samples currently used to calculate the loss. This is the actual label for AKI. Predict probabilities for the model; It is the sample index, from 1 to... Iterate through all samples in sequence; The average loss of all samples is obtained by averaging the loss values ​​of the entire sample, thus defining the subnetwork. Training parameters: The sub-network optimizer is Adam, the learning rate is 1e-4, the sample size is 32, the number of training epochs is 100, and the early stopping patience is 20.

[0091] This embodiment verifies the model's recognition performance in the subclinical phase of AP-AKI (i.e., not reaching the KDIGO diagnostic criteria within 48 hours of onset but subsequently developing into AKI) under three modeling methods: static features, original LNN temporal features, and optimized adaptive τLNN temporal features, as shown in Tables 1-5. Here, 95% CI represents the 95% confidence interval.

[0092] Table 1 Comparison of Identification Performance of Different Modeling Methods in AP-AKI Subclinical Phase

[0093]

[0094] Table 2. Comparison of AUC of ROC curves at different time scales

[0095]

[0096] Table 3 Comparison of recognition rates of temporal and static features in the subclinical phase

[0097]

[0098] Table 4. Comparison of Training Convergence Curves of the τ Subnetwork

[0099]

[0100] Table 5. Box plot statistics of τ distribution for different signal types at different time scales

[0101]

[0102] As an example of the present invention, the liquid neural network (LNN) feature dimensionality reduction layer uses principal component analysis (PCA) to reduce the 256-dimensional high-dimensional temporal features output by the LNN liquid layer to 10 dimensions, retaining more than 95% of the information and reducing redundancy. The 10-dimensional features after dimensionality reduction are input into the LR output layer to output the AKI prediction probability. Based on the LR layer coefficients, the contribution of the 10-dimensional features after dimensionality reduction to AKI prediction is quantified and back-mapped to the original 6 temporal features, giving a ranking of the clinical importance of the original features at each time point. Based on SHAP interpretability analysis and combined with the importance of LR features, a two-dimensional interpretation report of temporal feature change trend and single feature contribution is generated.

[0103] As an example of this invention, based on the original LNN topology, a sparse topology (NCP) is implemented using weight pruning and sparse activation of neurons, preserving key connections, minimizing the number of parameters, and ensuring that the model performance loss is ≤5%. The steps for constructing the sparse topology are as follows:

[0104] After training, L1 regularization is applied to the weights of the LNN liquid layer and fully connected layer, selecting weights whose absolute values ​​are less than a threshold. The weights are reset to 0, and 20% of the key weights are retained after pruning;

[0105] A Dropout layer is introduced into the LNN liquid layer to ensure that 30% of neurons are activated during training, while maintaining a fixed sparse topology; the dropout value is 0.7.

[0106] The number of parameters in the original LNN model was compressed from approximately 100,000 to less than 10,000 to adapt to the low memory requirements of bedside tools.

[0107] As an example of this invention, the Neural ODE numerical solution of the original LNN is bypassed, and the neuron state changes are analytically solved using a closed-form solution (CFC), eliminating the computational time of numerical solution and improving inference speed. The process is as follows:

[0108] For any time interval The changes in neuron state are as follows:

[0109] ;

[0110] By using numerical integration to approximate the integral term, the original iterative numerical solution is replaced, reducing the single-sample inference time from milliseconds to microseconds.

[0111] The closed-form LNN model is deployed in a lightweight manner, supports ONNX model format conversion, and is adapted to the Python / embedded C++ runtime environment of bedside tools.

[0112] This embodiment verifies the performance and deployability of the lightweight model based on five core metrics: number of detection parameters, single-sample inference time, AUC, sensitivity, and specificity, for different architecture models, as shown in Table 6. Specific metric requirements are as follows:

[0113] Number of parameters: ≤10000 after NCP+CfC optimization;

[0114] Inference time: ≤50ms per sample on edge devices;

[0115] Performance loss: Compared with the original LNN model, AUC decreased by ≤0.03.

[0116] Table 6 Performance and Deployability Metrics for Different Architecture Models

[0117]

[0118] As an example of this invention, the Youden exponent is calculated as follows:

[0119] ;

[0120] This refers to the Youden exponent, which has a value range of [−1, 1].

