Major trauma treatment decision artificial intelligence (AI) support system and method thereof

Abstract

The invention provides a major trauma treatment decision artificial intelligence (AI) support system and a method thereof, and relates to the technical field of artificial intelligence. And a multi-task deep learning model comprising a transformer encoder structure is adopted to extract time sequence features and inter-modal correlation features, an organ function failure risk score, a bleeding amount prediction value and a clinical intervention suggestion are output, and then a treatment decision scheme is generated. According to the method, deep fusion of multi-modal data can be realized, the prediction accuracy is improved, and timely and accurate treatment decision support is provided for clinicians.

Description

Technical Field

[0001] The present invention relates to artificial intelligence technology, and in particular to an artificial intelligence (AI) support system and method for major trauma treatment decision-making. Background Art

[0002] Major trauma refers to severe injuries caused by external force factors such as violence or accidental accidents, which may lead to multiple organ dysfunction in patients, and is characterized by rapid disease progression and high mortality. During clinical treatment, medical staff need to quickly evaluate the patient's condition and make timely intervention decisions to improve the patient's survival rate. Currently, clinical treatment decisions mainly rely on doctors' experience and existing trauma scoring systems. Commonly used scoring systems include Injury Severity Score, Revised Trauma Score, etc. With the development of artificial intelligence technology, its application in major trauma treatment decision support has become a research hotspot.

[0003] Traditional trauma scoring systems mainly evaluate based on single or limited clinical indicators, and fail to fully utilize the patient's multimodal physiological data, resulting in insufficient accuracy and comprehensiveness of the evaluation results. When making treatment decisions, doctors often need to pay attention to the dynamic changes of multiple clinical indicators at the same time, and existing decision support systems are difficult to effectively integrate and analyze these complex time-series data.

[0004] Most existing artificial intelligence-assisted decision-making methods are aimed at single prediction tasks, such as only predicting the risk of death or organ failure, lacking the ability to comprehensively evaluate the overall condition of the patient, and unable to give prediction results and intervention suggestions in multiple aspects at the same time.

[0005] Currently, the existing trauma treatment decision-making systems generally have the problem of insufficient real-time performance, being difficult to capture the dynamic change characteristics of the patient's condition in a timely manner, and lacking specific guidance on the optimal intervention time when giving intervention suggestions, which affects the clinical treatment effect. These systems usually adopt static decision rules and fail to fully consider the evolution law of the patient's state at different time points. Summary of the Invention

[0006] An embodiment of the present invention provides an artificial intelligence (AI) support system and method for major trauma treatment decision-making, which can solve the problems in the prior art.

[0007] In the first aspect of the embodiment of the present invention, there is provided an artificial intelligence (AI) support method for major trauma treatment decision-making, including:

[0008] Obtain multimodal physiological data of critically injured patients, where the multimodal physiological data includes vital sign data, blood biochemical index data, medical imaging data, and clinical manifestation data; preprocess the multimodal physiological data, convert the multimodal physiological data into a standardized data format, and perform time alignment on the standardized data based on temporal characteristics;

[0009] Construct a multi-task deep learning model, where the multi-task deep learning model includes a feature extraction sub-network and a multi-branch prediction sub-network. Among them, the feature extraction sub-network uses a Transformer encoder structure to extract the temporal characteristics and inter-modal correlation characteristics of the multimodal physiological data, and the multi-branch prediction sub-network outputs the organ failure risk score, blood loss prediction value, and clinical intervention suggestions respectively based on the extracted features;

[0010] Generate a critical injury treatment decision-making plan based on the output results of the multi-task deep learning model, where the critical injury treatment decision-making plan includes a sequence of clinical intervention measures and intervention timing.

[0011] The feature extraction sub-network uses a Transformer encoder structure to extract the temporal characteristics and inter-modal correlation characteristics of the multimodal physiological data, which includes:

[0012] Construct a multi-layer Transformer encoder network, where the multi-layer Transformer encoder network includes an intra-modal feature extraction layer and an inter-modal feature fusion layer. Among them, the intra-modal feature extraction layer uses a self-attention mechanism to extract the temporal characteristics of each modal data respectively, and the inter-modal feature fusion layer uses a cross-attention mechanism to achieve dynamic association of different modal features;

[0013] Perform feature combination on the outputs of the intra-modal feature extraction layer and the inter-modal feature fusion layer based on a residual connection structure, and adaptively fuse different-level features through learnable feature weight coefficients to obtain a fused multimodal feature representation; output the fused multimodal feature representation to the downstream task module.

