Special disease emergency whole process auxiliary decision method based on multi-modal perception and AI

Abstract

The application provides a disease-specific emergency whole-process auxiliary decision-making method based on multi-modal perception and AI, relates to the technical field of medical emergency, and comprises the following steps: acquiring multi-modal data of an emergency scene, constructing a feature dependency graph, and identifying disease types and symptom characteristics; indexing disease diagnosis and treatment data based on a symptom characteristic vector, determining reference cases and treatment paths; combining clinical norms to construct an action space, applying Monte Carlo search to predict expected returns, and generating an optimal treatment plan, which can improve the accuracy of emergency disease identification, shorten the treatment decision-making time, and improve the emergency intervention effect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical first aid, and in particular to a disease-specific first aid whole-process auxiliary decision-making method based on multi-modal perception and AI. BACKGROUND

[0002] First aid medicine is an important part of the medical service system, and rapid and accurate treatment of disease-specific first aid plays a decisive role in patient prognosis. With the development of information technology, multi-modal perception devices and artificial intelligence technology are increasingly widely used in the medical field, bringing new technical support to first aid medicine.

[0003] With the application of data-driven decision support systems in first aid scenes, multi-modal information such as patient physiological data and environmental information is collected, and artificial intelligence algorithms can provide assistance for first aid decision-making.

[0004] However, the prior art still has problems such as only focusing on single modal data, lacking comprehensive analysis of multi-modal data, being difficult to fully grasp the patient's state, being difficult to effectively capture the temporal correlation between symptom evolution rules and characteristics, being unable to provide targeted support for first aid decision-making at different stages, lacking deep learning and optimization of historical cases, being unable to adaptively adjust treatment plans according to individual differences and dynamic changes of patients, lacking intelligent decision optimization mechanisms, and being difficult to provide precise personalized treatment decisions in complex first aid scenarios. SUMMARY

[0005] The embodiments of the present application provide a disease-specific first aid whole-process auxiliary decision-making method based on multi-modal perception and AI, which can at least solve some of the problems in the prior art.

[0006] In a first aspect, the embodiments of the present application provide a disease-specific first aid whole-process auxiliary decision-making method based on multi-modal perception and AI, comprising:

[0007] Obtaining multi-modal data of a first aid scene and performing feature extraction to obtain an initial feature vector, calculating the temporal changes of the initial feature vector and constructing a feature dependency graph combining conditional probability, calculating the temporal correlation between feature nodes in the feature dependency graph, determining a key feature sequence based on the temporal correlation and identifying a disease type and a corresponding symptom feature vector;

[0008] Based on the symptom feature vector, the pre-set disease diagnosis and treatment data is indexed, a to-be-indexed subset is determined, and the initial similarity of the symptom feature vector in the to-be-indexed subset and the historical cases is calculated, a candidate case is determined, and the candidate case is analyzed to determine the symptom evolution rule, the treatment matching degree of the current physiological data and the symptom evolution rule is calculated, the reference case is determined based on the treatment matching degree, and the treatment path corresponding to the reference case is extracted as a reference path.

[0009] generating an initial treatment scheme based on the reference path and constructing an initial action space combined with clinical norms, constructing a state vector combined with the current physiological data, selecting a treatment action in the initial action space and calculating an immediate reward value, predicting an expected return combined with Monte Carlo search, updating the initial treatment scheme based on the predicted expected return to obtain an optimal treatment scheme.

[0010] In an optional implementation,

[0011] obtaining multi-modal data of the emergency scene and performing feature extraction to obtain an initial feature vector, calculating a time sequence change of the initial feature vector and constructing a feature dependency graph combined with a conditional probability include:

[0012] performing feature extraction on the multi-modal data of the emergency scene to obtain an initial feature vector, calculating a first-order difference value of the initial feature vector, and determining an inflection point based on the first-order difference value and a preset adaptive threshold;

[0013] dividing the initial feature vector into a plurality of feature segments according to the inflection point, modeling each feature segment and constructing a segmented feature function, and constructing a transition connection function of time-varying weight between adjacent feature segments based on the segmented feature function to obtain a time sequence change sequence;

[0014] calculating a conditional probability based on the time sequence change sequence, establishing a directed edge between corresponding feature nodes when the conditional probability exceeds a preset dynamic threshold, and repeatedly constructing the directed edge to obtain the feature dependency graph.

[0015] In an optional implementation,

[0016] calculating a time sequence correlation degree between feature nodes in the feature dependency graph, determining a key feature sequence based on the time sequence correlation degree, and identifying a special disease type and a corresponding symptom feature vector include:

[0017] For each feature node in the feature dependency graph, calculating a PageRank centrality combined with neighbor nodes of the current feature node, calculating a connection probability between the current feature node and the neighbor nodes to determine an information entropy, and calculating an importance weight of the current feature node based on the PageRank centrality and the information entropy;

[0018] calculating distances between different feature nodes and determining inter-node similarity, calculating an initial correlation degree based on the inter-node similarity and the importance weight, and performing time sequence adjustment based on the initial correlation degree and a preset time sequence decay factor to obtain a time sequence correlation degree;

[0019] extracting a feature mapping relationship corresponding to each special disease type from a pre-set special disease type knowledge base, taking a feature node with a time sequence correlation degree greater than a pre-set correlation degree threshold as a key node, analyzing combination correlation between the key nodes and constructing a key feature sequence in combination with the feature mapping relationship, calculating a matching degree of each feature in the key feature sequence with a special disease type to obtain a special disease type probability value and determining a special disease type corresponding to the current feature, and extracting a time sequence correlation degree corresponding to the special disease type to obtain a symptom feature vector.

[0020] In an optional implementation,

[0021] performing partition indexing on pre-set special disease diagnosis and treatment data based on the symptom feature vector, determining a to-be-indexed subset and calculating an initial similarity between the symptom feature vector in the to-be-indexed subset and a historical case include:

[0022] calculating a difference between each symptom feature value in the symptom feature vector and a pre-set normal reference mean value to determine an abnormality degree of each symptom, and determining a danger level corresponding to each symptom based on the abnormality degree;

[0023] dividing the symptom feature vector into multiple priority partitions based on the danger level, solving a first early warning threshold based on the danger level and a pre-set first balance parameter, for the symptom feature vector in each priority partition, solving a second early warning threshold by calculating a change trend and combining a pre-set second balance parameter, obtaining a combined early warning value based on the first early warning threshold and the second early warning threshold, and constructing an index structure based on the combined early warning value and the symptom feature vector;

[0024] locating a to-be-matched partition from the special disease diagnosis and treatment data based on the index structure, calculating a partition matching degree between the symptom feature vector and each to-be-matched partition, and selecting three to-be-matched partitions with the highest partition matching degrees to obtain a to-be-indexed subset;

[0025] calculating a Euclidean distance between the symptom feature vector and each historical case in the to-be-indexed subset to determine a feature similarity, calculating a timeliness weight based on time information of the historical case and combining the feature similarity to calculate an initial similarity.

[0026] In an optional implementation,

[0027] determining a candidate case and performing time sequence analysis on the candidate case to determine a symptom evolution rule, calculating a treatment matching degree of the current physiological data with the symptom evolution rule, and determining a reference case based on the treatment matching degree and extracting a treatment path corresponding to the reference case as a reference path include:

[0028] Based on the initial similarity, candidate cases are identified in the subset to be indexed. The candidate cases are then subjected to time series analysis using a dynamic time warping method. The time series distance for each candidate case is calculated, and the symptom evolution pattern is determined based on the time series distance.