[0121] YI=1 indicates that the model perfectly distinguishes between positive and negative samples;

[0122] YI=0 indicates that the model's predictive ability is no different from random guessing;

[0123] YI<0 indicates that the model's prediction is worse than random guessing;

[0124] Sensitivity: True positive rate, which is the proportion of patients who actually have AKI but are correctly predicted as positive by the model.

[0125] Specificity: True negative rate, which is the proportion of patients who are actually not AKI but are correctly predicted as negative by the model.

[0126] This invention trains five models—LR, DCT, RF, XGBoost, and the improved Liquid Neural Network (LNN)—using datasets before and after SMOTE processing. All models show improved AUC values ​​after using the SMOTE technique, see [link to relevant documentation]. Figure 3 . Figure 3 In this context, A represents the logistic regression (LR) model. Figure 3 In this context, B represents the decision tree (DCT) model. Figure 3 In this context, C represents the Extreme Gradient Boosting (XGBoost) model. Figure 3 In this context, D represents the Random Forest (RF) model. Figure 3 E in the diagram represents the Liquid Neural Network (LNN) model. Among them, the LNN model performed best (AUC=0.91, 95%CI=0.86~0.95), indicating that the SMOTE technique can effectively improve the class imbalance problem and is beneficial to improving the model's predictive performance.

[0127] To further evaluate the clinical applicability of the model, Figure 4 The calibration curves and decision curve analysis (DCA) results for each model are presented. Figure 4 In this context, A represents the calibration curve. Figure 4 In the figure, B represents the DCA curve. Figure 4 C in the diagram represents the histogram of the predicted probability distribution. Figure 4 The A in the data shows that the LR model has the best Brier score (0.111), followed by the LNN (0.116), indicating that these models have good calibration. Figure 4The DCA results of B showed that all machine learning models exhibited positive net benefits when the threshold probability was between 0.1 and 0.8, with LNN and DCT models showing relatively higher net benefits, indicating that these models have practical application value in clinical decision-making. Figure 4 The histogram of the predicted probability distribution of C in the DCT model shows that the model exhibits extreme prediction characteristics, with a large number of samples concentrated in the ranges of 0–0.1 and 0.9–1.0. This "polarized" prediction pattern lacks sufficient predictive granularity. In contrast, the LNN model has a more uniform predicted probability distribution and can provide more detailed risk stratification.

[0128] Visualizing the LNN model on the test set using SHAP value scatter plots. Figure 5 In the graph, each row represents one feature, with redder colors indicating higher feature values ​​and bluer colors indicating lower feature values. Points closer to the right side of the graph indicate a greater contribution of that feature to the risk of AKI, thus promoting its development. In this study, the six features that contributed most to AKI prediction were: NLR, serum creatinine, blood urea nitrogen, alkaline phosphatase, serum potassium, and albumin. For example, regarding NLR, the red area is mainly distributed on the right side of the graph, indicating that elevated NLR values ​​are closely related to an increased risk of AKI, reflecting the important role of inflammatory response in the progression of AP to AKI. Conversely, certain protective factors, such as decreased albumin levels (blue areas concentrated on the left side), are associated with an increased risk of AKI.

[0129] Two patients were randomly selected from the test set and SHAP force maps were plotted. In the force map, each feature is represented by an arrow. The arrow color and length reflect the direction and degree of influence, respectively: red arrows indicate positive influence (increased risk of AKI), and blue arrows indicate negative influence (reduced risk of AKI). The length of the arrow is proportional to the intensity of the influence. Figure 6 In case A, serum creatinine (73.9 μmol / L), alkaline phosphatase (68.2 U / L), and NLR (9.3) are shown in red with long arrows, significantly increasing the risk of AKI. Blood urea nitrogen (5.5 mg / dL) is shown in blue, decreasing the risk. Based on the combined data, the model predicts an 86% probability of AKI in this case. Figure 6 After considering the comprehensive characteristics of case B, the model predicts an 82% probability that AKI will not occur.

[0130] Based on the Python-Streamlit framework, the best-performing LNN model was developed into an AI bedside tool with a visual interface, aiming to provide clinicians with a convenient AKI risk assessment solution. The tool has a simple and intuitive interface design, and medical staff can operate it without any programming knowledge. After selecting from the drop-down menu or entering the patient's clinical parameters in the text box, the user can obtain the following by clicking the "Predict" button: (1) AKI risk assessment results; (2) specific risk probability values; and (3) a feature contribution visualization chart generated based on the SHAP (SHapley Additive exPlanations) method. Figure 7 The tool's user interface was demonstrated. Taking an AP patient as an example, after inputting their clinical parameters, the model predicted that the patient's risk of developing AKI was 68.8%. The SHAP graph was used to visually demonstrate the contribution of each indicator to the prediction result, while also providing personalized health management suggestions.