[0014] The multi-branch prediction sub-network outputs the organ failure risk score, blood loss prediction value, and clinical intervention suggestions respectively based on the extracted features, which includes:

[0015] The multi-branch prediction network structure includes a feature sharing layer and a task-specific layer. Among them, the feature sharing layer screens and enhances the input features through an attention gating mechanism, and the task-specific layer includes an organ function assessment branch, a blood loss prediction branch, and a clinical recommendation generation branch;

[0016] The organ function assessment branch uses a temporal convolutional network to calculate the organ failure risk score;

[0017] The predicted blood loss branch predicts the trend of blood loss change within a future time window based on a recurrent neural network;

[0018] The clinical recommendation generation branch adopts a hierarchical and progressive decision-making mechanism. First, it identifies abnormal patterns of key clinical indicators based on a multi-head attention network, then generates clinical intervention recommendations by combining the diagnosis and treatment rules in the medical knowledge graph, and finally scores the reliability of the generated recommendations through a confidence evaluation module.

[0019] The organ function evaluation branch uses a temporal convolutional network to calculate the organ failure risk score, including:

[0020] Receiving organ function-related index data of the patient, where the organ function-related index data includes vital sign data and blood biochemical index data;

[0021] Performing temporal alignment and normalization processing on the organ function-related index data;

[0022] Constructing a multi-layer temporal convolutional network, where the multi-layer temporal convolutional network includes multiple parallel one-dimensional convolutional layers. Each one-dimensional convolutional layer uses convolutional kernels of different sizes to extract feature patterns of different time scales, and the output of the one-dimensional convolutional layer undergoes feature selection through a gating mechanism to achieve adaptive capture of short-term fluctuations and long-term trends;

[0023] Designing a residual attention module to enhance the features extracted by the multi-layer temporal convolutional network. The residual attention module includes a self-attention mechanism in the time dimension and a cross-attention mechanism in the index dimension, and realizes prominent enhancement of key temporal features by calculating the correlation weights between different time points and different indexes;

[0024] Based on the enhanced features output by the residual attention module, calculate the organ failure risk scores of multiple organs through a fully connected layer.

[0025] The clinical recommendation generation branch adopts a hierarchical and progressive decision-making mechanism. First, it identifies abnormal patterns of key clinical indicators based on a multi-head attention network, then generates clinical intervention recommendations by combining the diagnosis and treatment rules in the medical knowledge graph, and finally scores the reliability of the generated recommendations through a confidence evaluation module, including:

[0026] The multi-head attention network includes multiple parallel attention sub-networks. Each attention sub-network is responsible for abnormal detection of specific types of clinical indicators. The attention sub-network identifies the abnormal patterns of clinical indicators and outputs the abnormal degree score by calculating the deviation degree between the clinical indicators and the preset normal value range and combining the temporal change trend;

[0027] Retrieve the diagnosis and treatment rules related to the abnormal pattern from the medical knowledge graph, where the diagnosis and treatment rules include the correspondence between clinical indicators and intervention measures;

[0028] Generate clinical intervention suggestions based on the deep reinforcement learning framework, specifically including: constructing a policy network with the abnormal pattern of clinical indicators as the state space, the set of candidate intervention measures as the action space, and the degree of improvement of clinical indicators as the reward function. The policy network takes the candidate intervention measures as input, combines the treatment effects of similar cases in the historical diagnosis and treatment data, and outputs the rationality score of each intervention measure.

[0029] Generate a major trauma treatment decision-making plan based on the output results of the multi-task deep learning model. The major trauma treatment decision-making plan includes a sequence of clinical intervention measures and intervention timing, including:

[0030] Receive the output results of the multi-task deep learning model, where the output results include the risk score of organ failure, the predicted value of blood loss, and clinical intervention suggestions;

[0031] Construct a time-series decision tree network. The time-series decision tree network organizes the intervention measures using a hierarchical tree structure. Each node represents an intervention decision point, and the transition probability between nodes is dynamically updated based on the patient's state. The time-series decision tree network determines the optimal sequence of intervention measures by recursively calculating the expected rewards of each decision path.

[0032] In the second aspect of the embodiments of the present invention, a major trauma treatment decision-making artificial intelligence (AI) support system is provided, including:

[0033] A first unit for acquiring multi-modal physiological data of major trauma patients. The multi-modal physiological data includes vital sign data, blood biochemical index data, medical image data, and clinical manifestation data; preprocessing the multi-modal physiological data, converting the multi-modal physiological data into a standardized data format, and performing time alignment on the standardized data based on time series features;

[0034] A second unit for constructing a multi-task deep learning model. The multi-task deep learning model includes a feature extraction sub-network and a multi-branch prediction sub-network. The feature extraction sub-network uses a transformer encoder structure to extract the time series features and inter-modal association features of the multi-modal physiological data, and the multi-branch prediction sub-network outputs the risk score of organ failure, the predicted value of blood loss, and clinical intervention suggestions based on the extracted features respectively;

[0035] A third unit for generating a major trauma treatment decision-making plan based on the output results of the multi-task deep learning model. The major trauma treatment decision-making plan includes a sequence of clinical intervention measures and intervention timing.

[0036] The third aspect of the embodiments of the present invention

[0037] provides an electronic device, comprising:

[0038] a processor;

[0039] a memory for storing instructions executable by the processor;

[0040] wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0041] The fourth aspect of the embodiments of the present invention

[0042] provides a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0043] The beneficial effects of the present application are as follows:

[0044] By constructing a multi-task deep learning model based on multi-modal physiological data, the present invention can accurately predict the risk of organ failure and the amount of bleeding in critically injured patients, thereby providing timely and accurate decision-making support for clinicians and improving the efficiency and accuracy of the treatment of critical injuries.

