A risk protection factor-based trend summary ICU criticality prediction method

Abstract

This invention discloses a knowledge-enhanced, time-series hierarchical summary-prediction method for predicting ICU mortality rates, belonging to the field of medical information technology. This method aims to address the problems of poor interpretability and failure to fully utilize time-series dynamic information and medical knowledge in existing ICU mortality prediction models. The method includes: acquiring patient medical data and enhancing it with clinical knowledge; identifying abnormal features in the time-series data; analyzing the dynamic trends of the same abnormal feature and assessing its clinical significance. Its core features are: first, by summarizing a thought chain framework, integrating medical knowledge, abnormal feature sets, and their dynamic trends to generate a structured natural language summary of the patient's condition; then, by using a predictive thought chain framework, extracting core risk factors and protective factors from the summary, and performing weighted logical reasoning based on their priority and causal relationships, ultimately calculating the mortality rate of critically ill ICU patients. This invention, by simulating the hierarchical reasoning process of clinicians, deeply integrates time-series data analysis with medical knowledge, significantly improving the accuracy, interpretability, and clinical applicability of the prediction results.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing technology, specifically to the field of medical prediction technology, and more specifically, to a method for predicting ICU critical illness based on trend summarization of risk protection factors. Background Technology

[0002] In clinical practice, accurately predicting patient prognosis is crucial for developing personalized treatment plans, optimizing the allocation of medical resources, and improving doctor-patient communication. Doctors typically rely on their professional knowledge and clinical experience, combined with various examination indicators and medical history, to make a comprehensive judgment. However, this approach depends to some extent on the doctor's subjective experience, and when faced with massive, high-dimensional, and dynamically changing medical data, the human brain's processing capacity is limited, which may lead to the overlooking of potential risk factors.

[0003] Existing computer-aided diagnostic systems, especially prediction methods based on traditional statistical models or some black-box artificial intelligence models, while achieving high prediction accuracy in certain tasks, often suffer from the following drawbacks:

[0004] 1) Lack of transparency and poor interpretability: The model is like a "black box," making it difficult to explain the basis of its predictions to doctors and patients, which limits its credibility and application in key medical decisions.

[0005] 2) Static data processing: Many models mainly process static data at specific points in time, which does not make full use of time series data that reflects the evolution of the disease and cannot effectively capture the dynamic trend of the disease.

[0006] 3) Insufficient knowledge integration: The model mainly learns correlations from data, making it difficult to effectively integrate and utilize the causal logic and diagnostic thinking summarized by doctors in long-term practice.

[0007] Therefore, how to simulate the clinical diagnostic thinking process of doctors, combine the dynamic changes of data with medical causal knowledge, and develop a prognostic prediction method that is both accurate and well interpretable is a technical problem that urgently needs to be solved in the field of medical artificial intelligence. Summary of the Invention

[0008] This invention aims to overcome the shortcomings of existing technologies and provide a trend summary prediction method for critically ill patients in ICU based on risk protection factors. This method simulates the doctor's diagnosis and treatment logic, integrates feature extraction, trend analysis, disease summary and risk prediction, and improves the accuracy and interpretability of prediction.

[0009] To achieve the above objectives, this invention provides a method for predicting ICU critical illness based on trend summarization of risk protection factors, characterized by the following steps:

[0010] S1: Medical Data Input and Knowledge Enhancement: Acquire medical data of patients to be analyzed and supplement the clinical significance of each feature in the data;

[0011] S2: Based on the preset clinical normal range or the knowledge recall within the model, analyze the time series data, identify the abnormal features, and form an abnormal feature set;

[0012] S3: For the same abnormal feature spanning multiple time points, calculate its trend over time, and judge the clinical significance of the trend by combining the medical knowledge from step S1.

[0013] S4: Based on the summary thinking chain framework, integrate and reason about the knowledge in step S1, the abnormal feature set identified in step S2, and the dynamic change trend analyzed in step S3 to generate a structured summary of the condition in natural language form.

[0014] S5: Extract the core risk factors and protective factors that are directly or indirectly related to the mortality rate of critically ill patients in the ICU from the disease summary generated in step S4.

