A method for constructing an early warning assessment model for acute respiratory distress syndrome

Abstract

The present invention relates to the field of early warning and evaluation technology for acute respiratory distress syndrome, and in particular to a method for constructing an early warning and evaluation model for acute respiratory distress syndrome. The method comprises the following steps: performing time series feature screening on a multidimensional heterogeneous database to obtain a preferred time series feature set; quantifying the severity of symptoms on the multidimensional heterogeneous database to obtain quantitative symptom feature data; performing symptom mutation point time series analysis on the quantitative symptom feature data to obtain symptom mutation point data; performing symptom interaction analysis on the quantitative symptom feature data to obtain symptom influence data; integrating symptom dynamic features based on symptom influence data to obtain dynamic symptom features; performing symptom time series feature fusion on dynamic symptom features and a preferred time series feature set to obtain a dynamic symptom feature set. The present invention solves the limitations of traditional models by constructing a hierarchical risk prediction model, and improves the accuracy and reliability of early warning of ARDS.

Description

Technical Field

[0001] The present invention relates to the technical field of early warning assessment of acute respiratory distress syndrome, and in particular to a method for constructing an early warning assessment model for acute respiratory distress syndrome. Background Art

[0002] Acute respiratory distress syndrome (ARDS) is a life-threatening respiratory failure with a high mortality rate. Early identification and intervention are crucial to improving the prognosis of ARDS patients. Therefore, building an effective ARDS early warning assessment model has always been a hot topic and difficulty in clinical research. However, a traditional method for building an early warning assessment model for acute respiratory distress syndrome (ARDS) has the disadvantages of insufficient data utilization, single feature extraction, and limited risk prediction accuracy.

[0003] Manifestations of insufficient data utilization: Traditional methods mainly rely on static indicators at a single time point or within a limited time period, such as laboratory test results and vital signs at admission, but ignore the dynamic changes in the patient's physiological indicators during the development of ARDS, such as the rate of decline of the oxygenation index and the dynamic changes in lung imaging.

[0004] Single manifestation of feature extraction: The feature selection of traditional methods mainly relies on expert experience, which is subjective and limited, and it is difficult to discover the complex patterns and associations hidden in the data. Traditional methods usually use simple statistical methods for feature extraction, such as univariate analysis, multivariate analysis, etc., which are difficult to capture the interactive relationship between multi-dimensional and multi-scale features under the complex pathophysiological mechanism of ARDS, such as the joint effect of inflammatory indicators and oxygenation indicators, and the interactive effect of gene polymorphism and mechanical ventilation.

[0005] Manifestations of limited risk prediction accuracy: Traditional methods mostly use a single model for risk prediction, lacking comprehensive consideration of risk factors at different time scales and biological scales, such as the combined effects of short-term risk factors (such as infection, surgery) and long-term risk factors (such as chronic diseases, genetic factors), resulting in limited prediction accuracy. Summary of the invention

[0006] Based on this, it is necessary to provide a method for constructing an early warning assessment model for acute respiratory distress syndrome to solve at least one of the above technical problems.

[0007] To achieve the above purpose, a method for constructing an early warning assessment model for acute respiratory distress syndrome comprises the following steps:

[0008] Step S1: Acquire ARDS early warning indicator data; collect multidimensional data according to the ARDS early warning indicator data through a medical information system to obtain an original multidimensional data set; construct a multidimensional heterogeneous database according to the original multidimensional data set to obtain a multidimensional heterogeneous database;

[0009] Step S2: Screen the time series features of the multidimensional heterogeneous database to obtain an optimal time series feature set; quantify the severity of symptoms in the multidimensional heterogeneous database to obtain quantitative symptom feature data; smooth the symptom time series of the quantitative symptom feature data to obtain a symptom time smoothed sequence; perform symptom mutation point time series analysis on the quantitative symptom feature data to obtain symptom mutation point data; perform symptom interaction analysis on the quantitative symptom feature data based on the symptom time smoothed sequence to obtain symptom influence data; integrate symptom dynamic features based on the symptom influence data to obtain dynamic symptom features; perform symptom time series feature fusion on the dynamic symptom features and the optimal time series feature set to obtain a dynamic symptom feature set;

[0010] Step S3: extract multi-scale interactive features according to the dynamic symptom feature set to obtain a multi-scale interactive feature set; perform time scale feature fusion and biological scale feature fusion according to the dynamic symptom feature set and the multi-scale interactive feature set to obtain a time scale interactive feature matrix and a biological scale interactive feature matrix; perform multi-scale interactive feature matrix integration on the time scale interactive feature matrix and the biological scale interactive feature matrix to obtain a multi-scale interactive feature matrix;

[0011] Step S4: performing hierarchical risk prediction modeling according to the multi-scale interactive feature matrix and the dynamic symptom feature set to obtain a hierarchical risk prediction model;

[0012] Step S5: Integrate the hierarchical risk prediction model into the medical information system, perform ARDS risk prediction on the target patient, and obtain the patient's ARDS risk data; perform early warning evaluation and early warning visualization based on the patient's ARDS risk data to obtain ARDS early warning visualization data.

[0013] The present invention can efficiently and comprehensively obtain high-quality ARDS early warning indicator data by integrating literature mining and medical information systems, and build a structured, standardized multi-dimensional heterogeneous database, laying a solid data foundation for subsequent feature engineering and model construction, overcoming the defects of scattered data and uneven quality of traditional methods. By deeply mining and analyzing multi-dimensional heterogeneous data, a dynamic symptom feature set containing rich information is extracted, such as time series characteristics, symptom severity, change trends, mutation points, interactive effects, etc., which can more comprehensively and accurately characterize the dynamic evolution of patient symptoms and provide more accurate feature expressions for ARDS risk prediction. By constructing multi-scale interactive features, the nonlinear interactive relationship between different features can be captured, and feature fusion is performed from the two dimensions of time and biology to construct a multi-scale interactive feature matrix, thereby more comprehensively and deeply describing the potential mechanism of the occurrence and development of ARDS and improving the prediction accuracy and interpretability of the model. By constructing a hierarchical risk prediction model system and adopting scientific model training, evaluation and optimization strategies, the advantages of different models, such as basic models, dynamic models, multi-scale models, etc., can be effectively integrated to build an integrated model with more powerful performance and more accurate prediction, thereby improving the accuracy and reliability of ARDS risk prediction. By integrating the model into the medical information system and realizing functions such as risk prediction, classification, evaluation and visualization, it can provide clinicians with timely, accurate and intuitive decision support, help doctors identify high-risk patients as early as possible, and take effective intervention measures, thereby improving the accuracy of early warning of ARDS and reducing morbidity and mortality. Therefore, the present invention provides a method for constructing an early warning assessment model for acute respiratory distress syndrome. It effectively solves the limitations of traditional ARDS early warning assessment models in data utilization, feature extraction and risk prediction. By making full use of time series information and multi-source heterogeneous data, as well as extracting dynamic symptom features and multi-scale interactive features, a hierarchical risk prediction model is constructed to improve the accuracy and reliability of early warning of ARDS.

[0014] Preferably, step S1 comprises the following steps:

[0015] Step S11: obtaining ARDS early warning indicator data through literature mining technology; constructing a candidate indicator pool according to the ARDS early warning indicator data to obtain a candidate indicator pool;

[0016] Step S12: Performing data quality assessment on the candidate indicator pool and screening the candidate indicators through the medical information system to obtain a preliminary indicator set;

[0017] Step S13: performing multidimensional data collection on the medical information system according to the preliminary selected indicator set to obtain an original multidimensional data set;

[0018] Step S14: performing an indicator correlation analysis based on the original multidimensional data set to obtain an optimal indicator set;

[0019] Step S15: integrating multidimensional heterogeneous data of the preferred indicator set and constructing a multidimensional heterogeneous database to obtain a multidimensional heterogeneous database.

[0020] The present invention can quickly and comprehensively obtain relevant research results of ARDS early warning indicators through literature mining technology, and extract potential warning indicators from them to build a candidate indicator pool. This effectively avoids the subjectivity and limitations of indicator screening based on experience or limited data alone, and lays a solid foundation for subsequent indicator screening and model construction. By conducting data quality assessment with the medical information system, candidate indicators with serious data missing and poor data quality can be eliminated to ensure that the indicators finally included in the model have a reliable data basis. This effectively avoids the decline in model performance or deviation in prediction results due to data quality problems, and improves the reliability and stability of the model. According to the preliminary indicator set, multidimensional data is collected from the medical information system in a targeted manner, which can avoid information redundancy and improve data collection efficiency. At the same time, the collection of multidimensional data can provide more extensive patient information, provide richer materials for subsequent feature extraction and model construction, and help improve the prediction performance of the model. By performing indicator correlation analysis on the original multidimensional data set, indicators closely related to the occurrence and development of ARDS can be screened out to build a preferred indicator set. This effectively reduces the data dimension, reduces the model calculation complexity, and avoids the interference of irrelevant indicators on the model, further improving the prediction efficiency and accuracy of the model. Integrating multi-dimensional heterogeneous data of the optimal indicator set and building a multi-dimensional heterogeneous database can break data silos and achieve centralized management and efficient use of data. This provides a unified and standardized data platform for subsequent model training, verification and application, which is conducive to iterative optimization and promotion of the model.