[0045] The present invention uses a transformer encoder structure to extract features from multi-modal physiological data, which can effectively capture the temporal correlations between different physiological indicators and the complex interaction relationships between modalities, improve the prediction ability of the development trend of the patient's condition, and provide a reliable basis for formulating personalized treatment plans.

[0046] The present invention outputs clinical intervention suggestions through a multi-branch prediction sub-network and generates a treatment decision plan including specific intervention measures and timings, realizing seamless connection from data analysis to clinical practice, helping to standardize and standardize the treatment process of critical injuries, and reducing the risk of medical errors. Description of the Drawings

[0047] Figure 1 is a schematic flowchart of the method for making a decision on the treatment of critical injuries based on artificial intelligence according to the embodiments of the present invention;

[0048] Figure 2 is a schematic diagram of the accuracy rate of clinical index abnormality detection according to the embodiments of the present invention. Detailed Embodiments

[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only a part rather than all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0050] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0051] Figure 1 The following is a schematic flowchart of a major trauma treatment decision-making method based on artificial intelligence according to an embodiment of the present invention. As Figure 1 shown, the method includes:

[0052] Obtain multi-modal physiological data of major trauma patients, where the multi-modal physiological data includes vital sign data, blood biochemical index data, medical image data, and clinical manifestation data; preprocess the multi-modal physiological data, convert the multi-modal physiological data into a standardized data format, and perform time alignment on the standardized data based on temporal features;

[0053] Construct a multi-task deep learning model, where the multi-task deep learning model includes a feature extraction sub-network and a multi-branch prediction sub-network. The feature extraction sub-network uses a Transformer encoder structure to extract the temporal features and inter-modal correlation features of the multi-modal physiological data, and the multi-branch prediction sub-network outputs the organ failure risk score, the predicted blood loss value, and clinical intervention suggestions respectively based on the extracted features;

[0054] Generate a major trauma treatment decision-making plan based on the output results of the multi-task deep learning model, where the major trauma treatment decision-making plan includes a sequence of clinical intervention measures and the intervention timing.

[0055] In an optional implementation manner, the feature extraction sub-network uses a Transformer encoder structure to extract the temporal features and inter-modal correlation features of the multi-modal physiological data, including:

[0056] Construct a multi-layer Transformer encoder network, where the multi-layer Transformer encoder network includes an intra-modal feature extraction layer and an inter-modal feature fusion layer. The intra-modal feature extraction layer uses a self-attention mechanism to extract the temporal features of each modal data respectively, and the inter-modal feature fusion layer uses a cross-attention mechanism to achieve dynamic association of different modal features;

[0057] Based on the residual connection structure, the outputs of the intra-modal feature extraction layer and the inter-modal feature fusion layer are combined, and different-level features are adaptively fused through learnable feature weight coefficients to obtain the fused multi-modal feature representation; the fused multi-modal feature representation is output to the downstream task module.

[0058] For the feature extraction of multi-modal physiological data, a feature extraction sub-network based on the transformer encoder is constructed. First, the input multi-modal physiological data is preprocessed, including operations such as data normalization, noise reduction, and segmentation, so that the data meets the network input requirements.

[0059] In the intra-modal feature extraction layer, the self-attention mechanism is used to process each modal data respectively. Taking electroencephalogram (EEG) and electrocardiogram (ECG) data as an example, for EEG data, it is divided into several time windows, and each window contains 128 sampling points; for ECG data, each window contains 256 sampling points. The temporal information is embedded into the features through position encoding, and then the self-attention mechanism is used to calculate the correlation weights between features at different time points within the same modality, realizing the effective extraction of single-modal temporal features.

[0060] In the inter-modal feature fusion layer, the cross-attention mechanism is used to achieve the dynamic association of different-modal features. The EEG features are used as the query vector, and the ECG features are used as the key-value vector. By calculating the attention weights between different-modal features, the dynamic dependence between modalities is captured. Specifically, 8 attention heads are set, and the feature dimension of each head is 64. The multi-head attention mechanism is used to enhance the robustness of feature extraction.

[0061] Based on the residual connection structure, the outputs of the intra-modal feature extraction layer and the inter-modal feature fusion layer are combined. For the output of each layer, learnable weight coefficients are introduced, and the initial values are all set to 0.5. The importance of different-level features is adaptively adjusted through the backpropagation process. Finally, the weighted features are concatenated and normalized to obtain the fused multi-modal feature representation.

[0062] The fused features are output to the downstream task module to achieve specific application goals, such as emotion recognition, disease diagnosis, etc. The entire feature extraction process can be trained end-to-end, and the network parameters are optimized through the backpropagation algorithm.