[0015] S6: Based on the predictive thinking chain framework, the core risk factors and protective factors extracted in step S5 are weighted and logically reasoned. At the same time, the priority and causal relationship between each risk factor and protective factor are considered to calculate the mortality rate of ICU critically ill patients.

[0016] Furthermore, step S1 specifically includes:

[0017] S11: Input medical data is time-series data, containing structured data at least one time point, including vital signs and laboratory test results, and unstructured data, including imaging reports and medical records. For patients... At a certain point in time The data is represented as a set ,in Representing patients A collection of data over a period of time, Represents the Tth time point. It is the electronic medical record data vector at time point t. ,in Indicates the first These features at a given time point The value;

[0018] S12: Internal Knowledge Retrieval and Activation: Based on the medical data vector input in step S11 This involves constructing a query request and activating the pre-trained medical knowledge parameters within the model to perform internal knowledge retrieval. This process can be formally represented as retrieving knowledge from the model parameter space. Knowledge vector related to recall Functions:

[0019]

[0020] in, The representative model uses a knowledge retrieval and activation mechanism based on its internal parameters. It is the electronic medical record data vector at time point t;

[0021] S13: External Knowledge Base Collaborative Supplementation: Setting a Threshold for Knowledge Completeness Assessment When the internal knowledge retrieved in step S12 The confidence level is below the threshold When a specific query fails to find the match, the system automatically redirects to an external knowledge base for collaborative supplementary retrieval. This external knowledge base is constructed from authoritative medical guidelines through vectorization and stored in a dedicated vector database. External knowledge retrieval is achieved by calculating the similarity between the query vector and the knowledge vector. The core retrieval process is as follows:

[0022]

[0023] in, This refers to knowledge obtained through external knowledge retrieval. This represents a text embedding function. Represents a knowledge set in an external vector database. At a certain point in time Electronic medical record data vectors, Represents the first in the knowledge set This knowledge The cosine similarity calculation function is used. Finally, the system integrates internal and external knowledge to form a complete data representation for knowledge enhancement in subsequent steps.

[0024] Furthermore, step S2 specifically includes:

[0025] S21: Regarding the time series data obtained in step S1 Features with different sampling frequencies and dimensions are preprocessed in a unified manner; through linear interpolation, time window smoothing and normalization, various features are converted into comparable sequences under a unified time scale, thereby ensuring the temporal consistency of the model input.

[0026] S22: Based on the model's internal knowledge parameters The provided clinical normal range and upper and lower limits of feature fluctuation constitute the knowledge constraint boundary; the large model automatically determines whether the feature at each time step exceeds the normal range through semantic parsing and knowledge retrieval of the input features.

[0027] S23: For anomalous events that are consecutive in time or similar in semantics, the model aggregates them through a self-attention mechanism and abstracts them into a higher-level anomalous pattern description. This step transforms the anomalous events from a numerical level to a medical semantic level, laying the foundation for subsequent reasoning.

[0028] S24: Anomaly Feature Set Formation: The model ultimately generates anomaly feature sets. ,in, A set representing anomalous features, each Indicates the first Each anomalous feature is represented by a large model in structured natural language, including feature type, anomalous direction, time span, anomalous intensity, and its potential physiological significance.

[0029] Furthermore, step S3 specifically includes:

[0030] S31: Employing a large model to analyze the same anomalous features over time. The semantic description is used for contextual modeling, and attention weights are used to automatically capture the direction of change over time, including increase, decrease, fluctuation, and persistence; the trend is inferred by the model through semantic consistency and temporal relationship.

[0031] S32: The large model calls upon its internal knowledge vectors Semantic matching was used to determine the correlation between trend changes and disease severity.

[0032] S33: Employing a cross-feature semantic alignment mechanism, the model identifies synergistic or antagonistic relationships between different abnormal trends; this process is achieved through multi-layered thought chain reasoning and outputs dynamic interaction patterns in natural language.

[0033] S34: Summary Table of Final Generation Trends for Large Models The data is presented in a structured natural language format, including abnormal features, trend direction, duration, clinical significance, and interaction patterns.

