CN-122004857-A - ICU patient agitation risk identification and early warning system and method based on intelligent agent driving

CN122004857ACN 122004857 ACN122004857 ACN 122004857ACN-122004857-A

Abstract

The invention relates to an ICU patient agitation risk identification and early warning system and method based on agent driving, comprising the steps of realizing offline modeling and classifier training of agitation state, grade and high risk of an ICU patient through multi-mode data fusion and layered machine learning model construction, collecting and processing physiological and behavior data at an edge end in real time, generating a structural semantic event through an integrated decision frame, forming a high-order event representation capable of providing recognition reasoning by combining original observation and context information, constructing a dynamic narrative thread at a central server end, organizing discrete events into coherent illness veins through multidimensional association intensity calculation, evaluating the belief state based on the narrative momentum and consistency, triggering intervention decision with high credibility, carrying out structural encapsulation on intervention advice generated by the agent, carrying out graded pushing to a clinical terminal according to the early warning grade, recording a full-link log and collecting manual feedback, and realizing closed-loop output and continuous optimization of a system.

Inventors

LI YUXUAN
LIU LIHONG
LIU ZHAO
CAI LINNING
LI BAOHUA
JIA CHENGBIN
ZHANG YU
FENG WENLU
LIU YANG
LIU HONGYU

Assignees

北京大学第三医院（北京大学第三临床医学院）
清华大学

Dates

Publication Date: 20260512
Application Date: 20251215

Claims (10)

1. The ICU patient agitation risk identification and early warning system based on intelligent agent driving is characterized by comprising a data acquisition module, an edge end, a server end and a clinical terminal, wherein the system can execute the following steps: s1, through multi-mode data fusion and layered machine learning model construction, offline modeling and classifier training on the agitation state, the grade and the high risk of an ICU patient are realized, and basic identification capacity with clinical interpretability is formed; S2, acquiring and processing physiological and behavioral data at the edge end in real time, generating a structural semantic event through an integrated decision framework, and forming a high-order event representation capable of providing recognition reasoning by combining original observation and context information; s3, constructing a dynamic narrative thread at a central server, organizing discrete events into coherent illness veins through multidimensional association intensity calculation, evaluating belief states based on the narrative momentum and consistency, and triggering high-credibility intervention decisions; And S4, carrying out structured encapsulation on intervention suggestions generated by the agent, carrying out hierarchical pushing to a clinical terminal according to the early warning level, recording a full-link log and collecting manual feedback, and realizing closed-loop output and continuous optimization of the system.
2. The system of claim 1, wherein the multi-modal data in S1 includes an overhead camera video stream, a bedside microphone audio signal, and monitor physiological parameters, the hierarchical machine learning model includes a fast screening classifier, a fine classification classifier, and a high risk early warning classifier for determining agitation, agitation level, and severe agitation, respectively, and each classifier employs RandomForestClassifier algorithm and configures class_weight parameters to cope with class imbalance problems.
3. The system according to claim 1 or 2, wherein the integrated decision framework in S2 invokes the fast screening classifier, the fine classification classifier and the high-risk early warning classifier through a cascading inference flow, and performs consistency verification based on a preset clinical logic rule, where the clinical logic rule includes a basic consistency rule and a high-risk consistency rule, and the basic consistency rule requires that the agitation state is sometimes greater than 0, and the high-risk consistency rule requires that the agitation level corresponding to the high-risk early warning state reaches 2 or more, and generates a comprehensive confidence score to quantify the reliability of the overall judgment.
4. The system according to any one of claims 1 to 3, wherein the structured semantic event generated in S2 includes a classification_results field, including a prediction tag, a confidence level, a probability distribution, and consistency evaluation information generated by a fusion decision module of each model, and constructs a context evidence chain by backtracking original observation data and a patient personalized baseline file, and combines a dynamic template filling mechanism to generate semantic _ narrative in a natural language form, and finally packages into JSON-format semantic event objects labeled as classification_event types and pushes to a stage three cognitive agent processing module.
5. The system of any one of claims 1-4, wherein in S3, structured semantic events pushed by edge devices are continuously consumed by subscribing to a specific topic in the Kafka message queue and temporarily stored in an event buffer in the form of a double-ended queue with a time window length of t minutes, all current active narrative threads are traversed, and the degree of association between new events and representative events in each thread is calculated based on a multi-dimensional event association strength function, wherein the multi-dimensional event association strength function comprises a time association, a spatial association, an entity association and a causality, wherein the causality is a scoring system based on clinical knowledge rules, and is assigned by analyzing whether trend changes in raw_ observations accord with a preset "pain escalation" or "agitation" causality pattern.
6. The system according to any one of claims 1 to 5, wherein the assessment of the state of the belief in S3 quantifies the risk assessment result by two core formulas of a narrative momentum for measuring acceleration of risk evolution and a narrative consistency, wherein the narrative momentum is calculated based on a second order differential approximation over time sequence, the input is an instantaneous risk value of the narrative thread at different time steps, the instantaneous risk value is a sum of prior risk weights of events in the thread, the narrative consistency is used for assessing the credibility of inter-modal evidence in the narrative thread, and the semantic vectors are obtained by calculating a cosine similarity weighted average between semantic vectors of different modalities, and the semantic vectors convert natural language descriptions of the events into vector representations in a high-dimensional embedding space through a semantic coding model.
7. The system of any one of claims 1-6, wherein when the narrative momentum of the narrative thread is greater than a preset momentum threshold and the narrative consistency is greater than a preset consistency threshold, triggering an action decision module to format the target narrative thread into a natural language abstract, inputting the natural language abstract into a future simulator based on a Transformer architecture to obtain a text prediction of the future trend, submitting the narrative abstract, the quantization index and the future prediction together to a LLM agent with domain expertise, and performing mental chain reasoning to generate a structured intervention suggestion.
8. The system of any one of claims 1-7, wherein the LLM agent is configured as a virtual clinical expert with intensive care experience, performs differential diagnostic analysis in conjunction with patient context information provided by electronic health records, and wherein the reasoning process includes fact analysis, etiology inference, risk assessment, and advice generation, outputting standardized JSON format early warning response objects comprising early warning levels, patient identification, event summaries, causal inference, and prioritized intervention lists, each advice labeling execution categories and specific operational instructions.
9. The system according to any one of claims 1 to 8, wherein the intervention advice structured in S4 is pushed in a hierarchical manner according to the early warning level, the high-level alarm is immediately pushed to the ICU central monitoring station large screen, the nurse station workstation and the medical staff mobile device through WebSocket or HTTP POST request with accompanying audio and visual cues, the medium-level alarm is displayed quietly in the nurse station workstation to-do list, and the low-level information is recorded only in the system log database for subsequent multi-disc analysis.
10. The ICU patient agitation risk identification and early warning method based on intelligent agent driving is characterized by comprising the following steps of: s1, through multi-mode data fusion and layered machine learning model construction, offline modeling and classifier training on the agitation state, the grade and the high risk of an ICU patient are realized, and basic identification capacity with clinical interpretability is formed; S2, acquiring and processing physiological and behavioral data at the edge end in real time, generating a structural semantic event through an integrated decision framework, and forming a high-order event representation capable of providing recognition reasoning by combining original observation and context information; s3, constructing a dynamic narrative thread at a central server, organizing discrete events into coherent illness veins through multidimensional association intensity calculation, evaluating belief states based on the narrative momentum and consistency, and triggering high-credibility intervention decisions; And S4, carrying out structured encapsulation on intervention suggestions generated by the agent, carrying out hierarchical pushing to a clinical terminal according to the early warning level, recording a full-link log and collecting manual feedback, and realizing closed-loop output and continuous optimization of the system.