[0063] Through a multi-layer Transformer encoder structure, the efficient extraction of temporal features of multi-modal physiological data is achieved, overcoming the problem that traditional methods are difficult to capture long-range dependence relationships, and improving the accuracy and robustness of feature extraction. The cross-attention mechanism is adopted to realize the dynamic association of features of different modalities, which can adaptively learn the interaction relationships between modalities, effectively fuse multi-modal information, and enhance the integrity and discriminability of feature representation. The residual connection and learnable weight coefficients are introduced to realize the adaptive fusion of features at different levels, avoid the loss of feature information, enhance the generalization ability and stability of the model, and make the feature extraction results more reliable and effective.

[0064] In an optional implementation manner, the multi-branch prediction sub-network respectively outputs an organ failure risk score, a predicted blood loss value, and clinical intervention suggestions based on the extracted features, including:

[0065] The multi-branch prediction network structure includes a feature sharing layer and a task-specific layer, wherein the feature sharing layer screens and enhances the input features through an attention gating mechanism, and the task-specific layer includes an organ function evaluation branch, a blood loss prediction branch, and a clinical recommendation generation branch;

[0066] The organ function evaluation branch uses a temporal convolutional network to calculate the organ failure risk score;

[0067] The blood loss prediction branch predicts the trend of blood loss change within a future time window based on a recurrent neural network;

[0068] The clinical recommendation generation branch adopts a hierarchical progressive decision-making mechanism. First, it identifies the abnormal patterns of key clinical indicators based on a multi-head attention network, then generates clinical intervention suggestions by combining the diagnosis and treatment rules in the medical knowledge graph, and finally scores the reliability of the generated suggestions through a confidence evaluation module.

[0069] The feature sharing layer first processes the input clinical indicator data. Through the attention gating mechanism, different weights are assigned to different types of clinical indicators. Specifically, when implementing, multiple attention heads are set, and each attention head is responsible for paying attention to feature information in different dimensions. For example, higher weights are assigned to vital sign indicators such as blood pressure and heart rate, and relatively lower weights are assigned to general laboratory test results. The features screened and enhanced by the attention mechanism are input into the three task-specific branches.

[0070] The organ function assessment branch adopts a temporal convolutional network structure. First, the input features are segmented according to time windows, and the length of each time window is 4 hours. For the feature sequence within each time window, multi-layer one-dimensional convolutions are used for feature extraction, and the kernel sizes of the convolutions are 3, 5, and 7 respectively, which are used to capture temporal patterns at different scales. After being processed by the convolutional layer and the pooling layer, a risk score between 0 and 1 is output. The closer the score is to 1, the higher the risk of organ failure.

[0071] The blood loss prediction branch is based on a gated recurrent unit network structure. This branch takes the blood loss data within the past 24 hours as the input sequence and predicts the blood loss in the next 4 hours, 8 hours, and 12 hours. The network contains two layers of recurrent units, and the hidden layer dimension is 128. A predicted value is output at each time step, indicating the expected blood loss at that moment. For example, by inputting the blood loss data recorded every hour within the past 24 hours, the changing trend of the blood loss at several future time points can be predicted.

[0072] The clinical recommendation generation branch adopts a hierarchical progressive decision-making mechanism. First, an 8-head attention network is used to identify abnormal indicators, and each attention head focuses on a combination of indicators of different systems. Then, a pre-constructed medical knowledge graph is queried, which contains standard intervention plans for common clinical situations. Finally, a confidence evaluation module scores the recommendations, and the score is between 0 and 1. Only recommendations with a confidence greater than 0.8 will be output for clinicians' reference.

[0073] Through the attention mechanism of the feature sharing layer, an adaptive weight assignment for different types of clinical indicators is realized, which improves the pertinence and accuracy of feature extraction. By combining the temporal convolutional network and the recurrent neural network, the short-term and long-term temporal dependencies can be captured simultaneously, enhancing the prediction model's perception ability of the changing trend of the condition. Based on the hierarchical progressive clinical recommendation generation mechanism, combined with the medical knowledge graph and confidence evaluation, the reliability and practicality of the output recommendations are ensured, providing effective auxiliary support for clinical decision-making.

[0074] In an alternative implementation, the organ function assessment branch using the temporal convolutional network to calculate the organ failure risk score includes:

[0075] Receiving the organ function-related index data of the patient, where the organ function-related index data includes vital sign data and blood biochemical index data;

[0076] Performing temporal alignment and normalization processing on the organ function-related index data;

[0077] Construct a multi-layer temporal convolutional network, which includes multiple parallel one-dimensional convolutional layers. Each one-dimensional convolutional layer uses convolutional kernels of different sizes to extract feature patterns of different time scales. The output of the one-dimensional convolutional layer is subjected to feature selection through a gating mechanism to achieve adaptive capture of short-term fluctuations and long-term trends;

[0078] Design a residual attention module to enhance the features extracted by the multi-layer temporal convolutional network. The residual attention module includes a self-attention mechanism in the time dimension and a cross-attention mechanism in the metric dimension. By calculating the correlation weights between different time points and different metrics, the prominent reinforcement of key temporal features is achieved;

[0079] Based on the enhanced features output by the residual attention module, calculate the functional failure risk scores of multiple organs through a fully connected layer.