[0021] Preferably, step S2 comprises the following steps:

[0022] Step S21: screening the time series features of the multi-dimensional heterogeneous database to obtain an optimal time series feature set; quantifying the severity of symptoms of the multi-dimensional heterogeneous database to obtain quantitative symptom feature data;

[0023] Step S22: performing symptom time series smoothing on the quantitative symptom feature data to obtain a symptom time smoothing sequence; performing symptom mutation point time series analysis on the quantitative symptom feature data to obtain symptom mutation point data;

[0024] Step S23: performing symptom trend fitting on the symptom time smoothing sequence to obtain a symptom trend function; performing symptom trend deviation calculation based on the symptom trend function and the symptom time smoothing sequence to obtain a symptom trend deviation time series;

[0025] Step S24: dividing the symptom evolution stages according to the symptom mutation point data and the symptom trend deviation time series to obtain symptom evolution stage data; extracting symptom evolution stage features from the symptom evolution stage data to obtain symptom evolution stage features;

[0026] Step S25: performing symptom interaction analysis on the quantitative symptom feature data according to the symptom time smoothing sequence to obtain symptom influence data; integrating symptom dynamic features based on the symptom evolution stage characteristics and the symptom influence data to obtain dynamic symptom features;

[0027] Step S26: Perform symptom-time series feature fusion on the dynamic symptom features and the preferred time series feature set to obtain a dynamic symptom feature set.

[0028] The present invention can extract the key time series features that best reflect the development trend of ARDS, such as a continuous increase in heart rate or a continuous decrease in blood oxygen saturation, through time series feature screening, thereby improving the sensitivity and specificity of the model for early warning of ARDS. At the same time, by quantifying the severity of symptoms, the text description information that was originally difficult to use directly for model training can be converted into quantifiable numerical features, providing data support for building a more accurate prediction model. Time series smoothing of quantified symptom feature data can effectively reduce the interference of noise and outliers on subsequent analysis, making the extracted features more stable and reliable. At the same time, through the time series analysis of symptom mutation points, the key time points when the patient's condition suddenly worsens can be captured, providing more timely and accurate signals for early warning of ARDS. By trend fitting the symptom time series, the overall development trend of the patient's symptoms can be more intuitively understood, and it can be judged whether the condition is improving or worsening. By calculating the symptom trend deviation, the time period when the patient's symptoms deviate from the expected trend can be further identified, providing more fine-grained reference information for early warning of ARDS. According to the symptom mutation point and trend deviation, the patient's ARDS development process can be divided into different stages, such as stable period, exacerbation period, recovery period, etc. This stage division can more clearly characterize the development law of ARDS and extract more targeted features for each stage, thereby improving the model's ability to predict the risk of ARDS at different stages. By performing interactive impact analysis on quantitative symptom feature data, the relationship between different symptoms can be identified, for example, whether the aggravation of one symptom will lead to the appearance or deterioration of another symptom. Calculating symptom influence data can help us understand which symptoms play a more critical role in the occurrence and development of ARDS, thereby providing more accurate targets for clinical intervention. Fusion of dynamic symptom features with preferred time series features can effectively integrate the patient's symptom information and physiological indicator information to construct a more comprehensive feature set. This multi-source information fusion can help the model capture the patient's condition changes more comprehensively, thereby improving the accuracy and reliability of ARDS risk prediction.

[0029] Preferably, step S21 includes the following steps:

[0030] Step S211: extracting time series data from a multidimensional heterogeneous database to obtain a multidimensional time series data set; extracting time series features from the multidimensional time series data set to obtain a time series feature library; and mining deep time series features from the time series feature library to obtain a deep time series feature set;

[0031] Step S212: evaluating the feature prediction capability and discrimination of the time series feature library and the deep time series feature set and screening the time series features to obtain a preferred time series feature set;

[0032] Step S213: Structuring the electronic medical records of the multi-dimensional heterogeneous database to obtain structured symptom data;

[0033] Step S214: mapping the structured symptom data to a preset ARDS-related symptom standard dictionary, standardizing the symptom representation, and obtaining a standardized symptom library;

[0034] Step S215: Acquire clinical evaluation standard data; quantify the severity of symptoms in the standardized symptom library according to the clinical evaluation standard data to obtain quantitative symptom characteristic data.

[0035] The present invention can more comprehensively capture the laws of changes in patients' physiological indicators over time, such as trends, cycles, fluctuations, etc., by extracting time series data from a multi-dimensional heterogeneous database and performing multi-level time series feature extraction. In particular, deep time series feature mining can utilize the powerful feature expression capabilities of deep learning models to extract complex time patterns that are difficult to capture with traditional methods, providing richer and more accurate information for early warning of ARDS. By evaluating the predictive ability and discrimination of time series features and screening them, redundant or irrelevant features can be removed, and key features with high contribution to ARDS prediction can be retained, thereby simplifying the model structure, improving the generalization ability and prediction efficiency of the model, and avoiding the occurrence of overfitting. Structuring electronic medical records can convert unstructured text information into structured data forms, such as extracting key information such as patients' symptoms, signs, and test results, which is convenient for subsequent analysis and utilization, and provides the necessary data foundation for building an ARDS early warning model based on multi-source heterogeneous data. Mapping structured symptom data to a standard dictionary can unify symptom information from different sources and in different ways of expression. For example, "dyspnea", "shortness of breath", "gasping for breath" and so on can be normalized under the standard term "respiratory distress", thereby eliminating data inconsistencies and improving data comparability and the reliability of analysis results. Quantifying the standardized symptom library according to clinical evaluation criteria can convert subjective symptom descriptions into objective numerical indicators. For example, assigning a value of 1 to "mild cough", 2 to "moderate cough", and 3 to "severe cough" can facilitate quantitative analysis and model training, and improve the model's prediction accuracy for ARDS risk.

[0036] Preferably, step S22 comprises the following steps:

[0037] Step S221: smoothing the symptom time series of the quantitative symptom feature data to obtain a symptom time smoothed series;

[0038] Step S222: Calculate the symptom change rate for the symptom time smoothing sequence to obtain a symptom change rate time series;

[0039] Step S223: Calculate the symptom change acceleration for the symptom change rate time series to obtain the symptom change acceleration time series;

[0040] Step S224: Perform symptom mutation point analysis based on the symptom change rate time series and the symptom change acceleration time series to obtain symptom mutation point data.

[0041] The present invention can effectively filter out random fluctuations and noise in the data by smoothing the quantitative symptom feature data in time series, making the trend of symptom changes smoother and easier to identify, thereby improving the reliability and stability of subsequent analysis results, avoiding the model from being disturbed by noise, and making false warnings. Calculating the symptom change rate can more sensitively capture subtle changes in the patient's symptoms, such as whether the symptom worsens slowly or rapidly, which is of great significance for judging the development trend of the disease and warning the occurrence of ARDS, and can help the model identify the signal of rising risk earlier. Further calculating the acceleration of symptom change can more accurately identify the turning point of the symptom change trend, such as from slow deterioration to rapid deterioration, which has important reference value for predicting the occurrence time and severity of ARDS, and can help the model more accurately predict the turning point of risk change. Combining the symptom change rate and the acceleration of change to analyze the symptom mutation point, it is possible to more accurately identify the time point when the patient's condition changes significantly, such as sudden deterioration or improvement of symptoms, which is crucial for taking timely intervention measures to avoid further deterioration of the condition, and can provide clinicians with clearer warning signals.

[0042] Preferably, step S25 comprises the following steps:

[0043] Step S251: Calculate the time dimension correlation coefficient of the quantitative symptom feature data according to the symptom time smoothing sequence and construct a symptom interaction relationship matrix to obtain the symptom interaction relationship matrix;

[0044] Step S252: constructing a symptom impact network according to the symptom interaction relationship matrix to obtain symptom impact network data;

[0045] Step S253: Calculate the symptom influence according to the symptom influence network data to obtain symptom influence data;

[0046] Step S254: Integrate the dynamic characteristics of the symptoms based on the symptom evolution stage characteristics and the symptom influence data to obtain dynamic symptom characteristics.

[0047] By calculating the correlation coefficient between the symptom time smoothing series, the present invention can quantify the degree of association between different symptoms over time, such as whether certain symptoms tend to appear at the same time or have a consistent change trend, thereby constructing a symptom interaction matrix, providing a data basis for revealing the complex symptom association mechanism in the development of ARDS. Converting the symptom interaction matrix into a symptom influence network can more intuitively display the mutual influence relationship between different symptoms, such as which symptoms are key nodes and which symptoms are easily affected by other symptoms, thereby helping doctors better understand the pathophysiological mechanism of ARDS. By analyzing the symptom influence network, the influence of each symptom can be calculated, such as whether changes in certain symptoms will cause a chain reaction of multiple other symptoms, thereby identifying the core symptoms that play a key role in the development of ARDS and providing a clearer target for early warning. By integrating the symptom evolution stage characteristics with the symptom influence data, a more comprehensive dynamic symptom feature can be constructed. For example, the core symptoms at different stages and their influence change trends can be used as the input features of the model, thereby more comprehensively describing the dynamic process of the development of ARDS and improving the prediction accuracy and personalized warning capabilities of the model.