[0034] Furthermore, step S4 specifically includes:

[0035] S41: The enhanced medical knowledge in S1, the abnormal feature set identified in S2, and the trend features summarized in S3 are uniformly input into the summary thinking chain prompt framework using a large model; this framework generates a logically coherent description of the disease through multi-step reasoning and automatically identifies the key causal chain of the disease's evolution.

[0036] S42: The summary of the large model output is presented in a structured natural language format and includes the following parts: the current overall disease status; major abnormalities and their evolution trends; possible pathophysiological explanations; and inferred high-risk signs or improvement signals.

[0037] S43: To ensure that the summary conforms to clinical logic, a large model is used to perform an internal knowledge consistency check before output to ensure that the generated text does not conflict with its own knowledge content;

[0038] Furthermore, step S5 specifically includes:

[0039] S51: A large model based on semantic parsing mechanism is used to automatically identify key factors related to mortality from the disease summary text; wherein, the key factors include risk factors and protective factors, such as identifying "shock" as a risk factor and "improved oxygenation" as a protective factor, thereby realizing the automatic extraction of medical risk elements from the content generated by the large language model;

[0040] S52: A large model is used to determine the causal logic direction between factors through a causal reasoning module; the causal reasoning module automatically generates causal chains between factors by semantic path parsing and causal template matching, for example, deriving the causal logic relationship of "low blood pressure leads to decreased renal perfusion", thereby establishing a causal relationship network between factors;

[0041] S53: The identified risk and protective factors are structured according to the organ system dimension using a large model; the structured organization classifies the factors into circulatory system risk, respiratory system risk, renal system risk and other organ system risk and protective factors, thereby establishing a set of structured factors based on organ system classification, providing a systematic input basis for subsequent semantic weight allocation;

[0042] Furthermore, step S6 specifically includes:

[0043] S61: Employ a large model to assign semantic weights to each factor based on the severity and confidence of each risk factor and protective factor in clinical knowledge. ;in, Indicates the first The semantic weights of each factor are dynamically generated based on the severity and confidence of clinical semantics through the knowledge activation mechanism of the language model.

[0044] S62: A large model is used to connect different risks and protective factors and their interactions through multi-step logical reasoning, forming a hierarchical reasoning chain about mortality. This process is executed in the form of natural language logical deduction, realizing causal logical deduction from factors to outcomes.

[0045] S63: A large model is used to integrate the semantic weights of various risk and protection factors and the directionality of the inference chain to output hierarchical mortality prediction results; the output format includes probabilistic output and semantic interpretation, thereby providing quantitative and qualitative predictions based on semantic reasoning;

[0046] S64: Use a large model to generate an interpretable report containing mortality prediction conclusions and reasoning paths; the report includes the semantic weights of each core factor, causal reasoning chains, natural language summaries of model reasoning, and recommended potential intervention directions to ensure that the prediction results are interpretable and clinically actionable.

[0047] After adopting the above strategy, the positive effects of the present invention are:

[0048] (1) High interpretability: The methodology simulates the doctor's clinical thinking, and the generated disease summary and risk factor analysis process are clear and transparent, making the prediction results verifiable;

[0049] (2) Dynamic analysis: This invention not only focuses on static outliers, but also emphasizes the analysis of the dynamic change trend of key indicators, which can more accurately capture the process of disease development and evolution;

[0050] (3) Combining knowledge and data-driven approaches: This method combines data-driven feature analysis with medical knowledge-driven causal relationships and risk priority judgments, thereby improving the robustness and accuracy of the model.

[0051] (4) High flexibility: The thinking chain framework in the method can easily adapt to prediction tasks of different diseases and different prognostic goals by adjusting the prompts. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating an ICU critical care prediction method based on trend summarization of risk protection factors, as described in this invention. Detailed Implementation

[0053] To enable those skilled in the art to better understand the present invention and to make the above-mentioned objectives, technical solutions and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings.

