An intensive care difficult offline decision method, device, equipment and medium

Abstract

The application relates to the technical field of medical artificial intelligence, and discloses a critical care difficult offline decision method and device, equipment and medium, multi-modal clinical features of a target object are acquired; one-way correlation fusion is performed according to dynamic causal correlation relationships between the multi-modal clinical features to obtain multi-modal fusion features; index evaluation is performed on the multi-modal fusion features according to dynamic index conditions to obtain multi-modal index scores and abnormal index information; pathological and physiological mechanism reasoning is performed on the abnormal index information to obtain a pathological and physiological causal chain; offline risk evaluation is performed on the target object according to the multi-modal index scores, the abnormal index information and the pathological and physiological causal chain to obtain an offline risk conclusion. The method has the beneficial effect that multi-modal index fusion evaluation is utilized to accurately predict the offline risk of the target object and output quantitative index conclusions, the randomness and differentiation of human decision-making are reduced, and the potential harm of a spontaneous breathing test to the target object is reduced.

Description

Technical Field

[0001] This application relates to the field of medical artificial intelligence technology, and in particular to a method, device, equipment and medium for decision-making regarding difficult weaning in intensive care. Background Technology

[0002] Mechanical ventilation is a critical life-saving technique in the Intensive Care Unit (ICU). Prolonged mechanical ventilation can lead to secondary physiological impairment, making it difficult for patients to withstand the strain of spontaneous breathing after weaning, resulting in weaning failure. Attempts to wean from the ventilator typically account for 40% to 50% of the total ventilation time. Clinically, approximately 20% to 30% of patients experience difficulty or failure to wean, significantly increasing the risk of ventilator-associated pneumonia (VAP) and potentially increasing mortality. Among related techniques, weaning decisions often heavily rely on the physician's clinical experience and the spontaneous breathing trial (SBT). However, for patients with extremely poor cardiopulmonary reserve, blindly initiating a spontaneous breathing trial can cause iatrogenic harm. Summary of the Invention

[0003] This application provides a method, device, equipment, and medium for decision-making regarding difficult weaning in intensive care. By utilizing multimodal index fusion assessment, it accurately predicts the weaning risk of the target subject and outputs quantitative index conclusions, reducing the arbitrariness and variability of human decision-making, thereby reducing the potential harm of spontaneous breathing tests to the target subject.

[0004] To achieve the above objectives, the main technical solutions adopted in this application include: In a first aspect, embodiments of this application provide a method for decision-making regarding weaning from intensive care in cases of difficulty, the method comprising: Obtain multimodal clinical characteristics of the target subjects; Based on the dynamic causal relationships among the multimodal clinical features, the multimodal clinical features are fused by unidirectional causal association to obtain multimodal fused features; The multimodal fusion features are evaluated based on dynamic index conditions to obtain the multimodal index score and abnormal index information of the target object; the abnormal index information is used to infer the pathophysiological mechanism to obtain the pathophysiological causal chain corresponding to the abnormal index information. The target object's risk of disconnection is assessed based on the multimodal index score, the abnormal index information, and the pathophysiological causal chain, and a conclusion on the target object's risk of disconnection is obtained.

[0005] The intensive care unit (ICU) weaning decision-making method proposed in this application involves fusing multimodal clinical features based on dynamic causal relationships between multiple modalities to obtain multimodal fusion features of the target subject. Dynamic indicators are then evaluated on these multimodal fusion features to obtain multimodal indicator scores and abnormal indicator information. Pathophysiological mechanisms are inferred from the abnormal indicator information to obtain a pathophysiological causal chain. Finally, weaning risk assessment is performed based on the multimodal indicator scores, abnormal indicator information, and the pathophysiological causal chain to arrive at a conclusion regarding the target subject's weaning risk. Compared to related technologies, this application allows for weaning risk assessment of the target subject before initiating spontaneous breathing trials, pre-determining the target subject's weaning risk status. This allows for determining whether to initiate spontaneous breathing trials based on the weaning risk status, reducing the potential harm of spontaneous breathing trials to the target subject and ensuring their physiological and psychological safety. Furthermore, this application uses dynamic causal relationships to fuse multimodal clinical features of the target subject through unidirectional causal association, integrating contextual information across multiple modalities to enhance the semantic richness of multimodal indicators, thereby improving the accuracy and reliability of the weaning risk conclusion.

[0006] Optionally, the multimodal clinical features include multiple clinical features corresponding to different modalities; the step of fusing the multimodal clinical features based on the dynamic causal relationships between them to obtain multimodal fused features includes: Obtain the causal relationship between the multiple clinical features; The causal fusion weights among the multiple clinical features are obtained by evaluating the fusion weights of the multiple clinical features. Based on the causal influence direction, the multiple clinical features are sequentially causally fused according to the causal fusion weight to obtain the fusion information corresponding to each of the multiple clinical features, which serves as the multimodal fusion feature.

[0007] Optionally, the step of evaluating the fusion weights of the multiple clinical features to obtain the causal fusion weights among the multiple clinical features includes: The background characteristics and offline risk baseline score of the target object are obtained; wherein, the offline risk baseline score is obtained by multidimensional contribution weighted evaluation based on the physiological index data of the target object; The background features and the baseline score of offline risk are used as weighted prior constraints. The weights of the multiple clinical features are evaluated based on the weighted prior constraints to obtain the causal fusion weights.

[0008] Optionally, after obtaining the fusion information corresponding to each of the plurality of clinical features, the method further includes: Obtain the historical indicator features of the target object; Based on the multimodal clinical features, historical correlations are identified in the historical indicator features to obtain correlated indicator features; trend decomposition is performed on the correlated indicator features to obtain multi-scale trend features of the correlated indicator features. The multimodal fusion features are obtained by concatenating the fusion information corresponding to the multi-scale trend features and the multiple clinical features respectively.

[0009] Optionally, the dynamic index conditions can be obtained in the following way: Obtain the background features and real-time evolution data of the target object; Case matching is performed based on the background features to obtain similar background cases that are similar to the target object, and offline success cases are selected from the similar background cases. The actual indicator conditions corresponding to the successful offline cases are fused and statistically analyzed to obtain similar indicator conditions, and the similar indicator conditions are dynamically updated according to the real-time evolution data to obtain the dynamic indicator conditions.

[0010] Optionally, the step of performing pathophysiological mechanism inference on the abnormal indicator information to obtain the pathophysiological causal chain corresponding to the abnormal indicator information includes: Mechanism reasoning is performed on the abnormal indicator information to obtain an initial causal chain; If the number of initial causal chains exceeds one and the initial causal chains contradict each other, obtain the prior confidence of the initial causal chains and the real-time evolution data of the target object; The initial causal chain is compared for consistency based on the real-time evolution data, and the prior confidence is corrected based on the results of the consistency comparison to obtain the causal confidence of the initial causal chain. The pathophysiological causal chain is obtained by voting among the initial causal chains based on the causal confidence level.

[0011] Optionally, the method further includes: Obtain real-time evolution data of the target object; Based on the real-time evolution data, an offline risk assessment is performed on the target object to obtain a real-time risk conclusion for the target object; If the real-time risk conclusion and the offline risk conclusion are consistent, the confidence level of the offline risk conclusion is increased; otherwise, anomaly indicators are identified in the real-time evolution data to obtain real-time anomaly information, and the pathophysiological causal chain is updated according to the real-time anomaly information.

[0012] Secondly, embodiments of this application provide a decision-making device for difficult weaning in intensive care, the device comprising: The clinical feature acquisition module is used to acquire multimodal clinical features of the target object; The cross-modal causal fusion module is used to perform unidirectional causal association fusion on the multimodal clinical features based on the dynamic causal relationship between the multimodal clinical features to obtain multimodal fused features; The indicator evaluation and reasoning module is used to evaluate the multimodal fusion features based on dynamic indicator conditions to obtain the multimodal indicator score and abnormal indicator information of the target object; and to perform pathophysiological mechanism reasoning on the abnormal indicator information to obtain the pathophysiological causal chain corresponding to the abnormal indicator information. The decision suggestion generation module is used to assess the offline risk of the target object based on the multimodal index score, the abnormal index information and the pathophysiological causal chain, and to obtain the offline risk conclusion of the target object.

[0013] Thirdly, embodiments of this application provide a computer device, including: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the method described in any of the above embodiments.

[0014] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer instructions, which are used to cause a computer to perform the method described in any one of the above embodiments.