[0048] Preferably, step S3 comprises the following steps:

[0049] Step S31: constructing a feature interaction matrix according to the dynamic symptom feature set to obtain an original interaction feature matrix; performing a significance analysis on the original interaction feature matrix to obtain a significance interaction feature matrix;

[0050] Step S32: constructing an interactive feature network according to the significant interactive feature matrix to obtain interactive feature network data; extracting multi-scale interactive features according to the interactive feature network data to obtain a multi-scale interactive feature set;

[0051] Step S33: performing time scale feature fusion on the multi-scale interaction feature set according to the dynamic symptom feature set to obtain a time scale interaction feature matrix; performing biological scale feature fusion on the multi-scale interaction feature set according to the dynamic symptom feature set to obtain a biological scale interaction feature matrix;

[0052] Step S34: performing multi-scale interaction feature matrix integration on the time-scale interaction feature matrix and the biological-scale interaction feature matrix to obtain a multi-scale interaction feature matrix.

[0053] By constructing a feature interaction matrix, the present invention can explore the nonlinear interaction relationship between dynamic symptom features, for example, when certain symptoms are combined together, they will have a synergistic or inhibitory effect on the risk of ARDS. The significance analysis can help us screen out those statistically significant interaction feature pairs, eliminate the interference of random factors, and provide a reliable basis for building a more accurate prediction model. Converting the significant interaction feature matrix into an interactive feature network can more intuitively display the interaction relationship between features, and use network analysis methods to identify feature modules or subnetworks with important biological significance. Multi-scale interactive feature extraction can capture feature interaction information from different granularities (such as local feature combinations, global feature patterns, etc.), construct a more comprehensive feature representation, and further improve the prediction ability of the model. Performing feature fusion at the time scale and biological scale separately can link feature interaction information with the time dimension and biological mechanism, for example, some feature interaction patterns only appear in a specific time period, or are related to specific biological pathways or disease processes. This multi-angle feature fusion can help us more comprehensively understand the relationship between feature interaction and the occurrence and development of ARDS, and improve the interpretability and prediction accuracy of the model. By integrating the time-scale interaction feature matrix and the biological-scale interaction feature matrix, we can construct the final multi-scale interaction feature matrix, which comprehensively considers the feature interaction intensity, temporal correlation and biological relevance, and can more comprehensively and accurately characterize the potential mechanism of the occurrence and development of ARDS, providing more powerful feature support for building a high-performance ARDS risk prediction model.

[0054] Preferably, step S33 includes the following steps:

[0055] Step S331: extracting time window features from the dynamic symptom feature set using the preset time window size data to obtain a time window feature data set; constructing a time scale feature matrix for the multi-scale interaction feature set based on the time window feature data set to obtain a time scale feature matrix;

[0056] Step S332: grouping the dynamic symptom feature set into symptom groups to obtain symptom group data; aggregating symptom features within the group on the symptom group data to obtain symptom group features;

[0057] Step S333: performing interactive feature grouping on the multi-scale interactive feature set to obtain interactive feature grouping data; performing intra-group interactive feature aggregation on the interactive feature grouping data to obtain interactive feature group features;

[0058] Step S334: construct a biological scale interactive feature matrix for the symptom group features and the interactive feature group features to obtain a biological scale interactive feature matrix.

[0059] By setting a time window, the present invention can capture the changing trends and patterns of dynamic symptom features in different time periods, such as certain symptom combinations that only appear in the early or late stages of the disease. Constructing a time-scale feature matrix can incorporate this time information into the feature representation, enabling the model to more accurately identify key time windows and feature evolution patterns that are closely related to the occurrence and development of ARDS. By grouping symptoms according to their clinical manifestations or physiological mechanisms, symptoms with similar characteristics can be aggregated together, such as grouping symptoms related to respiratory function into one group and symptoms related to inflammatory response into another group. This grouping strategy can reduce feature dimensions, reduce noise interference, and extract more representative symptom group features, thereby improving the robustness and generalization ability of the model. Grouping multi-scale interactive features can aggregate interactive features with similar biological significance or functions, such as grouping interactive features reflecting lung inflammation into one group and interactive features reflecting oxygenation dysfunction into another group. This grouping strategy can help us better understand the impact of different biological processes on the occurrence and development of ARDS and extract more explanatory interactive feature group features. By constructing a biological scale interaction feature matrix, we can analyze the degree of association between different symptom groups and different interaction feature groups, such as whether the symptom group related to respiratory function is closely related to the interaction feature group related to lung inflammation. This analysis can help us reveal the interaction relationship between different biological processes in the development of ARDS.

[0060] Preferably, step S4 comprises the following steps:

[0061] Step S41: risk labeling the multi-scale interaction feature matrix through the medical information system to obtain a labeled multi-scale interaction feature matrix; dividing the labeled multi-scale interaction feature matrix into a data set to obtain a training set, a validation set, and a test set;

[0062] Step S42: constructing a basic risk prediction model based on the training set and evaluating the model using the validation set to obtain a basic risk prediction model;

[0063] Step S43: constructing a dynamic risk prediction model according to the dynamic symptom feature set and the basic risk prediction model to obtain a dynamic risk prediction model;

[0064] Step S44: constructing a multi-scale risk prediction model according to the multi-scale interactive feature matrix and the dynamic risk prediction model to obtain a multi-scale risk prediction model;

[0065] Step S45: performing hierarchical risk prediction model integration on the basic risk prediction model, the dynamic risk prediction model and the multi-scale risk prediction model to obtain an initial hierarchical risk prediction model;

[0066] Step S46: Use the test set to verify the initial hierarchical risk prediction model to obtain a hierarchical risk prediction model.

[0067] The present invention uses a medical information system to perform risk annotation on a multi-scale interactive feature matrix, that is, to correspond the feature data of each patient to the actual situation of whether ARDS will eventually occur, and to construct a labeled annotated data set. By dividing the data set into a training set, a validation set, and a test set, the overfitting problem in the model training process can be effectively avoided, and an independent data set can be provided for model evaluation and performance comparison to ensure the generalization ability and prediction effect of the model. By constructing a basic risk prediction model based on the training set, the correlation between the features and the ARDS risk can be preliminarily learned, and the validation set can be used for model evaluation and parameter adjustment, such as selecting a suitable model structure, learning rate, etc., so as to obtain a basic model with relatively good performance, providing a baseline for subsequent model optimization. On the basis of the basic risk prediction model, a dynamic risk prediction model is constructed in combination with a dynamic symptom feature set, which can more comprehensively capture the law of changes in patient symptoms over time and their impact on the risk of ARDS. For example, the continued deterioration or combined occurrence of certain symptoms is an important early warning sign for the occurrence of ARDS. Further combining the multiscale interactive feature matrix to construct a multiscale risk prediction model can further explore the nonlinear interactive relationship between features and their impact on ARDS risk, such as the synergistic or inhibitory effects of certain features when combined. Hierarchical integration of basic models, dynamic models and multiscale models can comprehensively utilize the advantages of different models. For example, the basic model can provide a stable baseline prediction, the dynamic model can capture time information, and the multiscale model can mine feature interaction information, thereby building an integrated model with more powerful performance and more accurate prediction. Using an independent test set to verify the final hierarchical risk prediction model can objectively evaluate the prediction performance of the model in real scenarios, such as calculating the accuracy, sensitivity, specificity and other indicators of the model, and further optimize and adjust the model based on the evaluation results to ensure that the model can achieve the expected results in practical applications.

[0068] Preferably, step S5 comprises the following steps:

[0069] Step S51: Integrate the hierarchical risk prediction model into the medical information system, extract patient features of the target patient, and obtain a patient feature set;

[0070] Step S52: inputting the patient feature set into the hierarchical risk prediction model to perform ARDS risk prediction and obtain the patient's ARDS risk data;

[0071] Step S53: performing risk grading processing on the patient's ARDS risk data using a preset risk threshold to obtain patient risk grade data;

[0072] Step S54: Perform early warning assessment according to the patient's risk level data to obtain the patient's ARDS early warning assessment data;

[0073] Step S55: Perform early warning visualization on the patient's ARDS risk data and the patient's ARDS early warning assessment data to obtain ARDS early warning visualization data.

