An acute pancreatitis complication evaluation method based on artificial intelligence

Abstract

The application discloses an acute pancreatitis complication evaluation method based on artificial intelligence and belongs to the medical information technology field, and specifically comprises the following steps: firstly, a clinical data set containing real-time clinical indexes and basic disease history records is acquired; subsequently, time series analysis is performed on the basic disease history records, long-term influence features are extracted, and basic disease influence factors are generated; then, the real-time clinical indexes and the basic disease influence factors are subjected to standardization treatment, a comprehensive feature vector is generated after the correlation weight between features is calculated; then, the vector is input into a trained artificial intelligence evaluation model, a complication risk probability is output through layer-by-layer nonlinear transformation; finally, the risk probability value is mapped into a risk grade, patient information is integrated to generate a structured risk evaluation report; and an acute pancreatitis complication evaluation scheme fusing acute phase indexes and chronic basic disease influences is established.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, specifically to an artificial intelligence-based method for assessing complications of acute pancreatitis. Background Technology

[0002] Acute pancreatitis is a common digestive emergency with a rapid progression. Some patients may develop severe illness and multiple local or systemic complications, posing a serious threat to their lives and health. In clinical practice, timely and accurate assessment of a patient's risk of complications is crucial for developing targeted treatment plans and allocating medical resources rationally. With the advancement of medical informatization, utilizing clinical data to build risk assessment models to assist physicians in clinical decision-making has become an important research direction in medical artificial intelligence technology.

[0003] Currently, existing AI-based risk assessment methods for acute pancreatitis typically focus on real-time clinical indicators upon patient admission, such as serum amylase and lipase levels, and imaging findings. These methods use machine learning or deep learning models to learn the mapping relationship between real-time indicators and complication outcomes. Furthermore, some improved models attempt to incorporate patient underlying disease information as one of the model's input features, treating the underlying disease as a static, categorical variable to improve the model's predictive performance.

[0004] However, existing risk assessment models do not adequately incorporate underlying disease variables. They are typically treated as simple binary categorical variables or only a few static baseline indicators are introduced. This approach fails to characterize underlying diseases as a chronic, dynamically evolving pathological state, neglecting their long-term, cumulative impact on the organ's physiological environment. Specifically, existing methods generally lack in-depth time-series analysis of the historical records of underlying diseases, such as long-term fluctuations in blood glucose levels or stable changes in blood pressure control. This dynamic information is precisely the key factor influencing the pancreatic microenvironment and the risk of complications. Consequently, the models cannot fully reflect the true risk status under the combined influence of the physiological basis established by the underlying disease and acute exacerbations. This superficial use of underlying disease variables constitutes a major limitation of existing technologies. Summary of the Invention

[0005] The purpose of this invention is to provide an artificial intelligence-based method for assessing complications of acute pancreatitis, addressing the following technical problems:

[0006] Underlying medical conditions can increase the risk of complications, but current methods do not adequately incorporate these variables.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] An artificial intelligence-based method for assessing complications of acute pancreatitis includes the following steps:

[0009] S1. Obtain the clinical dataset of patients with acute pancreatitis, which includes real-time clinical indicators during an acute pancreatitis attack and the patient's underlying disease history.

[0010] S2. Perform time-series analysis on the historical records of the underlying diseases, extract the long-term impact characteristics of the underlying diseases on the physiological state of the pancreas, and generate the underlying disease impact factors.

[0011] S3. Standardize the real-time clinical indicators and the underlying disease influencing factors and calculate the correlation weights between features. Generate a fixed-dimensional comprehensive feature vector based on the weighted correlation weights.

[0012] S4. Input the comprehensive feature vector into the trained artificial intelligence evaluation model, perform layer-by-layer nonlinear transformation and information extraction on the input features, and output the probability of complication risk.

[0013] S5. Map the risk probability value to a predefined risk level, and integrate patient identification information and key clinical data to form a structured document, generating a risk assessment report.

[0014] As a further aspect of the present invention: in step S1, the specific process of obtaining the clinical dataset of patients with acute pancreatitis is as follows:

[0015] Complete medical records of patients are extracted from the hospital information system through a medical data interface. The medical records include admission diagnosis information, laboratory test results and imaging reports. Clinical indicators related to the diagnosis of acute pancreatitis are screened from the medical records. The clinical indicators include serum amylase level and abdominal imaging features.

[0016] At the same time, the patient's basic disease history records are extracted from their past medical history. The basic disease history records include disease type, diagnostic basis, and disease course records.

[0017] As a further aspect of the present invention: in step S2, the specific process of the time series analysis is as follows:

[0018] Time-series data from the history of underlying diseases is read. This data includes regularly recorded physiological parameter measurements and medical event markers. The time-series data is preprocessed, including data cleaning and missing value imputation. The preprocessed data is divided into continuous time windows, and the statistical characteristics of physiological parameters and the density of medical events within each time window are calculated. Feature sequences are constructed based on the time window data, and a dynamic system modeling method is used to simulate the disease progression trajectory. Feature extraction is performed on the simulated trajectory data, and the main change patterns are extracted as long-term impact features. The long-term impact features are input into a neural network structure, and the underlying disease impact factor is output.

[0019] As a further aspect of the present invention, the specific process of the dynamic system modeling method is as follows:

[0020] A system state equation is established to describe the change of the characteristic sequence over time. The system state equation is constructed based on the correlation analysis of the characteristic changes of adjacent time windows. The equation parameters are determined by fitting historical data. The fitting process uses an optimization algorithm to find the optimal parameter combination. In the parameter optimization process, the density of medical events is introduced as an adjustment factor. The adjustment factor affects the step size and direction of parameter updates. The system state equation is solved using numerical calculation methods to calculate the state estimate for each time window. The state estimate forms a continuous disease progression trajectory.

[0021] As a further aspect of the present invention: in step S3, the specific process of feature fusion is as follows:

[0022] Data standardization was performed on real-time clinical indicators and underlying disease influencing factors to eliminate dimensional differences. A dual-path coding structure was adopted, with the first path extracting features from clinical indicators and the second path transforming the features of influencing factors.

[0023] We design a feature interaction layer based on an attention allocation mechanism, calculate the correlation strength between two types of features, generate dynamic weight coefficients based on the correlation strength, perform weighted combination of the two types of features, input the weighted features into a feature compression layer, and generate a fixed-dimensional comprehensive feature vector through dimensionality reduction.

