Cerebral hemorrhage postoperative gastrointestinal hemorrhage prediction method based on LGBM model

Abstract

The invention discloses a cerebral hemorrhage postoperative gastrointestinal hemorrhage prediction method based on an LGBM model. The method comprises the steps of obtaining a multi-dimensional clinical feature sequence of a target patient, screening out a stable feature subset by adopting a Boruta feature selection algorithm, performing nonlinear relation fitting and integrated decision by utilizing a pre-trained LightGBM machine learning model, and generating an individualized ATH risk probability value; and when the risk probability value exceeds a dynamic risk threshold value, triggering a high-risk early warning signal, and based on a Kaplan-Meier survival analysis model, carrying out association mapping on a prognosis track of poor long-term neural function recovery, and finally generating a comprehensive prediction report. According to the invention, accurate quantitative evaluation of ATH risk is realized, clinical intervention timeliness is improved through a dynamic threshold early warning mechanism, short-term complication risk and long-term function prognosis are organically combined, and a comprehensive and reliable prognosis basis is provided for individualized treatment decision.

Description

Technical Field

[0001] This invention relates to the field of medical prognosis prediction technology, specifically to a method for predicting gastrointestinal bleeding after cerebral hemorrhage based on the LGBM model. Background Technology

[0002] Cerebral hemorrhage is one of the most deadly and disabling subtypes of stroke, and early identification and intervention of postoperative complications are crucial for improving patient prognosis. Postoperative gastrointestinal bleeding, a common and serious complication after cerebral hemorrhage, not only significantly prolongs hospital stays and increases the medical burden but is also closely related to poor neurological function recovery. Currently, clinical risk assessment for postoperative gastrointestinal bleeding relies heavily on empirical judgment or traditional statistical models based on linear assumptions, which struggle to effectively capture the complex nonlinear relationships between multidimensional clinical characteristics, resulting in limited predictive accuracy and insufficient generalization ability. Furthermore, existing prediction methods often focus on identifying short-term complications, failing to systematically link complication risk with long-term functional outcomes, leaving clinical decision-making lacking a holistic understanding of the prognostic trajectory. Therefore, there is an urgent need to develop a precise prediction method that can adapt to nonlinear relationships and connect short-term risk with long-term prognosis to improve the proactiveness and targeted nature of postoperative management of cerebral hemorrhage. Summary of the Invention

[0003] In view of the above problems, the present invention provides a method for predicting gastrointestinal bleeding after cerebral hemorrhage based on the LGBM model. By integrating Boruta feature selection with the LightGBM model for decision-making, it can achieve a comprehensive prediction of short-term complication risk and long-term neurological function prognosis, thus solving the problems of insufficient accuracy and inability to correlate long-term prognosis of existing prediction methods.

[0004] To achieve the above objectives, this application provides a method for predicting gastrointestinal bleeding after cerebral hemorrhage based on the LGBM model, comprising: Obtain a multidimensional clinical feature sequence of the target patient. The multidimensional clinical feature sequence should include at least neurological function score, imaging hemorrhage volume, operation duration, serum nutritional markers and blood glucose parameters. Based on historical multidimensional clinical feature sequences and historical postoperative ATH outcome labels of historical patient cohorts, the Boruta feature selection algorithm is used to perform high-dimensional feature screening on the multidimensional clinical feature sequences to construct a stable feature subset for ATH risk prediction. A stable subset of features is input into a pre-trained LightGBM machine learning model, which performs nonlinear relationship fitting and ensemble decision-making through its built-in multiple decision trees to generate individualized ATH risk probability values ​​for the target patient. The individualized ATH risk probability value is compared with the preset dynamic risk threshold. When the risk probability value exceeds the dynamic risk threshold, a high-risk warning signal is automatically triggered. In response to high-risk warning signals, the Kaplan-Meier survival analysis model is used to map the individualized ATH risk probability value with the prognostic trajectory of poor long-term neurological function recovery associated with the target patient, generating and outputting a comprehensive prediction report that integrates short-term complication risk and long-term functional prognosis.

[0005] In some embodiments, based on historical multidimensional clinical feature sequences and historical postoperative ATH outcome labels from historical patient cohorts, the Boruta feature selection algorithm is used to perform high-dimensional feature screening on the multidimensional clinical feature sequences to construct a stable feature subset for ATH risk prediction, including: A primitive feature space containing historical patient cohorts is constructed. The primitive feature space is composed of historical multidimensional clinical feature sequences and composite physiological indicators derived from calculations. The Boruta feature selection algorithm is used to perform multiple rounds of random forest importance evaluation in the original feature space, generating a feature importance score distribution for each original feature; A random shadow feature set with the same dimension as the original feature space is created based on the feature importance score distribution, and the maximum importance score of each shadow feature is calculated as the benchmark threshold for feature selection. The importance score distribution of each original feature is statistically compared with the benchmark threshold to select the core feature subset whose importance is consistently better than the benchmark threshold; Multicollinearity diagnosis is performed on the core feature subset, and feature variables with redundant information are eliminated through variance inflation factor analysis to form a stable feature subset.

[0006] In some embodiments, the Boruta feature selection algorithm is used to perform multiple rounds of random forest importance evaluation in the original feature space, generating a feature importance score distribution for each original feature, including: In each evaluation iteration, the Bootstrap resampling algorithm is used to construct a differentiated training subset, and the split contribution of each original feature in the random forest decision tree is calculated by the change in Gini impurity. Perform a preset number of iterative evaluations, record the importance score of each original feature in each iteration, and form a feature importance score sequence based on statistical distribution; The kernel density estimation algorithm is used to smooth the feature importance score sequence to generate a feature importance density curve with probability distribution characteristics; The confidence interval for each original feature is calculated based on the feature importance density curve, and the statistical significance of the feature importance score is verified by hypothesis testing. The statistically validated feature importance score distribution is dynamically compared with the random shadow feature set to identify feature variables that consistently outperform the noise benchmark in multiple iterations, thus obtaining the final feature importance score distribution.

[0007] In some embodiments, a stable subset of features is input into a pre-trained LightGBM machine learning model, which performs nonlinear relationship fitting and ensemble decision-making through its built-in multiple decision trees to generate an individualized ATH risk probability value for the target patient, including: A stable subset of features is fed in parallel into multiple decision trees of the LightGBM model. Each decision tree independently judges the input features based on its pre-trained structure and outputs a preliminary prediction score. Based on the predefined weights of each decision tree in the ensemble, the initial prediction scores output by all decision trees are weighted and summed to obtain an ensemble's original prediction score. The integrated raw prediction scores are input into the Sigmoid activation function for probability transformation, mapped to individualized ATH risk probability values ​​in the range of 0 to 1, and then output.

[0008] In some embodiments, the integrated raw predicted score is input into a Sigmoid activation function for probability transformation, mapped to an individualized ATH risk probability value in the interval of 0 to 1, and then output, including: The initial continuous probability value output by the Sigmoid function is calibrated using an ordinal-preserving regression algorithm to match the distribution of ATH incidence observed in real clinical settings, thus obtaining the calibrated probability value. The calibration probability values ​​are mapped to clinically interpretable risk level intervals using a piecewise linear transformation algorithm, and standardized based on the risk level intervals to obtain standardized probability values. The Monte Carlo Dropout algorithm is used to randomly discard some decision paths, generating multiple perturbation estimates of standardized probability values. The variance of these estimates is then used to obtain a quantitative index of the reliability of the standardized probability values. When the standardized probability value and its corresponding reliability quantification index jointly determine that it is near the clinical decision boundary, the ensemble model reconstruction mechanism is activated. By adjusting the ensemble weights of the decision tree in the LightGBM model, the standardized probability value is optimized and corrected, and the final individualized ATH risk probability value is generated and output.

[0009] In some embodiments, an individualized ATH risk probability value is compared with a preset dynamic risk threshold. When the risk probability value exceeds the dynamic risk threshold, a high-risk warning signal is automatically triggered, including: Based on the historical ATH risk probability value distribution of historical patient samples, a dynamic risk threshold matching different clinical scenarios is constructed using quantile statistical algorithms; When the individualized ATH risk probability value is detected to be continuously located in the high-risk quantile range of the dynamic risk threshold, a primary warning signal is generated. A retrospective analysis of clinical characteristics was performed on historical patient samples that triggered primary warning signals, and the co-occurrence pattern of high-risk characteristics was confirmed by logistic regression. Based on the confirmed co-occurrence patterns of high-risk features, a multi-level early warning escalation mechanism is activated to generate high-risk early warning signals that include feature contribution analysis. The high-risk early warning signal is exchanged with the hospital information system in real time, which automatically triggers the priority ranking and resource pre-allocation instructions for the clinical intervention queue.