[0029] Collect current physiological data, calculate the disease complexity matching value and medical resource matching degree between the current physiological data and the symptom evolution pattern, and sum them to obtain the treatment matching degree;

[0030] The candidate case with the highest treatment matching degree is selected as the reference case. A treatment optimization objective function is constructed based on the complication risk value. The optimal solution is obtained by solving the treatment optimization objective function through the differential evolution algorithm. The feasibility score corresponding to the optimal solution is extracted. If the feasibility score is greater than the preset feasibility threshold, the treatment path is extracted and output as the reference path.

[0031] In one alternative implementation,

[0032] An initial treatment plan is generated based on the reference path, and an initial action space is constructed in conjunction with clinical guidelines. A state vector is constructed using the current physiological data, including:

[0033] The reference path is decomposed into multiple treatment nodes by time sequence. Treatment measures in each treatment node are extracted to construct an initial treatment plan. An initial action space is constructed by combining the pre-acquired clinical guidelines. The constraint parameters in the initial action space are optimized by an adaptive fuzzy inference method to obtain optimized constraint parameters.

[0034] A fuzzy decision tree is constructed based on the optimization constraint parameters. The rationality of the initial action space is evaluated based on the fuzzy decision tree to obtain a rationality score. The feature mapping matrix is ​​then initialized based on the rationality score.

[0035] Based on the feature mapping matrix, the pre-acquired current physiological data is mapped to the feature space corresponding to the optimization constraint parameters to obtain the state vector.

[0036] In one alternative implementation,

[0037] Selecting a treatment action from the initial action space and calculating the immediate reward value, combined with Monte Carlo search to predict the expected return, and updating the initial treatment plan based on the predicted expected return to obtain the optimal treatment plan includes:

[0038] Collect treatment actions in the initial action space and determine the temporal dependencies between different treatment actions; construct an action transition graph based on the temporal dependencies and the treatment actions.

[0039] determine an action vertex set and a dependency edge set based on the action transition graph, calculate a timing constraint function according to the dependency edge set, add the state vector to the timing constraint function to obtain a predicted state transition result;

[0040] calculate a state improvement degree of the predicted state transition result relative to the state vector, calculate an immediate reward value in combination with treatment resource consumption, perform Monte Carlo search on the immediate reward value, calculate a confidence upper limit interval and select a search node to obtain the predicted expected return;

[0041] add the predicted state transition result and the predicted expected return to a preset state-action value function and perform iterative optimization through a timing difference algorithm, update the initial treatment scheme according to an optimization result to obtain an optimal treatment scheme.

[0042] A second aspect of the embodiment of the present application provides a disease-specific emergency first-aid whole-process auxiliary decision system based on multi-modal perception and AI, comprising:

[0043] A first unit is configured to acquire multi-modal data of an emergency scene and perform feature extraction to obtain an initial feature vector, calculate a timing change of the initial feature vector and construct a feature dependency graph in combination with a conditional probability, calculate a timing correlation degree between feature nodes in the feature dependency graph, determine a key feature sequence based on the timing correlation degree and identify a disease-specific type and a corresponding symptom feature vector;

[0044] A second unit is configured to perform partition indexing on preset disease-specific diagnosis and treatment data based on the symptom feature vector, determine a to-be-indexed subset and calculate an initial similarity between the symptom feature vector in the to-be-indexed subset and a historical case, determine a candidate case and perform timing analysis on the candidate case to determine a symptom evolution rule, calculate a treatment matching degree of current physiological data and the symptom evolution rule, determine a reference case based on the treatment matching degree and extract a treatment path corresponding to the reference case as a reference path;

[0045] A third unit is configured to generate an initial treatment scheme based on the reference path and construct an initial action space in combination with clinical norms, construct a state vector in combination with the current physiological data, select a treatment action in the initial action space and calculate an immediate reward value, predict an expected return in combination with Monte Carlo search, and update the initial treatment scheme based on the predicted expected return to obtain an optimal treatment scheme.

[0046] A third aspect of the embodiment of the present application provides an electronic device, comprising:

[0047] a processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0048] In a fourth aspect, the present application provides a computer readable storage medium having stored thereon computer program instructions which, when executed by a processor, implement the method described above.

[0049] In the present application, through multi-modal data perception and timing feature analysis, accurate identification of specific disease types and effective extraction of symptom features are realized, effectively improving the diagnosis accuracy of emergency scenes, solving the problem that traditional methods cannot comprehensively capture complex symptom manifestations, adopting a partition index strategy based on symptom feature vectors and a timing evolution analysis method, which can quickly locate similar historical cases and accurately evaluate the treatment matching degree, greatly shortening the emergency decision-making time and improving the emergency efficiency, while ensuring the clinical relevance and individual applicability of the treatment scheme, through instant reward calculation and Monte Carlo search to predict the expected income, the treatment scheme can be optimized and adapted to the patient's state changes, significantly improving the timeliness and accuracy of emergency intervention, and improving the survival rate and prognosis quality of patients. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 A flowchart of the method for assisting in decision-making of the whole process of specific disease emergency based on multi-modal perception and AI of the embodiment of the present application is shown in

[0051] Figure 2 A treatment decision generation flowchart based on a reference path for the method for assisting in decision-making of the whole process of specific disease emergency based on multi-modal perception and AI of the embodiment of the present application is shown in DETAILED DESCRIPTION

[0052] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions of the embodiments of the present application will be described in detail below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0053] The technical solutions of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes can not be described in some embodiments.

[0054] Figure 1 A flowchart of the method for assisting in decision-making of the whole process of specific disease emergency based on multi-modal perception and AI of the embodiment of the present application is shown in Figure 1 As shown, the method comprises:

[0055] The multi-modal data of the emergency scene is acquired and feature extraction is performed to obtain an initial feature vector, time sequence changes of the initial feature vector are calculated, and a feature dependency graph is constructed by combining conditional probabilities, time sequence correlation degrees between feature nodes in the feature dependency graph are calculated, key feature sequences are determined based on the time sequence correlation degrees, and a specific disease type and a corresponding symptom feature vector are identified;

[0056] The pre-set specific disease diagnosis and treatment data are indexed based on the symptom feature vector, a to-be-indexed subset is determined, initial similarity of the symptom feature vector in the to-be-indexed subset and historical cases is calculated, candidate cases are determined, time sequence analysis is performed on the candidate cases to determine a symptom evolution rule, a treatment matching degree of current physiological data and the symptom evolution rule is calculated, a reference case is determined based on the treatment matching degree, and a treatment path corresponding to the reference case is extracted as a reference path;

[0057] The initial treatment scheme is generated based on the reference path, an initial action space is constructed by combining clinical norms, a state vector is constructed by combining the current physiological data, a treatment action is selected in the initial action space, and an immediate reward value is calculated, an expected return is predicted by combining the Monte Carlo search, and the initial treatment scheme is updated based on the predicted expected return to obtain an optimal treatment scheme.

[0058] In an optional implementation,

[0059] The multi-modal data of the emergency scene is acquired and feature extraction is performed to obtain an initial feature vector, time sequence changes of the initial feature vector are calculated, and a feature dependency graph is constructed by combining conditional probabilities, time sequence correlation degrees between feature nodes in the feature dependency graph are calculated, key feature sequences are determined based on the time sequence correlation degrees, and a specific disease type and a corresponding symptom feature vector are identified;

[0060] The initial feature vector is obtained by performing feature extraction on the multi-modal data of the emergency scene, a first-order difference value of the initial feature vector is calculated, and a turning point is determined based on the first-order difference value and a pre-set adaptive threshold;

[0061] The initial feature vector is divided into a plurality of feature segments according to the turning point, a segmented feature function is constructed by modeling each feature segment, and a transition connection function of time-varying weights between adjacent feature segments is constructed based on the segmented feature function to obtain a time sequence change sequence;

[0062] The conditional probability is calculated based on the time sequence change sequence, a directed edge is established between corresponding feature nodes when the conditional probability exceeds a pre-set dynamic threshold, and the directed edge is repeatedly constructed to obtain the feature dependency graph.