[0024] As a further aspect of the present invention, the specific process of the attention allocation mechanism is as follows:

[0025] Construct a feature interaction matrix, where rows correspond to clinical indicator features and columns correspond to basic disease influencing factor features. Calculate the association strength value of each element in the feature interaction matrix. The association strength value is obtained by the inner product operation of the feature vectors. Normalize the association strength value to generate an attention weight distribution.

[0026] The original features are reweighted based on the attention weight distribution to highlight key feature dimensions. The weighted features are then passed through a non-linear transformation layer to enhance their expressive power. The calculation of attention weights adopts an iterative optimization method, and the number of iterations is dynamically adjusted according to the feature dimensions.

[0027] As a further aspect of the present invention: in step S4, the training process of the artificial intelligence evaluation model is as follows:

[0028] A deep neural network architecture is constructed, consisting of an input layer, a feature extraction layer, and an output layer. A training dataset is prepared, containing comprehensive feature vectors of historical cases and complication annotations. A stratified sampling method is used to divide the training and validation sets, maintaining the data distribution of the training and validation sets. A loss function is designed, including classification loss and a regularization term. An adaptive optimization algorithm is used to train the network parameters. The optimization algorithm adjusts the learning rate. The training error is monitored during training, and training is stopped when the validation set error increases. After training is completed, the model parameters are saved.

[0029] As a further aspect of the present invention, the specific process of the feature extraction layer is as follows:

[0030] Multiple cascaded feature transformation modules are constructed. Each module contains linear transformation and nonlinear activation function. The first feature transformation module receives the feature vector of the input layer and performs feature extraction. Subsequent modules perform feature transformation on the output of the previous module. Residual connections are added during the feature transformation process to pass low-level feature information. The output dimension of each feature transformation module is gradually reduced, and the feature dimension is gradually reduced. The output of the last feature transformation module is passed to the output layer. The parameters of the feature extraction layer are updated through the backpropagation algorithm.

[0031] As a further aspect of the present invention: in step S5, the specific process of generating the risk assessment report is as follows:

[0032] Obtain the mapping table between risk probability values ​​and risk levels. The mapping table defines the risk levels corresponding to different probability ranges. Match the complication risk probability values ​​with the mapping table to determine the current patient's risk level.

[0033] Patient identification information and key clinical data are extracted from the clinical dataset. The key clinical data includes major abnormal indicators and a summary of the underlying disease. Following a predefined structured document template, the patient identification information, key clinical data, risk probability values, and determined risk levels are combined to generate a complete risk assessment report document.

[0034] The beneficial effects of this invention are:

[0035] This invention quantifies the long-term dynamic impact of underlying diseases on pancreatic physiological status by constructing underlying disease influencing factors, thus addressing the problem of insufficient utilization of underlying disease variables in existing models. It employs time-series analysis to process historical records of underlying diseases, extracting long-term impact features from their time-series data, overcoming the limitation of simplifying underlying diseases as static variables. Through an attention allocation mechanism in the feature fusion process, it effectively integrates real-time clinical indicators with underlying disease influencing factors, establishing deep correlations between features at different time scales. A feature extraction layer based on a deep neural network achieves nonlinear transformation and deep feature extraction of the comprehensive feature vector, ultimately generating more accurate complication risk probabilities. The entire scheme enables a comprehensive assessment of complication risks in patients with acute pancreatitis, providing more reliable data support for clinical decision-making. Attached Figure Description

[0036] The invention will now be further described with reference to the accompanying drawings.

[0037] Figure 1 This is a flowchart illustrating the present invention. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0039] Please see Figure 1 As shown, this invention is an artificial intelligence-based method for assessing complications of acute pancreatitis, comprising the following steps:

[0040] S1. Obtain the clinical dataset of patients with acute pancreatitis, which includes real-time clinical indicators during an acute pancreatitis attack and the patient's underlying disease history.

[0041] S2. Perform time-series analysis on the historical records of the underlying diseases, extract the long-term impact characteristics of the underlying diseases on the physiological state of the pancreas, and generate the underlying disease impact factors.

[0042] S3. Standardize the real-time clinical indicators and the underlying disease influencing factors and calculate the correlation weights between features. Generate a fixed-dimensional comprehensive feature vector based on the weighted correlation weights.

[0043] S4. Input the comprehensive feature vector into the trained artificial intelligence evaluation model, perform layer-by-layer nonlinear transformation and information extraction on the input features, and output the probability of complication risk.

[0044] S5. Map the risk probability value to a predefined risk level, and integrate patient identification information and key clinical data to form a structured document, generating a risk assessment report.

[0045] In S1, the specific process of obtaining the clinical dataset of patients with acute pancreatitis is as follows:

[0046] A standardized medical data interface is adopted to connect with the hospital information system. The interface follows the HL7FHIR protocol specification and supports the establishment of real-time data interaction links with the inpatient information management subsystem, laboratory information management subsystem, and medical image archiving and communication subsystem within the hospital information system. This interface first extracts the patient's complete medical record. The admission diagnosis information is obtained from the patient record homepage module of the inpatient information management subsystem, including the patient's primary and secondary diagnoses and corresponding ICD-10 codes. Records with acute pancreatitis as the primary diagnosis are selected through code matching. Laboratory test results are extracted from the laboratory information management subsystem's test data warehouse, covering all biochemical, immunological, and hematological tests completed after the patient's admission, including test name, result value, unit, reference range, and test timestamp. Imaging reports are extracted from the report management module of the medical image archiving and communication subsystem, including the type of imaging examination (e.g., abdominal CT, abdominal MRI, abdominal ultrasound), examination time, structured report text issued by the radiologist, and image metadata.

[0047] After obtaining complete medical records, a clinical indicator screening process was initiated. Based on the clinical diagnosis and treatment guidelines for acute pancreatitis, screening rules were established. First, laboratory test results were screened, extracting serum amylase level data directly related to the diagnosis of acute pancreatitis. This was limited to venous blood samples collected within 24 hours of the onset of acute pancreatitis symptoms, while data with abnormalities such as hemolysis or lipemia during sample collection were excluded. Natural language processing was used to parse the imaging reports. Through a pre-trained medical text segmentation and entity recognition model, abdominal imaging features were extracted from the report text, including pancreatic morphology (e.g., pancreatic enlargement, blurred pancreatic outline), surrounding conditions (e.g., surrounding effusion, fluid accumulation), and changes in pancreatic parenchyma (e.g., parenchymal necrosis, liquefaction), forming a structured abdominal imaging feature dataset.