[0010] In some embodiments, based on the historical ATH risk probability distribution of historical patient samples, a dynamic risk threshold matching different clinical scenarios is constructed using a quantile statistical algorithm, including: Obtain clinical scenario characteristic data from historical patient samples. The clinical scenario characteristic data should include at least the type of surgery, the degree of neurological deficit at admission, and the comorbidity index. For each combination of clinical scenario features, the historical ATH risk probability value distribution of the corresponding historical patient subset is extracted, and the kernel density estimation algorithm is used to construct the risk probability density function under each clinical scenario. Based on the risk probability density function, dynamic risk thresholds for different clinical scenarios are calculated using a quantile regression model. The dynamic risk thresholds include warning thresholds and high-risk thresholds. In addition, when the newly added historical patient data reaches a preset scale, the risk probability density function for each clinical scenario is recalculated, and the dynamic risk threshold is adaptively adjusted.

[0011] In some embodiments, in response to a high-risk warning signal, the individualized ATH risk probability value is mapped to the prognostic trajectory of poor long-term neurological function recovery associated with the target patient based on the Kaplan-Meier survival analysis model, generating and outputting a comprehensive prediction report integrating short-term complication risk and long-term functional prognosis, including: A time-series-based database of neural function recovery trajectories was constructed, and functional recovery survival curves under different ATH risk stratifications were calculated using the Kaplan-Meier survival analysis model. Cox proportional hazards regression analysis was used to quantify the impact of ATH risk on long-term prognosis, and a prognostic trajectory map with time dimension prediction was generated. The key time nodes and expected recovery inflection points in the functional recovery process were identified by the time series pattern recognition algorithm. In the prognostic trajectory map, the short-term ATH risk warning, long-term functional recovery trajectory and real-time clinicopathological features are fused into multi-dimensional data to obtain the fusion result; A comprehensive prediction report is obtained based on the fusion results.

[0012] In some embodiments, a time-series pattern recognition algorithm is used to identify key time points and expected recovery inflection points during the functional recovery process, including: Time-series data of neural function recovery were extracted based on prognostic trajectory maps, and the recovery curves were smoothed using Savitzky-Golay filters to obtain denoised recovery trajectories. The first-order difference calculation is performed on the denoised recovered trajectory to construct the recovered velocity sequence, and the second-order difference calculation is used to construct the recovered acceleration sequence; A change point detection algorithm based on Bayesian inference is used to analyze the recovery acceleration sequence and identify statistical abrupt changes in the recovery trend. Statistical mutation points were mapped to the original timeline and combined with clinically defined milestone events in neurological function recovery to verify the clinical significance of statistical mutation points, which were then denoted as clinically significant mutation points. Based on clinically significant mutation points, the potential inflection points in the subsequent recovery path are predicted using an autoregressive integral moving average model, and key time nodes and expected recovery inflection points are output.

[0013] In some embodiments, the prognostic trajectory atlas integrates short-term ATH risk warning, long-term functional recovery trajectory, and real-time clinicopathological features to obtain a multi-dimensional data fusion result, including: Using the surgical time as a benchmark, short-term ATH risk warnings, long-term functional recovery trajectories, and real-time clinicopathological features are uniformly mapped onto the same time coordinate system; A feature fusion network based on attention weights is constructed. By calculating the correlation strength between various features and prognostic outcomes at different time points, the contribution weight of each feature in the data fusion process is dynamically allocated. Tensor decomposition is used to perform joint dimensionality reduction on temporal multimodal data, extracting potential prognostic correlation features across time scales and forming a unified multidimensional fusion feature representation; A heterogeneous information graph of patient prognosis is constructed by graph neural network, with clinical events, pathological features and recovery status at different time points as nodes, and the temporal relationship and physiological correlation between them as edges. Based on heterogeneous information graphs, a graph attention network is used to aggregate and update node features, generating a comprehensive prognostic feature representation that simultaneously includes short-term ATH risk warning, long-term functional recovery trajectory, and real-time clinicopathological features as the fusion result.

[0014] Unlike existing technologies, the above-mentioned technical solution provides a method for predicting gastrointestinal bleeding after cerebral hemorrhage based on the LGBM model. It acquires multidimensional clinical feature sequences of the target patient, uses the Boruta feature selection algorithm to screen a stable feature subset, and utilizes a pre-trained LightGBM machine learning model for nonlinear relationship fitting and ensemble decision-making to generate an individualized ATH risk probability value. When the risk probability value exceeds a dynamic risk threshold, a high-risk warning signal is triggered, and a prognostic trajectory of poor long-term neurological function recovery is mapped based on the Kaplan-Meier survival analysis model, ultimately generating a comprehensive prediction report. This invention achieves accurate quantitative assessment of ATH risk, improves the timeliness of clinical intervention through a dynamic threshold warning mechanism, and organically combines short-term complication risk with long-term functional prognosis, providing a comprehensive and reliable prognostic basis for individualized treatment decisions.

[0015] The above description of the invention is merely an overview of the technical solution of this application. In order to enable those skilled in the art to better understand the technical solution of this application and to implement it based on the description and drawings, and to make the above-mentioned objectives and other objectives, features and advantages of this application easier to understand, the following description is provided in conjunction with the specific embodiments and drawings of this application. Attached Figure Description

[0016] The accompanying drawings are only used to illustrate the principles, implementation methods, applications, features, and effects of specific embodiments of the present invention and other related contents, and should not be considered as limitations on this application.

[0017] In the accompanying drawings of the instruction manual: Figure 1 This is a schematic diagram illustrating steps S101 to S105 of the prediction method described in the specific implementation embodiment; Figure 2 The following is a schematic diagram illustrating steps S201 to S205 of the prediction method described in a specific implementation. Detailed Implementation

[0018] To illustrate the possible application scenarios, technical principles, implementable specific solutions, and achievable objectives and effects of this application in detail, the following description, in conjunction with the listed specific embodiments and accompanying drawings, provides a detailed explanation. The embodiments described herein are merely illustrative of the technical solutions of this application and are therefore intended to limit the scope of protection of this application.

[0019] In this document, the term "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The term "embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment, nor does it specifically limit its independence or connection with other embodiments. In principle, in this application, as long as there are no technical contradictions or conflicts, the technical features mentioned in each embodiment can be combined in any way to form corresponding implementable technical solutions.

[0020] Unless otherwise defined, the technical terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the use of related terms herein is merely for the purpose of describing particular embodiments and is not intended to limit this application.

[0021] In the description of this application, the term "and / or" is used to describe the logical relationship between objects, indicating that three relationships can exist. For example, A and / or B means: A exists, B exists, and A and B exist simultaneously. Additionally, the character " / " in this document generally indicates that the preceding and following objects have an "or" logical relationship.

[0022] In this application, terms such as “first” and “second” are used only to distinguish one entity or operation from another, and do not necessarily require or imply any actual quantity, hierarchy or order relationship between these entities or operations.

[0023] Without further limitations, the use of terms such as “comprising,” “including,” “having,” or other similar open-ended expressions in this application is intended to cover non-exclusive inclusion, which does not exclude the presence of additional elements in a process, method, or product that includes the stated elements, such that a process, method, or product that includes a list of elements may include not only those defined elements but also other elements not expressly listed, or elements inherent to such a process, method, or product.

[0024] As understood in the Examination Guidelines, in this application, expressions such as "greater than," "less than," and "exceeding" are understood to exclude the stated number; expressions such as "above," "below," and "within" are understood to include the stated number. Furthermore, in the description of the embodiments in this application, "multiple" means two or more (including two), and similar expressions related to "multiple" are also understood in this way, such as "multiple groups" and "multiple times," unless otherwise explicitly specified.

[0025] In the description of the embodiments of this application, the space-related expressions used, such as "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "vertical," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential," indicate the orientation or positional relationship based on the orientation or positional relationship shown in the specific embodiments or drawings. They are only for the purpose of describing the specific embodiments of this application or for the reader's understanding, and do not indicate or imply that the device or component referred to must have a specific position, a specific orientation, or be constructed or operated in a specific orientation. Therefore, they should not be construed as limitations on the embodiments of this application.

[0026] The processor described in the embodiments of this application can be implemented by hardware, firmware, software, or a combination thereof. It can be a circuit, one or more of an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field-programmable gate array (FPGA), a central processing unit (CPU), a controller, a microcontroller, or a microprocessor. It also includes other physical, biological, or chemical structures that can implement the same or equivalent functions as the processors listed above, such as biological neurons, quantum computing units, DNA computing units, etc., so that the processor can execute some or all of the steps in the computer program or method involved in the various embodiments of this application, or any combination of the steps mentioned therein.