[0015] Fifthly, embodiments of this application provide a computer program product, including computer instructions, which are used to cause a computer to perform the method described in any of the above embodiments. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating the steps of the decision-making method for difficult weaning in intensive care provided in this application embodiment; Figure 2 This is a flowchart illustrating the steps of causal one-way association fusion in the embodiments of this application; Figure 3This is a flowchart illustrating the steps of fusion weight evaluation in an embodiment of this application; Figure 4 This is a diagram illustrating the steps of trend feature decomposition and splicing in the embodiments of this application; Figure 5 This is a flowchart illustrating the steps involved in obtaining the dynamic index conditions in an embodiment of this application. Figure 6 This is a flowchart illustrating the steps of reasoning about the pathophysiological mechanism in the embodiments of this application; Figure 7 This is a flowchart illustrating the steps involved in comparing and correcting risk conclusions in the embodiments of this application. Figure 8 A block diagram of the decision-making device for difficult weaning in intensive care provided in the embodiments of this application; Figure 9 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] Mechanical ventilation is a crucial life-saving technique in intensive care units. Prolonged mechanical ventilation can lead to secondary physiological impairment, making it difficult for patients to withstand the strain of spontaneous breathing after weaning, resulting in weaning failure. Attempts to wean from the ventilator typically account for 40% to 50% of the total ventilation time. Clinically, approximately 20% to 30% of patients experience difficulty or failure to wean, significantly increasing the risk of ventilator-associated pneumonia and potentially raising mortality rates.

[0020] In related technologies, common methods for weaning decisions include traditional assessment systems based on clinical guidelines, predictive models based on structured data, and multimodal decision support systems based on high-frequency waveform variability analysis. Among these, the traditional assessment system based on clinical guidelines is currently the most commonly used standard weaning procedure. It combines weaning readiness assessment with a spontaneous breathing test. First, the clinician assesses the patient's weaning readiness based on their clinical presentation and objective indicators. Second, if the patient passes the weaning readiness assessment, physiological parameters are obtained, and a spontaneous breathing test is initiated to obtain the test results. Finally, based on the test results and physiological parameters, a comprehensive prediction of the patient's weaning success rate is made to obtain the patient's weaning success rate.

[0021] Predictive models based on structured data use statistical or machine learning methods to assist in prediction, such as predictive models based on logistic regression (LR), random forest, or support vector machine (SVM). These models integrate multi-dimensional numerical variables such as demographic characteristics, clinical scores, respiratory mechanics parameters, and laboratory test values ​​to make predictions and output binary prediction results or risk probability values.

[0022] A multimodal decision support system based on high-frequency waveform variability analysis acquires continuous monitoring waveforms during a patient's spontaneous breathing test in real time, and uses signal processing and machine learning algorithms to extract variability features to predict the patient's failure risk.

[0023] It can be seen that related technologies often rely heavily on the doctor's clinical experience and spontaneous breathing tests. However, for patients with extremely poor cardiopulmonary reserve, blindly initiating spontaneous breathing tests may cause iatrogenic harm. Furthermore, predictive models based on structured data typically use a purely data-driven approach for prediction, lacking interpretability.

[0024] To address the aforementioned issues, this application provides a method, device, equipment, and medium for decision-making regarding weaning from intensive care units in cases of difficulty. The method involves acquiring multimodal clinical characteristics of the target individual; performing unidirectional causal fusion based on the dynamic causal relationships between these characteristics to obtain multimodal fusion features; evaluating these features based on dynamic indicator conditions to obtain multimodal indicator scores and abnormal indicator information; inferring pathophysiological mechanisms from the abnormal indicator information to obtain a pathophysiological causal chain; and assessing the weaning risk of the target individual based on the multimodal indicator scores, abnormal indicator information, and the pathophysiological causal chain to obtain a weaning risk conclusion.

[0025] The intensive care unit (ICU) decision-making method for difficult weaning provided in this application performs unidirectional causal association fusion of multimodal clinical characteristics based on the dynamic causal relationship between multimodalities to obtain the multimodal fusion characteristics of the target subject; dynamically evaluates the multimodal fusion characteristics to obtain multimodal indicator scores and abnormal indicator information, and infers the pathophysiological mechanism of the abnormal indicator information to obtain the pathophysiological causal chain; and assesses the weaning risk based on the multimodal indicator scores, abnormal indicator information, and pathophysiological causal chain to obtain the weaning risk conclusion of the target subject.

[0026] Compared with related technologies, this application can conduct a weaning risk assessment on the target subject before initiating the spontaneous breathing test, pre-determine the weaning risk of the target subject, and then determine whether to initiate the spontaneous breathing test on the target subject based on the weaning risk, thereby reducing the potential harm of the spontaneous breathing test to the target subject and ensuring the physiological and psychological safety of the target subject.

[0027] Furthermore, this application performs unidirectional causal association fusion on the multimodal clinical characteristics of the target object based on dynamic causal relationships, so as to fuse the contextual information between multiple modalities across modalities, improve the semantic richness of multimodal indicators, and thus improve the accuracy and authenticity of the offline risk conclusion.

[0028] According to an embodiment of this application, an embodiment of a method for decision-making on difficult weaning in intensive care is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0029] Reference Figure 1 As shown, this embodiment provides a method for decision-making regarding difficult weaning in intensive care, the method comprising: S100. Obtain the multimodal clinical characteristics of the target object.

[0030] S200. Based on the dynamic causal relationship between multimodal clinical features, perform causal unidirectional association fusion on the multimodal clinical features to obtain multimodal fused features.

[0031] S300. Evaluate the multimodal fusion features based on dynamic indicator conditions to obtain the multimodal indicator scores and abnormal indicator information of the target object; infer the pathophysiological mechanism of the abnormal indicator information to obtain the pathophysiological causal chain corresponding to the abnormal indicator information.

[0032] S400. Based on multimodal index scores, abnormal index information, and pathophysiological causal chains, assess the target subject's risk of disconnection and obtain a conclusion on the target subject's risk of disconnection.

[0033] Specifically, real-time information is collected from the target subject to obtain multimodal clinical information of the target subject at the current moment. The target subject can be, for example, a patient receiving mechanical ventilation. Multimodal clinical information can include structured data, unstructured data, and modal data of the target subject. Structured data can include physiological parameters, test results, and clinical scores from electronic medical records. Unstructured data can include unstructured text information such as nursing records, ward round logs, and impact reports of the target subject. Modal data can include chest images, respiratory / ECG waveforms, and voice and breath sound signals of the target subject. It is understood that multimodal clinical information can be collected in real time through medical information systems and sensors, reducing human intervention and thus lowering the probability of missing high-value information, improving the objectivity and standardization of multimodal clinical information.

[0034] It's important to note that different modalities inherently differ in temporal granularity. For example, the temporal granularity of text modality can be irregular, driven by events; that of video modality can be hourly discrete frames; and that of speech modality can be a millisecond-level continuous stream. If feature extraction is performed on all modalities at the same temporal granularity, some modal features may become blurred, failing to extract effective features. To address this issue, independent temporal encoders are configured for each modality within the multimodal clinical features. Each independent temporal encoder has the same temporal granularity as its corresponding modality, enabling the modeling of multimodal clinical information at its respective temporal resolution and reducing the information distortion caused by forcing the same time window.

[0035] After time encoding, clinical features are extracted from the multimodal clinical information to obtain the corresponding clinical features for each modality, which are then used as multimodal clinical features. It is understood that the modalities in multimodal clinical information can include text modalities, image modalities, and speech modalities. For the text modality, semantic features can be extracted using the Qwen3 32B large language model, and key textual features of the target object (e.g., respiratory status, mental state, and airway secretions) can be captured based on these semantic features. For the image modality, image features can be extracted using the MedSegNet large visual model to obtain the clinical image features of the target object. For the speech modality, information can be collected and parsed using the RespVoiceNet clinical speech analysis model to obtain the clinical speech features of the target object. By combining the clinical image features and clinical speech features of the target object, potential anomaly detection is performed to obtain potential abnormal clues (e.g., respiratory muscle fatigue and airway obstruction). It is understood that multimodal clinical features of the target object can be obtained based on key textual features and potential abnormal clues.

[0036] In some embodiments, the Qwen3 32B large language model can be pre-trained. During the training process of the Qwen3 32B large language model, multiple reference case data of difficult weaning in intensive care can be obtained as fine-tuning predictions. The parameter efficient fine-tuning method (LoRA) is adopted to perform domain adaptation on the Qwen3 32B large language model according to the fine-tuning predictions, so as to fine-tune the temporal reasoning and clinical semantic understanding capabilities of the Qwen3 32B large language model. This enables the Qwen3 32B large language model to accurately understand the specialist terminology and pathophysiological logic of difficult weaning scenarios in intensive care, and strengthens the Qwen3 32B large language model's ability to model trends of long-term physiological index sequences.