[0074] The present invention integrates the hierarchical risk prediction model into the medical information system, so as to realize the seamless connection between the model and clinical data, automatically obtain the real-time data of the patient, and perform feature extraction, so as to provide timely and accurate input information for the model, avoid delays and errors caused by manual operation, and improve the practicality and efficiency of the model. By inputting the patient feature set into the trained hierarchical risk prediction model, the risk probability of the patient developing ARDS in the future can be calculated in real time, and an objective and quantitative risk assessment result can be provided to the clinician. By using the preset risk threshold to classify the patient's risk, the complex risk probability value can be converted into an intuitive and easy-to-understand risk level, such as high risk, medium risk, and low risk, so that the doctor can quickly understand the patient's risk level. According to the patient's risk level data, the early warning assessment can be combined with the patient's individual characteristics and the actual clinical situation, such as age and underlying diseases, to conduct a more comprehensive and accurate risk assessment, and issue early warning information in time to remind doctors to pay attention to high-risk patients and take preventive measures as soon as possible. Visualizing the patient's ARDS risk data and early warning assessment data, such as using charts and curves, can more intuitively and clearly present the patient's risk level and changing trend, helping doctors better understand the model prediction results, communicate effectively with patients, and improve patients' understanding and compliance with the disease. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 A schematic diagram of the steps of a method for constructing an early warning assessment model for acute respiratory distress syndrome;

[0076] Figure 2 for Figure 1 Detailed implementation steps of step S2 in the flowchart;

[0077] Figure 3 for Figure 1 Detailed implementation steps of step S3 in FIG.

[0078] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings in conjunction with the embodiments. DETAILED DESCRIPTION

[0079] The technical method of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0080] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0081] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0082] To achieve this, please refer to Figures 1 to 3 A method for constructing an early warning assessment model for acute respiratory distress syndrome comprises the following steps:

[0083] Step S1: Acquire ARDS early warning indicator data; collect multidimensional data according to the ARDS early warning indicator data through a medical information system to obtain an original multidimensional data set; construct a multidimensional heterogeneous database according to the original multidimensional data set to obtain a multidimensional heterogeneous database;

[0084] Step S2: Screen the time series features of the multidimensional heterogeneous database to obtain an optimal time series feature set; quantify the severity of symptoms in the multidimensional heterogeneous database to obtain quantitative symptom feature data; smooth the symptom time series of the quantitative symptom feature data to obtain a symptom time smoothed sequence; perform symptom mutation point time series analysis on the quantitative symptom feature data to obtain symptom mutation point data; perform symptom interaction analysis on the quantitative symptom feature data based on the symptom time smoothed sequence to obtain symptom influence data; integrate symptom dynamic features based on the symptom influence data to obtain dynamic symptom features; perform symptom time series feature fusion on the dynamic symptom features and the optimal time series feature set to obtain a dynamic symptom feature set;

[0085] Step S3: extract multi-scale interactive features according to the dynamic symptom feature set to obtain a multi-scale interactive feature set; perform time scale feature fusion and biological scale feature fusion according to the dynamic symptom feature set and the multi-scale interactive feature set to obtain a time scale interactive feature matrix and a biological scale interactive feature matrix; perform multi-scale interactive feature matrix integration on the time scale interactive feature matrix and the biological scale interactive feature matrix to obtain a multi-scale interactive feature matrix;

[0086] Step S4: performing hierarchical risk prediction modeling according to the multi-scale interactive feature matrix and the dynamic symptom feature set to obtain a hierarchical risk prediction model;

[0087] Step S5: Integrate the hierarchical risk prediction model into the medical information system, perform ARDS risk prediction on the target patient, and obtain the patient's ARDS risk data; perform early warning evaluation and early warning visualization based on the patient's ARDS risk data to obtain ARDS early warning visualization data.

[0088] In the embodiment of the present invention, reference Figure 1 The above is a schematic flow chart of the steps of the method for constructing an early warning evaluation model for acute respiratory distress syndrome of the present invention. In this example, the method for constructing an early warning evaluation model for acute respiratory distress syndrome includes the following steps:

[0089] Step S1: Acquire ARDS early warning indicator data; collect multidimensional data according to the ARDS early warning indicator data through a medical information system to obtain an original multidimensional data set; construct a multidimensional heterogeneous database according to the original multidimensional data set to obtain a multidimensional heterogeneous database;

[0090] In the embodiment of the present invention, first, the ARDS early warning indicator data is obtained through literature mining technology, and the data quality is evaluated and screened in combination with the medical information system to finally determine the preferred indicator set. Then, according to the preferred indicator set, multi-source heterogeneous data such as the patient's physiological indicators, laboratory test results, and text medical records are collected from the medical information system to form an original multidimensional data set. Finally, the original multidimensional data set is pre-processed by data cleaning, conversion, standardization, etc. to build a multidimensional heterogeneous database with unified structure and rich content.

[0091] Step S2: Screen the time series features of the multidimensional heterogeneous database to obtain an optimal time series feature set; quantify the severity of symptoms in the multidimensional heterogeneous database to obtain quantitative symptom feature data; smooth the symptom time series of the quantitative symptom feature data to obtain a symptom time smoothed sequence; perform symptom mutation point time series analysis on the quantitative symptom feature data to obtain symptom mutation point data; perform symptom interaction analysis on the quantitative symptom feature data based on the symptom time smoothed sequence to obtain symptom influence data; integrate symptom dynamic features based on the symptom influence data to obtain dynamic symptom features; perform symptom time series feature fusion on the dynamic symptom features and the optimal time series feature set to obtain a dynamic symptom feature set;

[0092] In the embodiment of the present invention, time series data is extracted from a multidimensional heterogeneous database, and feature extraction and screening are performed to obtain an optimal time series feature set that can effectively predict ARDS. At the same time, the symptom information in the database is structured, standardized and quantified to obtain quantitative symptom feature data. Then, the quantitative symptom feature data is subjected to time series smoothing, mutation point analysis, trend fitting and interaction analysis, and combined with the characteristics of the symptom evolution stage, a dynamic symptom feature set that can comprehensively describe the dynamic changes of the patient's symptoms is finally integrated.

[0093] Step S3: extract multi-scale interactive features according to the dynamic symptom feature set to obtain a multi-scale interactive feature set; perform time scale feature fusion and biological scale feature fusion according to the dynamic symptom feature set and the multi-scale interactive feature set to obtain a time scale interactive feature matrix and a biological scale interactive feature matrix; perform multi-scale interactive feature matrix integration on the time scale interactive feature matrix and the biological scale interactive feature matrix to obtain a multi-scale interactive feature matrix;

[0094] In the embodiment of the present invention, a feature interaction matrix is ​​constructed based on a dynamic symptom feature set, and a significance analysis is performed to screen out statistically significant interaction feature pairs. Then, an interaction feature network is constructed using a network analysis method, and multi-scale community division is performed to extract a multi-scale interaction feature set. Finally, the multi-scale interaction feature set is feature fused according to the two dimensions of time and biology to construct a multi-scale interaction feature matrix containing both the time scale and the biological scale.

[0095] Step S4: performing hierarchical risk prediction modeling according to the multi-scale interactive feature matrix and the dynamic symptom feature set to obtain a hierarchical risk prediction model;

[0096] In the embodiment of the present invention, the multi-scale interactive feature matrix is ​​matched with the occurrence of ARDS in the patient, and a labeled data set is constructed and divided into a training set, a validation set and a test set. Then, the basic risk prediction model, the dynamic risk prediction model and the multi-scale risk prediction model are constructed respectively using the training set data, and the three models are hierarchically integrated to obtain the initial hierarchical risk prediction model. Finally, the model performance is evaluated and optimized using the test set data to obtain the final hierarchical risk prediction model.

[0097] Step S5: Integrate the hierarchical risk prediction model into the medical information system, perform ARDS risk prediction on the target patient, and obtain the patient's ARDS risk data; perform early warning evaluation and early warning visualization based on the patient's ARDS risk data, and obtain ARDS early warning visualization data;

[0098] In an embodiment of the present invention, the trained hierarchical risk prediction model is deployed into the medical information system and connected to the system for real-time data. For the target patient, relevant features are extracted from the medical information system and input into the model for ARDS risk prediction to obtain the patient's risk probability value. Then, the patient is risk graded according to the preset risk threshold, and a comprehensive early warning assessment is performed in combination with other clinical information. Finally, the patient's risk information and early warning information are visualized.

[0099] Preferably, step S1 comprises the following steps:

[0100] Step S11: obtaining ARDS early warning indicator data through literature mining technology; constructing a candidate indicator pool according to the ARDS early warning indicator data to obtain a candidate indicator pool;

[0101] Step S12: Performing data quality assessment on the candidate indicator pool and screening the candidate indicators through the medical information system to obtain a preliminary indicator set;

[0102] Step S13: performing multidimensional data collection on the medical information system according to the preliminary selected indicator set to obtain an original multidimensional data set;

[0103] Step S14: performing an indicator correlation analysis based on the original multidimensional data set to obtain an optimal indicator set;

[0104] Step S15: integrating multidimensional heterogeneous data of the preferred indicator set and constructing a multidimensional heterogeneous database to obtain a multidimensional heterogeneous database.