[0027] The computer program involved in the embodiments can be stored in a computer device readable storage medium, which includes, but is not limited to, disks, magnetic tapes, magnetic cards, floppy disks, flash memory, optical disks, optical cards, read-only memory (ROM), random access memory (RAM), erasable programmable ROM (EPROM), and electrically erasable programmable ROM (EEPROM), etc., and also includes other biological, physical, or chemical structures that can achieve the same or equivalent functions as the storage media listed above, such as DNA, RNA, proteins, and other units with information storage capabilities. In specific embodiments, the storage medium involved can be one of the above-mentioned media types, or a combination of the above-mentioned media types. In different embodiments, the computer program involved in the embodiments can be centrally stored in a single medium, or distributed and stored in multiple media. The memory containing the computer device readable storage medium can be non-volatile memory or random access memory. These computer device readable storage media can be built into the device, or can be connected to the device involved in the embodiments as an external device or part of an external device. In some embodiments, the memory having a computer device readable storage medium is deployed locally; in other embodiments, the memory may be deployed remotely from the processor, for example, as a network-attached memory accessed via RF circuitry or an external port and a communication network, wherein the communication network may be the Internet, one or more intranets, a local area network (LAN), a wide area network (WLAN), a storage area network (SAN), or a suitable combination thereof, as long as computer device access to the memory is enabled. Furthermore, the computer program involved in the embodiments may be stored in plaintext / ciphertext form, or it may be designed as training data, integrated and recombined through model training and implicitly stored in the parameter states of a deep neural network or other machine learning model.

[0028] Please see Figure 1 This embodiment provides a method for predicting postoperative gastrointestinal bleeding after cerebral hemorrhage based on the LGBM model, including: S101. Obtain the multidimensional clinical feature sequence of the target patient. The multidimensional clinical feature sequence shall include at least the neurological function score, imaging hemorrhage volume, operation duration, serum nutritional markers and blood glucose parameters. S102. Based on the historical multidimensional clinical feature sequences and historical postoperative ATH outcome labels of historical patient cohorts, the Boruta feature selection algorithm is used to perform high-dimensional feature screening on the multidimensional clinical feature sequences to construct a stable feature subset for ATH risk prediction. S103. Input the stable feature subset into the pre-trained LightGBM machine learning model, and use its built-in multiple decision trees to perform nonlinear relationship fitting and ensemble decision-making to generate individualized ATH risk probability values ​​for the target patient. S104. Compare the individualized ATH risk probability value with the preset dynamic risk threshold. When the risk probability value exceeds the dynamic risk threshold, automatically trigger a high-risk warning signal. S105. In response to high-risk warning signals, based on the Kaplan-Meier survival analysis model, the individualized ATH risk probability value is correlated and mapped with the prognostic trajectory of poor long-term neurological function recovery associated with the target patient, generating and outputting a comprehensive prediction report integrating short-term complication risk and long-term functional prognosis.

[0029] In step S101, the multidimensional clinical feature sequence refers to the set of clinical data extracted from the target patient's electronic health record system. This includes neurological function scoring using the Glasgow Coma Scale to assess consciousness, imaging hemorrhage volume calculated from head CT images using the ABC / 2 method, surgical duration recording the total time from skin incision to suturing completion, serum nutritional markers including laboratory indicators reflecting nutritional status such as albumin, and blood glucose parameters including fasting blood glucose and stress-induced hyperglycemia indicators. These features are automatically collected and standardized into a data sequence through the hospital information system interface, providing multidimensional input data for subsequent predictive models. This step establishes a quantitative basis for comprehensively assessing the patient's physiological state by integrating key indicators reflecting the degree of nerve damage, surgical trauma stress, and metabolic status.

[0030] In step S102, the Boruta feature selection algorithm constructs a random shadow feature set with the same dimensions as the original features. Using multi-round random forest importance evaluation, it compares the importance scores of the original features with the maximum importance scores of the shadow features, selecting feature variables whose importance consistently outweighs random noise. Historical postoperative ATH outcome labels are derived from clinical case data confirmed by standardized diagnosis and are used as supervised learning labels for constructing the predictive model. The historical multidimensional clinical feature sequences of the historical patient cohort and historical postoperative ATH outcome labels constitute the training dataset, where the ATH outcome labels are determined according to clinical diagnostic criteria. The stable feature subset selected by this algorithm eliminates redundant features and retains core features significantly correlated with the occurrence of ATH, ensuring the simplicity and predictive power of the feature set. This step, through a fully enclosed feature selection mechanism, overcomes the subjectivity of traditional feature selection methods and improves the robustness and repeatability of feature selection.

[0031] In step S103, the pre-trained LightGBM machine learning model employs a gradient boosting decision tree framework. Multiple decision trees are constructed using a leaf-splitting strategy, with each tree learning different aspects of the data. Finally, the predictions from each tree are integrated through a weighted voting mechanism. It should be noted that the pre-trained LightGBM machine learning model is based on historical patient cohort data, and this training process includes a feature processing flow similar to the aforementioned feature selection step. In practical applications, a stable subset of features from new patients is directly input into the pre-trained model for prediction, ensuring efficiency and consistency in clinical applications.

[0032] To further enhance the interpretability of the model and enable clinicians to understand and trust its decision-making process, this embodiment introduces the Shapley Additive exPlanations (SHAP) method to interpret and visualize the output of the LightGBM model. SHAP analysis, based on Shapley values ​​in game theory, assigns an importance value (SHAP value) to each feature in each prediction sample, thereby quantifying the feature's contribution to the final prediction result. In practice, the system generates two types of SHAP visualization charts: a global importance chart and an individualized force chart. The global importance chart ranks and displays features by calculating the average absolute value of the SHAP values ​​for each feature across all prediction samples. This chart visually reveals which clinical features contribute most to the prediction of ATH risk in the overall patient cohort. For example, the analysis results may show that surgical intervention, serum albumin level, Glasgow Coma Scale (GCS) score on admission, and hemorrhage volume are the core variables with the strongest predictive value.

[0033] The individualized force map interprets the prediction results for a single target patient. This graph mechanically illustrates how each feature "pushes" the model's baseline prediction to the final individualized ATH risk probability. In the graph, the arrows representing each feature show the direction (red arrows indicate a "positive contribution" increasing ATH risk, and blue arrows indicate a "negative contribution" decreasing risk) and intensity (arrow length) of their contribution. For a specific patient, if their model prediction is significantly higher than the baseline, it suggests they are at high risk for ATH, allowing clinicians to visually identify the key drivers leading to this high risk.

[0034] After stabilizing the feature subset input into the model, the complex relationship between features and ATH risk is captured through nonlinear transformations and feature interactions within a tree structure, outputting an individualized ATH risk probability value between 0 and 1. During the training phase, the model uses historical patient data to learn the mapping relationship between features and outcomes; during the prediction phase, it infers from new patient data. This step leverages the advantages of ensemble learning, effectively handling the nonlinear relationships and interaction effects between clinical features through hierarchical decision-making across multiple trees.

[0035] In step S104, the dynamic risk threshold is determined based on the risk distribution of ATH occurrence in historical data. By analyzing the correspondence between risk probability values ​​and actual ATH occurrences, a critical value that balances sensitivity and specificity is selected. When the individualized ATH risk probability value exceeds this threshold, the system automatically generates a high-risk warning signal, which includes information such as patient identification, risk level, and trigger time. The dynamic risk threshold can be adjusted according to the characteristics of different patient groups to ensure the accuracy and clinical applicability of the warning. This step, by quantifying the risk threshold, realizes the transformation from continuous probability to classification decision-making, providing a clear time point for clinical intervention.

[0036] In step S105, the Kaplan-Meier survival analysis model constructs a prognostic trajectory for long-term functional recovery by calculating the survival probability at different time points. This model correlates individualized ATH risk probabilities with modified Rankin Scale scores and compares differences in functional recovery among patients in different risk stratifications using a log-rank test. The comprehensive prediction report integrates short-term ATH risk warning and long-term neurological functional prognosis prediction, using a standardized template to generate a structured document containing risk level, prognostic trends, and clinical recommendations. This step, by linking complication risk with functional outcomes, expands from single complication prediction to comprehensive prognostic assessment, providing a more comprehensive reference for clinical decision-making.

[0037] This embodiment achieves accurate quantitative assessment of acute brain hemorrhage (ATH) risk after surgery by constructing a complete prediction process from feature acquisition to risk warning. The Boruta algorithm is used for feature selection to ensure the quality of the model input features. The LightGBM model effectively captures the complex nonlinear relationships between clinical features. The dynamic risk threshold mechanism enables risk grading and timely warning. Combined with prognostic correlation analysis using the Kaplan-Meier model, the clinical value of the prediction is expanded. This approach overcomes the linear assumptions of traditional prediction methods, improves prediction accuracy through machine learning technology, and organically combines short-term complication risk with long-term functional prognosis, providing clinical support throughout the entire process from risk identification to intervention decision-making.