[0080] First, receive the organ function-related index data of the patient. The vital sign data includes indicators such as heart rate, blood pressure, body temperature, respiratory rate, and blood oxygen saturation, and the sampling frequency is once per hour; the blood biochemical index data includes test results such as blood routine, liver function, and kidney function, and the sampling frequency is once per day. For each patient, establish a unified time axis to align the index data of different frequencies to the same time point.

[0081] Perform standardization processing on the data, and normalize each index value to the 0-1 interval. The sliding window method is used to process missing values, the window size is set to 24 hours, and the nearest valid value within the window is used for filling. Outliers are truncated using the three-standard-deviation principle.

[0082] Construct a multi-layer temporal convolutional network to extract temporal features. Design 3 parallel one-dimensional convolutional layers, and the convolutional kernel sizes are 3, 6, and 12 respectively, corresponding to time scales of 3 hours, 6 hours, and 12 hours. Each layer contains 64 convolutional kernels, and the stride is 1. The output of the convolutional layer passes through the ReLU activation function and then through the sigmoid gating unit for feature selection to achieve adaptive weight allocation for features of different time scales.

[0083] Design a residual attention module to enhance feature expression. In the time dimension, calculate the similarity matrix between feature vectors at different time points within 24 hours to generate time attention weights. In the metric dimension, calculate the correlation matrix between different metric features to generate metric attention weights. Add the original features and the weighted features to form a residual connection.

[0084] Finally, through two fully connected layers, the output layer has a dimension of 5, corresponding to the risk scores of functional failure of the five important organ systems of the heart, liver, kidney, respiration, and blood coagulation respectively. The scoring range is 0 - 100, and the higher the score, the greater the risk of failure. When the score of an organ exceeds 80 points, the system issues a warning prompt.

[0085] Data case: Partial index data of a patient within 24 hours: Heart rate (beats per minute): 65 - 78 - 82 - 75; Systolic blood pressure (mmHg): 125 - 118 - 130 - 122; Blood oxygen saturation (%): 98 - 97 - 96 - 98; White blood cell count (10^9 / L): 6.8; Serum creatinine (umol / L): 75. After processing, the predicted risk scores of the patient's organ function failures are: Heart 35 points, Liver 28 points, Kidney 42 points, Respiration 40 points, Coagulation 25 points, and no high-risk organs are found for the time being.

[0086] This solution adaptively captures the short-term fluctuations and long-term change trends of physiological indicators through a multi-scale temporal convolutional network, improving the ability to depict the dynamic changes of organ function. The residual attention mechanism is used to prominently strengthen the key temporal features, enhancing the sensitivity of the model to changes in important physiological indicators and improving the accuracy of risk assessment. It realizes the real-time quantitative assessment of the risk of functional failure of multiple important organ systems, can timely detect potential risks, provides decision support for clinical intervention, and has important practical value.

[0087] In an alternative implementation, the clinical recommendation generation branch adopts a hierarchical progressive decision-making mechanism. First, it identifies the abnormal patterns of key clinical indicators based on a multi-head attention network, then generates clinical intervention recommendations by combining the diagnosis and treatment rules in the medical knowledge graph, and finally scores the reliability of the generated recommendations through a confidence evaluation module, including:

[0088] The multi-head attention network includes multiple parallel attention sub-networks. Each attention sub-network is responsible for the abnormal detection of specific types of clinical indicators. The attention sub-network calculates the deviation degree between the clinical indicators and the preset normal value range, combines the temporal change trend, identifies the abnormal patterns of the clinical indicators, and outputs the abnormal degree score;

[0089] Retrieve the diagnosis and treatment rules related to the abnormal pattern from the medical knowledge graph. The diagnosis and treatment rules include the corresponding relationship between clinical indicators and intervention measures;

[0090] Generate clinical intervention recommendations based on the deep reinforcement learning framework, specifically including: constructing a policy network with the abnormal pattern of clinical indicators as the state space, the set of candidate intervention measures as the action space, and the improvement degree of clinical indicators as the reward function. The policy network takes the candidate intervention measures as input, combines the treatment effects of similar cases in historical diagnosis and treatment data, and outputs the rationality score of each intervention measure.

[0091] The clinical recommendation generation branch adopts a hierarchical and progressive decision-making mechanism to realize the intelligent generation of clinical recommendations through three core modules: a multi-head attention network, a medical knowledge graph, and a confidence evaluation.

[0092] First, the multi-head attention network contains multiple parallel attention sub-networks. Each sub-network is specifically responsible for the anomaly detection of a certain type of clinical indicator, such as vital sign indicators like blood pressure, blood sugar, and blood oxygen, or test indicators like blood routine and biochemistry. For each clinical indicator, the attention sub-network evaluates the degree of anomaly by calculating its deviation from the normal reference value range. For example, for the blood pressure indicator, the normal range of systolic blood pressure is 90 - 140 mmHg. If the measured value is 160 mmHg, then the deviation value exceeding the normal upper limit is calculated as 20 mmHg. At the same time, combined with the temporal change trend of this indicator, if the measured values for 3 consecutive times show an upward trend, the anomaly score is further increased. The anomaly scores output by each sub-network are weighted and aggregated to form an overall anomaly pattern representation of the clinical indicators.