[0105] In an embodiment of the present invention, first, a literature search is performed using a medical literature database (such as PubMed, Embase, etc.) and relevant keywords (such as "acute respiratory distress syndrome", "early warning indicators", etc.) to obtain research literature related to early warning of ARDS. Then, the retrieved literature is screened, the warning indicators involved therein are extracted, and they are organized into a candidate indicator pool. The candidate indicator pool is matched with the data in the medical information system to evaluate the data quality indicators such as the integrity, consistency, and accuracy of the candidate indicator data. According to the pre-set data quality standards, the candidate indicators are screened, and indicators with poor data quality are eliminated to obtain a preliminary indicator set with data quality that meets the requirements. Based on the preliminary indicator set, relevant data is collected from the medical information system. For example, multi-dimensional information such as physiological indicators, laboratory test results, and imaging data of patients can be obtained from multiple sources such as electronic medical records, inspection systems, and monitoring systems to form an original multidimensional data set. The original multidimensional data set is subjected to indicator correlation analysis, such as calculating the Pearson correlation coefficient and the Spearman correlation coefficient between the indicators to evaluate the degree of association between the indicators. According to the results of correlation analysis, the indicators closely related to the occurrence and development of ARDS were screened out to form a preferred indicator set. The preferred indicator set was integrated with multidimensional heterogeneous data, that is, data from different sources and in different formats were cleaned, converted, and standardized so that they could be uniformly stored and managed. Finally, the integrated data was used to construct a multidimensional heterogeneous database to provide a data basis for subsequent model construction.

[0106] Preferably, step S2 comprises the following steps:

[0107] Step S21: screening the time series features of the multi-dimensional heterogeneous database to obtain an optimal time series feature set; quantifying the severity of symptoms of the multi-dimensional heterogeneous database to obtain quantitative symptom feature data;

[0108] Step S22: performing symptom time series smoothing on the quantitative symptom feature data to obtain a symptom time smoothing sequence; performing symptom mutation point time series analysis on the quantitative symptom feature data to obtain symptom mutation point data;

[0109] Step S23: performing symptom trend fitting on the symptom time smoothing sequence to obtain a symptom trend function; performing symptom trend deviation calculation based on the symptom trend function and the symptom time smoothing sequence to obtain a symptom trend deviation time series;

[0110] Step S24: dividing the symptom evolution stages according to the symptom mutation point data and the symptom trend deviation time series to obtain symptom evolution stage data; extracting symptom evolution stage features from the symptom evolution stage data to obtain symptom evolution stage features;

[0111] Step S25: performing symptom interaction analysis on the quantitative symptom feature data according to the symptom time smoothing sequence to obtain symptom influence data; integrating symptom dynamic features based on the symptom evolution stage characteristics and the symptom influence data to obtain dynamic symptom features;

[0112] Step S26: Perform symptom-time series feature fusion on the dynamic symptom features and the preferred time series feature set to obtain a dynamic symptom feature set.

[0113] As an example of the present invention, refer to Figure 2 As shown, in this example, step S2 includes:

[0114] Step S21: screening the time series features of the multi-dimensional heterogeneous database to obtain an optimal time series feature set; quantifying the severity of symptoms of the multi-dimensional heterogeneous database to obtain quantitative symptom feature data;

[0115] In the embodiment of the present invention, for time series data in a multidimensional heterogeneous database, such as physiological indicators such as heart rate, respiratory rate, and blood oxygen saturation of patients, time series features such as mean, variance, and kurtosis are extracted using methods such as sliding windows, and statistical methods or machine learning models are used for screening to obtain a preferred time series feature set that can effectively predict ARDS. At the same time, the text information describing the patient's symptoms in the database is structured, and the severity of the symptoms is quantified according to pre-defined scoring standards or medical guidelines to obtain quantitative symptom feature data.

[0116] Step S22: performing symptom time series smoothing on the quantitative symptom feature data to obtain a symptom time smoothing sequence; performing symptom mutation point time series analysis on the quantitative symptom feature data to obtain symptom mutation point data;

[0117] In the embodiment of the present invention, the quantitative symptom feature data is smoothed using methods such as moving average and exponential smoothing to reduce the impact of noise and obtain a smoother symptom time series. At the same time, the smoothed time series is analyzed using methods such as sliding window and CUSUM algorithm to identify the mutation points of symptom changes, such as the time point when the symptoms suddenly worsen or improve, to obtain symptom mutation point data.

[0118] Step S23: performing symptom trend fitting on the symptom time smoothing sequence to obtain a symptom trend function; performing symptom trend deviation calculation based on the symptom trend function and the symptom time smoothing sequence to obtain a symptom trend deviation time series;

[0119] In the embodiment of the present invention, linear regression, polynomial regression and other methods are used to fit the symptom time smoothing series to obtain a function that can reflect the overall trend of symptom changes. Then, the difference between each time point on the time series and the fitting function is calculated to obtain the symptom trend deviation time series, which is used to characterize the degree to which the symptoms deviate from the overall trend at each time point.

[0120] Step S24: dividing the symptom evolution stages according to the symptom mutation point data and the symptom trend deviation time series to obtain symptom evolution stage data; extracting symptom evolution stage features from the symptom evolution stage data to obtain symptom evolution stage features;

[0121] In the embodiment of the present invention, the symptom mutation point data and the symptom trend deviation time series are combined to divide the patient's symptom evolution process into different stages, such as symptom stability stage, symptom exacerbation stage, symptom recovery stage, etc. Then, corresponding features are extracted for each stage, such as stage duration, average symptom value within the stage, symptom change rate within the stage, etc., to obtain symptom evolution stage features.

[0122] Step S25: performing symptom interaction analysis on the quantitative symptom feature data according to the symptom time smoothing sequence to obtain symptom influence data; integrating symptom dynamic features based on the symptom evolution stage characteristics and the symptom influence data to obtain dynamic symptom features;

[0123] In the embodiment of the present invention, based on the symptom time smoothing sequence, Granger causality test, mutual information and other methods are used to analyze the interactive influence relationship between different symptoms, such as whether the deterioration of one symptom will lead to the change of another symptom, and the degree of this influence is quantified to obtain symptom influence data. Finally, combined with the symptom evolution stage characteristics and symptom influence data, multimodal feature fusion methods such as splicing and attention mechanism are used to integrate features from different sources to obtain dynamic symptom features that can fully describe the dynamic changes of patient symptoms.

[0124] Step S26: performing symptom-time series feature fusion on the dynamic symptom features and the optimal time series feature set to obtain a dynamic symptom feature set;

[0125] In an embodiment of the present invention, the dynamic symptom features are fused with the preferred temporal feature set, for example, by using methods such as feature concatenation and feature crossover, to obtain a dynamic symptom feature set containing richer information.

[0126] Preferably, step S21 includes the following steps:

[0127] Step S211: extracting time series data from a multidimensional heterogeneous database to obtain a multidimensional time series data set; extracting time series features from the multidimensional time series data set to obtain a time series feature library; and mining deep time series features from the time series feature library to obtain a deep time series feature set;

[0128] Step S212: evaluating the feature prediction capability and discrimination of the time series feature library and the deep time series feature set and screening the time series features to obtain a preferred time series feature set;

[0129] Step S213: Structuring the electronic medical records of the multi-dimensional heterogeneous database to obtain structured symptom data;

[0130] Step S214: mapping the structured symptom data to a preset ARDS-related symptom standard dictionary, standardizing the symptom representation, and obtaining a standardized symptom library;

[0131] Step S215: Acquire clinical evaluation standard data; quantify the severity of symptoms in the standardized symptom library according to the clinical evaluation standard data to obtain quantitative symptom characteristic data.

[0132] In an embodiment of the present invention, first, time-related patient data, such as body temperature, heart rate, respiratory rate, blood oxygen saturation, etc. recorded every hour, are extracted from a multidimensional heterogeneous database to form a multidimensional time series data set. Then, feature extraction is performed on the time series data of each indicator, for example, the mean, variance, maximum value, minimum value, trend change, etc. in each time window are calculated to construct an initial time series feature library. Finally, deep learning models such as recurrent neural networks (RNN) and long short-term memory networks (LSTM) are used to further extract features from the time series feature library to capture the complex time patterns hidden in the data, and obtain a more abstract and expressive deep time series feature set. Statistical methods or machine learning models are used to evaluate the contribution of each feature in the time series feature library and the deep time series feature set to ARDS prediction, for example, the correlation coefficient, information gain, AUC value, etc. between the feature and the occurrence of ARDS are calculated. At the same time, the ability of each feature to distinguish between ARDS patients and non-ARDS patients is evaluated, for example, the difference significance of the feature in the two groups of patients, the classification accuracy, etc. are calculated. According to the evaluation results, features with strong predictive ability and high discrimination are screened out to form a preferred time series feature set. Use natural language processing (NLP) technology to structure electronic medical records in multidimensional heterogeneous databases. For example, use named entity recognition, relation extraction and other methods to identify ARDS-related symptom information from text, such as cough, dyspnea, fever, etc., and convert it into a structured data format, such as a triple (patient ID, symptom, time of onset), to form structured symptom data. Build a standard dictionary containing ARDS-related symptoms in advance, such as SNOMED CT, ICD-10, etc., and match the structured symptom data with the standard dictionary, map symptoms with different expressions to unified standard terms, for example, map "shortness of breath" and "labor of breathing" to "dyspnea", so as to achieve standardization of symptom representation and build a standardized symptom library. Refer to medical guidelines, expert consensus or clinical research results to formulate clinical evaluation criteria for ARDS-related symptoms, for example, formulate scoring criteria based on factors such as the severity and duration of symptoms. Then, each symptom in the standardized symptom library was quantitatively scored according to the established evaluation criteria. For example, a mild cough was scored as 1 point and a severe cough was scored as 3 points. Finally, quantitative symptom characteristic data were obtained for subsequent analysis.