[0037] Furthermore, an attention mechanism is employed to perform cross-modal alignment of multimodal clinical features, and the aligned multimodal clinical features are mapped to the same semantic space, completing the preprocessing of multimodal clinical features. Modal causal association analysis is then performed on the preprocessed multimodal clinical features to obtain the dynamic causal relationships between them. The process of modal causal association analysis may include: performing causal relationship priors on the multimodal clinical features to obtain the causal influence direction between them, and generating a causal mask between the multimodal clinical features based on the causal influence direction; evaluating the fusion weights of multiple clinical features in the multimodal clinical features to obtain the causal fusion weights between them; and obtaining the dynamic causal relationships based on the causal mask and the causal fusion weights. In some embodiments, the dynamic causal relationships can be represented as a directed attention graph, where each attention node in the directed attention graph corresponds to multiple clinical features in the multimodal clinical features, the connecting edges between attention nodes have attention directions representing the causal influence direction between them, and the weights of the connecting edges represent the causal fusion weights. It is understandable that generating a causal mask based on the direction of causal influence can block the leakage of reverse information between modalities, ensuring that the fusion of multimodal clinical features conforms to the logic of clinical information generation.

[0038] In some embodiments, after obtaining the causal fusion weights, modal credibility assessments can be performed on the multimodal clinical information separately. A fusion weight correction factor is obtained based on the information quality of each modality in the multimodal clinical information, and the causal fusion weights are then adjusted according to the fusion weight correction factor to obtain corrected fusion weights for use in causal unidirectional association fusion. It is understood that when the information quality of any modality is poor, it indicates that the clinical information of that modality lacks reliability, and the clinical features of that modality should be prevented from affecting the clinical features of other modalities. Therefore, when the information quality of any modality is poor, the causal fusion weight between that modality and other modalities can be reduced by using the fusion weight correction factor to improve the feature quality of the multimodal fusion features.

[0039] After obtaining the dynamic causal relationships, multimodal clinical features are fused using unidirectional causal associations based on these relationships. The multimodal clinical features are then fused sequentially along the causal influence direction to obtain multimodal fused features. The unidirectional causal association fusion can be performed using any of the following methods: weighted fusion or splicing projection. It is understandable that the dynamic causal relationships represent the direction of information flow and the degree of influence between multimodal clinical features. Feature fusion based on dynamic causal relationships can utilize contextual information from other modalities to enhance features in any modality, improving the semantic richness of multimodal indicators and thus enhancing the accuracy and reliability of offline risk conclusions.

[0040] Furthermore, dynamic indicator conditions are obtained, and the multimodal fusion features are evaluated based on these conditions to obtain the multimodal indicator score of the target subject. The dynamic indicator conditions can be multiple indicators of the target subject during the weaning process, including but not limited to respiratory mechanics indicators, oxygenation status, airway protection ability, and level of consciousness. Respiratory mechanics indicators may include the shallow and rapid breathing index, maximum inspiratory pressure, and minute ventilation; oxygenation status may include the oxygenation index and blood oxygen saturation trend; and airway protection ability may include cough characteristics and airway secretion volume.

[0041] For example, when the dynamic indicator condition is a threshold, the indicator evaluation method may include: comparing the threshold of the dynamic indicator condition with the corresponding modality in the multimodal fusion features to determine the threshold range in which the corresponding modality falls, and obtaining the indicator score of the corresponding modality based on the threshold range. If the corresponding modality deviates significantly from the threshold range of the dynamic indicator condition, it is determined that the corresponding modality is abnormal, and abnormal indicator information is generated based on the corresponding modality and its difference from the dynamic indicator condition.

[0042] In some embodiments, when evaluating the indicators of multimodal fusion features, a clinical natural language summary can be simultaneously input to assist in the indicator evaluation process. In this embodiment, the clinical natural language summary can be a qualitative clinical description in text form, used to describe the abnormal state of the target object from the perspective of text modality, thereby accurately identifying abnormal indicator information. The clinical natural language summary can be extracted from multimodal clinical information in text modality. For example, the form of the clinical natural language summary can be semantic information such as "patient complains of difficulty breathing" recorded in medical orders or "weak cough" in nursing records.

[0043] When a target object has abnormal indicator information, the pathophysiological mechanism of the abnormal indicator information is inferred to determine the pathophysiological logic causing the abnormal indicator information, thus obtaining the pathophysiological causal chain corresponding to the abnormal indicator information. In some embodiments, the pathophysiological mechanism inference process can be performed through a Transformer large language model. The Transformer large language model corresponds to a pre-built clinical pathophysiological knowledge graph and has pre-learned clinical pathophysiological rules related to intensive care unit (ICU) difficulties in weaning. The abnormal indicator information in text form is input into the Transformer large language model, which performs semantic understanding and logical inference on the abnormal indicator information. Based on the clinical pathophysiological knowledge graph, the abnormal indicator information is mapped to the corresponding pathophysiological mechanism, establishing a symptom-mechanism-effect causal chain (Chain-of-Thought) as the pathophysiological causal chain corresponding to the abnormal indicator information. For example, the pathophysiological causal chain could be hypomagnesemia-diaphragmatic weakness-weaning difficulty, indicating that the target object clinically presents with hypomagnesemia, causing diaphragmatic relaxation disorders and subsequent contractile weakness, which is highly likely to lead to weaning difficulties.

[0044] Understandably, reasoning through pathophysiological mechanisms to obtain a pathophysiological causal chain that can represent pathophysiological logic breaks the limitation of related technologies relying solely on the correlation of clinical data. This allows for a clear explanation of the pathophysiological logic behind conclusions about the risk of weaning off the ventilator, thus improving the interpretability of weaning decisions.

[0045] Furthermore, the target subjects are assessed for their risk of being taken off the ventilator based on multimodal index scores, abnormal index information, and pathophysiological causal chains, resulting in conclusions about the target subjects' risk of being taken off the ventilator and their corresponding confidence scores. The process of assessing the risk of weaning from ventilator ventilator malfunction may include: In the absence of abnormal indicators and pathophysiological causal chains, comparing multimodal indicator scores with indicator scoring thresholds; if the multimodal indicator scores meet the thresholds, the target subject is deemed capable of withstanding the risk of weaning, and further spontaneous breathing tests can be conducted to determine whether weaning is necessary; if the multimodal indicator scores do not meet the thresholds, the target subject is deemed unable to withstand the risk of weaning, and weaning is recommended to be delayed; In the presence of abnormal indicators and pathophysiological causal chains, if the number of abnormal indicators does not exceed the indicator quantity threshold, and there are no potential risk factors affecting the weaning process in the pathophysiological causal chain, weaning is recommended to be delayed, and clinical intervention recommendations are given based on the pathophysiological causal chain; If the number of abnormal indicators exceeds the indicator quantity threshold or there are potential risk factors in the pathophysiological causal chain, such as the target subject exhibiting hypomagnesemia or latent heart failure, the target subject is deemed unable to withstand the risk of weaning, and weaning is not recommended.

[0046] In some embodiments, the process of providing clinical intervention recommendations to a target subject based on a pathophysiological causal chain may include: obtaining background characteristics of the target subject, including but not limited to information such as the target subject's disease type, severity, and medical history; generating intervention recommendations for the target subject based on abnormal indicator information and background characteristics, thus obtaining clinical intervention recommendations. For example, abnormal indicator information may include reversible physiological abnormalities and non-numerical risk factors. Reversible physiological abnormalities may include phenomena such as low protein levels or electrolyte imbalances, while non-numerical risk factors may include phenomena such as excessive airway secretions. When the abnormal indicator information is a reversible physiological abnormality, the value corresponding to the abnormal indicator information can be used as a target value to generate clinical intervention recommendations for correcting the target value, such as recommending intravenous albumin supplementation to the target subject to achieve an albumin level ≥35 g / L. When the abnormal indicator information is a non-numerical risk factor, targeted nursing prompts can be generated based on the abnormal indicator information as clinical intervention recommendations, such as prompting thorough suctioning before extubation and assessing the target subject's coughing ability for excessive airway secretions.

[0047] In some embodiments, during the assessment of ventilator discontinuation risk, key features can be screened based on the causal fusion weights of each modality in multimodal clinical features to obtain one or more key modalities that significantly contribute to the conclusion on ventilator discontinuation risk. Feature dimensions are located within the key modalities based on the pathophysiological causal chain to obtain one or more key driving features that significantly contribute to the conclusion on ventilator discontinuation risk within the key modalities. These key driving features are then output synchronously with the ventilator discontinuation risk conclusion to assist clinicians in quickly understanding the decision-making basis and improve the interpretability of the ventilator discontinuation risk conclusion. It is understood that modified fusion weights can also be used for key feature screening in the above process. For example, key driving features may include cough frequency characteristics in the speech modality or diaphragmatic displacement amplitude in the imaging modality.

[0048] It should be noted that the weaning risk conclusion indicates whether the target subject can tolerate the burden of spontaneous breathing caused by weaning, and is used to determine whether a spontaneous breathing trial can be initiated for the target subject without causing potential harm. If a weaning risk conclusion is obtained, and the target subject can tolerate the weaning risk, a spontaneous breathing trial can be initiated for the target subject.