[0131] To facilitate clinical application, this tool supports access from multiple platforms. Through SHAP interpretability analysis, clinicians can not only obtain predictive results but also gain a deeper understanding of the model's decision-making rationale, enhancing the credibility and clinical acceptance of AI-assisted diagnosis. Furthermore, the tool's built-in sharing function facilitates collaboration and communication among medical teams, helping to promote the application of this predictive model in a wider range of clinical scenarios.

[0132] In a test set of 144 AP patients, the AI-based bedside tool developed in this study correctly classified 136 cases, with an overall accuracy of 94.4% and an AUC of 0.91 (95% CI = 0.86–0.95). See [link to study]. Figure 8 In the A category, specifically, 25 out of 28 patients with AKI were correctly identified; and 111 out of 116 patients without AKI were accurately excluded. The model's sensitivity for predicting AKI was 89.29%, and its specificity was 95.69%. The corresponding confusion matrix is ​​shown below. Figure 8 B in the context of feature importance analysis ( Figure 8 As can be seen from C), NLR is the most important predictor (SHAP value = 0.334), followed by serum creatinine (SHAP value = 0.219) and serum potassium (SHAP value = 0.159).

[0133] This invention also discloses a bedside tool for AP-AKI prediction based on a liquid neural network, using the Python-Streamlit framework. This bedside tool performs the aforementioned AP-AKI prediction method based on a liquid neural network, including:

[0134] The temporal feature input module exports the liquid neural network (LNN) and loads it in the Streamlit environment. Clinical medical staff can input the patient's key clinical features through numerical input boxes or drop-down options in the interactive interface. This bedside tool automatically performs format verification and standardization on the input data to ensure that it is consistent with the data structure in the training phase.

[0135] The intervention strategy recommendation module uses the KDIGO AKI diagnostic criteria as the gold standard for training labels to classify patients into AKI positive / negative and subclinical stages. The risk stratification threshold is determined by the Youden index, and the prediction probability of the liquid neural network LNN is linked to the KDIGO classification to obtain a clinical closed-loop strategy of prediction-classification-intervention-feedback.

[0136] The risk level display module presents the calculated risk probability on the interface and simultaneously displays the corresponding risk level. This bedside tool utilizes SHAP interpretability analysis to visualize the contribution of each feature to the prediction results, enabling clinicians to quickly understand the decision-making basis of the model and assist in personalized diagnosis and treatment.

[0137] The feedback data upload module adopts an incremental learning approach, incorporating clinical feedback data to iteratively update the parameters of the liquid neural network (LNN).

[0138] This bedside tool converts liquid neural networks (LNNs) into ONNX format and supports dual Python / embedded C++ runtime environments.

[0139] The incremental learning optimizer is AdamW, with a learning rate of 5e-5, a batch size of 16, and the weights of the original model's underlying layers are frozen. Only the parameters of the top-level classifier and the τ sub-network are fine-tuned, and the number of training rounds is 50.

[0140] As an example of the present invention, the tool reserves a multimodal data fusion interface, which supports the following data sources: pancreatic CT classification, real-time vital signs (heart rate, blood oxygen, blood pressure), medication history, and urine output monitoring data; the interface data format is HL7 / FHIR, which supports seamless integration with medical information systems and reserves space for subsequent feature dimension expansion; the interface transmission rate is ≥10MB / s, the data latency is ≤1s, it supports concurrent access ≥100, and it is compatible with Python 3.8+ and C++11+ development environments.

[0141] This embodiment compares the clinical effects of the closed-loop system before application (conventional treatment only) and after application (model prediction + stratified intervention). It uses a retrospective historical control dataset design, and the specific scenario is as follows:

[0142] Dataset composition:

[0143] Pre-application group (historical control group): Hospitalized patients with acute pancreatitis (AP) admitted to the Department of Gastroenterology of our hospital from January 2023 to June 2024 were selected. They were managed using routine clinical pathways (without model assistance), and their AKI occurrence and clinical outcomes were retrospectively collected.