[0038] Please see Figure 2In some embodiments, based on historical multidimensional clinical feature sequences and historical postoperative ATH outcome labels of historical patient cohorts, the Boruta feature selection algorithm is used to perform high-dimensional feature screening on the multidimensional clinical feature sequences to construct a stable feature subset for ATH risk prediction, including: S201. Construct an original feature space containing historical patient cohorts. The original feature space is composed of historical multidimensional clinical feature sequences and composite physiological indicators derived from calculations. S202. Perform multiple rounds of random forest importance evaluation in the original feature space using the Boruta feature selection algorithm to generate a feature importance score distribution for each original feature; S203. Create a random shadow feature set with the same dimension as the original feature space based on the feature importance score distribution, and calculate the maximum importance score of each shadow feature as the benchmark threshold for feature selection. S204. Compare the importance score distribution of each original feature with the benchmark threshold statistically to select the core feature subset whose importance is consistently better than the benchmark threshold. S205. Perform multicollinearity diagnosis on the core feature subset, and eliminate feature variables with redundant information through variance inflation factor analysis to form a stable feature subset.

[0039] In step S201, the original feature space is an initial feature set composed of clinical data from historical patient cohorts. This initial feature set includes historical multidimensional clinical feature sequences such as neurological function scores and imaging hemorrhage volume. Composite physiological indicators are derived from these basic features; for example, a neurological injury severity index can be constructed by multiplying the Glasgow Coma Scale score by the hemorrhage volume, or a nutritional metabolic balance indicator can be constructed using the albumin-to-glucose ratio. These composite indicators are constructed using clinical empirical formulas or statistical methods and can reflect the interactions between different physiological systems. This step, through the combination of basic and derived features, expands the representational dimensions of the features, providing richer input information for subsequent feature selection.

[0040] In step S202, the multi-round random forest importance evaluation constructs multiple training subsets through Bootstrap resampling. A random forest model is trained on each subset, and feature importance scores are calculated. Specifically, the Gini impurity reduction is used as an importance metric, recording the total impurity reduction for each feature during decision tree node splits. After a preset number of iterations, a feature importance score distribution is formed. Preferably, the number of iterations is set to 100-500 to ensure statistical stability. This step, through multiple resampling and model training, obtains a stable statistical distribution of feature importance, providing a reliable basis for subsequent screening.

[0041] In step S203, the random shadow feature set is generated by randomly shuffling the order of the original feature values, while maintaining the same dimensionality as the original feature space. The baseline threshold is determined by calculating the highest importance score obtained by all shadow features in multiple rounds of evaluation, specifically taking the maximum value of the shadow feature importance score in each round of evaluation as the screening benchmark. This threshold represents the potential importance level of random noise features, providing an objective standard for distinguishing truly important features from random noise. This step establishes a statistical significance benchmark for feature screening by introducing a random control mechanism.

[0042] In step S204, the statistical significance comparison uses a non-parametric test method to compare the importance score distribution of each original feature with a benchmark threshold. Specifically, the proportion of original features whose importance scores exceed the benchmark threshold in each round of evaluation is calculated. When this proportion reaches a preset significance level (e.g., 95%), the feature is considered to have consistently higher importance than random noise. The selected core feature subset contains feature variables that consistently demonstrate significant importance in multiple evaluations.

[0043] In step S205, multicollinearity diagnosis assesses the linear correlation between features by calculating the variance inflation factor. The variance inflation factor is calculated by establishing a linear regression model of each feature against other features; a larger value indicates more severe collinearity. Preferably, features with a variance inflation factor greater than 5 are considered to have severe collinearity and are removed. The features retained after screening constitute a stable feature subset, ensuring feature importance while avoiding information redundancy.

[0044] This embodiment achieves refinement from the original feature space to a stable feature subset through a systematic feature selection process. The Boruta algorithm's all-encompassing selection mechanism ensures comprehensive feature selection, and the combination with multicollinearity diagnosis further enhances the purity of the feature set. This embodiment overcomes the subjectivity and instability of traditional feature selection methods. Through the dual guarantees of statistical significance testing and collinearity analysis, the selected feature subset is both clinically significant and meets statistical requirements, providing high-quality input features for subsequent predictive model construction.

[0045] In some embodiments, the Boruta feature selection algorithm is used to perform multiple rounds of random forest importance evaluation in the original feature space, generating a feature importance score distribution for each original feature, including: In each evaluation iteration, the Bootstrap resampling algorithm is used to construct a differentiated training subset, and the split contribution of each original feature in the random forest decision tree is calculated by the change in Gini impurity. Perform a preset number of iterative evaluations, record the importance score of each original feature in each iteration, and form a feature importance score sequence based on statistical distribution; The kernel density estimation algorithm is used to smooth the feature importance score sequence to generate a feature importance density curve with probability distribution characteristics; The confidence interval for each original feature is calculated based on the feature importance density curve, and the statistical significance of the feature importance score is verified by hypothesis testing. The statistically validated feature importance score distribution is dynamically compared with the random shadow feature set to identify feature variables that consistently outperform the noise benchmark in multiple iterations, thus obtaining the final feature importance score distribution.

[0046] In this embodiment, the Bootstrap resampling algorithm constructs training subsets by randomly sampling historical patient samples with replacement. Each subset has the same sample size as the original dataset but differs in its composition. The change in Gini impurity is measured by the reduction in the Gini coefficient of a feature before and after splitting a decision tree node; specifically, it is the impurity of the node before splitting minus the weighted sum of the impurities of each child node after splitting according to that feature. This calculation method accurately reflects the contribution of each feature to improving data purity during decision tree construction. This step introduces data variability and impurity-based contribution quantification through resampling, providing a stable and reliable metric foundation for feature importance assessment.

[0047] The feature importance score sequence is formed by recording the importance score obtained by each original feature in a predetermined number of iterations. Preferably, the number of iterations is set to a sufficiently large value to ensure statistical stability, and the feature importance score in each iteration is calculated based on the average reduction in Gini impurity across all decision trees in that round. The score sequence obtained through multiple iterations constitutes an empirical distribution of feature importance, providing sufficient data support for subsequent statistical analysis. This step constructs a statistical distribution of feature importance through multiple independent evaluations, overcoming the randomness effect of a single evaluation.

[0048] The kernel density estimation algorithm fits a discrete sequence of feature importance scores to a smooth kernel function. By placing a kernel function at each data point and superimposing all kernel functions, a continuous probability density curve is generated. In practice, a Gaussian kernel function can be used, and the smoothness can be controlled by adjusting the bandwidth parameter. The generated density curve can visually display the distribution and central tendency of the feature importance scores, providing a probabilistic basis for statistical inference. This step achieves the transformation from discrete scores to a continuous probability distribution through a non-parametric density estimation method.

[0049] Confidence intervals are calculated based on the feature importance density curve, obtained by determining the boundary values ​​of the tail regions on both sides of the density curve at specific proportions. Hypothesis testing employs a one-tailed test method, comparing the distribution of feature importance scores with the benchmark distribution of random shadow features to test the null hypothesis that the feature's importance is no higher than random noise. The statistical significance of feature importance is determined by calculating the p-value and comparing it with the significance level.

[0050] The dynamic comparison process compares the statistically validated feature importance score distribution with the importance distribution of a random shadow feature set in real time, identifying feature variables whose importance scores consistently exceed the maximum values ​​of the shadow features in most iterations. These features are considered to have stable and significant importance, reliably distinguishing them from random noise. The final feature importance score distribution includes not only numerical information but also statistical significance markers. This step ensures the robustness and reliability of the feature selection results through continuous multi-round comparisons.

[0051] This embodiment achieves precise quantification and rigorous verification of feature importance through a systematic statistical evaluation process. Bootstrap resampling ensures evaluation stability, kernel density estimation provides the probability distribution basis, hypothesis testing guarantees statistical significance, and a dynamic comparison mechanism ensures the rigor of the selection criteria. This embodiment constructs a reliable feature selection system through multi-level statistical verification, providing a solid statistical foundation for feature selection in subsequent prediction models.

[0052] In some embodiments, a stable subset of features is input into a pre-trained LightGBM machine learning model, which performs nonlinear relationship fitting and ensemble decision-making through its built-in multiple decision trees to generate an individualized ATH risk probability value for the target patient, including: A stable subset of features is fed in parallel into multiple decision trees of the LightGBM model. Each decision tree independently judges the input features based on its pre-trained structure and outputs a preliminary prediction score. Based on the predefined weights of each decision tree in the ensemble, the initial prediction scores output by all decision trees are weighted and summed to obtain an ensemble's original prediction score. The integrated raw prediction scores are input into the Sigmoid activation function for probability transformation, mapped to individualized ATH risk probability values ​​in the range of 0 to 1, and then output.