[0054] S1: Medical Data Input and Knowledge Enhancement

[0055] Data input: Obtain continuous monitoring data of ICU patient p. ,Include:

[0056] (1) Renal function characteristics: serum creatinine, blood urea nitrogen, estimated glomerular filtration rate;

[0057] (2) Characteristics of muscle damage: creatine kinase, myoglobin;

[0058] (3) Circulatory system characteristics: arterial blood pressure, heart rate, and dosage of vasoactive drugs;

[0059] (4) Inflammation / infection characteristics: white blood cell count, proportion of immature granulocytes, C-reactive protein, early warning score;

[0060] (5) Liver function characteristics: aspartate aminotransferase, alanine aminotransferase;

[0061] (6) Coagulation function characteristics: prothrombin time, activated partial thromboplastin time, prothrombin activity;

[0062] (7) Respiratory function characteristics: blood oxygen saturation and ventilator support parameters.

[0063] Knowledge enhancement: Through internal knowledge retrieval and collaborative supplementation with external knowledge bases, the clinical significance of the above features is enhanced with medical knowledge to form a complete data representation.

[0064] S2: Anomaly Feature Recognition

[0065] Anomaly detection: The system identifies a set of abnormal features that deviate from the normal range. ,include:

[0066] (1) eGFR consistently below 15 ml / min / 1.73 m²;

[0067] (2) Creatine kinase > 10000 U / L;

[0068] (3) Norepinephrine dose > 0.5 μg / kg / min;

[0069] (4) White blood cell count > 20 x 10^9 / L;

[0070] (5) EWS score consistently > 7 points;

[0071] (6) AST / ALT > 500 U / L;

[0072] S3: Dynamic Trend Analysis

[0073] Trend identification and analysis, and generation of a trend summary table. :

[0074] (1) eGFR trend: "remains at an extremely low level with no meaningful recovery observed";

[0075] (2) Creatine kinase trend: "Persistently high level";

[0076] (3) Trend of vasoactive drugs: "Continued increase";

[0077] (4) Trend of inflammatory markers: "Continuously elevated or repeatedly high risk";

[0078] (5) SpO2 trend: "Stability depends on high-intensity invasive ventilation support";

[0079] S4: Disease Summary Generation

[0080] Summary of thought chain integration: Generating a structured disease summary based on the summary of thought chain framework:

[0081] "This patient presented with a classic cascade of multiple organ dysfunction syndrome. The primary and most critical factor was the persistently worsening acute renal failure, characterized by a persistently low estimated glomerular filtration rate with no significant recovery. The immediate cause of the renal failure was severe rhabdomyolysis, evidenced by persistently high levels of myoglobin and creatine kinase, constituting an irreversible core injury. Secondly, the patient was in a state of persistent shock, requiring vasoactive drugs such as norepinephrine to maintain blood pressure, indicating circulatory system collapse. The underlying driving factor appears to be an uncontrolled systemic inflammatory response or infection, evidenced by persistently elevated white blood cell and immature granulocyte counts and a recurring high-risk EWS score. This systemic process had been complicated by severe liver damage and coagulation dysfunction, further exacerbating the condition. Although individual indicators such as SpO2 stabilized with invasive ventilation, this reflected the maintenance effect of the life support system rather than an improvement in the patient's own physiological function."

[0082] S5: Core Factor Extraction

[0083] Risk and protective factor identification, extraction of a structured factor set:

[0084] (1) Circulatory system risks: intractable shock due to drug dependence;

[0085] (2) Renal system risks: irreversible renal failure;

[0086] (3) Systemic risks: Uncontrolled systemic inflammatory response;

[0087] (4) Systemic risks: severe liver damage and coagulation disorders;

[0088] S6: Mortality Prediction

[0089] Reasoning is performed using predictive thought chain cues, based on semantic weighting and multi-step logical reasoning:

[0090] Reasoning process: "The patient's core risk factors—irreversible kidney failure, drug-dependent shock, and uncontrolled systemic inflammation—combine to create an irreversible death process. Kidney failure leads to toxin accumulation and internal environment disturbance, exacerbating inflammation and circulatory burden; refractory shock results in insufficient perfusion of all organs, further damaging kidney and liver function; and uncontrolled inflammation is the fundamental driving force behind all of this."