[0144] Post-application group (intervention group): AP inpatients admitted between July 2024 and December 2025 (or the period after the system went online) were selected and managed using the LNN prediction model + hierarchical intervention closed-loop system of this invention.

[0145] Inclusion / exclusion criteria (consistent for both groups):

[0146] Inclusion: Adult hospitalized patients who meet the AP diagnostic criteria (such as the Atlanta revision).

[0147] Exclusion criteria: Patients with CKD stage 4-5 at admission, those transferred to another hospital or died within 24 hours of admission, and those with severely incomplete medical records.

[0148] Sample size reference: 300 cases per group, as shown in Table 7. Core indicators include:

[0149] Key metrics: incidence of adverse events in AKI (such as renal replacement therapy rate and mortality rate), number of days spent in the ICU, and total number of days spent in the hospital;

[0150] Soft indicators: intervention implementation rate at each risk level, clinician satisfaction, and bedside tool usage rate.

[0151] Table 7 Comparison of Clinical Effects Before and After Application of the Closed-Loop System

[0152]

[0153] Based on the AKI prediction probability of the optimized model and combined with the KDIGO AKI classification standard, AP patients are divided into four risk levels: low, intermediate, high, and very high risk. Corresponding intervention thresholds and clinical management strategies are set. The thresholds are determined by a combination of clinical threshold method and Youden index. The specific classification is shown in Table 8.

[0154] Table 8 Intervention Thresholds and Clinical Management Strategies

[0155]

[0156] This invention improves the sensitivity of AP-AKI subclinical phase identification to ≥92%, the specificity to ≥96%, and the overall AUC to ≥0.93;

[0157] Applications of this invention:

[0158] The number of deployment parameters for bedside tools is reduced to below 10,000, and the inference time per sample at the edge device is ≤50ms, meeting the needs of real-time prediction at the bedside. It enables multi-dimensional interpretation of temporal feature changes, LR coefficients, and SHAP values, allowing clinicians to directly understand the basis for model decision-making. The incidence of AKI adverse events in AP patients is reduced by ≥15%, and the length of ICU stay is shortened by ≥2 days, realizing the implementation of the entire process from risk prediction to clinical intervention.

[0159] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. An AP-AKI prediction method based on liquid neural networks, characterized in that: Includes the following steps: Using the admission time as the baseline, the clinical latent characteristic variables of AP patients were dynamically collected as input features to construct a time-series feature sequence and preprocess the time-series feature sequence as the basic data for the input layer of the liquid neural network (LNN). A dynamic time constant containing the temporal feature sequence of the input features is introduced into the liquid layer of the liquid neural network (LNN) to achieve an adaptive response to the fast / slow changing clinical signals of AP patients. Weight pruning and sparse activation of neurons are introduced into the liquid layer of the liquid neural network (LNN) to construct a sparse topology. The numerical solution of the neural ordinary differential equations of the liquid layer of the liquid neural network (LNN) is changed to a closed-form analytical solution. After feature dimensionality reduction, the output model predicts probabilities linearly. The Liquid Neural Network (LNN) uses the KDIGO AKI diagnostic criteria as the gold standard for training labels, classifying patients into AKI positive / negative and subclinical stages. By determining the risk stratification threshold through the Youden index, the predicted probability of the Liquid Neural Network (LNN) is linked to the KDIGO classification, forming a clinical closed-loop strategy of prediction-classification-intervention-feedback.

2. The AP-AKI prediction method based on liquid neural network according to claim 1, characterized in that: Preprocessing of time-series feature sequences specifically includes: Using time series interpolation, missing values ​​at individual time points are filled in based on the data trends of two adjacent time points, preserving the time series variation characteristics; The time series sequences of each feature are Z-score standardized, using the following formula: ,in , For the first The feature is the mean and standard deviation of the data at all time points; These are the original feature values, i.e., the latent clinical feature variables of AP patients; It is the first The standardized eigenvalues ​​of the original eigenvalues ​​at time t; Outliers were removed using the 3σ principle combined with clinical common sense, and the original data were retained and labeled as characteristic points of fast-changing signals.