[0053] In this embodiment, a stable subset of features is simultaneously distributed to all decision trees of the LightGBM model for processing. Each decision tree has learned specific feature splitting rules and leaf node values ​​during the pre-training phase. When patient features are input, the tree traverses the path from the root node to the leaf node based on the feature values. The value stored in the final leaf node is the tree's initial predicted score. This parallel processing mechanism fully utilizes the parallel computing capabilities of modern computing hardware, significantly improving model inference efficiency. This step achieves efficient distributed processing of input features through multi-tree parallel inference.

[0054] Predefined weights are determined during model training using a gradient boosting algorithm, reflecting the relative importance of each decision tree in the overall ensemble model. The weighted summation process multiplies the initial prediction scores of each tree by their corresponding weights and then sums them, mathematically representing a linear combination. This ensemble process effectively integrates the different perspectives of multiple decision trees, reducing the prediction bias of individual trees through collective decision-making and improving the model's generalization ability. This step achieves the organic fusion of multi-model predictions through a weighted ensemble mechanism.

[0055] The Sigmoid activation function uses a standard logistic function form to map the integrated raw predicted scores to the 0-1 interval. The specific calculation employs an exponential function to achieve a non-linear transformation, with the output value representing the predicted probability of the target patient developing acute trichomonas (ATH). This probability transformation makes the model output clinically interpretable, facilitating physician understanding and use of the prediction results. Preferably, the raw predicted scores are appropriately scaled before being input into the Sigmoid function to ensure computational stability. This step converts the model output into a standardized risk score that meets clinical needs through probability calibration.

[0056] This embodiment achieves end-to-end prediction from feature input to risk probability output by constructing a complete model inference pipeline. Multi-tree parallel processing ensures computational efficiency, weighted ensemble improves prediction accuracy, and probability transformation enhances the clinical applicability of the results. This embodiment fully leverages the advantages of the LightGBM model in high-dimensional feature processing and nonlinear relationship modeling, providing clinicians with an accurate and interpretable personalized risk assessment tool.

[0057] In some embodiments, the integrated raw predicted score is input into a Sigmoid activation function for probability transformation, mapped to an individualized ATH risk probability value in the interval of 0 to 1, and then output, including: The initial continuous probability value output by the Sigmoid function is calibrated using an ordinal-preserving regression algorithm to match the distribution of ATH incidence observed in real clinical settings, thus obtaining the calibrated probability value. The calibration probability values ​​are mapped to clinically interpretable risk level intervals using a piecewise linear transformation algorithm, and standardized based on the risk level intervals to obtain standardized probability values. The Monte Carlo Dropout algorithm is used to randomly discard some decision paths, generating multiple perturbation estimates of standardized probability values. The variance of these estimates is then used to obtain a quantitative index of the reliability of the standardized probability values. When the standardized probability value and its corresponding reliability quantification index jointly determine that it is near the clinical decision boundary, the ensemble model reconstruction mechanism is activated. By adjusting the ensemble weights of the decision tree in the LightGBM model, the standardized probability value is optimized and corrected, and the final individualized ATH risk probability value is generated and output.

[0058] In this embodiment, the ordination-preserving regression algorithm adjusts the numerical distribution of probabilities while maintaining the original probability order, making the initial continuous probability values ​​output by the model more consistent with the incidence distribution of ATH observed in real clinical data. Specifically, the algorithm sorts the predicted probabilities by magnitude and then performs a smooth adjustment by minimizing the difference between the predicted probabilities and the true frequencies, ensuring that the probability calibration process does not change the relative risk order among samples. This step improves the accuracy of the model's output probabilities through statistical calibration, making them more consistent with actual clinical situations.

[0059] The piecewise linear transformation algorithm maps calibrated probability values ​​to preset clinical risk level ranges, such as low risk, medium risk, and high risk. Each risk level corresponds to a specific probability range, and continuous probability values ​​are transformed to a standardized range using linear interpolation. This transformation gives the probability output clear clinical semantics, making it easier for healthcare professionals to understand and apply the risk assessment results.

[0060] The Monte Carlo Dropout algorithm randomly masks some decision paths during the model inference phase, obtaining the probability estimate distribution through repeated inference. Reliability is quantified by calculating the variance of these perturbation estimates; a smaller variance indicates a more stable and reliable prediction for that sample. This step improves the reliability assessment of the prediction results by introducing uncertainty quantification.

[0061] The ensemble model reconstruction mechanism is automatically triggered when the predicted result is detected to be at the clinical decision boundary. It optimizes the prediction result by readjusting the ensemble weights of each decision tree. In practice, the system dynamically updates the weight allocation based on the predictive performance of recent similar cases, assigning higher weights to decision trees with better predictive performance, thus ensuring the model's predictive accuracy in critical situations. This step improves the model's performance at key decision points through dynamic weight adjustment.

[0062] This embodiment significantly improves the accuracy and reliability of the model output by constructing a multi-level probabilistic optimization process. Probabilistic calibration ensures the consistency between the predicted results and the actual clinical distribution, risk level mapping enhances the interpretability of the results, uncertainty quantification provides a confidence reference for decision-making, and the dynamic reconstruction mechanism guarantees the prediction accuracy under critical conditions. Through systematic post-processing optimization, this embodiment makes the output of the machine learning model more in line with the needs of clinical practice, providing a more reliable risk assessment basis for medical decision-making.

[0063] In some embodiments, an individualized ATH risk probability value is compared with a preset dynamic risk threshold. When the risk probability value exceeds the dynamic risk threshold, a high-risk warning signal is automatically triggered, including: Based on the historical ATH risk probability value distribution of historical patient samples, a dynamic risk threshold matching different clinical scenarios is constructed using quantile statistical algorithms; When the individualized ATH risk probability value is detected to be continuously located in the high-risk quantile range of the dynamic risk threshold, a primary warning signal is generated. A retrospective analysis of clinical characteristics was performed on historical patient samples that triggered primary warning signals, and the co-occurrence pattern of high-risk characteristics was confirmed by logistic regression. Based on the confirmed co-occurrence patterns of high-risk features, a multi-level early warning escalation mechanism is activated to generate high-risk early warning signals that include feature contribution analysis. The high-risk early warning signal is exchanged with the hospital information system in real time, which automatically triggers the priority ranking and resource pre-allocation instructions for the clinical intervention queue.

[0064] In this embodiment, the quantile statistics algorithm calculates thresholds corresponding to different percentiles by analyzing the risk probability distribution of historical patient samples. Specifically, for different clinical scenarios such as surgical type and age grouping, the corresponding risk distribution is calculated separately, and a specific quantile is taken as the dynamic risk threshold for that scenario. This scenario-adaptive threshold setting method can more accurately reflect the risk characteristics of different patient groups, avoiding the applicability problem of a single threshold in different clinical situations. This step achieves personalized dynamic setting of the risk threshold through quantile statistics.

[0065] High-risk quantile intervals are typically defined as the higher quantile range within a dynamic risk threshold, such as the upper quartile. When an individualized ATH risk probability value consistently falls within this interval, it indicates that the patient is in a relatively high-risk state. The generation of the primary warning signal is based on time-series analysis, requiring the risk probability value to remain above the threshold for multiple consecutive time points before triggering, thus avoiding false alarms from single fluctuations. This step improves the reliability of the warning signal through continuous assessment.

[0066] Clinical feature retrospective analysis extracted complete clinical data from historical patient samples that triggered the initial warning, and used a logistic regression model to analyze the association strength between each feature and the warning signal. High-risk feature co-occurrence patterns refer to combinations of multiple high-risk features appearing simultaneously; their stable association with the warning signal was confirmed through regression coefficients and significance tests. This step statistically validated the clinical rationale of the warning signal.

[0067] The multi-level early warning escalation mechanism is activated after confirming the co-occurrence pattern of high-risk features, adjusting the warning level based on the severity and frequency of feature combinations. Feature contribution analysis is achieved by calculating the standardized coefficients of each feature in the logistic regression model, quantifying the relative importance of each feature to the warning signal. The upgraded high-risk warning signal includes specific risk factor analysis, providing targeted guidance for clinical intervention. This step improves the accuracy of early warning through a tiered response mechanism.

[0068] The hospital information system interacts through standardized data interfaces, transmitting key information from high-risk early warning signals to the hospital's patient management system in real time. A priority ranking algorithm, considering factors such as risk probability, feature contribution, and resource availability, generates a queue sequence for clinical intervention. Resource pre-allocation commands automatically allocate appropriate monitoring equipment and nursing resources based on the early warning level, ensuring that high-risk patients receive timely and necessary medical intervention. This step achieves a seamless transition from early warning to intervention through system integration.