[0133] Preferably, step S22 comprises the following steps:

[0134] Step S221: smoothing the symptom time series of the quantitative symptom feature data to obtain a symptom time smoothed series;

[0135] Step S222: Calculate the symptom change rate for the symptom time smoothing sequence to obtain a symptom change rate time series;

[0136] Step S223: Calculate the symptom change acceleration for the symptom change rate time series to obtain the symptom change acceleration time series;

[0137] Step S224: Perform symptom mutation point analysis based on the symptom change rate time series and the symptom change acceleration time series to obtain symptom mutation point data.

[0138] The present invention can effectively filter out random fluctuations and noise in the data by smoothing the quantitative symptom feature data in time series, making the trend of symptom changes smoother and easier to identify, thereby improving the reliability and stability of subsequent analysis results, avoiding the model from being disturbed by noise, and making false warnings. Calculating the symptom change rate can more sensitively capture subtle changes in the patient's symptoms, such as whether the symptom worsens slowly or rapidly, which is of great significance for judging the development trend of the disease and warning the occurrence of ARDS, and can help the model identify the signal of rising risk earlier. Further calculating the acceleration of symptom change can more accurately identify the turning point of the symptom change trend, such as from slow deterioration to rapid deterioration, which has important reference value for predicting the occurrence time and severity of ARDS, and can help the model more accurately predict the turning point of risk change. Combining the symptom change rate and the acceleration of change to analyze the symptom mutation point, it is possible to more accurately identify the time point when the patient's condition changes significantly, such as sudden deterioration or improvement of symptoms, which is crucial for taking timely intervention measures to avoid further deterioration of the condition, and can provide clinicians with clearer warning signals.

[0139] In an embodiment of the present invention, for quantified symptom feature data, such as cough scores and dyspnea scores recorded every hour, smoothing techniques such as moving average method and exponential smoothing method are used to process the time series. For example, a moving average method with a window size of 3 is used to replace the value of each time point with the average value of the time point and the two time points before and after it to reduce data fluctuations and obtain a smoother symptom time smoothing sequence. Based on the symptom time smoothing sequence, the symptom change rate at each time point is calculated. For example, the difference between the symptom scores between two adjacent time points can be calculated, or the derivative of the time series can be calculated using a more complex calculus method to obtain a symptom change rate time series that can reflect the speed of symptom changes. The symptom change rate time series is derivatized again to calculate the acceleration of symptom changes at each time point. For example, the difference between the symptom change rates between two adjacent time points can be calculated, or the second-order derivative of the symptom change rate time series can be calculated using a more complex calculus method to obtain a symptom change acceleration time series that can reflect the change in the symptom change trend. Combined with the symptom change rate time series and the symptom change acceleration time series, a mutation point detection algorithm is used to identify the mutation point of symptom change. For example, a sliding window-based method can be used to set a threshold, and when the symptom change rate or symptom change acceleration exceeds the threshold, it is considered that there is a mutation point at that time point. Finally, symptom mutation point data recording the time point when each symptom has a mutation is obtained.

[0140] Preferably, step S25 comprises the following steps:

[0141] Step S251: Calculate the time dimension correlation coefficient of the quantitative symptom feature data according to the symptom time smoothing sequence and construct a symptom interaction relationship matrix to obtain the symptom interaction relationship matrix;

[0142] Step S252: constructing a symptom impact network according to the symptom interaction relationship matrix to obtain symptom impact network data;

[0143] Step S253: Calculate the symptom influence according to the symptom influence network data to obtain symptom influence data;

[0144] Step S254: Integrate the dynamic characteristics of the symptoms based on the symptom evolution stage characteristics and the symptom influence data to obtain dynamic symptom characteristics.

[0145] In an embodiment of the present invention, based on each pair of symptom time smoothing sequences, the time dimension correlation coefficient between them is calculated to quantify the degree of association between symptoms over time. The calculated correlation coefficient is used to construct a symptom interaction matrix, and each element in the matrix represents the strength of the correlation between the corresponding symptom pairs. The symptom interaction matrix is ​​used to construct a symptom influence network. For example, each symptom can be regarded as a node in the network, and an edge is connected between the symptom pairs whose correlation coefficient exceeds the preset threshold, and the weight of the edge represents the size of the correlation. The network constructed in this way can intuitively show the mutual influence relationship between symptoms. Based on the symptom influence network, the influence of each symptom is calculated using a network analysis method. For example, node centrality indicators such as degree centrality, betweenness centrality, closeness centrality, etc. can be used to measure the importance of each symptom in the network, and use it as a metric for the influence of the symptom. The symptom evolution stage characteristics are integrated with the symptom influence data to construct a more comprehensive dynamic symptom feature. For example, the two can be combined into a new feature vector by feature splicing, or the features of different stages can be weighted and fused according to the symptom influence using methods such as attention mechanisms, and finally a dynamic symptom feature that can more comprehensively describe the dynamic changes of symptoms is obtained.

[0146] Preferably, step S3 comprises the following steps:

[0147] Step S31: constructing a feature interaction matrix according to the dynamic symptom feature set to obtain an original interaction feature matrix; performing a significance analysis on the original interaction feature matrix to obtain a significance interaction feature matrix;

[0148] Step S32: constructing an interactive feature network according to the significant interactive feature matrix to obtain interactive feature network data; extracting multi-scale interactive features according to the interactive feature network data to obtain a multi-scale interactive feature set;

[0149] Step S33: performing time scale feature fusion on the multi-scale interaction feature set according to the dynamic symptom feature set to obtain a time scale interaction feature matrix; performing biological scale feature fusion on the multi-scale interaction feature set according to the dynamic symptom feature set to obtain a biological scale interaction feature matrix;

[0150] Step S34: performing multi-scale interaction feature matrix integration on the time-scale interaction feature matrix and the biological-scale interaction feature matrix to obtain a multi-scale interaction feature matrix.

[0151] As an example of the present invention, refer to Figure 3 As shown, in this example, step S3 includes:

[0152] Step S31: constructing a feature interaction matrix according to the dynamic symptom feature set to obtain an original interaction feature matrix; performing a significance analysis on the original interaction feature matrix to obtain a significance interaction feature matrix;

[0153] In the embodiment of the present invention, based on all the features in the dynamic symptom feature set, their interaction information or other correlation indicators are calculated in pairs to construct an initial feature interaction matrix. Each element in the matrix represents the interaction intensity between the corresponding feature pairs. Then, a statistical test method (e.g., chi-square test, t-test) is used to perform a significance analysis on the original interaction feature matrix, filter out statistically significant interaction feature pairs, and construct a significant interaction feature matrix.

[0154] Step S32: constructing an interactive feature network according to the significant interactive feature matrix to obtain interactive feature network data; extracting multi-scale interactive features according to the interactive feature network data to obtain a multi-scale interactive feature set;

[0155] In an embodiment of the present invention, the significant interaction feature matrix is ​​converted into an interaction feature network. For example, each feature is regarded as a node in the network, and an edge is connected between the significant interaction feature pairs, and the weight of the edge represents the interaction strength. Then, a community discovery algorithm (for example, Louvain algorithm, Leiden algorithm) is used to analyze the interaction feature network, and the network is divided into multiple communities, and the features in each community have a high interaction strength. Finally, for each community, its representative features are extracted or combined features are constructed to form a multi-scale interaction feature set.

[0156] Step S33: performing time scale feature fusion on the multi-scale interaction feature set according to the dynamic symptom feature set to obtain a time scale interaction feature matrix; performing biological scale feature fusion on the multi-scale interaction feature set according to the dynamic symptom feature set to obtain a biological scale interaction feature matrix;

[0157] In an embodiment of the present invention, according to the time attribute of each feature in the dynamic symptom feature set (for example, the time of symptom onset, duration, etc.), the multi-scale interactive feature set is subjected to time-scale feature fusion. For example, the features can be grouped according to the time window, the statistical indicators of each interactive feature in each time window are calculated, and the time-scale interactive feature matrix is ​​constructed. On the other hand, according to the biological significance of each feature in the dynamic symptom feature set (for example, body temperature is related to inflammatory response, respiratory rate is related to lung function, etc.), the multi-scale interactive feature set is subjected to biological-scale feature fusion. For example, the features can be grouped according to predefined biological pathways or networks, the statistical indicators of each interactive feature in each pathway or network are calculated, and the biological-scale interactive feature matrix is ​​constructed.

[0158] Step S34: performing multi-scale interaction feature matrix integration on the time scale interaction feature matrix and the biological scale interaction feature matrix to obtain a multi-scale interaction feature matrix;

[0159] In the embodiment of the present invention, the time scale interaction feature matrix and the biological scale interaction feature matrix are integrated to construct a final multi-scale interaction feature matrix. The matrix comprehensively considers the interaction intensity, time correlation and biological correlation between features, and can more comprehensively characterize the potential mechanism of the occurrence and development of ARDS.