[0049] During the spontaneous breathing trial, real-time indicators of the target subjects were monitored. Multimodal real-time data from the target subjects at multiple time steps were continuously input at fixed time steps to track the dynamic trends of various indicators. Modal features were extracted from the multimodal real-time data to obtain multimodal real-time features. Modal causal association analysis was performed on these features to determine the real-time causal relationships between them. Based on these real-time causal relationships, unidirectional causal association fusion was performed on the multimodal real-time features to obtain real-time fused features. Autoregressive prediction was performed on the real-time fused features to obtain the indicator data change curves of the target subjects at future time steps. Risk trend assessment was conducted based on these indicator data change curves to obtain continuous risk score curves for the target subjects at future time steps. After the spontaneous breathing trial was completed and continuous risk score curves were obtained, the feasibility of weaning the target subjects off the ventilator was assessed based on the continuous risk score curves, abnormal indicator information, and pathophysiological causal chains to arrive at a conclusion on the feasibility of weaning the target subjects off the ventilator. For example, the feasibility assessment of disconnection may include an assessment of the trend anomalies in the continuous risk score curve, identifying whether there are abnormal slopes or risk inflection points in the continuous risk score curve, in order to determine whether the target subject has a situation of respiratory muscle fatigue accumulation or a downward trend in tolerance.

[0050] It is understood that this embodiment obtains the conclusion of the weaning risk in advance before the spontaneous breathing test, and initiates the spontaneous breathing test on the target subject when the conclusion of the weaning risk indicates that the target subject can withstand the weaning risk. The feasibility of weaning the target subject is then assessed by combining the results of the spontaneous breathing test, thereby reducing the potential harm of the spontaneous breathing test to the target subject, reducing the probability of causing iatrogenic injury to the target subject, and ensuring the physiological and psychological safety of the target subject.

[0051] In some embodiments, cross-validation of the conclusions regarding the feasibility of weaning off the ventilator can be performed to ensure the rigor and clinical safety of the conclusions, and to improve the accuracy and reliability of the conclusions. The process of cross-validation of conclusions may include: comparing the feasibility conclusions of weaning off ventilator use with rules based on the clinical weaning guidelines rule base, checking whether there are any conclusions that contradict evidence-based medicine standards, and obtaining rule validation results; using multiple preset reasoning paths to perform feasibility reasoning on ventilator use based on multimodal clinical characteristics, obtaining parallel reasoning conclusions, where each preset reasoning path differs from the process of obtaining the feasibility conclusions on ventilator use, and may include, but is not limited to, rule engine reasoning; comparing the consistency of the parallel reasoning conclusions and the feasibility conclusions on ventilator use, and obtaining consistency validation results; based on the background characteristics and multimodal clinical information of the target object, performing historical case matching in the historical ventilator use case database, obtaining historical similar cases similar to the target object, and comparing and validating the ventilator use outcomes of historical similar cases with the feasibility conclusions on ventilator use, and obtaining rationality validation results for the feasibility conclusions on ventilator use; performing contradiction verification on the feasibility conclusions on ventilator use based on the rule validation results, consistency validation results, and rationality validation results, outputting the feasibility conclusions on ventilator use if no risky or contradictory conclusions appear in the feasibility conclusions on ventilator use, otherwise intercepting the feasibility conclusions on ventilator use, and returning to re-evaluate the dynamic indicators of the multimodal fusion features to obtain the feasibility conclusions on ventilator use again.

[0052] It is understood that this embodiment can be executed based on a multi-agent collaborative reasoning framework. This framework can include a data processing agent, a reasoning and decision-making agent, and a reflective verification agent. The data processing agent acquires multimodal clinical information of the target object; extracts clinical features from the multimodal clinical information to obtain multimodal clinical features; and performs unidirectional causal fusion of the multimodal clinical features based on the dynamic causal relationships between them to obtain multimodal fused features. The reasoning and decision-making agent evaluates the multimodal fused features based on dynamic indicator conditions to obtain the target object's multimodal indicator score and abnormal indicator information; it infers the pathophysiological mechanism of the abnormal indicator information to obtain the pathophysiological causal chain corresponding to the abnormal indicator information; and it assesses the target object's offline risk based on the multimodal indicator score, abnormal indicator information, and pathophysiological causal chain to obtain an offline risk conclusion for the target object. The reflective verification agent performs cross-validation of the offline feasibility conclusion to determine whether to output an offline feasibility conclusion.

[0053] The intensive care unit (ICU) decision-making method for difficult weaning provided in this embodiment performs unidirectional causal association fusion of multimodal clinical features based on the dynamic causal relationship between multimodal features to obtain the multimodal fusion features of the target object; dynamically evaluates the multimodal fusion features to obtain multimodal indicator scores and abnormal indicator information, and infers the pathophysiological mechanism of the abnormal indicator information to obtain the pathophysiological causal chain; and assesses the weaning risk based on the multimodal indicator scores, abnormal indicator information, and pathophysiological causal chain to obtain the weaning risk conclusion of the target object.

[0054] Compared with related technologies, this application can conduct a weaning risk assessment on the target subject before initiating the spontaneous breathing test, pre-determine the weaning risk of the target subject, and then determine whether to initiate the spontaneous breathing test on the target subject based on the weaning risk, thereby reducing the potential harm of the spontaneous breathing test to the target subject and ensuring the physiological and psychological safety of the target subject.

[0055] Furthermore, this application performs unidirectional causal association fusion on the multimodal clinical characteristics of the target object based on dynamic causal relationships, so as to fuse the contextual information between multiple modalities across modalities, improve the semantic richness of multimodal indicators, and thus improve the accuracy and authenticity of the offline risk conclusion.

[0056] Reference Figure 2 As shown, in one embodiment of this application, multimodal clinical features include multiple clinical features corresponding to different modalities; based on the dynamic causal relationship between multimodal clinical features, causal unidirectional association fusion is performed on the multimodal clinical features to obtain multimodal fused features, including: S210. Obtain the causal direction among multiple clinical features.

[0057] S220. Evaluate the fusion weights of multiple clinical features to obtain the causal fusion weights among the multiple clinical features.

[0058] S230. Based on the causal influence direction, multiple clinical features are sequentially causally fused according to the causal fusion weight to obtain the fusion information corresponding to each of the multiple clinical features, which serves as the multimodal fusion feature.

[0059] Specifically, the causal influence direction can be a fixed direction representing the contextual dependencies between different modalities, or it can be obtained by prioritizing causal relationships among multimodal clinical features based on a pre-defined knowledge graph or medical knowledge base. It is understandable that anomalies in the target object typically first appear in the modality with the smallest temporal granularity, and then sequentially appear in other modalities as the temporal granularity increases. Therefore, the causal influence direction can be that the speech modality precedes the image modality, and the image modality precedes the text modality. For example, the target object may first exhibit abnormal breath sounds in the speech modality, followed by visible diaphragmatic displacement in the image modality, while the medical record text in the text modality lags behind the speech and image modalities. A causal mask is generated along the causal influence direction to block reverse information leakage between modalities, ensuring that the fusion method between multimodal clinical features conforms to the logic of clinical information generation.

[0060] Furthermore, a fusion weight assessment is performed on multiple clinical features within the multimodal clinical features to obtain the association strength between them, and the causal fusion weights between the multiple clinical features are determined based on the association strength. The fusion weight assessment of multimodal clinical features can be performed using an attention mechanism, calculating attention weights between any two modalities to obtain initial weights between those two modalities. After obtaining the initial weights, a causal mask is applied to correct the initial weights, obtaining intermediate weights along the causal influence direction, and these intermediate weights are then normalized to obtain the causal fusion weights.

[0061] Furthermore, in terms of causal influence, multiple clinical features are sequentially causally fused according to causal fusion weights. This involves fusing each clinical feature with clinical features that have a causal influence on it, resulting in fusion information corresponding to each clinical feature, which serves as the multimodal fusion feature. The sequential causal fusion process is illustrated exemplarily. The multimodal clinical features include a first feature, a second feature, and a third feature. The causal influence direction is that the first feature precedes the second feature, and the second feature precedes the third feature. A first fusion weight corresponds to the first feature and the second feature, and a second fusion weight corresponds to the second feature and the third feature. In the sequential causal fusion process, firstly, the first feature is fused to the second feature according to the first fusion weight, resulting in a second fusion feature enhanced by the first feature; secondly, the second fusion feature is fused to the third feature according to the second fusion weight, resulting in a third fusion feature enhanced by the second fusion feature; and finally, the multimodal fusion feature is obtained based on the first feature, the second fusion feature, and the third fusion feature.