[0069] This embodiment achieves closed-loop management from risk identification to clinical intervention by constructing a complete early warning triggering and response process. Dynamic threshold setting ensures the accuracy of early warnings, a multi-level escalation mechanism improves the adaptability of the response, and system integration guarantees the timeliness of intervention. This embodiment transforms the output of the predictive model into actionable clinical decision support, and effectively improves the efficiency of identifying and treating high-risk patients through systematic early warning management.

[0070] In some embodiments, based on the historical ATH risk probability distribution of historical patient samples, a dynamic risk threshold matching different clinical scenarios is constructed using a quantile statistical algorithm, including: Obtain clinical scenario characteristic data from historical patient samples. The clinical scenario characteristic data should include at least the type of surgery, the degree of neurological deficit at admission, and the comorbidity index. For each combination of clinical scenario features, the historical ATH risk probability value distribution of the corresponding historical patient subset is extracted, and the kernel density estimation algorithm is used to construct the risk probability density function under each clinical scenario. Based on the risk probability density function, dynamic risk thresholds for different clinical scenarios are calculated using a quantile regression model. The dynamic risk thresholds include warning thresholds and high-risk thresholds. In addition, when the newly added historical patient data reaches a preset scale, the risk probability density function for each clinical scenario is recalculated, and the dynamic risk threshold is adaptively adjusted.

[0071] In this embodiment, clinical scenario feature data is extracted through a hospital information system. Surgical types include classifications such as craniotomy for hematoma evacuation and minimally invasive surgery; the degree of neurological deficit is quantified using the Glasgow Coma Scale; and the comorbidity index is calculated using the Charlson Comorbidity Index. These features collectively define different combinations of patient clinical states, providing a classification basis for setting scenario-based thresholds. This step achieves refined segmentation of clinical scenarios through multi-dimensional feature combinations.

[0072] The kernel density estimation algorithm uses a smoothed kernel function to estimate the probability density of discrete historical ATH risk probabilities, controlling the smoothing degree by selecting an appropriate bandwidth parameter. For each subset of patients corresponding to each clinical scenario feature combination, a continuous distribution function of their risk probabilities is constructed. This function more accurately reflects the true distribution of risk values ​​in that scenario. This step obtains a complete characterization of the risk distribution for each scenario through nonparametric density estimation.

[0073] Quantile regression models estimate the risk probability density function at different quantiles by minimizing the weighted absolute error loss function. Preferably, the warning threshold is set at a mid-to-high quantile, such as the 75th percentile, and the high-risk threshold is set at a higher quantile, such as the 90th percentile. Quantile regression directly models conditional quantiles, avoiding the limitations of distribution assumptions and making it more suitable for the heterogeneous nature of medical data. This step achieves robust estimation of the risk threshold through quantile regression.

[0074] The preset threshold size is determined based on statistical power analysis and is typically set as the minimum data volume that ensures sufficient sample size for each clinical scenario. When new data reaches this size, the system automatically triggers a threshold update process, recalculating the risk probability density function and corresponding quantile thresholds for all scenarios, ensuring that the thresholds are continuously optimized as clinical practice changes. This step maintains the timeliness of the threshold settings through a periodic update mechanism.

[0075] This embodiment achieves precise and personalized risk warning by constructing a scenario-adaptive dynamic threshold setting system. Stratification based on clinical characteristics ensures the targeted nature of threshold settings, kernel density estimation provides accurate risk distribution modeling, quantile regression guarantees the statistical robustness of the thresholds, and the adaptive update mechanism maintains the system's continuous optimization capability. This scheme effectively solves the problem of poor applicability of traditional fixed thresholds across different patient groups, providing a more scientific and reasonable judgment benchmark for clinical risk warning.

[0076] In some embodiments, in response to a high-risk warning signal, the individualized ATH risk probability value is mapped to the prognostic trajectory of poor long-term neurological function recovery associated with the target patient based on the Kaplan-Meier survival analysis model, generating and outputting a comprehensive prediction report integrating short-term complication risk and long-term functional prognosis, including: A time-series-based database of neural function recovery trajectories was constructed, and functional recovery survival curves under different ATH risk stratifications were calculated using the Kaplan-Meier survival analysis model. Cox proportional hazards regression analysis was used to quantify the impact of ATH risk on long-term prognosis, and a prognostic trajectory map with time dimension prediction was generated. The key time nodes and expected recovery inflection points in the functional recovery process were identified by the time series pattern recognition algorithm. In the prognostic trajectory map, the short-term ATH risk warning, long-term functional recovery trajectory and real-time clinicopathological features are fused into multi-dimensional data to obtain the fusion result; A comprehensive prediction report is obtained based on the fusion results.

[0077] In this embodiment, the neurological function recovery trajectory database is constructed by periodically collecting patients' modified Rankin Scale scores, forming a time-series record of functional status. The Kaplan-Meier survival analysis model uses good functional recovery as the endpoint event, calculating the probability of functional recovery during follow-up for patients in different ATH risk stratifications and plotting corresponding survival curves. This analytical method can intuitively display the differences in functional recovery among different risk groups over time, providing a temporal perspective for prognostic assessment. This step achieves dynamic monitoring of long-term functional outcomes through survival analysis.

[0078] Cox proportional hazards regression analysis quantifies the impact of ATH risk probability values ​​on neurological function recovery by establishing a hazard function model, calculating the hazard ratio and its confidence interval. The prognostic trajectory atlas, based on survival curves and combined with regression analysis results, predicts the future functional recovery path of individual patients. The temporal pattern recognition algorithm identifies time points where significant changes in recovery speed occur by analyzing changes in the derivative of the recovery curve; these inflection points often correspond to important clinical outcome stages. This step achieves quantitative prognostic prediction through regression analysis and pattern recognition.

[0079] Multidimensional data fusion integrates ATH risk warning information, functional recovery trajectory prediction, and real-time clinical characteristics. In practice, a feature-weighted fusion method is used, assigning fusion weights based on the correlation strength between each data source and prognosis to form a unified prognostic feature representation. This ensures the organic connection between short-term risk and long-term prognosis, providing a comprehensive information foundation for integrated assessment. This step constructs a complete prognostic assessment framework through multi-source information fusion.

[0080] The comprehensive predictive report, generated based on the fusion results, includes structured information such as risk stratification, prognostic trends, and early warnings of key time points. The report uses a format easily understood by clinicians, highlighting key decision points and intervention recommendations to provide data support for the development of individualized treatment plans. This step, through standardized report output, effectively translates predictive results into clinical practice.

[0081] This embodiment constructs a complete analytical chain from risk warning to long-term prognostic assessment, realizing the correlation mapping between complication risk and functional outcomes. Survival analysis provides a statistical basis for long-term follow-up, regression models quantify the impact of risk factors, data fusion ensures the comprehensiveness of the assessment, and standardized reporting promotes the clinical application of the results. This embodiment effectively connects short-term complication management with long-term functional prognosis, providing scientific decision support for the comprehensive management of patients with cerebral hemorrhage.

[0082] In some embodiments, a time-series pattern recognition algorithm is used to identify key time points and expected recovery inflection points during the functional recovery process, including: Time-series data of neural function recovery were extracted based on prognostic trajectory maps, and the recovery curves were smoothed using Savitzky-Golay filters to obtain denoised recovery trajectories. The first-order difference calculation is performed on the denoised recovered trajectory to construct the recovered velocity sequence, and the second-order difference calculation is used to construct the recovered acceleration sequence; A change point detection algorithm based on Bayesian inference is used to analyze the recovery acceleration sequence and identify statistical abrupt changes in the recovery trend. Statistical mutation points were mapped to the original timeline and combined with clinically defined milestone events in neurological function recovery to verify the clinical significance of statistical mutation points, which were then denoted as clinically significant mutation points. Based on clinically significant mutation points, the potential inflection points in the subsequent recovery path are predicted using an autoregressive integral moving average model, and key time nodes and expected recovery inflection points are output.

[0083] In this embodiment, the Savitzky-Golay filter smooths the neural function recovery time series through local polynomial fitting, effectively filtering out random noise while preserving the curve's morphological characteristics. This filter uses a fixed-length sliding window, performing polynomial least-squares fitting within the window and replacing the original observations with the fitted values. This maintains the overall trend characteristics of the recovery trajectory, providing clean input data for subsequent analysis. This step improves the quality of the time series data through signal smoothing techniques.

[0084] First-order differencing constructs a recovery velocity sequence by comparing functional scores at adjacent time points, reflecting the rate of functional improvement. Second-order differencing constructs a recovery acceleration sequence by further differencing the velocity sequence, characterizing the trend of recovery velocity changes. This step reveals the dynamic characteristics of the recovery process through differential transformation.