[0048] While extracting clinical indicators related to acute pancreatitis, the process of extracting historical records of underlying diseases was initiated. The patient's unique identifier (such as ID card number or medical insurance card number) was used to link the patient's past medical records in the electronic medical record system, including outpatient records, summaries of past inpatient medical histories, health check-up reports, and chronic disease management files. Historical information on underlying diseases was extracted from these records. The disease type was determined using the ICD-10 underlying disease code, covering common chronic underlying diseases such as hypertension, type 2 diabetes, chronic kidney disease, and coronary heart disease. The extracted diagnostic evidence included laboratory test evidence at the time of the initial diagnosis of the underlying disease (such as blood glucose and blood pressure measurements), imaging diagnostic results (such as echocardiogram reports and kidney CT reports), and textual descriptions of clinical symptoms. The extracted disease progression records included the time of the initial diagnosis of the underlying disease, the duration of the disease, the diagnostic and treatment interventions received during disease progression (such as drug treatment plans, dosage adjustment records, and records of interventional or surgical treatments), and data on changes in key indicators from regular follow-ups. These were then sorted by timestamps to form a continuous time-series record of the underlying disease progression.

[0049] In S2, the specific process of the time series analysis is as follows:

[0050] I. Time Series Data Reading and Preprocessing

[0051] The time-series data parsing module is activated to read time-stamped structured data from the history of underlying diseases. Physiological parameter measurements include regularly monitored data such as blood pressure (systolic and diastolic), fasting blood glucose, glycated hemoglobin, four lipid profiles (total cholesterol, triglycerides, LDL cholesterol, and HDL cholesterol), and liver and kidney function indicators (serum creatinine and alanine aminotransferase). Data timestamps are accurate to the collection time (year-month-day-hour-minute). Medical event markers include interventional events related to underlying diseases, such as adjustments to antihypertensive / hypoglycemic drug regimens (changes in drug type and dosage), records of complications (e.g., diagnosis of diabetic nephropathy, hypertensive heart attack), and interventional procedures (e.g., coronary stent implantation, initiation of renal dialysis). Each event marker is associated with a unique event type code and the time of occurrence.

[0052] The preprocessing stage consists of two steps: data cleaning and missing value imputation. The data cleaning step employs a multi-rule verification mechanism: for physiological parameter measurements, outliers are identified using the 3σ principle (calculating the mean and standard deviation of the parameter, and removing values ​​exceeding the mean ± 3 times the standard deviation). Simultaneously, abnormal data (such as records with systolic blood pressure less than 80 mmHg or greater than 220 mmHg) are filtered using clinical common sense. Data deemed abnormal is marked as invalid, and the original records are retained for traceability. For medical event markers, records with time discrepancies (such as an event occurring earlier than the diagnosis of the underlying disease) or incorrect coding (such as non-existent event type codes) are removed through logical verification using event type coding and timestamps. The missing value imputation process employs a scenario-specific strategy: For continuous physiological parameters (such as blood glucose and blood pressure), if the time interval between missing data is less than 7 days, linear interpolation is used to calculate the imputation value based on the effective measurements before and after the missing data; if the missing interval is greater than 7 days and there are other related parameters in the same time period (such as glycated hemoglobin and fasting blood glucose), the multivariate adaptive imputation (MICE) method is used to build a predictive model based on the clinical correlation between parameters to generate the imputation value; for missing medical event markers, they are imputed by associating with indirect data such as outpatient prescription records and laboratory report conclusions. Missing event markers that cannot be imputed are marked as "not recorded" and annotated in the data log.

[0053] II. Time Window Division and Feature Calculation

[0054] After preprocessing, the time window segmentation module is activated, employing a fixed window length and sliding step strategy. The window length is determined based on the clinical monitoring cycle of the underlying disease; for example, the routine follow-up cycle for patients with hypertension / diabetes is 7 days, so the window length is set to 7 days. The sliding step is set to 1 day to ensure continuous coverage of the time series. During the segmentation process, starting from the time of the initial diagnosis of the underlying disease, data from 7 consecutive days are extracted sequentially to form windows, up to 24 hours before the onset of acute pancreatitis. Each window is assigned a unique window number and time range identifier (e.g., "Window 1: 2023-01-01 to 2023-01-07").

[0055] For each time window, the statistical characteristics of physiological parameters and the density of medical events were calculated separately. The calculation of physiological parameter statistical characteristics employed a multi-dimensional extraction method: for each continuous parameter, the mean (reflecting the average level of the parameter), standard deviation (reflecting the degree of parameter fluctuation), trend slope (fitting the relationship between the measured value and the collection time through linear regression; the slope sign indicates an upward / downward trend, and the absolute value indicates the trend strength), and coefficient of variation (the ratio of the standard deviation to the mean, a fluctuation index after eliminating the influence of dimensions) of all valid measurements within the window were calculated; for categorical parameters (such as drug usage type), the frequency and proportion of each category within the window were calculated. The calculation of medical event density used the window duration (7 days) as a baseline, counting the number of occurrences of various types of medical events within the window, and combining this with the clinical weight of the events (e.g., interventional treatment events were weighted at 3, medication adjustment events at 2, and routine follow-up events at 1) to calculate the weighted total number of events, which was then divided by the window duration to obtain the weighted event density per unit time (day), reflecting the activity level of basic disease intervention within the window.

[0056] III. Feature Sequence Construction and Dynamic System Modeling

[0057] Based on the calculation results of each time window, a feature sequence is constructed. Each time window corresponds to a feature vector, and the vector dimension is determined by the sum of the number of statistical features of physiological parameters and the number of medical event densities (e.g., extracting 4 statistical features for each of the 8 physiological parameters and 2 types of medical event densities, the total dimension is 8×4+2=34). The feature vectors of all windows are arranged in chronological order (from early to late) to form a feature sequence with the dimension of "number of windows × feature dimension". This sequence fully describes the change process of the underlying disease over time.

[0058] The dynamic system modeling phase begins with establishing the system state equations. Based on the feature changes between two adjacent time windows in the feature sequence, Pearson correlation analysis is used to calculate the correlation coefficients between each feature in the preceding window and each feature in the following window. Feature pairs with an absolute correlation coefficient greater than 0.6 are selected to determine the system state variables (features from the preceding window that are strongly correlated with the features of the subsequent window, such as the slope of the blood glucose trend, the standard deviation of blood pressure, and the density of hypoglycemic drug adjustment events). The state equation structure can be linear or nonlinear. If the relationship between the state variables and time in the feature sequence is linearly distributed, a linear state equation is used (change in state variable = coefficient matrix × preceding state variable + constant term). If the distribution is nonlinear, a nonlinear state equation is used (introducing quadratic or exponential terms of the state variables).