[0062] It is understandable that the third fusion feature in the above example implicitly includes the associated features of the first, second and third features. These associated features are used to enhance the third feature, thereby providing a rich cross-modal semantic background for the third feature, improving the feature richness of the modality corresponding to the third feature, and making the offline risk conclusion closer to the actual clinical situation of the target object.

[0063] Reference Figure 3 As shown, in one embodiment of this application, a fusion weight evaluation is performed on multiple clinical features to obtain the causal fusion weights among the multiple clinical features, including: S222. Obtain the background characteristics and offline risk baseline score of the target subject; wherein, the offline risk baseline score is obtained by multidimensional contribution weighting assessment based on the target subject's physiological index data.

[0064] S224. Using background features and baseline scores for off-line risk as weighted prior constraints, multiple clinical features are weighted according to the weighted prior constraints to obtain causal fusion weights.

[0065] Specifically, the background characteristics and baseline score for weaning risk of the target subject are obtained. Background characteristics may include, but are not limited to, information such as the target subject's disease type, severity, medical history, and organ function status. The baseline score for weaning risk can be obtained through a multidimensional contribution weighted assessment of the target subject's physiological indicators. These physiological indicators may include, but are not limited to, the target subject's baseline respiratory mechanics indicators, oxygenation reserve indicators, electrolyte status, neuromuscular function assessment, and comorbidity burden. Baseline respiratory mechanics indicators may include the shallow and rapid breathing index and maximum inspiratory pressure, while oxygenation reserve indicators may include the oxygenation index and blood oxygen saturation trend. The physiological indicators can be extracted from the target subject's multimodal clinical information. For example, baseline respiratory mechanics indicators and oxygenation reserve indicators may be obtained from respiratory sound signals in the speech modality and monitoring data in the imaging modality; electrolyte status and neuromuscular function assessment may be obtained from laboratory test records and medical orders in the text modality; and comorbidity burden may be obtained from diagnostic records and underlying disease annotations in the text modality.

[0066] In some embodiments, the process of obtaining the offline risk baseline score may include: using the causal fusion weight corresponding to the modality of the physiological indicator data as the risk fusion weight of the physiological indicator data; and performing a weighted calculation on the physiological indicator data according to the risk fusion weight to obtain the offline risk baseline score. It is understood that the modified fusion weight corresponding to the modality can also be used as the risk fusion weight of the physiological indicator data.

[0067] Furthermore, factors leading to weaning difficulties in different target subjects can exist in different modalities, and the same target subject may also experience weaning difficulties due to different modalities at different time periods. Considering these factors, this embodiment uses the target subject's background characteristics and baseline weaning risk score as weighted prior constraints. Multiple clinical characteristics are weighted according to these background characteristics and baseline weaning risk scores to obtain the causal fusion weights corresponding to each clinical characteristic. For example, for target subjects with chronic obstructive pulmonary disease (COPD), the weighted prior constraint could be to increase the causal fusion weights of the target subject on respiratory mechanics indicators (such as the shallow and rapid breathing index and maximum inspiratory pressure). For target subjects with heart failure, the weighted prior constraint could be to increase the causal fusion weights of the target subject on oxygenation and fluid load-related indicators.

[0068] In some embodiments, the weight evaluation process based on weight prior constraints can be performed using a pre-trained indicator weight allocation network. The weight prior constraints are input into the indicator weight allocation network, and causal fusion weights are output based on these constraints. The indicator weight allocation network can be trained using a multi-task learning framework. The training data for the network can be historical offline assessment data, including historical indicator data and historical offline outcomes. The offline outcomes corresponding to the historical offline assessment data are used as supervision signals for network training, enabling the indicator weight allocation network to learn the contribution of historical indicator data to historical offline outcomes. This allows it to output causal fusion weights corresponding to each clinical feature during the actual weight evaluation process.

[0069] The indicator weight allocation network can also be connected to the clinical rule hard constraint layer to verify the output results of the indicator weight allocation network according to the basic requirements of medical guidelines, ensuring that the output results of the indicator weight allocation network do not violate the basic requirements of evidence-based medicine guidelines, and achieving dual protection of data-driven and rule-constrained approaches.

[0070] Understandably, weighted prior constraints can indicate which clinical characteristics of the target subject are dominant and contribute significantly to the target subject's offline risk assessment. By introducing weighted prior constraints and evaluating the weights based on these constraints to obtain causal fusion weights, the relevance of the causal fusion weights to the target subject is improved. This allows for unidirectional causal association fusion of multimodal clinical characteristics based on the target subject's individual features, significantly enhancing the fit between multimodal fusion features and the target subject, as well as the ability to express the target subject's current state.

[0071] Reference Figure 4As shown in one embodiment of this application, after obtaining the fusion information corresponding to multiple clinical features, the method further includes: S240. Obtain historical indicator characteristics of the target object.

[0072] S250. Based on the multimodal clinical features, historical correlations are identified in the historical indicator features to obtain the associated indicator features; trend decomposition is performed on the associated indicator features to obtain the multi-scale trend features of the associated indicator features.

[0073] S260. Feature splicing is performed on the fusion information corresponding to the multi-scale trend features and multiple clinical features to obtain multimodal fusion features.

[0074] The historical indicator features can be obtained by feature extraction from historical indicator data. This historical indicator data can be a complete physiological indicator sequence of the target object during intensive care, including but not limited to time-series data such as respiratory rate, blood oxygen saturation, blood pressure, and heart rate. The historical indicator data can be stored in an external memory matrix and updated in real time according to changes in the target object's state. In some embodiments, before storing the historical indicator data in the external memory matrix, a memory compression network can be used to reduce the dimensionality of the historical indicator data, obtaining an implicit state representation of the historical indicator data, and extracting the key temporal states of the target object from the implicit state representation for storage in the external memory matrix. The memory compression network can be a network architecture that improves the efficiency of temporal data processing by transforming long-term temporal data into temporal features through mechanisms such as compression, filtering, or selective storage. This can include, but is not limited to, explicit memory enhancement architectures, compressed memory frameworks, and generative adversarial network compression. Methods for extracting key temporal states can include, but are not limited to, gating networks, attention mechanisms, or similarity calculations. Understandably, by performing dimensionality reduction encoding and key state extraction on historical indicator data, the storage resources required for historical indicator data are effectively reduced, the processing efficiency of historical indicator data is improved, and the computational bottleneck caused by long sequence input is alleviated.

[0075] Specifically, based on the current moment corresponding to the multimodal clinical features, historical indicator data of the target object from multiple historical moments prior to the current moment are retrieved from the external memory matrix, and historical indicator features corresponding to the historical indicator features are extracted. After obtaining the historical indicator features, the multimodal clinical features of the target object are used as query vectors, and the historical indicator features are used as key and value vectors. A temporal attention mechanism is used to calculate the correlation between the multimodal clinical features and the historical indicator features, resulting in a correlation weight matrix between the multimodal clinical features and the historical indicator features. The historical indicator features are then filtered for correlation based on the correlation weight matrix, and the filtered historical indicator features are used as associated indicator features. For example, associated indicator features can be multiple historical indicator features that contribute significantly to the multimodal clinical features at the current moment, such as electrolyte abnormalities that occurred in the target object several days ago, or previous failed spontaneous breathing tests of the target object.

[0076] Furthermore, after obtaining the characteristics of the correlated indicators, trend decomposition is performed on these characteristics to obtain multi-scale trend features. This allows for the determination of the evolution direction of the correlated indicators at different time scales, enabling refined modeling of the indicator evolution direction. It is understood that multi-scale trend features can include long-term trend terms, cyclical fluctuation terms, and short-term residual terms of the correlated indicators, where the time scale of the long-term trend term is longer than that of the cyclical fluctuation term, and the time scale of the cyclical fluctuation term is longer than that of the short-term residual term. For example, specific methods of trend decomposition may include, but are not limited to, seasonal-trend decomposition using LOESS (STL), wavelet transform, and empirical mode decomposition.

[0077] Furthermore, for the fusion information corresponding to each of the multiple clinical features obtained through sequential causal fusion, the multi-scale trend features are concatenated with the aforementioned fusion information to perform historical enhancement on the fusion information of each modality, thereby obtaining multimodal fusion features. It is understood that in this embodiment, multimodal fusion features are obtained through historical enhancement, enabling them to reflect not only the clinical indicator features at any given time but also the changing trends of the clinical indicator features, thus achieving early prediction of the offline risk of the target subject.

[0078] Reference Figure 5 As shown, in one embodiment of this application, dynamic index conditions are obtained in the following manner: S310. Obtain the background features and real-time evolution data of the target object.

[0079] S320. Perform case matching based on background features to obtain similar background cases that are similar to the target object, and filter out successful offline cases from the similar background cases.

[0080] S330. Perform fusion statistics on the actual indicator conditions corresponding to successful offline cases to obtain similar indicator conditions, and dynamically update the similar indicator conditions based on real-time evolution data to obtain dynamic indicator conditions.