[0063] Multi-modal data is collected at the scene of emergency, including physiological signals such as electrocardiogram, blood pressure, respiration, image data such as ultrasound, tomography, speech data such as patient symptom description, and video data such as patient expression, body state, etc. Initial feature vectors are obtained by using feature extraction algorithms. For physiological signal data, time-frequency features are extracted by wavelet transform. For image data, spatial features are extracted by convolutional neural network. For speech data, mel-frequency cepstral coefficient features are extracted. For video data, dynamic features are obtained by combining optical flow method and deep feature extraction technology. The extracted features are integrated into initial feature vectors, represented as a multi-dimensional array, with each dimension corresponding to a specific feature attribute.

[0064] After obtaining the initial feature vector, the first-order difference value of the vector is calculated. The first-order difference value reflects the rate of change of the feature in the time dimension, and the calculation method is the difference between the feature values of adjacent time points. For example, for the feature vector representing the patient's heart rate [120, 123, 128, 125, 122, 118, 115], the first-order difference value is [3, 5, -3, -3, -4, -3]. An adaptive threshold mechanism is introduced here, which dynamically adjusts the threshold value according to the historical statistical distribution of feature changes. The adaptive threshold is determined by calculating the weighted combination of the standard deviation and mean of the historical data, and is updated as the data continues to be input. When the first-order difference value exceeds the adaptive threshold, the point is marked as an inflection point. In the aforementioned example, if the adaptive threshold is ±4, the positions with difference values of 5 and -4 are marked as inflection points, corresponding to the 3rd and 5th elements in the original feature vector.

[0065] According to the determined inflection points, the initial feature vector is divided into multiple feature segments, and the original feature vector is divided into three segments: [120, 123, 128], [128, 125, 122, 118], and [118, 115]. Each feature segment is modeled and a segment feature function is constructed. The construction of the segment feature function uses piecewise linear regression or polynomial fitting method, and the most suitable modeling method is selected according to the data characteristics in the segment. For heart rate data, assuming that the piecewise linear regression method is used, the feature function of the first segment can be represented as an increasing relationship, the second segment as a decreasing relationship, and the third segment also as a decreasing relationship.

[0066] A time-varying weight transition connection function is constructed between adjacent feature segments to achieve smooth transition. The transition connection function uses a weight decay strategy to make the adjacent segments transition smoothly at the connection point. In specific implementation, a time window size is defined, and within the window, the weight distribution ratio is gradually adjusted over time. For example, when transitioning from the first segment to the second segment, within the window near the inflection point, the closer the time is to the inflection point, the greater the weight of the second segment feature function and the smaller the weight of the first segment feature function. The time-varying weight transition connection function ensures the continuity of feature change when transitioning between segments, resulting in a complete time series change sequence.

[0067] The conditional probability is calculated based on the obtained time series of changes. In a first aid scenario, for example, given the condition that the patient's heart rate is rapidly decreasing, the probability that the blood pressure will also decrease. The calculation of the conditional probability considers the trend of feature changes in multiple time windows, and extracts the probability distribution from the historical data through statistical learning methods. For patient data of acute myocardial infarction, it can be concluded that the conditional probability between heart rate decrease and blood pressure decrease is 0.85, indicating that in the case of rapid heart rate decrease, there is an 85% chance that blood pressure will also decrease.

[0068] A dynamic threshold is set to determine the significance of the conditional probability. The dynamic threshold is dynamically adjusted according to the first aid situation and the importance of the features, and a lower threshold is set for key vital sign indicators and a higher threshold is set for secondary indicators. When the calculated conditional probability exceeds the preset dynamic threshold, a directed edge is established between the corresponding feature nodes. If the conditional probability between heart rate and blood pressure is 0.85, which exceeds the set dynamic threshold of 0.75, a directed edge is established between the heart rate node and the blood pressure node, indicating that heart rate changes may cause blood pressure changes. By repeatedly constructing directed edges between multiple feature nodes, a complete feature dependency graph is formed.

[0069] In this embodiment, by extracting features from multi-modal data in the first aid scene and introducing first-order difference and adaptive threshold to identify inflection points, the mutation points of features over time can be more sensitively captured, and key abnormalities can be avoided. By dividing the initial feature vector into multiple segments and constructing a segmented feature function, and then using a time-varying weight transition connection function for modeling, the dynamic transition relationship between different feature stages can be more accurately reflected, and the time series adaptability is higher. By calculating the conditional probability based on the time series of changes and combining the dynamic threshold to establish the feature dependency graph, the dynamic dependency relationship between features can be effectively identified, and the accuracy and reliability of multi-modal feature correlation mining in complex scenarios are improved.

[0070] In an alternative embodiment,

[0071] The time series correlation between the feature nodes in the feature dependency graph is calculated, and the key feature sequence is determined based on the time series correlation to identify the special disease type and the corresponding symptom feature vector, including:

[0072] For each feature node in the feature dependency graph, the PageRank centrality is calculated in combination with the neighbor nodes of the current feature node, the connection probability between the current feature node and the neighbor nodes is calculated to determine the information entropy, and the importance weight of the current feature node is calculated based on the PageRank centrality and the information entropy.

[0073] The distance between different feature nodes is calculated, and the similarity between nodes is determined, the initial correlation degree is calculated based on the similarity between nodes and the importance weight, and the time sequence correlation degree is obtained by time sequence adjustment based on the initial correlation degree and the preset time sequence attenuation factor;

[0074] The feature mapping relationship corresponding to each special disease type is extracted in the preset special disease type knowledge base, the feature nodes with a time sequence correlation degree greater than a preset correlation degree threshold are taken as key nodes, the combination correlation between the key nodes is analyzed, and the key feature sequence is constructed in combination with the feature mapping relationship, the matching degree of each feature in the key feature sequence with the special disease type is calculated to obtain a special disease type probability value, and the special disease type corresponding to the current feature is determined, and the symptom feature vector is obtained by extracting the time sequence correlation degree corresponding to the special disease type.

[0075] For each feature node in the feature dependency graph, the PageRank algorithm is used to calculate the centrality, and the number of neighbor nodes and their own centrality values are considered comprehensively. Taking the feature dependency graph of acute myocardial infarction patients as an example, the electrocardiogram ST segment elevation node is connected to multiple important physiological indicator nodes, such as heart rate, blood pressure, etc., and the PageRank centrality value calculation result is 0.28, while the ordinary pain indicator node is only connected to two secondary nodes, and the centrality value is only 0.05.

[0076] The connection probability between the current feature node and its neighbor nodes is calculated, reflecting the degree of uncertainty of the relationship between nodes, and then the information entropy is determined. The connection probability is obtained by normalizing the weight of the edge between nodes, and the information entropy is calculated by the weighted sum of the connection probability. For example, the connection probabilities of the heart rate node and its three neighbor nodes are 0.5, 0.3 and 0.2, and the information entropy is calculated to be 1.03. The higher the information entropy, the more complex and variable the connection relationship of the node, and there is more uncertainty in clinical judgment.