[0059] The equation parameters were determined through fitting historical data. Feature sequences from historical cases (sample size no less than 500 cases) with the same underlying disease type and similar disease duration as the current patient were selected as training data. The objective function was to minimize the error between the predicted and actual state variable values. The L-BFGS optimization algorithm was used iteratively to find the optimal parameter combination (coefficient matrix, constant term, etc.). During parameter optimization, medical event density was introduced as an adjustment factor: when the medical event density within the window was higher than the historical average density (the average density of similar events calculated based on historical cases), the parameter update step size was reduced (step size coefficient set to 0.8) to avoid over-adjustment of parameters due to dense events; when the event density was lower than the historical average density, the parameter update step size was increased (step size coefficient set to 1.2) to improve the sensitivity of parameters to slow changes. Simultaneously, the parameter update direction was adjusted according to the event type. For example, the update direction of the blood glucose trend slope parameter corresponding to an event of increased hypoglycemic drug dosage was negative (inhibiting blood glucose rise), and the update direction of the blood pressure standard deviation parameter corresponding to hypertension complication events was positive (reflecting increased volatility).

[0060] Numerical methods are used to solve the system state equations. For linear state equations, the Euler method is employed: the state variables of the initial window (the first time window) are used as initial conditions, substituted into the state equation to calculate the predicted state value for the next window, and this predicted value is then used as input to calculate the predicted value for the next window after that, iterating in this manner until all windows are solved. For nonlinear state equations, the fourth-order Runge-Kutta method is used, which improves the accuracy of the numerical solution for nonlinear functions by setting four intermediate calculation points within each window time step. During the solution process, a state estimate (containing the predicted values ​​of all state variables) is output for each time window. The state estimates of all windows are arranged in chronological order to form a continuous disease progression trajectory, where each node corresponds to the basic disease state of a time window.

[0061] IV. Long-term impact feature extraction and generation of underlying disease impact factors

[0062] For the simulated disease progression trajectory data, the feature extraction module is activated. Three methods—trend analysis, peak detection, and cycle identification—are used to extract the main change patterns: Trend analysis involves linear regression of the estimated values ​​of each state variable in the trajectory to obtain the slope and goodness of fit (R²) of the long-term trend (e.g., a continuous increase in mean blood glucose and a gradual decrease in blood pressure fluctuations), and significant trends with R² greater than 0.7 are selected as features; Peak detection calculates local maxima using a sliding window (with a window length of 3 time windows) to identify abnormal peaks in the state variables in the trajectory (e.g., a significantly higher mean blood glucose in a certain window than in the preceding and following windows), recording the peak occurrence time, peak size, and peak duration window number; Cycle identification converts the time-domain data of the trajectory into frequency-domain data using Fast Fourier Transform, identifying frequency components with an energy proportion exceeding 10% in the frequency domain, corresponding to the fluctuation cycle of the state variables (e.g., a small fluctuation in blood pressure every 14 days). The extracted trend features, peak features, and cycle features are integrated to form a long-term impact feature vector with fixed dimensions (e.g., containing 5 trend features, 3 peak features, and 2 cycle features, with a total dimension of 10).

[0063] Finally, the long-term impact feature vector is input into a pre-defined neural network structure. This neural network employs a three-layer fully connected architecture: the input layer dimension is consistent with the long-term impact feature vector dimension; two hidden layers are set, with the first hidden layer having twice the number of neurons as the input layer dimension, and the second hidden layer having half the number of neurons as the first layer, both using ReLU activation functions; the number of neurons in the output layer is determined according to the type of underlying disease (e.g., for hypertension, diabetes, and chronic kidney disease, the output layer dimension is 3), and the sigmoid activation function is used to ensure that the output value is in the 0-1 range. The neural network updates its parameters through a backpropagation algorithm (using the cross-entropy loss function to calculate the error between the predicted value and the actual impact label of the underlying disease in historical cases, and adjusting the weights through the Adam optimizer), and the final output vector is the underlying disease impact factor, with each dimension value reflecting the long-term impact strength of the corresponding underlying disease on the physiological state of the pancreas.

[0064] In S3, the specific process of feature fusion is as follows:

[0065] I. Data Standardization Processing

[0066] First, standardization operations were performed on real-time clinical indicators and underlying disease influencing factors, with the core objective of eliminating feature weight bias caused by differences in the units and value ranges of the two types of data. For real-time clinical indicators, the Z-score standardization method was adopted: first, the clinical indicator dataset of patients with the same type of acute pancreatitis in the historical case database was retrieved, and the global mean and standard deviation of each indicator (such as serum amylase, lipase, systolic blood pressure, and heart rate) were calculated; then, each real-time indicator value of the current patient was substituted into the transformation process, so that the mean of the transformed indicators approached 0 and the standard deviation approached 1, ensuring that indicators with different units (such as serum amylase in U / L and blood pressure in mmHg) were in the same numerical order of magnitude.

[0067] For the underlying disease influencing factors, a Min-Max standardization method was adopted. Since the influencing factors were already in the 0-1 range after the initial neural network output, but the distribution density of different dimensions (such as hypertension influencing factors and diabetes influencing factors) varied, further compression to a uniform value range was necessary. Specifically, the minimum and maximum values ​​of each influencing factor dimension in the historical training set were first calculated. Then, a linear transformation was used to map the current patient's influencing factor values ​​to the [0,1] range, avoiding an imbalance in feature contributions caused by dense values ​​in some dimensions and sparse values ​​in others. During the standardization process, all transformation parameters (mean, standard deviation, minimum, and maximum) were calculated offline from the historical case dataset and stored. These parameters were directly retrieved each time new patient data was processed, ensuring the consistency of the standardization rules.

[0068] II. Design and Implementation of Dual-Path Encoding Structure

[0069] After standardization, a dual-path encoding structure is used to process the features of the two types of data separately to adapt to the subsequent interactive fusion requirements. The first path targets real-time clinical indicators and focuses on "high-dimensional feature extraction": this path adopts a multi-layer fully connected network architecture, with the input layer dimension consistent with the number of real-time clinical indicators (e.g., when including 8 indicators such as serum amylase, lipase, abdominal imaging feature quantification value, heart rate, and blood oxygen saturation, the input layer dimension is set to 8); the number of neurons in the first fully connected layer is set to twice the input layer dimension, and the activation function is ReLU, which is used to initially expand the feature expression space; the number of neurons in the second fully connected layer is consistent with the first layer, and a BatchNorm layer is added for batch normalization to alleviate the gradient vanishing problem; the last fully connected layer compresses the feature dimension to a preset intermediate dimension (e.g., 16 dimensions) and outputs the clinical indicator encoded feature vector.