3. The AP-AKI prediction method based on liquid neural network according to claim 2, characterized in that: The improved liquid neural network (LNN) is constructed as follows: Dynamic time constant The calculation formula is as follows: ； in, The Sigmoid activation function maps the output to the (0,1) interval; and The weights and biases for the first fully connected layer; and The weights and biases of the second fully connected layer together constitute the generation. Subnetwork parameters ={ , , , }; , These are the upper and lower limits of the time constant, set according to the rate of change of the clinical signal; fast-changing signals are selected as... Slow-changing signal acquisition ; These are the latent clinical characteristic variables corresponding to time point t; It is the hyperbolic tangent function; Dynamic time constant The neural network ordinary differential equation embedded in a liquid neural network (LNN), and the state update formula for the liquid neural network LNN: ； in, The state of neurons in the liquid layer of a liquid neural network (LNN); It is a non-linear activation function; These are the weights of the main network of a liquid neural network (LNN). During training, Subnetwork parameters With LNN main network weights As a whole parameter set, it is jointly optimized through backpropagation (BP), and the loss function adopts binary classification cross-entropy (BCE): ； in, This represents the total loss value. This indicates the sample size, i.e., the number of samples currently used to calculate the loss. This is the actual label for AKI. Predict probabilities for the model; It is the sample index, from 1 to... Iterate through all samples in sequence; The average loss is obtained by averaging the loss values ​​of all samples.

4. The AP-AKI prediction method based on liquid neural network according to claim 3, characterized in that: The liquid neural network (LNN) feature dimensionality reduction layer uses principal component analysis (PCA) to reduce the 256-dimensional high-dimensional temporal features of the LNN liquid layer to 10 dimensions; the 10-dimensional reduced features are then input into the LR output layer to output the AKI prediction probability. Based on the LR layer coefficient, the contribution of the 10-dimensional features after dimensionality reduction to AKI prediction is quantified and back-mapped to the original 6 time-series features. The clinical importance ranking of the original features at each time point is given. Based on SHAP interpretability analysis and combined with the importance of LR features, a two-dimensional interpretation report of time-series feature change trend and single feature contribution is generated.

5. The AP-AKI prediction method based on liquid neural network according to claim 4, characterized in that: The steps to construct a sparse topology are as follows: After training, L1 regularization is applied to the weights of the LNN liquid layer and fully connected layer, selecting weights whose absolute values ​​are less than a threshold. The weights are reset to 0, and 20% of the key weights are retained after pruning; A Dropout layer is introduced into the LNN liquid layer to ensure that 30% of neurons are activated during training, while maintaining a fixed sparse topology; the dropout value is 0.

7. The number of parameters in the original LNN model was compressed to suit the low memory requirements of bedside tools.

6. The AP-AKI prediction method based on liquid neural network according to claim 5, characterized in that: Closed-form analytical solutions are used to solve for neuron state changes, eliminating the computational time of numerical solutions and improving inference speed. The process is as follows: For any time interval The changes in neuron state are as follows: ； in, This indicates the autonomous decay term, which refers to the natural decay of the neuron's state; The integral input term refers to the cumulative effect of historical input signals; By approximating the integral term through numerical integration, the single-sample inference time is reduced from milliseconds to microseconds.

7. An AP-AKI prediction bedside tool based on liquid neural networks, characterized in that: Based on the Python-Streamlit framework, this bedside tool performs the AP-AKI prediction method based on liquid neural networks as described in any one of claims 1-6, including: The temporal feature input module exports the liquid neural network (LNN) and loads it in the Streamlit environment. Clinical medical staff can input the patient's key clinical features through numerical input boxes or drop-down options in the interactive interface. This bedside tool automatically performs format verification and standardization on the input data to ensure that it is consistent with the data structure in the training phase. The intervention strategy recommendation module uses the KDIGO AKI diagnostic criteria as the gold standard for training labels to classify patients into AKI positive / negative and subclinical stages. The risk stratification threshold is determined by the Youden index, and the prediction probability of the liquid neural network LNN is linked to the KDIGO classification to obtain a clinical closed-loop strategy of prediction-classification-intervention-feedback. The risk level display module presents the calculated risk probability on the interface and simultaneously displays the corresponding risk level. This bedside tool utilizes SHAP interpretability analysis to visualize the contribution of each feature to the prediction results, enabling clinicians to quickly understand the decision-making basis of the model and assist in personalized diagnosis and treatment. The feedback data upload module adopts an incremental learning approach, incorporating clinical feedback data to iteratively update the parameters of the liquid neural network (LNN).

Images