[0077] The importance weight of the feature node is calculated based on the PageRank centrality and the information entropy, the centrality value and the reciprocal of the information entropy are weighted and combined, and the weight coefficient is set according to the specific needs of the emergency scene. For myocardial infarction situation, assuming that the weight coefficients are 0.7 and 0.3 respectively, the importance weight of the electrocardiogram ST segment elevation node is 0.28x0.7+1 / 1.03x0.3≈0.486, indicating that it has high importance in myocardial infarction judgment.

[0078] The distance between different feature nodes is calculated and the similarity between nodes is determined. The shortest path algorithm is used to calculate the distance between nodes, considering the weight influence of the edge. The similarity is a function of the distance, the shorter the distance, the higher the similarity. In the feature dependency graph, the path distance between the ST segment elevation node of electrocardiogram and the myocardial enzyme index node is 1.2, and the similarity is 0.833. The similarity between nodes reflects the closeness of the clinical association between two features, and the features with high similarity often point to the same pathological state.

[0079] The initial association degree is calculated based on the similarity and importance weight between nodes. The importance weight of two nodes is multiplied by the similarity between nodes. Taking the ST segment elevation node of electrocardiogram and the myocardial enzyme index node as an example, assuming that the importance weight of the myocardial enzyme index node is 0.420, then the initial association degree is 0.486 x 0.420 x 0.833 ≈ 0.170.

[0080] A time decay factor is introduced to adjust the initial association degree in time, and the time association degree is obtained. For rapidly changing physiological indicators in emergency scenes, a small decay factor such as 0.95 is set; for relatively stable indicators, a larger decay factor such as 0.99 is set. For the detected ST segment elevation and myocardial enzyme index, if the time interval is 10 minutes and the decay factor is 0.95, then the time association degree is 0.170 x 0.95^10 ≈ 0.104.

[0081] The feature mapping relationship corresponding to each special disease type is extracted in the pre-set special disease type knowledge base. The special disease knowledge base stores various feature patterns of acute diseases, such as the ST segment elevation, myocardial enzyme elevation, chest pain, and other features and their typical performance value range, time sequence, and mutual association of acute myocardial infarction knowledge entries. The feature mapping relationship is stored in a structured form, which is convenient for quick retrieval and matching. The feature mapping relationship of acute myocardial infarction includes: ST segment elevation → myocardial enzyme elevation (time delay 30-60 minutes), chest pain → ST segment change (occur simultaneously or delay 5-10 minutes), etc.

[0082] The feature nodes with a time association degree greater than a pre-set association degree threshold are marked as key nodes. If the pre-set threshold is 0.10, then the time association degree of the ST segment elevation and myocardial enzyme feature nodes 0.104 is greater than the threshold, and they are marked as key nodes.

[0083] The combination association between key nodes is analyzed and the key feature sequence is constructed in combination with the feature mapping relationship. The combination association is calculated by combining the time association degrees between multiple key nodes, reflecting the clinical significance of the co-occurrence of multiple features. In the case of myocardial infarction, the key feature sequence may be: chest pain → ST segment elevation → myocardial enzyme elevation → heart rate instability → blood pressure drop. This sequence is matched with the feature mapping relationship of myocardial infarction in the special disease knowledge base, and the matching degree is calculated.

[0084] The matching degree of each feature in the key feature sequence with the special disease type is calculated to obtain a special disease type probability value. The matching degree calculation considers multiple factors such as feature values, occurrence time sequence, duration, and the like, and adopts a weighted similarity calculation method. For example, the ST segment elevation of a patient is 2 mm, and the duration is 15 minutes, and the matching degree with the typical feature (ST segment elevation ≥ 1.5 mm, and duration ≥ 10 minutes) of central myocardial infarction in the knowledge base is 0.92; the myocardial enzyme index is elevated by 35%, and the matching degree with the typical feature (elevation ≥ 30%) is 0.88. The matching degrees of all key features are comprehensively calculated to obtain the special disease type probability value of the symptoms of the patient with acute myocardial infarction as 0.85.

[0085] The special disease type corresponding to the current feature is determined according to the special disease type probability value. When the probability value exceeds a preset threshold (such as 0.80), it is confirmed as the corresponding special disease type. After the special disease type is determined, the time sequence correlation degree of all feature nodes corresponding to the special disease type is extracted from the feature dependency graph to form a symptom feature vector. For acute myocardial infarction, the symptom feature vector contains the time sequence correlation degree values between the key feature nodes.

[0086] In this embodiment, the importance weight of the feature node is calculated by combining the PageRank center degree and the information entropy in the feature dependency graph, which can improve the accuracy of feature node screening, avoid the deviation caused by a single index, and reflect the time sequence evolution relationship between the features by simultaneously considering the node similarity and the importance weight in the calculation of the node correlation and introducing a time sequence attenuation factor for dynamic adjustment, thereby improving the timeliness and rationality of the correlation modeling. By combining the time sequence correlation degree with the feature mapping relationship in the special disease type knowledge base, the key nodes can be effectively screened and the key feature sequence can be constructed, and the matching precision between the features and the special disease types can be improved.

[0087] In an alternative embodiment,

[0088] The pre-set special disease diagnosis and treatment data are partitioned and indexed based on the symptom feature vector, a to-be-indexed subset is determined, and the initial similarity between the symptom feature vector in the to-be-indexed subset and the historical cases is calculated, which includes:

[0089] The difference between each symptom feature value in the symptom feature vector and the pre-set normal reference mean value is calculated to determine the abnormality degree of each symptom, and the danger level corresponding to each symptom is determined based on the abnormality degree;

[0090] partitioning the symptom feature vector into multiple priority partitions based on the risk level, solving a first early warning threshold based on the risk level and a pre-set first balance parameter, for each symptom feature vector in each priority partition, solving a second early warning threshold by calculating a change trend and combining a pre-set second balance parameter, obtaining a combined early warning value based on the first early warning threshold and the second early warning threshold, and constructing an index structure based on the combined early warning value and the symptom feature vector;

[0091] locating a to-be-matched partition from the special disease diagnosis and treatment data based on the index structure, calculating a partition matching degree between the symptom feature vector and each to-be-matched partition, and selecting three to-be-matched partitions with the highest partition matching degrees to obtain a to-be-indexed subset;

[0092] calculating the Euclidean distance between the symptom feature vector and each historical case in the to-be-indexed subset and determining a feature similarity, calculating a timeliness weight based on time information of the historical cases and combining the feature similarity to obtain an initial similarity.

[0093] calculating the difference between each symptom feature value in the symptom feature vector and a pre-set normal reference mean value to determine the abnormality degree of each symptom. The normal reference mean value is statistically obtained based on a large number of healthy population data, and different reference values are set for different genders and age groups. For an acute myocardial infarction patient, the ST segment elevation in the symptom feature vector is 2.5 millimeters, while the normal reference mean value is 0.1 millimeter, the difference is 2.4 millimeters, and the standardized abnormality degree is 0.96; the myocardial enzyme index is 3.8 times the normal value, corresponding to an abnormality degree of 0.92; the blood pressure is 85 / 50 millimeters of mercury, which is lower than the normal reference value of 120 / 80 millimeters of mercury, and the abnormality degree is 0.74. The abnormality degree is represented by a standardized value of 0 to 1, and the value closer to 1 indicates a higher abnormality degree.