[0085] The change point detection algorithm based on Bayesian inference identifies time points where the statistical properties of the sequence change significantly by calculating the posterior probability distribution of the recovered acceleration sequence. The algorithm assumes that the sequence follows different probability distributions in different segments and estimates the posterior distribution of the change point location using the Markov chain Monte Carlo method, selecting the point with the highest probability density as the candidate change point. This step achieves probabilistic identification of change points through Bayesian statistics.

[0086] Clinically significant mutation points are identified by time-aligning statistical mutation points with clinical milestone events. Milestone events in neurological function recovery include clinically significant rehabilitation milestones such as progressing from bedridden to sitting up, and from assisted walking to independent walking. When a statistical mutation point closely matches the timing of these events, it is identified as a clinically significant mutation point. This step ensures the medical relevance of the identification results through clinical validation.

[0087] The autoregressive integral moving average model builds a time-series prediction model based on historical recovery trajectories, and combines this with identified clinically significant aberrations to predict potential trend inflection points during the future recovery process. The model captures sequence autocorrelation through an autoregressive term, handles random fluctuations through a moving average term, and ensures sequence stationarity through a differencing term, thus generating reliable inflection point predictions. This step achieves a prospective assessment of the recovery process through time-series prediction.

[0088] This embodiment achieves accurate identification of key nodes in the neurological function recovery process by constructing a complete time-series analysis workflow. Signal preprocessing ensures data quality, differential transformation reveals dynamic characteristics, Bayesian change point detection provides statistical basis, clinical validation guarantees medical significance, and time series prediction extends the analytical perspective. This approach organically combines mathematical analysis methods with clinical knowledge, providing a scientific basis for monitoring the rehabilitation process and selecting the timing of interventions.

[0089] In some embodiments, the prognostic trajectory atlas integrates short-term ATH risk warning, long-term functional recovery trajectory, and real-time clinicopathological features to obtain a multi-dimensional data fusion result, including: Using the surgical time as a benchmark, short-term ATH risk warnings, long-term functional recovery trajectories, and real-time clinicopathological features are uniformly mapped onto the same time coordinate system; A feature fusion network based on attention weights is constructed. By calculating the correlation strength between various features and prognostic outcomes at different time points, the contribution weight of each feature in the data fusion process is dynamically allocated. Tensor decomposition is used to perform joint dimensionality reduction on temporal multimodal data, extracting potential prognostic correlation features across time scales and forming a unified multidimensional fusion feature representation; A heterogeneous information graph of patient prognosis is constructed by graph neural network, with clinical events, pathological features and recovery status at different time points as nodes, and the temporal relationship and physiological correlation between them as edges. Based on heterogeneous information graphs, a graph attention network is used to aggregate and update node features, generating a comprehensive prognostic feature representation that simultaneously includes short-term ATH risk warning, long-term functional recovery trajectory, and real-time clinicopathological features as the fusion result.

[0090] In this embodiment, the time coordinate system uses the operation time as the zero point, and arranges the ATH risk warning, functional recovery score, and clinicopathological features at different postoperative time points in chronological order. This unified time benchmark ensures the time alignment of data from different sources, providing a unified time reference framework for subsequent time series analysis and feature fusion. This step solves the time series alignment problem of multi-source data through time standardization.

[0091] The attention-weighted feature fusion network automatically calculates the fusion weight of each feature by learning the correlation between various features and prognostic outcomes at different time points. In practice, the network dynamically adjusts the contribution of each feature in the fusion process by calculating the mutual information or correlation coefficient between features and prognostic labels, giving higher weights to features more strongly associated with the prognosis. This step optimizes the feature fusion effect through adaptive weight allocation.

[0092] Tensor decomposition organizes temporal multimodal data into a three-dimensional tensor structure and extracts latent features using methods such as Tucker decomposition or CP decomposition. This decomposition can simultaneously capture the interaction relationships between features, time points, and modes, generating a low-dimensional fused feature representation that retains the main information of the original data while reducing computational complexity. This step achieves a compact representation of multimodal data through tensor operations.

[0093] Heterogeneous infographics abstract clinical events, pathological features, and recovery states into nodes, constructing a graph structure that reflects the complex relationships between them. Edges between nodes can represent temporal sequence, physiological correlation, or statistical correlation; this graph structure naturally characterizes the complex interaction network between multi-source data. This step achieves the structured integration of multi-source information through graph modeling.

[0094] Graph attention networks distinguish the importance of different neighboring nodes during information propagation by calculating attention coefficients between nodes. During feature aggregation, the features of each node's neighbors are weighted and summed, with the weights dynamically calculated by the attention mechanism, thereby generating new node features that contain global information. This step achieves deep feature fusion through graph neural networks, effectively integrating multi-source features from heterogeneous information graphs.

[0095] This embodiment achieves effective integration of multi-source heterogeneous information by constructing a multi-layered data fusion framework. Time alignment provides the foundation for fusion, the attention mechanism optimizes feature selection, tensor decomposition achieves dimensionality reduction representation, graph structure models complex relationships, and graph attention network completes deep feature fusion, making full use of complementary information from different data sources and providing a comprehensive and accurate feature foundation for prognostic evaluation.

[0096] By adopting the above technical solutions, the present invention differs from the prior art and has the following beneficial effects: The Boruta feature selection algorithm was used to perform high-dimensional feature screening on multidimensional clinical feature sequences, constructing a stable feature subset for ATH risk prediction. This effectively eliminated redundant features while retaining core features significantly correlated with ATH occurrence, overcoming the subjectivity and instability of traditional feature selection methods. A pre-trained LightGBM machine learning model was used to fit nonlinear relationships and perform ensemble decision-making through its built-in multiple decision trees. This accurately captured the complex nonlinear relationships and interaction effects between clinical features, generating individualized ATH risk probability values ​​for target patients and significantly improving prediction accuracy. By comparing individualized ATH risk probability values ​​with preset dynamic risk thresholds and automatically triggering high-risk warning signals, risk grading and timely warnings were achieved, providing clear time points for clinical intervention. Based on the Kaplan-Meier survival analysis model, the individualized ATH risk probability values ​​were mapped to the prognostic trajectory of poor long-term neurological function recovery associated with target patients, generating a comprehensive prediction report integrating short-term complication risk and long-term functional prognosis. This effectively connected complication risk warning and functional outcome prediction, providing comprehensive decision support for individualized treatment and full-process management of patients with cerebral hemorrhage. The above-mentioned technical solution overcomes the limitations of the linear assumptions of traditional prediction methods and achieves accurate prediction across the entire chain from risk identification to prognosis assessment through a systematic machine learning process.

[0097] Finally, it should be noted that although the above embodiments have been described in the text and drawings of this application, this should not limit the scope of patent protection of this application. Any technical solutions that are based on the essential concept of this application and utilize the content described in the text and drawings of this application, resulting in equivalent structural or procedural substitutions or modifications, as well as the direct or indirect application of the technical solutions of the above embodiments to other related technical fields, are all included within the scope of patent protection of this application.

Claims

1. A method for predicting gastrointestinal bleeding after cerebral hemorrhage based on the LGBM model, characterized in that, include: Obtain a multidimensional clinical feature sequence of the target patient, wherein the multidimensional clinical feature sequence includes at least neurological function score, imaging hemorrhage volume, operation duration, serum nutritional markers and blood glucose parameters; Based on historical multidimensional clinical feature sequences and historical postoperative ATH outcome labels of historical patient cohorts, the Boruta feature selection algorithm is used to perform high-dimensional feature screening on the multidimensional clinical feature sequences to construct a stable feature subset for ATH risk prediction. The stable feature subset is input into the pre-trained LightGBM machine learning model, and nonlinear relationship fitting and ensemble decision-making are performed through its built-in multiple decision trees to generate individualized ATH risk probability values ​​for the target patient. The individualized ATH risk probability value is compared with a preset dynamic risk threshold. When the risk probability value exceeds the dynamic risk threshold, a high-risk warning signal is automatically triggered. In response to the high-risk warning signal, the individualized ATH risk probability value is correlated with the prognostic trajectory of poor long-term neurological function recovery associated with the target patient based on the Kaplan-Meier survival analysis model, generating and outputting a comprehensive prediction report integrating short-term complication risk and long-term functional prognosis.

2. The method for predicting postoperative gastrointestinal bleeding based on the LGBM model according to claim 1, characterized in that, Based on historical multidimensional clinical feature sequences and historical postoperative ATH outcome labels from historical patient cohorts, the Boruta feature selection algorithm is used to perform high-dimensional feature screening on the multidimensional clinical feature sequences to construct a stable feature subset for ATH risk prediction, including: Construct an original feature space containing a historical patient cohort, wherein the original feature space is composed of the historical multidimensional clinical feature sequence and the composite physiological indicators derived from the calculation; The Boruta feature selection algorithm is used to perform multiple rounds of random forest importance evaluation in the original feature space, generating a feature importance score distribution for each original feature; Based on the feature importance score distribution, a random shadow feature set with the same dimension as the original feature space is created, and the maximum importance score of each shadow feature is calculated as the benchmark threshold for feature selection. The importance score distribution of each original feature is statistically compared with the benchmark threshold to select a subset of core features whose importance is consistently better than the benchmark threshold. Multicollinearity diagnosis is performed on the core feature subset, and feature variables with redundant information are removed by variance inflation factor analysis to form the stable feature subset.