[0091] Predicted output: Mortality rate R = 0.99

[0092] Specific embodiments of the present invention have been described above with reference to the accompanying drawings. However, those skilled in the art will understand that various modifications and substitutions can be made to the specific embodiments of the present invention without departing from the spirit and scope of the invention. All such modifications and substitutions fall within the scope defined by the claims of the present invention.

Claims

1. A trend summary method for predicting ICU critical illness based on risk protection factors, characterized in that... Includes the following steps: S1: Medical Data Input and Knowledge Enhancement: Acquire medical data of patients to be analyzed and supplement the clinical significance of each feature in the data; S2: Based on the preset clinical normal range or the knowledge recall within the model, analyze the time series data, identify the abnormal features, and form an abnormal feature set; S3: For the same abnormal feature spanning multiple time points, calculate its trend over time, and judge the clinical significance of the trend by combining the medical knowledge from step S1. S4: Based on the summary thinking chain framework, integrate and reason about the knowledge in step S1, the abnormal feature set identified in step S2, and the dynamic change trend analyzed in step S3 to generate a structured summary of the condition in natural language form. S5: Extract the core risk factors and protective factors that are directly or indirectly related to the mortality rate of critically ill patients in the ICU from the disease summary generated in step S4. S6: Based on the predictive thinking chain framework, the core risk factors and protective factors extracted in step S5 are weighted and logically reasoned. At the same time, the priority and causal relationship between each risk factor and protective factor are considered to calculate the mortality rate of critically ill patients in the ICU.

2. The method for predicting ICU critical illness based on trend summarization of risk protection factors according to claim 1, characterized in that, Step S1 specifically includes: S11: Input medical data is time-series data, containing structured data at least one time point, including vital signs and laboratory test results, and unstructured data, including imaging reports and medical records. For patients... At a certain point in time The data is represented as a set ,in Representing patients A collection of data over a period of time, Represents the Tth time point. It is the electronic medical record data vector at time point t. ,in Indicates the first These features at a given time point The value; S12: Internal Knowledge Retrieval and Activation: Based on the medical data vector input in step S11 This involves constructing a query request and activating the pre-trained medical knowledge parameters within the model to perform internal knowledge retrieval. This process can be formally represented as retrieving knowledge from the model parameter space. Knowledge vector related to recall Functions: in, The representative model uses a knowledge retrieval and activation mechanism based on its internal parameters. It is the electronic medical record data vector at time point t; S13: External Knowledge Base Collaborative Supplementation: Setting a Threshold for Knowledge Completeness Assessment When the internal knowledge retrieved in step S12 The confidence level is below the threshold When a specific query fails to find the match, the system automatically redirects to an external knowledge base for collaborative supplementary retrieval. This external knowledge base is constructed from authoritative medical guidelines through vectorization and stored in a dedicated vector database. External knowledge retrieval is achieved by calculating the similarity between the query vector and the knowledge vector. The core retrieval process is as follows: in, This refers to knowledge obtained through external knowledge retrieval. This represents a text embedding function. Represents a knowledge set in an external vector database. At a certain point in time Electronic medical record data vectors, Represents the first in the knowledge set This knowledge The cosine similarity calculation function is used. Finally, the system integrates internal and external knowledge to form a complete data representation for knowledge enhancement in subsequent steps.

3. The method for predicting ICU critical illness based on trend summarization of risk protection factors according to claim 1, characterized in that, Step S2 specifically includes: S21: Regarding the time series data obtained in step S1 Features with different sampling frequencies and dimensions are preprocessed in a unified manner; through linear interpolation, time window smoothing and normalization, various features are converted into comparable sequences under a unified time scale, thereby ensuring the temporal consistency of the model input. S22: Based on the model's internal knowledge parameters The provided clinical normal range and upper and lower limits of feature fluctuation constitute the knowledge constraint boundary; the large model automatically determines whether the feature at each time step exceeds the normal range through semantic parsing and knowledge retrieval of the input features. S23: For anomalous events that are consecutive in time or similar in semantics, the model aggregates them through a self-attention mechanism and abstracts them into a higher-level anomalous pattern description. This step transforms the anomalous events from a numerical level to a medical semantic level, laying the foundation for subsequent reasoning. S24: Anomaly Feature Set Formation: The model ultimately generates anomaly feature sets. ,in, A set representing anomalous features, each Indicates the first Each anomalous feature is represented by a large model in structured natural language, including feature type, anomalous direction, time span, anomalous intensity, and its potential physiological significance.