Description

ICU patient agitation risk identification and early warning system and method based on intelligent agent driving Technical Field The invention relates to the technical field of agitation risk identification and early warning, in particular to an ICU patient agitation risk identification and early warning system and method based on intelligent agent driving. Background In an Intensive Care Unit (ICU), patients are prone to agitation behavior due to various factors such as pain, anxiety, delirium, mechanical ventilation discomfort or metabolic disorder, and the like, so that physiological stress response of the patients is aggravated, risks of unplanned tube drawing, falling and other adverse events are increased, workload of medical staff is remarkably increased, and overall nursing quality is affected. To achieve an objective assessment of patient sedation-agitation status, the Richmond agitation-sedation score (RASS) scale is commonly used clinically, which scales by observing the level of consciousness and the degree of physical activity of the patient. However, the traditional RASS scoring relies on the regular manual observation of medical staff, has the limitations of strong subjectivity, limited frequency, easy influence of workload and the like, and is difficult to realize continuous and real-time dynamic monitoring. Therefore, developing a technical means capable of automatically and accurately identifying the agitation risk of the ICU patient has become an important research direction in the field of intelligent monitoring. In the prior art, there have been some attempts to assist in agitation identification using artificial intelligence methods. For example, patent application publication No. CN119541047A proposes a method for identifying agitation video of an ICU patient based on spatiotemporal motion detection. According to the method, through improving SlowFast deep learning models, video streams of ICU patients are analyzed, complex agitation behaviors are decomposed into predefined atomic actions such as head movement, hand movement, shoulder movement, lifting up, leg bending and the like, the identified atomic actions are weighted and summed based on a RASS scoring table, and finally a total score is obtained to judge agitation levels. The scheme improves the accuracy of video motion recognition and provides a technical path for automatic agitation evaluation. However, this prior art approach has several inherent drawbacks that limit its practical application in a complex, dynamic ICU clinical setting. First, the prior art relies heavily on single video modality information, lacking multi-source information fusion capabilities. ICU environments are complex and variable, and patient's somatic movements are often blocked by quilts, bed curtains, or ongoing medical procedures (e.g., turning over, sputum aspiration). When a critical atomic motion cannot be captured by a video due to occlusion, the system fails, resulting in a missing report. In addition, the scheme completely ignores the acoustic information (such as groan and shouting) of the patient and the continuous physiological signals (such as heart rate and dynamic change of respiratory frequency) on the monitor, and the information is important for comprehensively judging the state of the patient and distinguishing painful agitation from delirium agitation. This singleness of the information dimension results in poor system robustness and difficulty in coping with real world complex disturbances. Second, the prior art processing logic is static and responsive in nature, lacking the ability to dynamically understand and predict the evolution of patient state. The mode of operation is "detect-score-alarm", i.e., the system can only identify and score after the patient has completed one or more obvious atomic actions. This prevents the system from warning before agitation occurs or when early signs occur. For example, the patient may first develop signs of a slight rise in heart rate, increased respiration, slight facial distortion, etc., which cannot be captured and correlated in this regimen, resulting in missing the best early intervention opportunity. The intelligence of the method is only realized by utilizing a deep network to automatically identify actions, but the understanding of the 'narrative context' of the disease is not realized. Finally, the decision mechanism of the prior art is stiff and lacks intelligent reasoning and personalized adaptation capability. The scoring rule is a fixed linear weighted summation formula, and complex nonlinear relations in clinic cannot be captured. For example, it cannot understand the clinical meaning of "progressive exacerbation of pain" which is implied by the combination of "continuous rise of heart rate, incremental increase of movement amplitude, and increase of volume". The output is only a simple level of agitation, lacking inferences about the underlying cause (e.g., pain, hypoxia, del