[0081] Among them, real-time evolution data can be the physiological data of the target object after the current moment, which is continuously accumulated as the intensive care time increases, indicating the trend of the target object's condition changes.

[0082] Specifically, background features of the target object are obtained, and a subgroup matching model is established based on these features. This model may include disease type matching, severity matching, and comorbidity background matching, used to retrieve cases with similar disease backgrounds to the target object. The subgroup matching model is then used to match cases in a historical case database, resulting in multiple historical cases with high similarity to the target object, which are considered similar background cases. In some embodiments, the specific method for case matching may be K-nearest neighbor matching.

[0083] Furthermore, cases with successful offline outcomes are selected from similar background cases to obtain successful offline cases. For successful offline cases, the corresponding actual indicator conditions are extracted, and the actual indicator conditions are fused and statistically analyzed to obtain similar indicator conditions applicable to the target object. In some embodiments, the specific method of fusion and statistical analysis may include: constructing an actual achievement distribution interval based on the actual indicator conditions corresponding to multiple successful offline cases, and extracting feature values ​​within the actual achievement distribution interval to obtain similar indicator conditions. The maximum value of the actual achievement distribution interval can be the maximum value among the actual indicator conditions, the minimum value of the actual achievement distribution interval can be the minimum value among the actual indicator conditions, and feature value extraction may involve extracting special values ​​such as median or average values ​​from the actual achievement distribution interval. In other embodiments, when real-time evolution data is not available, similar indicator conditions can also be used as dynamic indicator conditions to evaluate the multimodal fusion features of the target object.

[0084] Understandably, the actual indicator conditions can be multiple indicator conditions used to evaluate the case subjects in successful offline cases. The case subjects meet the actual indicator conditions in successful offline cases, resulting in an offline risk conclusion indicating a recommendation to go offline. Therefore, the actual indicator conditions of successful offline cases have reference value for target subjects. When the target subjects meet similar indicator conditions obtained from the actual indicator conditions, the probability of successful offline operation is relatively high.

[0085] Furthermore, based on the similarity index conditions, real-time evolution data of the target object is obtained, and the posterior estimation of the similarity index conditions is performed using the real-time evolution data to obtain the posterior estimation distribution. Based on a preset acceptable risk level, an inverse equation for index update is constructed using the posterior estimation distribution and the acceptable risk level, and the inverse equation for index update is solved to obtain the dynamic index conditions, thereby updating the similarity index conditions in real time. In some embodiments, the specific method of posterior estimation may be to use a Bayesian online update mechanism, calculating the posterior distribution based on the real-time evolution data and the similarity index conditions to obtain the posterior estimation distribution, which is then used to adjust the similarity index conditions.

[0086] Compared with the evaluation method using fixed threshold conditions in related technologies, this embodiment generates similar indicator conditions based on the actual indicator conditions of successful offline cases with similar background characteristics, and performs posterior dynamic updates on the similar indicator conditions based on the real-time evolution data of the target object. This improves the relevance of the similar indicator conditions to the target object, enhances the adaptability of the indicator evaluation process to the target object, and thus improves the accuracy of the offline risk conclusion.

[0087] Reference Figure 6 As shown, in one embodiment of this application, pathophysiological mechanism inference is performed on abnormal indicator information to obtain the pathophysiological causal chain corresponding to the abnormal indicator information, including: S340. Perform mechanism reasoning on the abnormal indicator information to obtain the initial causal chain.

[0088] S350. When there is more than one initial causal chain and the initial causal chains contradict each other, obtain the prior confidence of the initial causal chains and the real-time evolution data of the target object.

[0089] S360. Perform a path consistency comparison on the initial causal chain based on the real-time evolution data, and adjust the prior confidence based on the results of the path consistency comparison to obtain the causal confidence of the initial causal chain.

[0090] S370. Based on the causal confidence level, a voting decision is made among the initial causal chains to obtain the pathophysiological causal chain.

[0091] Specifically, when the target object has abnormal indicator information, the pathophysiological mechanism of the abnormal indicator information is inferred to obtain the initial causal chain corresponding to the abnormal indicator information. It can be understood that the initial causal chain corresponds to the abnormal indicator information, representing the pathophysiological logic that caused the abnormal indicator information, and its number is the same as the number of abnormal indicator information.

[0092] In some embodiments, when inferring the pathophysiological mechanisms of abnormal indicator information, a clinical natural language summary can be simultaneously input to assist in the indicator evaluation process. In this embodiment, the clinical natural language summary can be a qualitative clinical description in text form, used to describe the clinical disease background of the target object from a text modality perspective, so that the mechanism inference process can be combined with the target object's medical history, comorbidities, and treatment process, thereby improving the completeness of the initial causal chain.

[0093] Furthermore, when there is more than one initial causal chain, and the offline outcomes pointed to by the initial causal chains contradict each other, the prior confidence of the initial causal chains and the real-time evolution data of the target object at subsequent time points are obtained, where the prior confidence can represent the reliability of the conclusions of each initial causal chain. In some embodiments, the prior confidence can be obtained by using the causal fusion weight of the modality corresponding to the anomaly index information corresponding to the initial causal chain as the prior confidence. If there is a modified fusion weight, the modified fusion weight can also be used as the prior confidence.

[0094] Furthermore, trend analysis is performed on the real-time evolution data to obtain the real-time evolution trend of the target object. A consistency comparison is then performed between the real-time evolution trend and the offline outcome pointed to by the initial causal chain to obtain the trend comparison result. For example, when the offline outcome pointed to by the initial causal chain is a suggested offline state, and the real-time evolution trend indicates that the real-time evolution data of the target object is changing in a direction that satisfies the dynamic indicator conditions, a trend comparison result indicating that the real-time evolution data and the initial causal chain have a consistent trend can be obtained. When the offline outcome pointed to by the initial causal chain is a suggested offline state, and the real-time evolution trend indicates that the real-time evolution data of the target object is changing in a direction that does not satisfy the dynamic indicator conditions, a trend comparison result indicating that the real-time evolution data and the initial causal chain have a contradictory trend can be obtained.

[0095] After obtaining the trend comparison results, the prior confidence of the initial causal chain is adjusted based on these results to obtain the causal confidence of the initial causal chain. It is understandable that when the trend comparison results indicate a consistent trend between the real-time evolution data and the initial causal chain, the conclusion regarding offline risk pointed to by the initial causal chain can be considered credible, and the prior confidence of the initial causal chain can be increased, giving it a larger voting weight in the voting decision. Conversely, when the trend comparison results indicate a contradictory trend between the real-time evolution data and the initial causal chain, the conclusion regarding offline risk pointed to by the initial causal chain can be considered unreliable, and the prior confidence of the initial causal chain can be decreased, giving it a smaller voting weight in the voting decision.

[0096] Furthermore, the causal confidence level is used as the voting weight of the initial causal chain in the voting decision. Based on the voting weight, the offline outcome pointed to by all initial causal chains is voted on to determine the offline outcome with higher confidence, and the initial causal chain that contradicts the offline outcome is screened out. The initial causal chain corresponding to the offline outcome is then used as the pathophysiological causal chain.

[0097] It should be noted that the initial causal chain may contain high-risk causal chains, which can represent a high risk of the target object going offline. If at least one high-risk causal chain exists in the initial causal chain, the confidence of the high-risk causal chain is verified based on the prior confidence level, resulting in a high-risk confidence verification result. If the high-risk confidence verification result indicates that the prior confidence of all high-risk causal chains does not meet the pre-set confidence requirement, then the high-risk causal chains are unreliable, and the target object has a low risk of going offline. If the prior confidence of any high-risk causal chain meets the pre-set confidence requirement, then at least one high-risk causal chain is reliable, and the target object is determined to have a high risk of going offline.

[0098] When a target subject faces a high risk of being weaned off ventilator infection, safety restrictions are imposed on the weaning outcome pointed to by the pathophysiological causal chain. This ensures that the weaning risk conclusion cannot be a recommendation to wean off ventilator infection, but only a recommendation to delay or not recommend weaning off ventilator infection. It is understandable that if high-risk causal chains are blindly eliminated through voting when a target subject may face a high risk of weaning off ventilator infection, erroneous conclusions regarding the risk of weaning off ventilator infection may be reached, posing a safety risk to the target subject. Therefore, when the prior confidence of a high-risk causal chain meets the pre-set confidence requirement, a priority weight is assigned to that high-risk causal chain to safety-restrict the weaning risk conclusion and safeguard the life safety of the target subject.

[0099] Reference Figure 7 As shown in one embodiment of this application, the method further includes: S510. Obtain real-time evolution data of the target object.

[0100] S520. Conduct an offline risk assessment of the target object based on real-time evolution data to obtain a real-time risk conclusion for the target object.