[0094] determining the risk level corresponding to each symptom based on the abnormality degree. The risk level is divided into four levels: mild abnormality, moderate abnormality, severe abnormality, and extremely severe abnormality, corresponding to abnormality degree intervals of 0-0.25, 0.25-0.50, 0.50-0.75, and 0.75-1. In the foregoing example, the ST segment elevation and the myocardial enzyme index belong to extremely severe abnormality, and the blood pressure belongs to severe abnormality. The risk level assessment not only considers the current abnormality degree, but also combines the symptom change rate, and the symptoms that worsen quickly increase the risk level. For example, the ST segment elevation increases from 1.5 millimeters to 2.5 millimeters within 10 minutes, the change rate is high, and it is confirmed as extremely severe abnormality; while the mild dizziness symptom lasts for 30 minutes without significant aggravation, it remains at the mild abnormality level.

[0095] The symptom feature vectors are divided into multiple priority partitions based on the risk level. The priority partitions are used to distinguish the urgency of therapeutic intervention. The extremely abnormal symptoms are classified into the highest priority partition, requiring immediate intervention; the severely abnormal symptoms are classified into the high priority partition, requiring close monitoring and preparation for intervention; the moderately and mildly abnormal symptoms are classified into the medium and low priority partitions, respectively. For patients with myocardial infarction, the ST segment elevation and myocardial enzyme indicators are classified into the highest priority partition, blood pressure is classified into the high priority partition, and other minor symptoms are classified into a lower priority partition.

[0096] The first early warning threshold is obtained based on the risk level and a pre-set first balance parameter. The first balance parameter reflects the medical resource status and the overall risk tolerance of the patient, and the value range is 0 to 1. In the case of sufficient resources, the balance parameter takes a smaller value such as 0.3, making the system more sensitive; in the case of resource shortage, a larger value such as 0.7 is taken to avoid excessive early warning. The first early warning threshold is calculated by weighted summation, with the highest weight given to the extremely abnormal, such as 0.5, the weight of the severe abnormal is 0.3, the weight of the moderate abnormal is 0.15, and the weight of the mild abnormal is 0.05.

[0097] For the symptom feature vector in each priority partition, the change trend is calculated and combined with a pre-set second balance parameter to obtain the second early warning threshold. The change trend is calculated by the symptom feature values at multiple consecutive time points, and is represented by the linear regression slope. For the ST segment elevation indicator, the change trend is positive growth from 1.5 mm to 2.5 mm within 10 minutes, with a slope of 0.1 mm / min; the myocardial enzyme indicator increases from 2.6 times to 3.8 times of the normal value within 30 minutes, with a change trend slope of 0.04 times / min. The second balance parameter represents the sensitivity to the change trend, and in the case of acute myocardial infarction, a smaller value such as 0.25 is taken, and the second early warning threshold of the highest priority partition is 0.78, and the second early warning threshold of the high priority partition is 0.65.

[0098] The combined early warning value is obtained based on the first early warning threshold and the second early warning threshold. For the highest priority partition, the first early warning threshold weight is 0.6 and the second early warning threshold weight is 0.4; for other priority partitions, the weights are 0.7 and 0.3, respectively, so the combined early warning value of the highest priority partition of the patient is 0.84 x 0.6 + 0.78 x 0.4 = 0.816, and the combined early warning value of the high priority partition is 0.84 x 0.7 + 0.65 x 0.3 = 0.783.

[0099] An index structure is constructed based on the combination of the warning value and the symptom feature vector. The index structure adopts a multi-level hash table design, with the first-level index being the disease type, the second-level index being the priority partition, and the third-level index being the combination of the warning value range. For the current patient, the index path is: acute myocardial infarction → highest priority partition → combination of warning value 0.80-0.85. This index structure supports fast positioning of similar cases and improves retrieval efficiency.

[0100] Based on the index structure, the matching partition is located from the disease diagnosis and treatment data. The disease diagnosis and treatment data contain a large number of historical cases and their diagnosis and treatment process and result data. The matching partition located by the index structure contains multiple historical cases, which are similar to the current patient in terms of disease type, symptom priority and warning value. For example, through the aforementioned index path, a matching partition containing 12 historical cases is located.

[0101] The partition matching degree between the symptom feature vector and each matching partition is calculated. The partition matching degree is calculated by cosine similarity, considering the distribution of the symptom feature vector in each partition. For the three possible matching partitions, the calculated partition matching degrees are 0.92, 0.87 and 0.83, respectively. The matching partition with the highest partition matching degree is selected to obtain the index subset, which contains a total of 35 historical cases.

[0102] The Euclidean distance between the symptom feature vector and each historical case in the index subset is calculated, and the feature similarity is determined. The Euclidean distance calculation considers the weight of each symptom feature, and gives higher weight to key symptoms such as ST segment elevation and myocardial enzyme indicators. The feature similarity is a normalized conversion of the Euclidean distance, and the smaller the distance, the higher the similarity. For a historical case in the index subset, the calculated Euclidean distance is 0.23, and the converted feature similarity is 0.88.

[0103] The timeliness weight is calculated based on the time information of the historical cases, and the initial similarity is calculated by combining the feature similarity. The timeliness weight reflects the time evolution of medical technology and treatment scheme, and the timeliness weight calculation adopts an exponential decay function, taking the current time as the benchmark, the closer the case time, the higher the weight. For example, the timeliness weight of a case one year ago is 0.95, the timeliness weight of a case three years ago is 0.86, and the timeliness weight of a case five years ago is 0.78. Multiply the feature similarity by the timeliness weight to obtain the initial similarity.

[0104] In this embodiment, the severity of the symptoms is finely distinguished by determining the abnormality degree by calculating the difference between the symptom characteristic value and the normal reference mean value, and dividing the risk level according to the abnormality degree, avoiding misjudgment caused by a single threshold value. The flexibility and accuracy of the early warning mechanism can be improved by dynamically solving the first and second early warning thresholds in different priority partitions combined with the risk level and the balance parameter, and generating a combined early warning value, and the adaptability to complex symptom change trends can be enhanced. The search range can be effectively reduced and the matching efficiency can be improved by constructing an index structure based on the combined early warning value and locating the matching partition. The real-time performance and reliability of the matching result can be improved while ensuring the matching accuracy by introducing partition matching degree screening and Euclidean distance calculation feature similarity in the to-be-indexed subset, and adjusting combined with the timeliness weight of the historical case.

[0105] In an alternative embodiment,

[0106] Determining candidate cases and performing time series analysis on the candidate cases to determine the symptom evolution rule, calculating the treatment matching degree of the current physiological data and the symptom evolution rule, determining the reference case based on the treatment matching degree and extracting the treatment path corresponding to the reference case as the reference path comprises:

[0107] Based on the initial similarity, determine the candidate cases in the to-be-indexed subset, perform time series analysis on the candidate cases by dynamic time warping method, calculate the time series distance corresponding to each candidate case and determine the symptom evolution rule based on the time series distance;

[0108] Acquire the current physiological data, calculate the disease complexity matching value and medical resource matching degree of the current physiological data and the symptom evolution rule and sum them to obtain the treatment matching degree;

[0109] Select the candidate case with the highest treatment matching degree as the reference case, construct a treatment optimization objective function according to the complication risk value, solve the treatment optimization objective function by differential evolution algorithm to obtain the optimal solution, extract the feasibility score corresponding to the optimal solution, and if the feasibility score is greater than the pre-set feasibility threshold, extract the treatment path and output it as the reference path.

[0110] An initial similarity is obtained, and historical cases with an initial similarity greater than a pre-set threshold value are selected as candidate cases in the to-be-indexed subset. For acute myocardial infarction patients, the initial similarity threshold value is set to 0.80, and 8 candidate cases are selected from 35 historical cases, with initial similarities of 0.836, 0.825, 0.818, 0.814, 0.809, 0.805, 0.802 and 0.801.