[0160] Preferably, step S33 includes the following steps:

[0161] Step S331: extracting time window features from the dynamic symptom feature set using the preset time window size data to obtain a time window feature data set; constructing a time scale feature matrix for the multi-scale interaction feature set based on the time window feature data set to obtain a time scale feature matrix;

[0162] Step S332: grouping the dynamic symptom feature set into symptom groups to obtain symptom group data; aggregating symptom features within the group on the symptom group data to obtain symptom group features;

[0163] Step S333: performing interactive feature grouping on the multi-scale interactive feature set to obtain interactive feature grouping data; performing intra-group interactive feature aggregation on the interactive feature grouping data to obtain interactive feature group features;

[0164] Step S334: construct a biological scale interactive feature matrix for the symptom group features and the interactive feature group features to obtain a biological scale interactive feature matrix.

[0165] In an embodiment of the present invention, each feature in the dynamic symptom feature set is divided into multiple time windows in chronological order according to a preset time window size (for example, 2 hours, 4 hours, etc.), and feature statistics such as mean, variance, maximum value, minimum value, etc. are extracted for each time window. Then, the multi-scale interaction features in each time window are combined with the corresponding time window features, such as concatenating or calculating interaction terms, to construct a time scale feature matrix for describing the changes in the feature interaction pattern in different time periods. The features in the dynamic symptom feature set are grouped according to the clinical manifestations or physiological mechanisms of the symptoms. For example, symptoms related to fever (such as body temperature, heart rate, etc.) can be grouped into one group, and symptoms related to dyspnea (such as respiratory rate, blood oxygen saturation, etc.) can be grouped into another group. Then, the features in each symptom group are aggregated, such as calculating statistical indicators such as mean, maximum value, minimum value, etc., or using dimensionality reduction methods such as principal component analysis to extract representative features, and obtain symptom group features that can summarize the symptom information of the group. The features in the multi-scale interaction feature set are grouped according to the biological significance or function of the interaction features. For example, the interactive features reflecting inflammatory response can be divided into one group, and the interactive features reflecting lung function damage can be divided into another group. Then, the features in each interactive feature group are aggregated, such as calculating statistical indicators such as mean, maximum, minimum, or using dimensionality reduction methods such as principal component analysis to extract representative features, and obtain interactive feature group features that can summarize the interactive feature information of this group. The symptom group features and the interactive feature group features are combined in pairs, and their correlation coefficients or other similarity indicators are calculated to construct a biological scale interactive feature matrix. This matrix reflects the degree of association between different symptom groups and different interactive feature groups, and can characterize the impact of feature interaction patterns on symptom development from a biological perspective.

[0166] Preferably, step S4 comprises the following steps:

[0167] Step S41: risk labeling the multi-scale interaction feature matrix through the medical information system to obtain a labeled multi-scale interaction feature matrix; dividing the labeled multi-scale interaction feature matrix into a data set to obtain a training set, a validation set, and a test set;

[0168] Step S42: constructing a basic risk prediction model based on the training set and evaluating the model using the validation set to obtain a basic risk prediction model;

[0169] Step S43: constructing a dynamic risk prediction model according to the dynamic symptom feature set and the basic risk prediction model to obtain a dynamic risk prediction model;

[0170] Step S44: constructing a multi-scale risk prediction model according to the multi-scale interactive feature matrix and the dynamic risk prediction model to obtain a multi-scale risk prediction model;

[0171] Step S45: performing hierarchical risk prediction model integration on the basic risk prediction model, the dynamic risk prediction model and the multi-scale risk prediction model to obtain an initial hierarchical risk prediction model;

[0172] Step S46: Use the test set to verify the initial hierarchical risk prediction model to obtain a hierarchical risk prediction model.

[0173] In an embodiment of the present invention, according to the actual diagnosis and treatment data of the patient, the occurrence of ARDS of the patient is obtained from the medical information system as label information, and it is matched with the samples in the multi-scale interactive feature matrix to obtain a labeled multi-scale interactive feature matrix. Then, the labeled data set is randomly divided into a training set, a validation set and a test set according to a certain ratio (for example, 7:1:2), which are used for model training, parameter adjustment and performance evaluation respectively. Using the training set data, a suitable machine learning model or a deep learning model is selected as the basic risk prediction model, and the model is trained and parameter optimized using methods such as cross-validation, and finally the model with the best performance on the validation set is selected as the basic risk prediction model. On the basis of the basic risk prediction model, combined with the dynamic symptom feature set, a dynamic risk prediction model that can capture the impact of dynamic changes in symptoms on ARDS risk is constructed. For example, a recurrent neural network (RNN) or its variants can be used to process time series data, and the model output is fused with the output of the basic risk prediction model, such as weighted averaging or splicing, to obtain the final dynamic risk prediction model. On the basis of the dynamic risk prediction model, the multi-scale interactive feature matrix is ​​further combined to construct a multi-scale risk prediction model that can comprehensively consider feature interaction information. For example, the multi-scale interaction feature matrix can be used as part of the model input and input into the model together with the dynamic symptom features for training, or the dynamic symptom features can be weighted and fused according to the multi-scale interaction feature matrix using methods such as the attention mechanism to enhance the model's ability to capture the interactive information of key features. The three models obtained in steps S42, S43, and S44 are integrated to construct a hierarchical risk prediction model. For example, the stacking method can be used to use the prediction results of the previous level model as the input features of the next level model to obtain more accurate and robust prediction results. The initial hierarchical risk prediction model obtained in step S45 is evaluated for performance using the test set data, such as calculating the model's AUC, accuracy, recall rate, and other indicators, and the model is further optimized and adjusted based on the evaluation results, and finally a hierarchical risk prediction model with satisfactory performance is obtained.

[0174] In the embodiment of the present invention,

[0175] Preferably, step S5 comprises the following steps:

[0176] Step S51: Integrate the hierarchical risk prediction model into the medical information system, extract patient features of the target patient, and obtain a patient feature set;

[0177] Step S52: inputting the patient feature set into the hierarchical risk prediction model to perform ARDS risk prediction and obtain the patient's ARDS risk data;

[0178] Step S53: performing risk grading processing on the patient's ARDS risk data using a preset risk threshold to obtain patient risk grade data;

[0179] Step S54: Perform early warning assessment according to the patient's risk level data to obtain the patient's ARDS early warning assessment data;

[0180] Step S55: Perform early warning visualization on the patient's ARDS risk data and the patient's ARDS early warning assessment data to obtain ARDS early warning visualization data.

[0181] In an embodiment of the present invention, the trained hierarchical risk prediction model is deployed into a medical information system, and a data interface between the model and the system is established. For the target patient, relevant physiological indicators, laboratory test results, text medical records and other data are obtained from the medical information system in real time, and preprocessed and feature extracted in the format required by the model to obtain a feature set containing patient information. The patient feature set is input into the hierarchical risk prediction model, and the model will calculate based on the input feature information and output the risk probability value of ARDS in the patient in the future (for example, the next 24 hours, 48 ​​hours, etc.), that is, the patient's ARDS risk data. According to clinical experience or relevant guidelines, different risk thresholds are pre-set, for example, the risk probability value below 0.3 is classified as a low risk level, 0.3 to 0.7 is classified as a medium risk level, and above 0.7 is classified as a high risk level. The patient ARDS risk data obtained in step S52 is graded according to the preset risk threshold to obtain patient risk level data, so that medical staff can quickly judge the patient's risk level. Based on the patient's risk level data, combined with other clinical information, such as the patient's age, underlying disease conditions, etc., a comprehensive early warning assessment is conducted. For example, for patients with high risk levels, the system can automatically issue early warning information to remind medical staff to take timely intervention measures. The patient's ARDS risk data and early warning assessment data are visualized in an intuitive way. For example, charts, curves, colors, etc. can be used to clearly display the patient's risk level, risk trend, early warning information, etc. to medical staff, so that they can better understand the changes in the patient's condition and take appropriate intervention measures in a timely manner.