[0093] Next, based on the detected anomaly pattern, relevant diagnosis and treatment rules are retrieved from a pre-constructed medical knowledge graph. This knowledge graph contains a large number of corresponding relationships between clinical indicators and intervention measures, such as "hypertension - antihypertensive drugs", "hyperglycemia - insulin", etc. When retrieving, a semantic similarity matching algorithm is used to find the set of diagnosis and treatment rules most relevant to the current anomaly pattern.

[0094] Then, a deep reinforcement learning framework is used to generate specific clinical intervention recommendations. The clinical indicator anomaly pattern is used as the state space, and the set of candidate intervention measures is used as the action space. The policy network takes the candidate intervention measures as input and also refers to the treatment effects of similar cases in historical diagnosis and treatment data. For example, for the case of elevated blood pressure, hypertensive cases with similar basic characteristics such as age and gender can be screened from historical data, and the effects of different antihypertensive regimens are analyzed. After comprehensive evaluation by the policy network, a rationality score between 0 and 1 is output for each candidate intervention measure.

[0095] Finally, the confidence evaluation module scores the reliability of the generated clinical recommendations. The evaluation dimensions include: the compliance with clinical guidelines, the treatment effects of historical similar cases, the risk of drug interactions, etc. The scores of each dimension are weighted and averaged to obtain the final confidence score.

[0096] Figure 2 Schematic diagram of the accuracy of clinical indicator anomaly detection in the embodiments of the present invention:

[0097] Based on a comprehensive analysis of the clinical indicator anomaly detection accuracy chart, which shows the performance comparison of three different technical solutions. Among them, "this technical solution" represents a hierarchical progressive decision-making mechanism based on a multi-head attention network, which includes multiple parallel attention sub-networks for anomaly detection and an intelligent diagnosis and treatment solution that combines a medical knowledge graph and deep reinforcement learning; "traditional method" represents a clinical indicator anomaly detection method based on a single attention mechanism, relying only on simple threshold judgment and statistical analysis; "benchmark method" represents traditional machine learning methods, such as clinical indicator analysis solutions using algorithms like support vector machine (SVM) and random forest.

[0098] Firstly, by designing multiple parallel attention sub-networks, specialized anomaly detection of different types of clinical indicators is achieved, improving the accuracy of detection; secondly, introducing a medical knowledge graph as decision support enables the system to generate more reasonable intervention suggestions based on professional medical knowledge; finally, adopting a deep reinforcement learning framework to optimize the selection of intervention measures realizes the intelligence and personalization of clinical decision-making.

[0099] Compared with the traditional method, this technical solution can reach a higher detection accuracy more quickly and can obtain ideal performance under smaller sample sizes; compared with the benchmark method, this technical solution has stronger data utilization efficiency and a more stable performance improvement trend. At the same time, this technical solution also shows obvious advantages in terms of model convergence speed and stability, and can provide more reliable anomaly detection and intervention suggestion support for clinical practice. These improvements effectively solve the problems existing in the prior art in clinical indicator anomaly detection, such as low accuracy, poor generalization ability, and insufficient decision support, and provide strong support for improving the intelligent level of medical diagnosis.

[0100] In an optional implementation manner, based on the output result of the multi-task deep learning model, a major trauma treatment decision-making plan is generated. The major trauma treatment decision-making plan includes a sequence of clinical intervention measures and intervention timing, including:

[0101] Receiving the output result of the multi-task deep learning model, the output result includes the risk score of organ failure, the predicted value of blood loss, and clinical intervention suggestions;

[0102] Constructing a temporal decision tree network, the temporal decision tree network organizes the intervention measures using a hierarchical tree structure. Each node represents an intervention decision point, and the transition probability between nodes is dynamically updated based on the patient's state. The temporal decision tree network determines the optimal sequence of intervention measures by recursively calculating the expected return of each decision path.

[0103] First, receive the result data output by the multi-task deep learning model. This result data includes the risk score of organ failure, specifically the probability values of organ failure in major organ systems such as liver function, kidney function, heart function, and respiratory function, with a value range of 0 - 1. It also includes the predicted blood loss value, predicting the possible blood loss within the next 4 hours in milliliters. In addition, it includes clinical intervention suggestions, such as specific treatment measure suggestions like blood transfusion, fluid replacement, and using vasopressors.

[0104] Next, construct a temporal decision tree network for treatment plan planning. This network adopts a multi-layer tree structure. The first-layer nodes are initial assessment decision points, including airway management, respiratory support, circulatory status assessment, etc.; the second-layer nodes are emergency treatment decision points, including hemostasis, fluid replacement, blood transfusion, etc.; the third-layer nodes are important organ function support decision points, including adjusting ventilator support parameters, using vasoactive drugs, etc. Each node is connected by a directed edge, and a transition probability value is attached to the edge.