[0111] The time series of the candidate cases are analyzed by dynamic time warping method, and the time series distance of symptom evolution is calculated. The time axis is stretched and aligned to the sequence of symptom development. Taking ST segment elevation as an example, the current patient takes 40 minutes from the onset of symptoms to the elevation of 2.5 mm, while a candidate case takes 55 minutes to reach a similar degree, and the time series distance calculated by dynamic time warping is 0.24. Similar calculations are made for all key symptoms, and the total time series distance of each candidate case is obtained, such as 0.28, 0.31, 0.25, etc. The smaller the time series distance, the more similar the symptom evolution process.

[0112] Determine the evolution rule of symptoms based on time series distance. Analyze the symptom change curve of each candidate case, extract key time nodes and turning points, and exemplarily, the evolution rule of ST segment elevation of a candidate case is: initial slight elevation (0.8 mm), rapid increase to 2.0 mm within 30 minutes, then slow increase to 2.8 mm within 60 minutes, and gradually decrease to 1.2 mm within 90 minutes after medication. Through clustering analysis of the symptom evolution rules of multiple candidate cases, three typical evolution patterns are obtained: rapid deterioration type, slow progression type and fluctuant progression type. The symptom evolution characteristics of the current patient are closest to the rapid deterioration type, indicating that the disease develops rapidly and needs emergency intervention.

[0113] Collect the physiological data of the current patient, including real-time indicators such as electrocardiogram, blood pressure, heart rate, respiratory rate, blood oxygen saturation, etc. The current patient's electrocardiogram shows that the ST segment is elevated by 2.5 mm and continues to increase, the heart rate is 110 beats per minute, the blood pressure is 85 / 50 mmHg, the respiratory rate is 22 times per minute, and the blood oxygen saturation is 92%.

[0114] Calculate the disease complexity matching value of the current physiological data and the symptom evolution rule. Disease complexity considers multiple dimensions: symptom severity, multi-system involvement, complication risk and patient's basic condition. The symptom severity score of the current patient is 8.5 points (full score 10 points), the multi-system involvement is moderate (heart, circulatory system), the complication risk is high (risk of arrhythmia), the patient has a history of hypertension and diabetes, and the basic condition score is 7 points (full score 10 points). Comprehensive calculation shows that the disease complexity is 8.2 points. Compare the disease complexity of each candidate case (such as 7.8 points, 8.4 points, 7.5 points, etc.), and calculate the disease complexity matching value, such as 0.95, 0.98, 0.91, etc. The closer the matching value is to 1, the closer the disease complexity is.

[0115] A medical resource matching degree is calculated, taking into account factors such as device availability, professional configuration, drug supply, and treatment space. The current emergency environment is a pre-hospital emergency vehicle equipped with basic electrocardiogram monitoring equipment, a defibrillator, an emergency medicine kit, and a senior first aid technician and a general first aid attendant. Compared with the required medical resources of the candidate cases, the medical resource matching degree is calculated, such as 0.65, 0.72, 0.58, etc. The higher the matching degree, the more the existing resources can meet the treatment needs.

[0116] The treatment matching degree is obtained by weighted sum of the illness complexity matching value and the medical resource matching degree. The weight setting is dynamically adjusted according to the emergency environment, and the weight of the illness complexity matching value is 0.6 and the weight of the medical resource matching degree is 0.4 in the pre-hospital emergency stage. The treatment matching degrees of the eight candidate cases are calculated as 0.83, 0.87, 0.78, 0.85, 0.81, 0.76, 0.79, and 0.74, respectively. The treatment matching degree reflects the applicability of the treatment plan of the candidate cases to the current patient.

[0117] The candidate case with the highest treatment matching degree is selected as the reference case. In the foregoing example, the treatment matching degree of the second candidate case is 0.87, which is the highest, so this case is selected as the reference case. The reference case is a 62-year-old male patient who also has acute myocardial infarction, similar ST segment elevation position and degree, and obtained good treatment effect under similar medical resource conditions.

[0118] A treatment optimization objective function is constructed according to the complication risk value. The complication risk value is based on the physiological data and underlying disease of the current patient, mainly considering arrhythmia risk, shock risk, and heart failure risk. The arrhythmia risk of the current patient is 0.75, the shock risk is 0.68, and the heart failure risk is 0.56. The treatment optimization objective function considers the maximization of treatment effect and the minimization of complication risk, and the function variables include drug selection, administration timing, administration dose, referral timing, and other treatment parameters.

[0119] The optimal solution is obtained by solving the treatment optimization objective function by differential evolution algorithm, with population size of 50, maximum iteration number of 100, crossover probability of 0.8, and mutation factor of 0.5. The algorithm converges after about 68 iterations, and the optimal solution is obtained: immediately give dual antiplatelet therapy (aspirin 300mg orally, clopidogrel 300mg orally), intravenous injection of nitroglycerin 0.4mg, and establish intravenous access for thrombolytic therapy, and immediately contact the nearest medical institution with catheter room for transfer preparation.

[0120] The feasibility score corresponding to the optimal solution is extracted. The feasibility score evaluates the implementation possibility of the optimal solution under the current conditions, considering factors such as drug accessibility, operation difficulty, time requirement, and risk control. The calculated feasibility score is 0.92, which is higher than the pre-set feasibility threshold of 0.85, indicating that the optimal solution has a high implementation possibility under the current conditions.

[0121] The optimal solution corresponding to the treatment path is extracted as the reference path output. The complete reference path includes: T0 time: dual antiplatelet therapy + nitroglycerin intravenous injection + establishment of venous access; T0+10 minutes: evaluate thrombolytic conditions, if satisfied, implement thrombolytic therapy; T0+15 minutes: start the transportation program, contact the receiving hospital; T0+20 minutes: start transportation, continuously monitor vital signs during transportation; T0+40 minutes: arrive at the receiving hospital, hand over the patient to the catheter room team; T0+60 minutes: complete coronary intervention.

[0122] In this embodiment, the dynamic time warping method is introduced in the candidate cases for time series analysis, and the symptom evolution rule is determined based on the time series distance, which can more accurately reflect the dynamic characteristics of the patient's condition development over time, avoid the deviation caused by static feature comparison, and calculate the disease complexity matching value and medical resource matching degree by combining the current physiological data with the symptom evolution rule. The reference value of the case can be comprehensively evaluated, the clinical applicability of the matching result can be improved, the reference path can be filtered and output through the feasibility score, and the scientificity and feasibility of the treatment path recommendation can be effectively improved.

[0123] In an alternative embodiment,

[0124] Based on the reference path, an initial treatment plan is generated and an initial action space is constructed in combination with clinical norms, and a state vector is constructed in combination with the current physiological data, which includes:

[0125] The reference path is time-decomposed to obtain a plurality of treatment nodes, the treatment measures in each treatment node are extracted to construct an initial treatment plan, the initial action space is constructed in combination with the pre-acquired clinical norms, and the constraint parameters in the initial action space are optimized by an adaptive fuzzy reasoning method to obtain optimized constraint parameters;

[0126] Based on the optimized constraint parameters, a fuzzy decision tree is constructed, the initial action space is reasonably evaluated based on the fuzzy decision tree, a reasonability score is obtained, and a feature mapping matrix is initialized based on the reasonability score;

[0127] The pre-acquired current physiological data is mapped into the feature space corresponding to the optimized constraint parameters based on the feature mapping matrix, and a state vector is obtained.