[0070] The second path targets the impact factors of underlying diseases, focusing on "low-dimensional feature transformation": This path adopts a lightweight multilayer perceptron architecture, with the input layer dimension consistent with the impact factor dimension (e.g., when there are 3 types of underlying diseases, the input layer dimension is set to 3); the first layer uses a 1×1 convolutional kernel to upscale the features, expanding the dimension to 4 times that of the input layer, and the activation function is LeakyReLU (with a negative slope of 0.1) to enhance the feature capture capability of weak impact factors; the second layer is a fully connected layer, compressing the dimension to an intermediate dimension (e.g., 16 dimensions) that is the same as the output dimension of the first path, and a BatchNorm layer is also added to stabilize the training process; finally, the output is the encoded feature vector of the underlying disease impact factors, ensuring that the dimensions of the two types of encoded feature vectors match, laying the foundation for subsequent interactive fusion.

[0071] III. Implementation of Feature Interaction Layer Based on Attention Allocation Mechanism

[0072] (1) Construction of feature interaction matrix

[0073] First, a feature interaction matrix is ​​constructed. The row dimension of the matrix is ​​consistent with the dimension of the clinical indicator coding features output by the first path, and the column dimension is consistent with the dimension of the impact factor coding features output by the second path (e.g., if both are 16-dimensional, the matrix size is 16×16). The position of each element in the matrix corresponds to a pairing of "single physiological indicator coding feature - single underlying disease impact factor coding feature". For example, the element in the i-th row and j-th column corresponds to the interaction relationship between the i-th feature of the clinical indicator coding and the j-th feature of the underlying disease impact factor coding. During the construction process, the two types of coding feature vectors are converted into a two-dimensional matrix form through tensor reshaping operations, and then the initial interaction matrix framework is generated through outer product operations to ensure that each feature pairing relationship is uniquely mapped to a matrix element.

[0074] (2) Correlation strength calculation and normalization

[0075] For each element in the feature interaction matrix, the association strength value is calculated: A vector dot product operation is used, taking the i-th feature vector (1-dimensional vector) of the clinical indicator encoding feature and the j-th feature vector (1-dimensional vector) of the underlying disease impact factor encoding feature, to obtain the original association strength value of that element. The core logic of this operation is: the larger the dot product result, the stronger the association between the two types of features in expressing the patient's physiological state (e.g., a high dot product value between "hyperglycemia impact factor feature" and "elevated serum amylase feature" reflects a strong association between underlying diabetes and the degree of inflammation in acute pancreatitis).

[0076] After calculating the original association strength, Softmax normalization is performed on each row of the matrix: the original association strength values ​​of all elements in each row are substituted into the Softmax function, so that the sum of the normalized results of each row's elements is 1, generating the attention weight distribution. The reason for normalization at the "row" level is to focus on "which underlying disease influencing factors should be primarily associated with a single physiological indicator feature." For example, for the "serum amylase feature" row, the column with the highest weight after normalization corresponds to the underlying disease factor that has the greatest impact on it, ensuring the targeted allocation of weights.

[0077] (3) Feature weighting and nonlinear transformation

[0078] The underlying disease impact factor coding features are reweighted based on the attention weight distribution: the attention weight matrix is ​​multiplied by the impact factor coding feature matrix output from the second path, adjusting each impact factor feature dimension according to its corresponding weight. Dimensions with higher weights retain more original information, while information in dimensions with lower weights is suppressed. After weighting, the adjusted impact factor feature matrix is ​​element-wise added to the clinical indicator coding feature matrix output from the first path to form a preliminary interactive feature matrix.

[0079] To enhance feature representation, a nonlinear transformation layer is introduced: This layer employs the LeakyReLU activation function (with a negative slope of 0.1) to perform a nonlinear transformation on each element of the initial interaction feature matrix, breaking the linear correlation between features. Simultaneously, a Dropout layer (with a dropout probability of 0.2) is added to randomly mask some feature dimensions, preventing the model from over-relying on local features. The output after the nonlinear transformation is an enhanced interaction feature matrix, with dimensions consistent with the initial interaction feature matrix.

[0080] (4) Iterative optimization of attention weights

[0081] The attention weights are calculated using an iterative optimization approach, with the core objective of gradually adjusting the weight distribution to better reflect clinical correlation patterns. The iteration trigger condition is that the difference between the current weight distribution and the previous iteration distribution exceeds a preset threshold (e.g., 0.05), calculated using KL divergence. The number of iterations is dynamically adjusted based on the total feature dimension: when the total feature dimension after dual-path encoding (clinical indicator encoding dimension + impact factor encoding dimension) is greater than 50, the maximum number of iterations is set to 3; when the total dimension is less than or equal to 50, the maximum number of iterations is set to 2. During each iteration, the attention weight matrix is ​​fine-tuned using a gradient descent algorithm, aiming to maximize the similarity between the interaction feature matrix and the historical best feature matrix, until the iteration termination condition is met or the maximum number of iterations is reached. The final output is the optimized attention weights and the enhanced interaction feature matrix.

[0082] IV. Feature Compression and Generation of Synthetic Feature Vectors

[0083] The enhanced interactive feature matrix is ​​input into the feature compression layer, which adopts a hybrid dimensionality reduction strategy of "fully connected + principal component analysis (PCA)": First, the dimension of the interactive feature matrix is ​​compressed to a preset intermediate dimension (such as 32 dimensions) through the fully connected layer to reduce the complexity of subsequent PCA calculations; then, PCA dimensionality reduction is performed on the feature vector output by the fully connected layer to retain the principal components whose cumulative variance contribution exceeds 95%, and finally, a comprehensive feature vector with fixed dimensions (such as 64 dimensions or 128 dimensions, the specific dimension is specified by the configuration file before model training).

[0084] During feature compression, the number of principal components in PCA is determined by offline analysis of historical case interaction feature data, ensuring that the compressed features retain the core information of the original interaction features. Simultaneously, L2 normalization is performed on the compressed feature vectors to ensure the Euclidean norm of the vectors is 1, preventing differences in vector magnitude from affecting the input stability of subsequent AI evaluation models. The final output fixed-dimensional comprehensive feature vector can be directly used as input data for subsequent artificial intelligence evaluation models, supporting the calculation of complication risk probabilities.