[0105] In practical applications, first input clinical data such as the patient's vital signs and laboratory tests. After analysis by the deep learning model, risk scores and prediction results are obtained. Taking a trauma patient as an example, the organ function scores show: respiratory function 0.7, circulatory function 0.8, and the predicted blood loss in 4 hours is 800 ml. Based on these data, the temporal decision tree automatically calculates the optimal path: first perform endotracheal intubation and mechanical ventilation, while replenish 3000 ml of fluid and transfuse 4 units of red blood cells, and then use norepinephrine to maintain blood pressure.

[0106] The system continuously monitors the changes in the patient's condition and dynamically updates the transition probabilities between nodes. When it is observed that the patient's blood pressure rises and oxygenation improves, the benefit values of the corresponding treatment measures increase, and the system accordingly adjusts the subsequent decision path, such as adjusting ventilator parameters and reducing the dosage of vasopressors.

[0107] This solution realizes the accurate assessment and prediction of the patient's condition through the deep learning model, provides an objective basis for clinical decision-making, and improves the accuracy and timeliness of treatment. The temporal decision tree network is used to systematically organize and optimize the intervention measures to ensure the scientificity and coherence of the treatment plan, and avoid blindness and randomness in clinical practice. The system can dynamically adjust the treatment plan according to the patient's condition to achieve personalized and precise treatment. At the same time, it improves the medical quality and reduces the medical risk through a standardized decision-making process.

[0108] In the second aspect of the embodiments of the present invention, an artificial intelligence (AI) support system for major trauma treatment decision-making is provided, including:

[0109] A first unit for acquiring multimodal physiological data of critically injured patients, where the multimodal physiological data includes vital sign data, blood biochemical index data, medical image data, and clinical manifestation data; preprocessing the multimodal physiological data, converting the multimodal physiological data into a standardized data format, and performing time alignment on the standardized data based on temporal characteristics;

[0110] A second unit for constructing a multi-task deep learning model, where the multi-task deep learning model includes a feature extraction sub-network and a multi-branch prediction sub-network. The feature extraction sub-network uses a transformer encoder structure to extract the temporal characteristics and inter-modal correlation characteristics of the multimodal physiological data, and the multi-branch prediction sub-network respectively outputs the organ failure risk score, the predicted blood loss value, and clinical intervention suggestions based on the extracted features;

[0111] A third unit for generating a decision-making plan for the treatment of critical trauma based on the output results of the multi-task deep learning model, where the decision-making plan for the treatment of critical trauma includes a sequence of clinical intervention measures and the timing of intervention.

[0112] The third aspect of the embodiments of the present invention

[0113] Provided is an electronic device, including:

[0114] A processor;

[0115] A memory for storing instructions executable by the processor;

[0116] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0117] The fourth aspect of the embodiments of the present invention,

[0118] Provided is a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0119] The present invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are uploaded.

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An artificial intelligence (AI) support method for major trauma treatment decision-making, characterized in that, Including: Obtain multimodal physiological data of severely injured patients, where the multimodal physiological data includes vital sign data, blood biochemical index data, medical imaging data, and clinical manifestation data; Preprocess the multimodal physiological data, convert the multimodal physiological data into a standardized data format, and perform time alignment on the standardized data based on temporal features; Construct a multi-task deep learning model, where the multi-task deep learning model includes a feature extraction sub-network and a multi-branch prediction sub-network. The feature extraction sub-network uses a Transformer encoder structure to extract the temporal features and inter-modal correlation features of the multimodal physiological data, and the multi-branch prediction sub-network outputs the organ failure risk score, blood loss prediction value, and clinical intervention suggestions based on the extracted features; Generate a decision-making plan for the treatment of severe trauma based on the output results of the multi-task deep learning model. The decision-making plan for the treatment of severe trauma includes a sequence of clinical intervention measures and intervention timing.

2. The method according to claim 1, wherein The feature extraction sub-network uses a Transformer encoder structure to extract the temporal features and inter-modal correlation features of the multimodal physiological data, including: Construct a multi-layer Transformer encoder network, where the multi-layer Transformer encoder network includes an intra-modal feature extraction layer and an inter-modal feature fusion layer. The intra-modal feature extraction layer uses a self-attention mechanism to extract the temporal features of each modal data, and the inter-modal feature fusion layer uses a cross-attention mechanism to achieve dynamic association of different modal features; Perform feature combination on the outputs of the intra-modal feature extraction layer and the inter-modal feature fusion layer based on a residual connection structure, and adaptively fuse different-level features through learnable feature weight coefficients to obtain a fused multi-modal feature representation; output the fused multi-modal feature representation to the downstream task module.