4. The method for predicting ICU critical illness based on trend summarization of risk protection factors according to claim 1, characterized in that, Step S3 specifically includes: S31: Employing a large model to analyze the same anomalous features over time. The semantic description is used for contextual modeling, and attention weights are used to automatically capture the direction of change over time, including increase, decrease, fluctuation, and persistence; the trend is inferred by the model through semantic consistency and temporal relationship. S32: The large model calls upon its internal knowledge vectors Semantic matching was used to determine the correlation between trend changes and disease severity. S33: Employing a cross-feature semantic alignment mechanism, the model identifies synergistic or antagonistic relationships between different abnormal trends; this process is achieved through multi-layered thought chain reasoning and outputs dynamic interaction patterns in natural language. S34: Summary Table of Final Generation Trends for Large Models It is presented in a structured natural language format, including abnormal features, trend direction, duration, clinical significance, and interaction patterns.

5. The method for predicting ICU critical illness based on trend summarization of risk protection factors according to claim 1, characterized in that, Step S4 specifically includes: S41: The enhanced medical knowledge in S1, the abnormal feature set identified in S2, and the trend features summarized in S3 are uniformly input into the summary thinking chain prompt framework using a large model; this framework generates a logically coherent description of the disease through multi-step reasoning and automatically identifies the key causal chain of the disease's evolution. S42: The summary of the large model output is presented in a structured natural language format and includes the following parts: the current overall disease status; major abnormalities and their evolution trends; possible pathophysiological explanations; and inferred high-risk signs or improvement signals. S43: To ensure that the summary conforms to clinical logic, a large model is used to perform an internal knowledge consistency check before output to ensure that the generated text does not conflict with its own knowledge content.

6. The method for predicting ICU critical illness based on trend summarization of risk protection factors according to claim 1, characterized in that, Step S5 specifically includes: S51: A large model based on semantic parsing mechanism is used to automatically identify key factors related to mortality from the disease summary text; wherein, the key factors include risk factors and protective factors, such as identifying "shock" as a risk factor and "improved oxygenation" as a protective factor, thereby realizing the automatic extraction of medical risk elements from the content generated by the large language model; S52: A large model is used to determine the causal logic direction between factors through a causal reasoning module; the causal reasoning module automatically generates causal chains between factors by semantic path parsing and causal template matching, for example, deriving the causal logic relationship of "low blood pressure leads to decreased renal perfusion", thereby establishing a causal relationship network between factors; S53: The identified risk and protective factors are structured according to the organ system dimension using a large model; the structured organization classifies the factors into circulatory system risk, respiratory system risk, renal system risk and other organ system risk and protective factors, thereby establishing a set of structured factors based on organ system classification, providing a systematic input basis for subsequent semantic weight allocation.

7. The method for predicting ICU critical illness based on trend summarization of risk protection factors according to claim 1, characterized in that, Step S6 specifically includes: S61: Employ a large model to assign semantic weights to each factor based on the severity and confidence of each risk factor and protective factor in clinical knowledge. ;in, Indicates the first The semantic weights of each factor are dynamically generated based on the severity and confidence of clinical semantics through the knowledge activation mechanism of the language model. S62: A large model is used to connect different risks and protective factors and their interactions through multi-step logical reasoning, forming a hierarchical reasoning chain about mortality. This process is executed in the form of natural language logical deduction, realizing causal logical deduction from factors to outcomes. S63: A large model is used to integrate the semantic weights of various risk and protection factors and the directionality of the inference chain to output hierarchical mortality prediction results; the output format includes probabilistic output and semantic interpretation, thereby providing quantitative and qualitative predictions based on semantic reasoning; S64: Use a large model to generate an interpretable report containing mortality prediction conclusions and reasoning paths; the report includes the semantic weights of each core factor, causal reasoning chains, natural language summaries of model reasoning, and recommended potential intervention directions to ensure that the prediction results are interpretable and clinically actionable.

Images