[0085] In step S4, the training process of the artificial intelligence evaluation model is as follows:

[0086] I. Training Process of Artificial Intelligence Evaluation Model

[0087] (1) Construction of deep neural network architecture

[0088] First, the architecture design and initialization of the deep neural network are completed. The overall architecture adopts a three-layer progressive structure of "input layer - feature extraction layer - output layer". The dimension of the input layer is consistent with the dimension of the comprehensive feature vector generated by S3. For example, when the comprehensive feature vector is 64-dimensional, the number of neurons in the input layer is set to 64. At the same time, a BatchNorm layer is connected after the input layer to perform batch normalization processing on the input features, so that the feature mean approaches 0 and the variance approaches 1, providing a stable input distribution for the subsequent feature extraction layer. The input layer and the feature extraction layer are connected by a fully connected layer. The weight parameters of this connection layer are initialized using the He normal method, and the bias parameters are initialized to 0 to ensure that the initial parameter distribution is adapted to the gradient propagation characteristics of the subsequent ReLU activation function.

[0089] The feature extraction layer, as the core of the architecture, adopts a multi-module cascade structure (the number of modules is determined by the dimension of the comprehensive feature vector, typically 3-5). Each module contains sub-layers such as linear transformation, non-linear activation, and regularization; specific implementation details will be discussed separately later. The design of the output layer is determined based on the type of complication assessment task: for risk assessment of a single complication (such as pancreatic abscess), the output layer has one neuron, using the Sigmoid activation function, and the output value directly corresponds to the risk probability of the corresponding complication; for joint assessment of multiple complications (such as pancreatic abscess, multiple organ failure, gastrointestinal bleeding), the number of neurons in the output layer is consistent with the number of complication types, with each neuron corresponding to one complication, also using the Sigmoid activation function, outputting the risk probabilities of each complication separately. The output layer weights are initialized using Xavier normality, and the bias is initialized to 0 to ensure that the output probability distribution is within a reasonable range during the initial training phase.

[0090] (2) Preparation of training dataset

[0091] The training dataset is constructed based on historical clinical case data, and its core consists of two types of data: "comprehensive feature vector" and "complication annotation information". The comprehensive feature vector comes from the output results of historical cases after processing through the S1-S3 process. That is, for each historical case of acute pancreatitis, a comprehensive feature vector with fixed dimensions is generated according to the same clinical data collection, underlying disease time series analysis and feature fusion process, forming a feature matrix (the dimension is "number of cases × feature dimension").

[0092] Complication labeling information is determined through a dual approach: manual review of historical case records and machine coding matching. First, descriptions of complications mentioned in the discharge diagnosis and progress notes are extracted, and a medical named entity recognition model (fine-tuned based on a BERT pre-trained model) is used to identify the complication type. Then, ICD-10 coding matching (e.g., pancreatic abscess corresponds to K85.802, multiple organ failure corresponds to R65.100) verifies the identification results to ensure labeling accuracy. Finally, the labeled information is converted into numerical labels; for example, in a single complication assessment, "with complications" is labeled as 1, and "no complications" as 0. In multiple complication assessments, one-hot coding is used (e.g., "pancreatic abscess only" is labeled as [1,0,0]). During dataset construction, cases with more than 10% missing comprehensive feature vectors or ambiguous complication labeling (e.g., only mentioning "complication" without specifying the type) are removed to ensure data quality.

[0093] (3) Division of training set and validation set

[0094] The dataset is partitioned using stratified sampling, with the core objective of maintaining consistency in the distribution of complication between the training and validation sets to avoid a decline in model generalization ability due to sample bias. The stratification is based on "complication type distribution" and "underlying disease combination type": First, the dataset is divided into different levels according to "presence of complications" or "specific complication type," for example, level 1 is "no complications," level 2 is "pancreatic abscess," and level 3 is "multiple organ failure." Then, within each level, sub-strata are further subdivided according to "underlying disease combination type" (e.g., "hypertension + diabetes," "chronic kidney disease only," "no underlying disease").

[0095] The sampling process is implemented using the StratifiedShuffleSplit algorithm. A predetermined ratio for the training and validation sets is set (typically 7:3). Samples are randomly drawn from each sub-layer according to this ratio, ensuring that the proportion of each sub-layer in the training and validation sets is completely consistent with the original dataset. After the split, data augmentation operations (such as random feature perturbation, adding small Gaussian noise to each feature value) are performed on the training set, while the validation set remains unchanged to avoid interference from data augmentation on the validation results.

[0096] (4) Loss function design and optimization algorithm selection

[0097] The loss function employs a composite structure of "classification loss + regularization term" to balance the model's fitting ability and resistance to overfitting. The classification loss is selected based on the task type: binary cross-entropy loss is used for single complication assessment, calculating the cross-entropy between the model's predicted probability and the true label (0 or 1); multivariate cross-entropy loss is used for joint assessment of multiple complications, calculating the cross-entropy between the predicted probability and the true label for each complication separately and then summing the results.

[0098] The regularization term employs L2 regularization (weight decay), multiplying the sum of squares of all trainable parameters (fully connected layer weights, BatchNorm layer parameters, etc.) by a preset weight decay coefficient, and adding it to the classification loss to suppress excessive parameter growth and reduce the risk of overfitting. The weight decay coefficient is determined through 5-fold cross-validation, selecting the coefficient value that minimizes the validation set loss within a preset candidate interval.

[0099] The optimization algorithm used is the Adam adaptive optimization algorithm, which combines momentum gradient descent with an adaptive learning rate adjustment mechanism. The initial learning rate is set according to the feature dimension (usually a preset baseline value determined based on previous experimental experience). The algorithm accumulates gradient directions using first-order momentum (β1, usually set to 0.9) and accumulates the squared gradient using second-order momentum (β2, usually set to 0.999), dynamically adjusting the learning rate of each parameter: the learning rate decreases for parameters with large gradients and increases for parameters with small gradients, achieving targeted parameter updates. A learning rate decay strategy is also introduced: if the validation set loss does not decrease for several consecutive epochs (e.g., 5 epochs), the current learning rate is multiplied by a decay factor (usually set to 0.5) until the learning rate drops to a preset minimum value.

[0100] (5) Training process monitoring and early stop mechanism

[0101] The training process iterates in epochs. Within each epoch, the training set is input into the model in batches (BatchSize is set according to the GPU memory capacity). The predicted probabilities are calculated through forward propagation, and the gradient of the loss function with respect to the parameters of each layer is calculated through backpropagation. The parameters are then updated by the optimization algorithm. After each epoch, a complete forward propagation is performed on the validation set to calculate the validation set loss and evaluation metrics (such as accuracy and F1 score), and the changing trends of the training set loss and validation set loss are recorded.