[0101] S530. If the real-time risk conclusion and the offline risk conclusion are consistent, enhance the confidence of the offline risk conclusion; otherwise, identify abnormal indicators in the real-time evolution data to obtain real-time abnormal information, and update the pathophysiological causal chain based on the real-time abnormal information.

[0102] Specifically, after determining the offline risk, data monitoring of the target object is continuously performed in subsequent timeframes to obtain real-time evolution data of the target object at those subsequent timeframes. An offline risk assessment is then conducted on the target object based on this real-time evolution data to obtain a real-time risk conclusion for that target object at that subsequent timeframe. Here, "subsequent timeframes" can be times after obtaining the offline risk conclusion but before the actual execution of the offline process. It is understood that the specific method for assessing the offline risk of the target object based on real-time evolution data can be the same as or different from the method used to obtain the offline risk conclusion.

[0103] In some embodiments, the specific process of obtaining a real-time risk conclusion may include: performing change detection on the real-time evolution data at subsequent time points; when the degree of change of any indicator in the real-time evolution data meets a preset change threshold, determining that the target object has undergone a significant state change at that subsequent time point, and the offline risk conclusion needs to be updated. Data features are extracted from the real-time evolution data to obtain real-time evolution features, and the multimodal fusion features are incrementally updated using the real-time evolution features to obtain updated fusion features. Specific methods for incrementally updating the multimodal fusion features using the real-time evolution features may include weighted averaging, feature concatenation, or temporal updates of the real-time evolution features and the multimodal fusion features. The offline risk conclusion is locally corrected based on the updated fusion features to obtain the real-time risk conclusion for the target object at that subsequent time point. Specific methods for locally correcting the offline risk conclusion based on the updated fusion features may include, but are not limited to, using residual correction models or Bayesian incremental updates.

[0104] Furthermore, the real-time risk conclusion and the offline risk conclusion are compared to determine whether they point to the same offline outcome. If the real-time and offline risk conclusions are consistent, it indicates that the real-time evolution data did not introduce substantial changes in the patient's condition, and the target subject remained stable in subsequent timeframes. This allows for confidence enhancement of the offline risk conclusion, ensuring its stability. If the real-time and offline risk conclusions deviate, it indicates that the real-time evolution data introduced substantial changes in the patient's condition, and the target subject experienced significant fluctuations in their condition in subsequent timeframes. In this case, anomaly indicators are identified in the real-time evolution features to obtain real-time anomaly information. Based on this anomaly information, incremental mechanism inference is performed to obtain a real-time causal chain used to update the pathophysiological causal chain. This real-time causal chain can be used to concatenate with the pathophysiological causal chain to supplement causal chain nodes not included in the original pathophysiological causal chain. The real-time causal chain can also be used to replace the pathophysiological causal chain to correct cognitive biases present in the original pathophysiological causal chain.

[0105] It is understandable that the updated offline risk conclusions or pathophysiological causal chains can be cross-validated again to form a complete closed loop of update-correction-verification, ensuring the timeliness and accuracy of the offline risk conclusions and pathophysiological causal chains. In this embodiment, the offline risk conclusions and pathophysiological causal chains are optimized and verified in real time based on the real-time evolution data of the target object at subsequent times, improving the timeliness of the offline risk conclusions and pathophysiological causal chains, ensuring that the offline risk conclusions and pathophysiological causal chains can reflect the real-time state of the target object, and guaranteeing the accuracy of the offline risk conclusions. In addition, this embodiment also uses real-time evolution data to locally correct the offline risk conclusions and pathophysiological causal chains, fully integrating the real-time data and historical data of the target object. On the one hand, this reduces the computational resource requirements and improves the timeliness of conclusion correction; on the other hand, it reduces the possibility of the offline risk conclusions being overturned due to short-term fluctuations in a single piece of real-time evolution data, ensuring the stability of the offline risk conclusions.

[0106] This application also provides an embodiment of a decision-making method for difficult weaning in intensive care. In this embodiment, the target subject is a 68-year-old male patient receiving mechanical ventilation. The target subject's BMI is 24.3 kg / m², and the duration of mechanical ventilation received is 72 hours.

[0107] Multimodal clinical information was obtained for this target group, including structured data, unstructured data, and modal data. The structured data included: shallow and rapid breathing index (RSBI): 82 breaths / min·L; Glasgow Coma Scale (GCS) score: 15 points; delirium score (CAM-ICU): negative (0 points); use of vasoactive drugs: none; lactate: 1.6 mmol / L; diaphragmatic range of motion (DE): 2.3 cm; serum albumin: 32 g / L; serum magnesium: 0.75 mmol / L; oxygenation index (PaO2 / FiO2): 280 mmHg; B-type natriuretic peptide (BNP): 180 pg / mL.

[0108] Unstructured data included: Nursing record: "The patient is alert, with regular spontaneous breathing rhythm, moderate airway secretions, and oxygen saturation maintained at 95%-98% after suctioning; urine output is normal today, and fluid balance is -150mL."; Ward rounds log: "The patient was admitted for severe community-acquired pneumonia. The underlying condition is chronic obstructive pulmonary disease (GOLD 2). After anti-infection and respiratory support treatment, the pulmonary infection symptoms have improved, respiratory function is gradually recovering, and there are no obvious signs of dyspnea."

[0109] Modal data can include imaging and audio data. The imaging data report summary includes: Chest imaging report: "Scattered inflammatory foci in both lungs, reduced in size compared to admission; lung ventilation is basically uniform, with no obvious atelectasis or pleural effusion, and the airway is patent."; Respiratory waveform analysis: Respiratory rate fluctuation range is 15 breaths / min, tidal volume fluctuation range is 500 mL, respiratory rate variability coefficient is 7%, with no significant abnormal fluctuations. The audio data report summary includes: Breath sound analysis: Breath sounds are symmetrical in both lungs during spontaneous breathing, with no obvious dry or wet rales heard; inspiratory breath sound intensity is stable, with no inspiratory wheezing; respiratory rate is consistent with waveform monitoring results; Speech assessment: The patient speaks clearly as instructed, without significant interruption of breathing during utterance, and with good speech coherence, indicating that respiratory muscle reserve can meet basic vocalization needs.

[0110] Based on the method provided in this application, the target subject was evaluated using the aforementioned multimodal clinical information to obtain multimodal indicator scores and abnormal indicator information. The multimodal indicator scores included: primary disease and acute triggers: acute trigger: severe community-acquired pneumonia; primary disease / major mechanism: chronic obstructive pulmonary disease (GOLD grade 2); strong contraindication evidence: no strong contraindication evidence such as late-stage irreversible primary disease or central respiratory dysfunction; core weaning indicator achievement: all extracted core weaning indicators met the preset achievement standards; information sufficiency: information was sufficient, no core weaning indicators were missing, and information on the primary disease and acute triggers was complete. Abnormal indicator information included: reversible nutritional abnormalities (low protein), reversible electrolyte abnormalities (low magnesium), and moderate amounts of airway secretions. Based on the above abnormal indicators, the following treatment recommendations can be given: ① Reversible nutritional abnormalities (low protein): albumin supplementation is recommended, with a target value ≥35g / L; ② Reversible electrolyte abnormalities (low magnesium): magnesium supplementation is recommended, with a target value ≥0.85mmol / L; ③ Moderate airway secretions: close monitoring is required, and airway clearance capacity should be fully assessed before extubation. Based on the above multimodal indicator scores and abnormal indicator information, a weaning risk assessment of the target subject can be conducted, resulting in a weaning risk conclusion indicating recommended weaning.

[0111] Accordingly, please refer to Figure 8 This application provides a decision-making device for difficult weaning in intensive care, the device comprising: The clinical feature acquisition module 810 is used to acquire multimodal clinical features of the target object.

[0112] The cross-modal causal fusion module 820 is used to perform unidirectional causal association fusion of multimodal clinical features based on the dynamic causal relationship between multimodal clinical features to obtain multimodal fused features.

[0113] The indicator evaluation and reasoning module 830 is used to evaluate the multimodal fusion features based on dynamic indicator conditions, obtain the multimodal indicator scores and abnormal indicator information of the target object, and perform pathophysiological mechanism reasoning on the abnormal indicator information to obtain the pathophysiological causal chain corresponding to the abnormal indicator information.

[0114] The decision suggestion generation module 840 is used to assess the offline risk of the target object based on multimodal index scores, abnormal index information and pathophysiological causal chain, and to obtain the offline risk conclusion of the target object.

[0115] In some alternative implementations, the cross-modal causal fusion module 820 includes: The causal direction acquisition unit is used to acquire the causal influence direction between multiple clinical features.

[0116] The fusion weight assessment unit is used to assess the fusion weight of multiple clinical features to obtain the causal fusion weight among the multiple clinical features.