3. The method for predicting postoperative gastrointestinal bleeding based on the LGBM model according to claim 2, characterized in that, The Boruta feature selection algorithm is used to perform multiple rounds of random forest importance evaluation in the original feature space, generating a feature importance score distribution for each original feature, including: In each evaluation iteration, the Bootstrap resampling algorithm is used to construct a differentiated training subset, and the split contribution of each original feature in the random forest decision tree is calculated by the change in Gini impurity. Perform a preset number of iterative evaluations, record the importance score of each original feature in each iteration, and form a feature importance score sequence based on statistical distribution; The feature importance score sequence is smoothed using a kernel density estimation algorithm to generate a feature importance density curve with probability distribution characteristics; The confidence interval for each original feature is calculated based on the feature importance density curve, and the statistical significance of the feature importance score is verified by hypothesis testing. The statistically validated feature importance score distribution is dynamically compared with the random shadow feature set to identify feature variables that consistently outperform the noise benchmark in multiple iterations, thus obtaining the final feature importance score distribution.

4. The method for predicting postoperative gastrointestinal bleeding based on the LGBM model according to claim 1, characterized in that, The stable feature subset is input into a pre-trained LightGBM machine learning model, which performs nonlinear relationship fitting and ensemble decision-making through its built-in multiple decision trees to generate an individualized ATH risk probability value for the target patient, including: The stable feature subset is input in parallel into multiple decision trees of the LightGBM model. Each decision tree independently judges the input features based on its pre-trained structure and outputs a preliminary prediction score. Based on the predefined weights of each decision tree in the ensemble, the initial prediction scores output by all decision trees are weighted and summed to obtain an ensemble's original prediction score. The integrated raw prediction score is input into the Sigmoid activation function for probability transformation, mapped to an individualized ATH risk probability value in the range of 0 to 1, and then output.

5. The method for predicting postoperative gastrointestinal bleeding based on the LGBM model according to claim 4, characterized in that, The integrated raw predicted score is input into a Sigmoid activation function for probability transformation, mapped to an individualized ATH risk probability value in the interval of 0 to 1, and then output, including: The initial continuous probability value output by the Sigmoid function is calibrated using an ordinal-preserving regression algorithm to match the distribution of ATH incidence observed in real clinical settings, thus obtaining the calibrated probability value. The calibration probability value is mapped to a clinically interpretable risk level interval using a piecewise linear transformation algorithm, and the calibration probability value is standardized based on the risk level interval to obtain a standardized probability value. The Monte Carlo Dropout algorithm is used to randomly discard a portion of the decision path, generating multiple perturbation estimates of the standardized probability value. The variance of these estimates is then used to obtain a quantitative index of the reliability of the standardized probability value. When the standardized probability value and its corresponding reliability quantification index jointly determine that it is near the clinical decision boundary, the ensemble model reconstruction mechanism is activated. By adjusting the ensemble weights of the decision tree in the LightGBM model, the standardized probability value is optimized and corrected, and the final individualized ATH risk probability value is generated and output.

6. The method for predicting postoperative gastrointestinal bleeding based on the LGBM model according to claim 1, characterized in that, The individualized ATH risk probability value is compared with a preset dynamic risk threshold. When the risk probability value exceeds the dynamic risk threshold, a high-risk warning signal is automatically triggered, including: Based on the historical ATH risk probability value distribution of historical patient samples, a dynamic risk threshold matching different clinical scenarios is constructed using quantile statistical algorithms; When the individualized ATH risk probability value is detected to be continuously located in the high-risk quantile range of the dynamic risk threshold, a primary warning signal is generated; A retrospective analysis of clinical characteristics was performed on historical patient samples that triggered primary warning signals, and the co-occurrence pattern of high-risk characteristics was confirmed by logistic regression. Based on the confirmed co-occurrence patterns of high-risk features, a multi-level early warning escalation mechanism is activated to generate high-risk early warning signals that include feature contribution analysis. The high-risk early warning signal is exchanged with the hospital information system in real time to automatically trigger the priority sorting and resource pre-allocation instructions for the clinical intervention queue.

7. The method for predicting postoperative gastrointestinal bleeding based on the LGBM model according to claim 6, characterized in that, Based on the historical ATH risk probability distribution of historical patient samples, dynamic risk thresholds matching different clinical scenarios are constructed using quantile statistical algorithms, including: Acquire clinical scenario feature data of historical patient samples, wherein the clinical scenario feature data includes at least the type of surgery, the degree of neurological deficit at admission, and the comorbidity index; For each combination of clinical scenario features, the historical ATH risk probability value distribution of the corresponding historical patient subset is extracted, and the kernel density estimation algorithm is used to construct the risk probability density function under each clinical scenario. Based on the risk probability density function, dynamic risk thresholds for different clinical scenarios are calculated using a quantile regression model. The dynamic risk thresholds include warning thresholds and high-risk thresholds. Furthermore, when the newly added historical patient data reaches a preset scale, the risk probability density function for each clinical scenario is recalculated, and the dynamic risk threshold is adaptively adjusted.

8. The method for predicting postoperative gastrointestinal bleeding based on the LGBM model according to claim 1, characterized in that, In response to the high-risk warning signal, based on the Kaplan-Meier survival analysis model, the individualized ATH risk probability value is correlated and mapped with the prognostic trajectory of poor long-term neurological function recovery associated with the target patient. A comprehensive prediction report integrating short-term complication risk and long-term functional prognosis is generated and output, including: A time-series-based database of neural function recovery trajectories was constructed, and functional recovery survival curves under different ATH risk stratifications were calculated using the Kaplan-Meier survival analysis model. Cox proportional hazards regression analysis was used to quantify the impact of ATH risk on long-term prognosis, and a prognostic trajectory map with time dimension prediction was generated. The key time nodes and expected recovery inflection points in the functional recovery process were identified by the time series pattern recognition algorithm. In the prognostic trajectory map, the short-term ATH risk warning, long-term functional recovery trajectory and real-time clinicopathological features are fused into multi-dimensional data to obtain the fusion result; The comprehensive prediction report is obtained based on the fusion results.

9. The method for predicting postoperative gastrointestinal bleeding based on the LGBM model according to claim 8, characterized in that, The time-series pattern recognition algorithm identifies key time points and expected recovery inflection points during the functional recovery process, including: Based on the prognostic trajectory map, time-series data of neural function recovery are extracted, and the recovery curve is smoothed by Savitzky-Golay filter to obtain the denoised recovery trajectory. The first-order difference calculation is performed on the denoised recovered trajectory to construct the recovered velocity sequence, and the second-order difference calculation is used to construct the recovered acceleration sequence; A change point detection algorithm based on Bayesian inference is used to analyze the recovery acceleration sequence and identify statistical abrupt changes in the recovery trend. The statistical mutation points were mapped to the original timeline and combined with clinically defined milestone events in neurological function recovery to verify their clinical significance, and were denoted as clinically significant mutation points. Based on clinically significant mutation points, the potential inflection points in the subsequent recovery path are predicted using an autoregressive integral moving average model, and key time nodes and expected recovery inflection points are output.

10. The method for predicting postoperative gastrointestinal bleeding based on the LGBM model according to claim 8, characterized in that, In the prognostic trajectory mapping, short-term ATH risk warning, long-term functional recovery trajectory, and real-time clinicopathological features are fused into multidimensional data to obtain the fusion results, including: Using the surgical time as a reference point, the short-term ATH risk warning, long-term functional recovery trajectory and real-time clinicopathological characteristics are uniformly mapped to the same time coordinate system; A feature fusion network based on attention weights is constructed. By calculating the correlation strength between various features and prognostic outcomes at different time points, the contribution weight of each feature in the data fusion process is dynamically allocated. Tensor decomposition is used to perform joint dimensionality reduction on temporal multimodal data, extracting potential prognostic correlation features across time scales and forming a unified multidimensional fusion feature representation; A heterogeneous information graph of patient prognosis is constructed by graph neural network, with clinical events, pathological features and recovery status at different time points as nodes, and the temporal relationship and physiological correlation between them as edges. Based on the heterogeneous information graph, a graph attention network is used to aggregate and update node features, generating a comprehensive prognostic feature representation that simultaneously includes short-term ATH risk warning, long-term functional recovery trajectory, and real-time clinicopathological features as the fusion result.

Images