[0128] The reference pathway is time-decomposed into multiple treatment nodes. For example, the reference pathway of an acute myocardial infarction patient is divided into six treatment nodes: initial assessment and medication node T0, thrombolytic evaluation node T0+10 minutes, transport preparation node T0+15 minutes, transport implementation node T0+20 minutes, hospital handover node T0+40 minutes, and coronary intervention treatment node T0+60 minutes. Each node has a clear time marker and corresponding treatment goals.

[0129] The treatment measures in each treatment node are extracted to construct an initial treatment plan. The treatment measures for the initial assessment and medication node T0 include establishing a venous access, administering aspirin 300 mg orally, clopidogrel 300 mg orally, nitroglycerin 0.4 mg intravenously, and continuous electrocardiogram monitoring. The treatment measures for the thrombolytic evaluation node T0+10 minutes include evaluating thrombolytic indications and contraindications, and preparing alteplase 50 mg. The treatment measures for the transport preparation node T0+15 minutes include contacting the receiving hospital, confirming the preparation of the catheter room, and preparing the transport equipment. The treatment measure extraction process takes into account the specific details of the operation, drug dosage, and administration route to ensure the completeness and executability of the treatment plan.

[0130] An initial action space is constructed in combination with pre-acquired clinical guidelines. The clinical guidelines are derived from authoritative first aid guidelines and expert consensus, providing standardized processing procedures for various emergency situations. For acute myocardial infarction, the clinical guidelines specify the order of drug use, dosage range, contraindications, and standard procedures for various operations. The initial action space is represented in the form of a multi-dimensional matrix, including dimensions such as action type, execution time, operator, required resources, and expected effect. For example, the action space of anti-platelet therapy includes parameters such as drug selection (e.g., aspirin, clopidogrel), dosage range (e.g., aspirin 100-300 mg, clopidogrel 75-300 mg), administration route (e.g., oral, intravenous), and administration timing (e.g., immediately, within 10 minutes).

[0131] The constraint parameters in the initial action space are optimized through an adaptive fuzzy reasoning method. The constraint parameters include dosage range, time window, operation sequence, and other limiting conditions. The adaptive fuzzy reasoning method is based on fuzzy set theory, representing the constraint parameters as fuzzy membership functions. For aspirin dosage, the membership function maps the dosage range of 100-300 mg to a suitability score of 0-1. The membership function is dynamically adjusted according to the specific circumstances of the patient, for example, for a patient with high bleeding risk, the membership function of aspirin dosage is shifted towards low dosage, making the membership value of 150 mg higher than that of 300 mg. Through a feedback mechanism, the membership function is continuously optimized, and when new patient data is obtained, the membership function is automatically adjusted to adapt to changes.

[0132] For the constraint parameter of thrombolytic time window, the initial setting is within 6 hours after the onset of symptoms, and after optimization by adaptive fuzzy reasoning, the window is reduced to within 4 hours for patients over 70 years old, and the membership value gradually decreases as time increases. The optimized constraint parameter is no longer a hard limit, but a fuzzy range with membership, which is more in line with the complexity of actual clinical decision-making. For example, the optimized aspirin dose constraint is: less than 100 mg with a membership of 0.2, 100 to 200 mg with a membership of 0.7, 200 to 300 mg with a membership of 0.9, and more than 300 mg with a membership of 0.1.

[0133] A fuzzy decision tree is constructed based on the optimized constraint parameters. The fuzzy decision tree is an extension of the traditional decision tree, with node splitting and path selection based on fuzzy logic, including feature selection, node splitting, and path connection. For acute myocardial infarction treatment decision, the root node is the initial evaluation result, the first layer branch node is the ST segment elevation degree and duration, the second layer branch node is the patient's age and underlying disease condition, and the leaf node is the specific treatment decision. For example, a path may be: ST segment elevation greater than 2 mm with a membership of 0.9, duration greater than 30 minutes with a membership of 0.8, age less than 65 years with a membership of 0.7, and no severe underlying disease with a membership of 0.9, recommending direct transport to the catheter room for coronary intervention treatment.

[0134] Each node of the fuzzy decision tree contains a fuzzy set defined based on the optimized constraint parameters, and the connection strength between nodes is determined by the combination of the membership values of each node. For a patient with ST segment elevation of 2.5 mm, duration of 40 minutes, history of hypertension and diabetes, the path in the fuzzy decision tree is represented as: ST segment elevation greater than 2 mm with a membership of 0.95, duration greater than 30 minutes with a membership of 0.85, and moderate underlying disease with a membership of 0.7. The comprehensive recommendation is to immediately administer anti-platelet therapy and prepare for transport to the catheter room.

[0135] Based on the fuzzy decision tree, the rationality of the initial action space is evaluated, and the rationality score is calculated by traversing the decision tree. During the evaluation process, the matching degree of the action and the patient's state, resource accessibility, time urgency, and risk-reward ratio are considered. For example, for the example patient, the rationality score of anti-platelet therapy with aspirin 300 mg and clopidogrel 300 mg is 0.88, the rationality score of thrombolytic therapy is 0.65, and the rationality score of direct transport to the catheter room is 0.92. The rationality score ranges from 0 to 1, and the higher the score, the more reasonable the action.

[0136] The feature mapping matrix is initialized based on the rationality score, and the matrix element represents the influence weight of a specific physiological feature on a specific treatment decision. The matrix initialization is weighted based on the rationality score, and the treatment decision with a high rationality score obtains a higher initial weight in the matrix. For example, the initial mapping weight of the ST segment elevation feature on the direct transport catheter room decision is 0.85, and the initial mapping weight of the thrombolytic therapy decision is 0.75. The dimension of the feature mapping matrix is determined by the number of physiological features and the number of possible treatment decisions, and for the current patient, the matrix dimension is 15 by 8, including 15 key physiological features and 8 possible treatment decisions.

[0137] The pre-acquired current physiological data is mapped into the feature space corresponding to the optimization constraint parameter based on the feature mapping matrix to obtain a state vector.

[0138] In the embodiment, the initial treatment scheme can be constructed by time sequence decomposition of the reference path and extraction of the treatment nodes, the fine-grained disassembly of the treatment path can be realized, the integrity and operability of the scheme generation are improved, the initial action space is constructed in combination with the clinical specification, the flexibility and adaptability of the scheme are enhanced under the premise of ensuring compliance with the medical specification by the adaptive fuzzy reasoning method for optimizing the constraint parameter, the fuzzy decision tree is constructed based on the optimization constraint parameter and the rationality is evaluated, the unreasonable action combination can be effectively eliminated, the scientificity and reliability of the scheme design are improved, the feature mapping matrix is initialized by the rationality score, and the current physiological data is mapped into the optimized feature space, the individual state of the patient can be accurately expressed, and the subsequent decision process can better match the individualized treatment demand.

[0139] Figure 2 The reference path-based treatment decision generation flowchart of the multi-modal perception and AI-based special disease emergency whole-process auxiliary decision method of the embodiment of the application.

[0140] In an optional implementation,

[0141] The treatment actions in the initial action space are selected and the immediate reward value is calculated, the expected return is predicted by combining the Monte Carlo search, and the optimal treatment scheme is obtained by updating the initial treatment scheme based on the predicted expected return, which includes:

[0142] The treatment actions in the initial action space are selected and the time sequence dependency relationship between different treatment actions is determined, and an action transition graph is constructed based on the time sequence dependency relationship and the treatment actions;

[0143] The action vertex set and the dependency relationship edge set are determined based on the action transition graph, the time sequence constraint function is calculated according to the dependency relationship edge set, and the state vector is added to the time sequence constraint function to calculate the predicted state transition result;

[0144] The state improvement degree of the predicted state transition result relative to the state vector is calculated, an immediate reward value is calculated in combination with treatment resource consumption, a Monte Carlo search is performed on the immediate reward value, a confidence upper limit interval is calculated and a search node is selected to obtain the predicted expected return;

[0145] The predicted state transition result and the predicted expected return are added to a preset state-action value function and iteratively optimized by a time difference algorithm, and the initial treatment scheme is updated according to the optimization result to obtain an optimal treatment scheme.