[0102] The early stopping mechanism is used to prevent model overfitting. The trigger condition is set as "within a consecutive preset number of epochs (e.g., 10 epochs), the validation set loss continuously increases or decreases by a rate less than a preset threshold". When this condition is triggered, training is immediately stopped, the currently trained model parameters are discarded, and the model parameters with the smallest validation set loss before stopping are loaded (saved through the model checkpoint mechanism during training). This parameter set is the final trained model parameters and can be directly called upon for subsequent actual evaluation tasks.

[0103] II. Specific Implementation Process of Feature Extraction Layer

[0104] (1) Structural design of cascaded feature transformation module

[0105] The feature extraction layer contains multiple cascaded feature transformation modules. The number of modules is determined by the initial dimension of the synthesized feature vector. For example, when the initial dimension is 64, four modules are set to achieve gradual compression of feature dimensions and information purification. The basic structure of each feature transformation module is "linear transformation layer - BatchNorm layer - nonlinear activation layer - Dropout layer":

[0106] The linear transformation layer is implemented using a fully connected layer. The number of neurons is set according to the dimensionality compression target of the current module. For example, when the output of the previous module is 64 dimensions, the number of neurons in the linear transformation layer of the current module is set to 48 dimensions to achieve initial dimensionality compression.

[0107] The BatchNorm layer performs batch normalization on the features after linear transformation, eliminating inter-layer distribution bias and accelerating training convergence.

[0108] The nonlinear activation layer uses the GELU activation function, which can retain more negative region feature information than ReLU, thus improving the ability to capture weak features.

[0109] Dropout layers reduce co-adaptation among neurons and enhance the model's generalization ability by randomly disabling some neurons (with a preset fixed probability).

[0110] (2) Implementation of residual connection

[0111] Each feature transformation module introduces residual connections, the core purpose of which is to solve the gradient vanishing problem in deep networks while preserving low-level feature information. The specific implementation logic of residual connections is as follows: the input feature of the module (denoted as X) is added element-wise to the output feature of the module after "linear transformation-BatchNorm-activation-Dropout" (denoted as F(X)) to obtain the final output of the module (denoted as H(X)=F(X)+X).

[0112] When the input dimension and output dimension of a module are inconsistent (i.e., the linear transformation layer compresses the dimension), a 1×1 convolutional layer needs to be added to the residual connection path to adjust the dimension of the input feature X: the number of output channels of the 1×1 convolutional layer is consistent with the final output dimension of the module, the convolutional kernel weights are initialized using He normality, and the dimension of X is converted to be the same as F(X) through this convolutional layer before performing element-wise addition to ensure the dimension compatibility of the residual connection.

[0113] (3) The logic of progressive compression of feature dimensions

[0114] The output dimensions of each feature transformation module are gradually reduced according to preset rules. For example, when the initial input dimension is 64, the output dimensions of the four modules are set to 48, 32, 16, and 8 respectively. The basis for dimensionality compression is feature importance: the approximate dimensional range of the core features is determined through preliminary feature analysis (such as feature variance analysis based on historical data and mutual information calculation), ensuring that the final module output dimension (e.g., 8 dimensions) can cover more than 95% of the core information, while eliminating redundant features.

[0115] During dimensionality compression, the weight parameters of the linear transformation layer automatically learn the importance of features, assigning larger values ​​to core features and smaller values ​​to redundant features, thus achieving adaptive feature selection. The output of the final feature transformation module (such as an 8-dimensional feature vector) is directly passed to the output layer as the core input feature for calculating the risk probability.

[0116] (4) Parameter update mechanism

[0117] The parameters of all modules in the feature extraction layer (including the weights and biases of the linear transformation layer, the parameters of the BatchNorm layer, and the weights of the 1×1 convolutional layer) are updated through the backpropagation algorithm. In each training batch, the gradient of the loss function is backpropagated from the output layer to the feature extraction layer, and the gradient values ​​of the parameters of each layer in each module are calculated in turn. Then, the Adam optimization algorithm adjusts the parameters according to the gradient values: the weight parameters are updated in the direction of the gradient, and the bias parameters and the parameters of the BatchNorm layer are updated synchronously according to the same optimization rules. This ensures that the parameters of the entire feature extraction layer are optimized in conjunction with the parameters of the input and output layers, so that the model as a whole converges to the optimal state.

[0118] In step S5, the specific process for generating the risk assessment report is as follows:

[0119] The mapping table between risk probability values ​​and risk levels is constructed jointly using clinical treatment guidelines and historical case outcome data. It is stored in the configuration repository of the medical data management system and uses a structured data table format. Fields include complication type, lower limit of probability range, upper limit of probability range, risk level name, and level description. When the system starts matching, it first reads the complication risk probability value and corresponding complication type output by the artificial intelligence assessment model. Then, it uses a structured query language to retrieve the mapping table corresponding to that complication type from the configuration repository. The risk probability value is compared numerically with the probability range in the table. If the probability value falls within a certain range, the risk level corresponding to that range is determined as the current patient's complication risk level. If multiple complication risk probability values ​​exist, the matching process is executed separately to generate the risk level corresponding to each complication.

[0120] The key data extraction module was then activated. Patient identification information was extracted from the patient master index database of the hospital information system, obtained through association with the patient's unique identifier. Specific information included name, gender, age, hospital number, admission date, and ward / bed number, ensuring consistency with the hospital's existing medical record system. Key clinical data extraction consisted of two parts: abnormal indicator screening and underlying disease summary generation. Abnormal indicator screening compared real-time clinical indicators such as serum amylase level, lipase level, and white blood cell count with the corresponding reference ranges by calling the reference range configuration table of the laboratory information management system, marking indicators that exceeded the normal range and the extent of the exceedance. The underlying disease summary was extracted from the historical records of underlying diseases, screening for the type of underlying disease, diagnosis time, and average level and fluctuation of key physiological parameters (such as blood glucose and blood pressure) in the past three years. The structured data was converted into concise summary text using a natural language generation algorithm, ensuring the information was complete and conformed to clinical reading habits.