3. The method according to claim 1, wherein The multi-branch prediction sub-network outputs the organ failure risk score, blood loss prediction value, and clinical intervention suggestions based on the extracted features, including: The multi-branch prediction network structure includes a feature sharing layer and a task-specific layer. The feature sharing layer screens and enhances the input features through an attention gating mechanism, and the task-specific layer includes an organ function assessment branch, a blood loss prediction branch, and a clinical recommendation generation branch; The organ function assessment branch uses a temporal convolutional network to calculate the organ failure risk score; The blood loss prediction branch predicts the trend of blood loss change within a future time window based on a recurrent neural network; The clinical recommendation generation branch adopts a hierarchical progressive decision-making mechanism. First, it identifies abnormal patterns of key clinical indicators based on a multi-head attention network, then generates clinical intervention suggestions by combining the diagnosis and treatment rules in the medical knowledge graph, and finally scores the reliability of the generated suggestions through a confidence assessment module.

4. The method according to claim 3, characterized in that, The organ function assessment branch uses a temporal convolutional network to calculate the organ failure risk score, including: Receive organ function-related index data of the patient, where the organ function-related index data includes vital sign data and blood biochemical index data; Perform temporal alignment and normalization processing on the data of the organ function-related indicators; Construct a multi-layer temporal convolutional network, which includes multiple parallel one-dimensional convolutional layers. Each one-dimensional convolutional layer uses convolutional kernels of different sizes to extract feature patterns of different time scales. The output of the one-dimensional convolutional layer is subjected to feature selection through a gating mechanism to achieve adaptive capture of short-term fluctuations and long-term trends; Design a residual attention module to enhance the features extracted by the multi-layer temporal convolutional network. The residual attention module includes a self-attention mechanism in the time dimension and a cross-attention mechanism in the indicator dimension. By calculating the correlation weights between different time points and different indicators, prominent enhancement of key temporal features is achieved; Based on the enhanced features output by the residual attention module, calculate the functional failure risk scores of multiple organs through a fully connected layer.

5. The method according to claim 3, characterized in that The clinical recommendation generation branch adopts a hierarchical progressive decision-making mechanism. First, identify the abnormal patterns of key clinical indicators based on a multi-head attention network, then generate clinical intervention recommendations in combination with the diagnosis and treatment rules in the medical knowledge graph, and finally, a confidence evaluation module scores the reliability of the generated recommendations, including: The multi-head attention network includes multiple parallel attention sub-networks. Each attention sub-network is responsible for the abnormal detection of specific types of clinical indicators. The attention sub-network calculates the deviation degree between the clinical indicators and the preset normal value range, and combines the temporal change trend to identify the abnormal patterns of the clinical indicators and output the abnormal degree scores; Retrieve the diagnosis and treatment rules related to the abnormal patterns from the medical knowledge graph. The diagnosis and treatment rules include the corresponding relationship between clinical indicators and intervention measures; Generate clinical intervention recommendations based on a deep reinforcement learning framework, specifically including: constructing a policy network with the abnormal patterns of clinical indicators as the state space, the set of candidate intervention measures as the action space, and the improvement degree of clinical indicators as the reward function. The policy network takes the candidate intervention measures as inputs and combines the treatment effects of similar cases in the historical diagnosis and treatment data to output the rationality scores of each intervention measure.

6. The method according to claim 1, characterized in that, Generate a major trauma treatment decision-making plan based on the output results of the multi-task deep learning model. The major trauma treatment decision-making plan includes a sequence of clinical intervention measures and intervention timing, including: Receive the output results of the multi-task deep learning model. The output results include the organ function failure risk scores, the predicted blood loss values, and the clinical intervention recommendations; Construct a temporal decision tree network. The temporal decision tree network organizes the intervention measures using a hierarchical tree structure. Each node represents an intervention decision point, and the transition probabilities between nodes are dynamically updated based on the patient's state. The temporal decision tree network determines the optimal sequence of intervention measures by recursively calculating the expected benefits of each decision path.

7. An artificial intelligence (AI) support system for major trauma treatment decision-making, which is used to implement the method described in any one of claims 1-6, characterized in that, Including: The first unit is used to obtain the multi-modal physiological data of major trauma patients. The multi-modal physiological data includes vital sign data, blood biochemical index data, medical image data, and clinical manifestation data; Preprocess the multimodal physiological data, convert the multimodal physiological data into a standardized data format, and perform time alignment on the standardized data based on temporal features; A second unit for constructing a multi-task deep learning model, the multi-task deep learning model including a feature extraction sub-network and a multi-branch prediction sub-network, wherein the feature extraction sub-network uses a transformer encoder structure to extract the temporal features and inter-modal correlation features of the multimodal physiological data, and the multi-branch prediction sub-network respectively outputs an organ failure risk score, a blood loss prediction value, and a clinical intervention recommendation based on the extracted features; A third unit for generating a major trauma treatment decision-making plan based on the output results of the multi-task deep learning model, the major trauma treatment decision-making plan including a sequence of clinical intervention measures and intervention timing.

8. An electronic device, characterized in that, Comprising: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, The computer program instructions, when executed by the processor, implement the method according to any one of claims 1 to 6.

Images