[0182] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0183] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A method for constructing an early warning assessment model for acute respiratory distress syndrome, characterized in that: The following steps are involved: Step S1: Acquire ARDS early warning indicator data; collect multidimensional data according to the ARDS early warning indicator data through a medical information system to obtain an original multidimensional data set; construct a multidimensional heterogeneous database according to the original multidimensional data set to obtain a multidimensional heterogeneous database; Step S2: Screening time series features of the multi-dimensional heterogeneous database to obtain an optimal time series feature set; The severity of symptoms in the multi-dimensional heterogeneous database is quantified to obtain quantitative symptom feature data; the symptom time series of the quantitative symptom feature data is smoothed to obtain the symptom time smoothed series; the symptom mutation point time series analysis is performed on the quantitative symptom feature data to obtain the symptom mutation point data; the symptom interaction impact analysis is performed on the quantitative symptom feature data according to the symptom time smoothed series to obtain the symptom influence data; Integrate the dynamic characteristics of symptoms based on the symptom influence data to obtain dynamic symptom characteristics; Perform symptom-time series feature fusion on the dynamic symptom features and the optimal time series feature set to obtain a dynamic symptom feature set; Step S3: extracting multi-scale interactive features according to the dynamic symptom feature set to obtain a multi-scale interactive feature set; performing time-scale feature fusion and biological-scale feature fusion according to the dynamic symptom feature set and the multi-scale interactive feature set to obtain a time-scale interactive feature matrix and a biological-scale interactive feature matrix; Perform multi-scale interaction feature matrix integration on the time scale interaction feature matrix and the biological scale interaction feature matrix to obtain a multi-scale interaction feature matrix; Step S4: performing hierarchical risk prediction modeling according to the multi-scale interactive feature matrix and the dynamic symptom feature set to obtain a hierarchical risk prediction model; Step S5: Integrate the hierarchical risk prediction model into the medical information system, perform ARDS risk prediction on the target patient, and obtain the patient's ARDS risk data; According to the patient's ARDS risk data, early warning assessment is performed and early warning visualization is performed to obtain ARDS early warning visualization data.

2. The method for constructing an early warning assessment model for acute respiratory distress syndrome according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: obtaining ARDS early warning indicator data through literature mining technology; constructing a candidate indicator pool according to the ARDS early warning indicator data to obtain a candidate indicator pool; Step S12: Performing data quality assessment on the candidate indicator pool and screening the candidate indicators through the medical information system to obtain a preliminary indicator set; Step S13: performing multidimensional data collection on the medical information system according to the preliminary selected indicator set to obtain an original multidimensional data set; Step S14: performing an indicator correlation analysis based on the original multidimensional data set to obtain an optimal indicator set; Step S15: integrating multidimensional heterogeneous data of the preferred indicator set and constructing a multidimensional heterogeneous database to obtain a multidimensional heterogeneous database.

3. The method for constructing an early warning assessment model for acute respiratory distress syndrome according to claim 1, characterized in that: Step S2 includes the following steps: Step S21: screening the time series features of the multi-dimensional heterogeneous database to obtain an optimal time series feature set; quantifying the severity of symptoms of the multi-dimensional heterogeneous database to obtain quantitative symptom feature data; Step S22: performing symptom time series smoothing on the quantitative symptom feature data to obtain a symptom time smoothing sequence; performing symptom mutation point time series analysis on the quantitative symptom feature data to obtain symptom mutation point data; Step S23: performing symptom trend fitting on the symptom time smoothing sequence to obtain a symptom trend function; performing symptom trend deviation calculation based on the symptom trend function and the symptom time smoothing sequence to obtain a symptom trend deviation time series; Step S24: dividing the symptom evolution stages according to the symptom mutation point data and the symptom trend deviation time series to obtain symptom evolution stage data; extracting symptom evolution stage features from the symptom evolution stage data to obtain symptom evolution stage features; Step S25: performing symptom interaction analysis on the quantitative symptom feature data according to the symptom time smoothing sequence to obtain symptom influence data; integrating symptom dynamic features based on the symptom evolution stage characteristics and the symptom influence data to obtain dynamic symptom features; Step S26: performing symptom-time series feature fusion on the dynamic symptom features and the preferred time series feature set to obtain a dynamic symptom feature set.

4. The method for constructing an early warning assessment model for acute respiratory distress syndrome according to claim 3, characterized in that: Step S21 includes the following steps: Step S211: extracting time series data from a multidimensional heterogeneous database to obtain a multidimensional time series data set; extracting time series features from the multidimensional time series data set to obtain a time series feature library; and mining deep time series features from the time series feature library to obtain a deep time series feature set; Step S212: evaluating the feature prediction capability and discrimination of the time series feature library and the deep time series feature set and screening the time series features to obtain a preferred time series feature set; Step S213: Structuring the electronic medical records of the multi-dimensional heterogeneous database to obtain structured symptom data; Step S214: mapping the structured symptom data to a preset ARDS-related symptom standard dictionary, standardizing the symptom representation, and obtaining a standardized symptom library; Step S215: Acquire clinical evaluation standard data; quantify the severity of symptoms in the standardized symptom library according to the clinical evaluation standard data to obtain quantitative symptom characteristic data.

5. The method for constructing an early warning assessment model for acute respiratory distress syndrome according to claim 3, characterized in that: Step S22 includes the following steps: Step S221: smoothing the symptom time series of the quantitative symptom feature data to obtain a symptom time smoothed series; Step S222: Calculate the symptom change rate for the symptom time smoothing sequence to obtain a symptom change rate time series; Step S223: Calculate the symptom change acceleration for the symptom change rate time series to obtain the symptom change acceleration time series; Step S224: Perform symptom mutation point analysis based on the symptom change rate time series and the symptom change acceleration time series to obtain symptom mutation point data.

6. The method for constructing an early warning assessment model for acute respiratory distress syndrome according to claim 3, characterized in that: Step S25 includes the following steps: Step S251: Calculate the time dimension correlation coefficient of the quantitative symptom feature data according to the symptom time smoothing sequence and construct a symptom interaction relationship matrix to obtain the symptom interaction relationship matrix; Step S252: constructing a symptom impact network according to the symptom interaction relationship matrix to obtain symptom impact network data; Step S253: Calculate the symptom influence according to the symptom influence network data to obtain symptom influence data; Step S254: Integrate the dynamic characteristics of the symptoms based on the symptom evolution stage characteristics and the symptom influence data to obtain dynamic symptom characteristics.

7. The method for constructing an early warning assessment model for acute respiratory distress syndrome according to claim 1, characterized in that: Step S3 includes the following steps: Step S31: constructing a feature interaction matrix according to the dynamic symptom feature set to obtain an original interaction feature matrix; performing a significance analysis on the original interaction feature matrix to obtain a significance interaction feature matrix; Step S32: constructing an interactive feature network according to the significant interactive feature matrix to obtain interactive feature network data; extracting multi-scale interactive features according to the interactive feature network data to obtain a multi-scale interactive feature set; Step S33: performing time scale feature fusion on the multi-scale interaction feature set according to the dynamic symptom feature set to obtain a time scale interaction feature matrix; performing biological scale feature fusion on the multi-scale interaction feature set according to the dynamic symptom feature set to obtain a biological scale interaction feature matrix; Step S34: performing multi-scale interaction feature matrix integration on the time-scale interaction feature matrix and the biological-scale interaction feature matrix to obtain a multi-scale interaction feature matrix.

8. The method for constructing an early warning assessment model for acute respiratory distress syndrome according to claim 7, characterized in that: Step S33 includes the following steps: Step S331: extracting time window features from the dynamic symptom feature set using the preset time window size data to obtain a time window feature data set; constructing a time scale feature matrix for the multi-scale interaction feature set based on the time window feature data set to obtain a time scale feature matrix; Step S332: grouping the dynamic symptom feature set into symptom groups to obtain symptom group data; aggregating symptom features within the group on the symptom group data to obtain symptom group features; Step S333: performing interactive feature grouping on the multi-scale interactive feature set to obtain interactive feature grouping data; performing intra-group interactive feature aggregation on the interactive feature grouping data to obtain interactive feature group features; Step S334: construct a biological scale interactive feature matrix for the symptom group features and the interactive feature group features to obtain a biological scale interactive feature matrix.

9. The method for constructing an early warning assessment model for acute respiratory distress syndrome according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: risk labeling the multi-scale interaction feature matrix through the medical information system to obtain a labeled multi-scale interaction feature matrix; dividing the labeled multi-scale interaction feature matrix into a data set to obtain a training set, a validation set, and a test set; Step S42: constructing a basic risk prediction model based on the training set and evaluating the model using the validation set to obtain a basic risk prediction model; Step S43: constructing a dynamic risk prediction model according to the dynamic symptom feature set and the basic risk prediction model to obtain a dynamic risk prediction model; Step S44: constructing a multi-scale risk prediction model according to the multi-scale interactive feature matrix and the dynamic risk prediction model to obtain a multi-scale risk prediction model; Step S45: performing hierarchical risk prediction model integration on the basic risk prediction model, the dynamic risk prediction model and the multi-scale risk prediction model to obtain an initial hierarchical risk prediction model; Step S46: Use the test set to verify the initial hierarchical risk prediction model to obtain a hierarchical risk prediction model.

10. The method for constructing an early warning assessment model for acute respiratory distress syndrome according to claim 1, characterized in that: Step S5 includes the following steps: Step S51: Integrate the hierarchical risk prediction model into the medical information system, extract patient features of the target patient, and obtain a patient feature set; Step S52: inputting the patient feature set into the hierarchical risk prediction model to perform ARDS risk prediction and obtain the patient's ARDS risk data; Step S53: performing risk grading processing on the patient's ARDS risk data using a preset risk threshold to obtain patient risk grade data; Step S54: Perform early warning assessment according to the patient's risk level data to obtain the patient's ARDS early warning assessment data; Step S55: Perform early warning visualization on the patient's ARDS risk data and the patient's ARDS early warning assessment data to obtain ARDS early warning visualization data.

Images