[0117] The sequential causal fusion unit is used to sequentially fuse multiple clinical features based on the causal influence direction and according to the causal fusion weight, so as to obtain the fusion information corresponding to each of the multiple clinical features as multimodal fusion features.

[0118] In some optional implementations, the fusion weight evaluation unit includes: The target information acquisition subunit is used to acquire the background characteristics and offline risk baseline score of the target object; the offline risk baseline score is obtained by multidimensional contribution weighting evaluation based on the target object's physiological index data.

[0119] The constraint weight allocation subunit is used to use background features and offline risk baseline scores as weight prior constraints, and to evaluate the weights of multiple clinical features based on the weight prior constraints to obtain causal fusion weights.

[0120] In some alternative implementations, the cross-modal causal fusion module 820 further includes: The physiological data acquisition unit is used to acquire historical indicator characteristics of the target object.

[0121] The historical association identification unit is used to identify historical associations in historical indicator features based on multimodal clinical features to obtain associated indicator features; and to perform trend decomposition on associated indicator features to obtain multi-scale trend features of associated indicator features.

[0122] The trend feature stitching unit is used to stitch together the fusion information corresponding to multi-scale trend features and multiple clinical features to obtain multimodal fusion features.

[0123] In some optional implementations, the indicator evaluation inference module 830 includes: The target data acquisition unit is used to acquire the background features and real-time evolution data of the target object.

[0124] The similar case matching unit is used to match cases based on background features, obtain similar background cases that are similar to the target object, and filter out offline success cases from the similar background cases.

[0125] The dynamic condition acquisition unit is used to integrate and statistically analyze the actual indicator conditions corresponding to successful offline cases to obtain similar indicator conditions, and dynamically update the similar indicator conditions based on real-time evolution data to obtain dynamic indicator conditions.

[0126] In some optional implementations, the indicator evaluation inference module 830 further includes: The anomaly mechanism reasoning unit is used to perform mechanism reasoning on anomaly indicator information to obtain the initial causal chain.

[0127] The causal information acquisition unit is used to acquire the prior confidence of the initial causal chain and the real-time evolution data of the target object when the number of initial causal chains exceeds one and the initial causal chains contradict each other.

[0128] The causal confidence correction unit is used to perform a direction consistency comparison on the initial causal chain based on real-time evolution data, and to correct the prior confidence based on the result of the direction consistency comparison to obtain the causal confidence of the initial causal chain.

[0129] The causal chain voting adjudication unit is used to vote and adjudicate among the initial causal chains based on causal confidence to obtain the pathophysiological causal chain.

[0130] In some alternative implementations, the device further includes a conclusion dynamic update module, comprising: The real-time data acquisition unit is used to acquire real-time evolution data of the target object.

[0131] The offline risk assessment unit is used to assess the offline risk of the target object based on real-time evolution data and obtain a real-time risk conclusion for the target object.

[0132] The conclusion comparison and correction unit is used to enhance the confidence of the offline risk conclusion when the real-time risk conclusion and the offline risk conclusion are consistent. Otherwise, it identifies abnormal indicators in the real-time evolution data, obtains real-time abnormal information, and updates the pathophysiological causal chain based on the real-time abnormal information.

[0133] Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0134] In this embodiment, the decision-making device for difficult weaning in intensive care is presented in the form of a functional unit. Here, a unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0135] Please see Figure 9 , Figure 9 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application, such as... Figure 9 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 9 Take a processor 10 as an example.

[0136] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0137] The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the method shown in the above embodiments.

[0138] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0139] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0140] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0141] This application also provides a computer-readable storage medium. The methods described in this application can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code downloaded over a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and subsequently stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the methods shown in the above embodiments are implemented.

[0142] This application provides a computer program product including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the method of any embodiment of this application.

[0143] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and all such modifications and variations fall within the scope defined by the appended claims.

[0144] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0145] For ease of description, the above devices are described separately by function as various units. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware.

[0146] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0147] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0148] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0149] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0150] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0151] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0152] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

[0153] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A decision-making method for difficult weaning in intensive care, characterized in that, The method includes: Obtain multimodal clinical characteristics of the target subjects; Based on the dynamic causal relationships among the multimodal clinical features, the multimodal clinical features are fused by unidirectional causal association to obtain multimodal fused features; The multimodal fusion features are evaluated based on dynamic index conditions to obtain the multimodal index score and abnormal index information of the target object; the abnormal index information is used to infer the pathophysiological mechanism to obtain the pathophysiological causal chain corresponding to the abnormal index information. The target object's risk of disconnection is assessed based on the multimodal index score, the abnormal index information, and the pathophysiological causal chain, and a conclusion on the target object's risk of disconnection is obtained.

2. The method according to claim 1, characterized in that, The multimodal clinical features include multiple clinical features corresponding to different modalities; the step of fusing the multimodal clinical features based on the dynamic causal relationships between them to obtain multimodal fused features includes: Obtain the causal relationship between the multiple clinical features; The causal fusion weights among the multiple clinical features are obtained by evaluating the fusion weights of the multiple clinical features. Based on the causal influence direction, the multiple clinical features are sequentially causally fused according to the causal fusion weight to obtain the fusion information corresponding to each of the multiple clinical features, which serves as the multimodal fusion feature.

3. The method according to claim 2, characterized in that, The step of evaluating the fusion weights of the multiple clinical features to obtain the causal fusion weights among the multiple clinical features includes: The background characteristics and offline risk baseline score of the target object are obtained; wherein, the offline risk baseline score is obtained by multidimensional contribution weighted evaluation based on the physiological index data of the target object; The background features and the baseline score of offline risk are used as weighted prior constraints. The weights of the multiple clinical features are evaluated based on the weighted prior constraints to obtain the causal fusion weights.

4. The method according to claim 2, characterized in that, After obtaining the fusion information corresponding to each of the multiple clinical features, the method further includes: Obtain the historical indicator features of the target object; Based on the multimodal clinical features, historical correlations are identified in the historical indicator features to obtain correlated indicator features; trend decomposition is performed on the correlated indicator features to obtain multi-scale trend features of the correlated indicator features. The multimodal fusion features are obtained by concatenating the fusion information corresponding to the multi-scale trend features and the multiple clinical features respectively.

5. The method according to claim 1, characterized in that, The dynamic index conditions are obtained in the following way: Obtain the background features and real-time evolution data of the target object; Case matching is performed based on the background features to obtain similar background cases that are similar to the target object, and offline success cases are selected from the similar background cases. The actual indicator conditions corresponding to the successful offline cases are fused and statistically analyzed to obtain similar indicator conditions, and the similar indicator conditions are dynamically updated according to the real-time evolution data to obtain the dynamic indicator conditions.

6. The method according to claim 1, characterized in that, The step of inferring the pathophysiological mechanism of the abnormal indicator information to obtain the pathophysiological causal chain corresponding to the abnormal indicator information includes: Mechanism reasoning is performed on the abnormal indicator information to obtain an initial causal chain; If the number of initial causal chains exceeds one and the initial causal chains contradict each other, obtain the prior confidence of the initial causal chains and the real-time evolution data of the target object; The initial causal chain is compared for consistency based on the real-time evolution data, and the prior confidence is corrected based on the results of the consistency comparison to obtain the causal confidence of the initial causal chain. The pathophysiological causal chain is obtained by voting among the initial causal chains based on the causal confidence level.

7. The method according to claim 1, characterized in that, The method further includes: Obtain real-time evolution data of the target object; Based on the real-time evolution data, an offline risk assessment is performed on the target object to obtain a real-time risk conclusion for the target object; If the real-time risk conclusion and the offline risk conclusion are consistent, the confidence level of the offline risk conclusion is increased; otherwise, anomaly indicators are identified in the real-time evolution data to obtain real-time anomaly information, and the pathophysiological causal chain is updated according to the real-time anomaly information.

8. A decision-making device for difficult weaning in intensive care, characterized in that, The device includes: The clinical feature acquisition module is used to acquire multimodal clinical features of the target object; The cross-modal causal fusion module is used to perform unidirectional causal association fusion on the multimodal clinical features based on the dynamic causal relationship between the multimodal clinical features to obtain multimodal fused features; The indicator evaluation and reasoning module is used to evaluate the multimodal fusion features based on dynamic indicator conditions to obtain the multimodal indicator score and abnormal indicator information of the target object; and to perform pathophysiological mechanism reasoning on the abnormal indicator information to obtain the pathophysiological causal chain corresponding to the abnormal indicator information. The decision suggestion generation module is used to assess the offline risk of the target object based on the multimodal index score, the abnormal index information and the pathophysiological causal chain, and to obtain the offline risk conclusion of the target object.

9. A computer device, characterized in that, include: A memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, the processor executing the computer instructions to perform the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the method of any one of claims 1 to 7.

Images