[0121] Finally, the structured report generation process is executed. The predefined structured document template uses Extensible Markup Language (XML) format. The template contains four fixed sections: patient basic information, key clinical data, risk assessment results, and clinical recommendations. Each section contains several data placeholders, each corresponding to one of the extracted data fields. The system reads the template through its template engine, automatically filling the corresponding placeholders with the extracted patient identification information, key clinical data, risk probability values ​​for each complication, and risk levels. Simultaneously, in the clinical recommendations section, it calls a pre-defined clinical recommendations library based on the risk level to generate treatment recommendations matching the risk level. After document generation, the XML format is converted to a portable document format or a document format supported by the hospital's electronic medical record system using a format conversion engine. Finally, it is stored in the electronic medical record system through a medical document interface, linked to the patient's unique identifier, forming a traceable and viewable structured risk assessment report document.

[0122] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for assessing complications of acute pancreatitis based on artificial intelligence, characterized in that, Includes the following steps: S1. Obtain the clinical dataset of patients with acute pancreatitis, which includes real-time clinical indicators during an acute pancreatitis attack and the patient's underlying disease history. S2. Perform time-series analysis on the historical records of the underlying diseases, extract the long-term impact characteristics of the underlying diseases on the physiological state of the pancreas, and generate the underlying disease impact factors. S3. Standardize the real-time clinical indicators and the underlying disease influencing factors and calculate the correlation weights between features. Generate a fixed-dimensional comprehensive feature vector based on the weighted correlation weights. S4. Input the comprehensive feature vector into the trained artificial intelligence evaluation model, perform layer-by-layer nonlinear transformation and information extraction on the input features, and output the probability value of complication risk. S5. Map the risk probability value to a predefined risk level, and integrate patient identification information and key clinical data to form a structured document to generate a risk assessment report; In step S3, the specific process of feature fusion is as follows: Data standardization was performed on real-time clinical indicators and underlying disease influencing factors to eliminate dimensional differences. A dual-path coding structure was adopted, with the first path extracting features from clinical indicators and the second path transforming the features of influencing factors. Design a feature interaction layer based on attention allocation mechanism, calculate the correlation strength between two types of features, generate dynamic weight coefficients based on the correlation strength, perform weighted combination of the two types of features, input the weighted features into the feature compression layer, and generate a fixed-dimensional comprehensive feature vector through dimensionality reduction operation; The specific process of the attention allocation mechanism is as follows: Construct a feature interaction matrix, where rows correspond to clinical indicator features and columns correspond to basic disease influencing factor features. Calculate the association strength value of each element in the feature interaction matrix. The association strength value is obtained by the inner product operation of the feature vectors. Normalize the association strength value to generate an attention weight distribution. The original features are reweighted based on the attention weight distribution to highlight key feature dimensions. The weighted features are then passed through a non-linear transformation layer to enhance their expressive power. The calculation of attention weights adopts an iterative optimization method, and the number of iterations is dynamically adjusted according to the feature dimensions.

2. The method for assessing complications of acute pancreatitis based on artificial intelligence according to claim 1, characterized in that, In step S1, the specific process of obtaining the clinical dataset of patients with acute pancreatitis is as follows: Complete medical records of patients are extracted from the hospital information system through a medical data interface. The medical records include admission diagnosis information, laboratory test results and imaging reports. Clinical indicators related to the diagnosis of acute pancreatitis are screened from the medical records. The clinical indicators include serum amylase level and abdominal imaging features. At the same time, the patient's basic disease history records are extracted from their past medical history. The basic disease history records include disease type, diagnostic basis, and disease course records.

3. The method for assessing complications of acute pancreatitis based on artificial intelligence according to claim 1, characterized in that, In S2, the specific process of the time series analysis is as follows: Time-series data from the history of underlying diseases is read. This data includes regularly recorded physiological parameter measurements and medical event markers. The time-series data is preprocessed, including data cleaning and missing value imputation. The preprocessed data is divided into continuous time windows, and the statistical characteristics of physiological parameters and the density of medical events within each time window are calculated. Feature sequences are constructed based on the time window data, and a dynamic system modeling method is used to simulate the disease progression trajectory. Feature extraction is performed on the simulated trajectory data, and the main change patterns are extracted as long-term impact features. The long-term impact features are input into a neural network structure, and the underlying disease impact factor is output.

4. The method for assessing complications of acute pancreatitis based on artificial intelligence according to claim 3, characterized in that, The specific process of the dynamic system modeling method is as follows: A system state equation is established to describe the change of the characteristic sequence over time. The system state equation is constructed based on the correlation analysis of the characteristic changes of adjacent time windows. The equation parameters are determined by fitting historical data. The fitting process uses an optimization algorithm to find the optimal parameter combination. In the parameter optimization process, the density of medical events is introduced as an adjustment factor. The adjustment factor affects the step size and direction of parameter updates. The system state equation is solved using numerical calculation methods to calculate the state estimate for each time window. The state estimate forms a continuous disease progression trajectory.

5. The method for assessing complications of acute pancreatitis based on artificial intelligence according to claim 1, characterized in that, In step S4, the training process of the artificial intelligence evaluation model is as follows: A deep neural network architecture is constructed, consisting of an input layer, a feature extraction layer, and an output layer. A training dataset is prepared, containing comprehensive feature vectors of historical cases and complication annotations. A stratified sampling method is used to divide the training and validation sets, maintaining the data distribution of the training and validation sets. A loss function is designed, including classification loss and a regularization term. An adaptive optimization algorithm is used to train the network parameters. The optimization algorithm adjusts the learning rate. The training error is monitored during training, and training is stopped when the validation set error increases. After training is completed, the model parameters are saved.

6. The method for assessing complications of acute pancreatitis based on artificial intelligence according to claim 5, characterized in that, The specific process of the feature extraction layer is as follows: Multiple cascaded feature transformation modules are constructed. Each module contains linear transformation and nonlinear activation function. The first feature transformation module receives the feature vector of the input layer and performs feature extraction. Subsequent modules perform feature transformation on the output of the previous module. Residual connections are added during the feature transformation process to pass low-level feature information. The output dimension of each feature transformation module is gradually reduced, and the feature dimension is gradually reduced. The output of the last feature transformation module is passed to the output layer. The parameters of the feature extraction layer are updated through the backpropagation algorithm.

7. The method for assessing complications of acute pancreatitis based on artificial intelligence according to claim 1, characterized in that, In step S5, the specific process for generating the risk assessment report is as follows: Obtain the mapping table between risk probability values ​​and risk levels. The mapping table defines the risk levels corresponding to different probability ranges. Match the complication risk probability values ​​with the mapping table to determine the current patient's risk level. Patient identification information and key clinical data are extracted from the clinical dataset. The key clinical data includes major abnormal indicators and a summary of the underlying disease. Following a predefined structured document template, the patient identification information, key clinical data, risk probability values, and determined risk levels are combined to generate a complete risk assessment report document.

Images