[0146] Treatment actions in the initial action space are collected and the time sequence dependency relationship between different treatment actions is determined. Taking acute myocardial infarction first aid as an example, the treatment actions extracted from the initial action space include establishing a venous access, giving an anti-platelet drug, intravenous injection of nitroglycerin, evaluating thrombolytic indications, performing thrombolytic therapy and transporting to a catheter room, etc. The time sequence dependency relationship describes the execution order constraint between the treatment actions, such as establishing a venous access must be performed before intravenous injection of nitroglycerin, and anti-platelet therapy can be performed in parallel with intravenous injection of nitroglycerin. By analyzing clinical norms and expert experience, the time sequence dependency relationship between all treatment actions is determined and recorded as a directed relationship pair.

[0147] An action transition graph is constructed based on the time sequence dependency relationship and the treatment actions, and the nodes in the graph represent the treatment actions and the directed edges represent the dependency relationship between the actions. In the action transition graph, a directed edge from node A to node B indicates that action A must be performed before action B. For the treatment process of acute myocardial infarction, the established action transition graph contains 13 action nodes and 21 dependency relationship edges. For example, there is a directed edge between the node "establishing a venous access" and the node "intravenous injection of nitroglycerin", indicating that the venous access must be established before the drug injection. The action transition graph also contains parallel execution paths, such as "giving aspirin" and "giving clopidogrel" can be performed in parallel, and there is no direct dependency edge between them.

[0148] The action vertex set and the dependency edge set are determined based on the action transition graph. The action vertex set is a collection of all the treatment action nodes in the transition graph, and the dependency edge set is a collection of all the directed edges representing the dependency relationship. For the treatment of acute myocardial infarction, the action vertex set includes 13 vertices such as "establishing venous access", "administering aspirin 300mg", "administering clopidogrel 300mg", "intravenous injection of nitroglycerin 0.4mg", "evaluating thrombolytic indications", "preparing thrombolytic drugs", "performing thrombolytic therapy", "contacting receiving hospitals", "preparing transport equipment", "starting transport", "monitoring vital signs", "handing over patients", and "coronary intervention". The dependency edge set includes 21 directed edges such as "establishing venous access → intravenous injection of nitroglycerin", "evaluating thrombolytic indications → preparing thrombolytic drugs", and "preparing thrombolytic drugs → performing thrombolytic therapy".

[0149] The timing constraint function is calculated according to the dependency edge set. The timing constraint function describes the time relationship constraints between treatment actions, including the earliest start time, the latest end time, and the execution duration. The timing constraint function is obtained by analyzing the dependency edge set. For each action node, the earliest start time is not earlier than the latest end time of all the preceding actions. For example, the earliest start time of "intravenous injection of nitroglycerin" is the latest end time of "establishing venous access" plus the time for preparing the drug. For the treatment of acute myocardial infarction, the execution duration of each action is set according to clinical experience, such as "establishing venous access" lasting for 2 minutes, "administering aspirin" lasting for 1 minute, and "evaluating thrombolytic indications" lasting for 5 minutes. The timing constraint function also takes into account the urgency and priority. For patients with unstable vital signs, the execution time of key actions will be shortened accordingly.

[0150] The state vector is added to the timing constraint function to calculate the predicted state transition result. The state vector describes the current physiological state of the patient, including indicators such as ST segment elevation, heart rate, blood pressure, etc. By combining the state vector with the timing constraint function, the change in the patient's state after executing a specific sequence of treatment actions can be predicted. For example, for a patient with a state vector of 0.92, 0.85, 0.78, 0.65, 0.88, 0.72, 0.81, 0.59, after executing "administering aspirin 300mg" and "administering clopidogrel 300mg", the predicted state vector becomes 0.92, 0.87, 0.81, 0.68, 0.90, 0.75, 0.83, 0.62, indicating the expected changes in the patient's physiological indicators.

[0151] The state improvement degree of the predicted state transition result relative to the state vector is calculated. The state improvement degree measures the degree of positive influence of the treatment action on the patient's state, which is calculated by comparing the state vectors before and after the execution of the action. The state improvement degree considers multiple dimensions, including vital sign stability, symptom relief degree, and disease progression control. For acute myocardial infarction patients, the improvement of indicators such as ST segment change, myocardial enzyme index, heart rate, blood pressure, and blood oxygen saturation is focused on. The state improvement degree is calculated in the form of weighted sum, and the weight of each indicator is determined according to the clinical importance. For example, the weight of ST segment regression is 0.35, the weight of heart rate normalization is 0.25, the weight of blood pressure stability is 0.20, and the weight of other indicators is 0.20. For the above example, the state improvement degree after executing anti-platelet therapy is calculated as 0.15.

[0152] The immediate reward value is calculated in combination with the treatment resource consumption. The treatment resource consumption includes drug cost, equipment use, manpower investment, and time cost, etc. The immediate reward value is a comprehensive evaluation of the state improvement degree and resource consumption, reflecting the cost-effectiveness of the treatment action. The calculation uses the state improvement degree minus the weighted value of resource consumption, and the weight is dynamically adjusted according to the resource tightness. For example, the resource consumption score of anti-platelet therapy is 0.05, and the calculated immediate reward value is 0.15-0.05=0.10. The immediate reward value greater than 0 indicates that the benefit of the treatment action is greater than the cost, and the greater the value, the more valuable the action.

[0153] Monte Carlo search is performed on the immediate reward value to calculate the upper confidence interval and select the search node, and the predicted expected return is obtained. Monte Carlo search evaluates the long-term return of different action sequences by simulating multiple possible treatment paths. During the search process, the upper confidence bound algorithm is used to select the most potential node for expansion, balancing exploration and utilization. For each action node, the upper confidence interval is calculated, and the node with the highest upper confidence bound is selected for search. The upper confidence bound takes into account the expected return and uncertainty of the node, encouraging the algorithm to explore actions that have not been fully evaluated. For example, through 1000 times of Monte Carlo simulation, the predicted expected return of the path "anti-platelet therapy first, then thrombolytic therapy, and finally transfer to catheter room" is 0.72, and the confidence interval is 0.68 to 0.76; while the predicted expected return of the path "anti-platelet therapy first, direct transfer to catheter room" is 0.85, and the confidence interval is 0.81 to 0.89. The optimal path is selected.

[0154] The predicted state transition result and the predicted expected return are added to the preset state-action value function and iteratively optimized by a time difference algorithm. The time difference algorithm updates the value function by the actually observed reward and the estimated future return. For each treatment decision, the current state, the executed action, the obtained immediate reward and the transferred new state are recorded, the time difference error is calculated and the value function is updated. The learning rate is set to 0.05 and the discount factor is 0.95. Through iterative optimization of a large amount of case data, the value function gradually converges and can accurately evaluate the long-term value of different treatment actions.

[0155] The initial treatment scheme is updated according to the optimization result to obtain an optimal treatment scheme. The optimal treatment scheme is a sequence of actions with the highest expected return under the current state. By querying the optimized state-action value function, the sequence of actions with the highest value is selected as the optimal scheme. For an acute myocardial infarction patient, the optimal treatment scheme is: immediately give aspirin 300mg and clopidogrel 300mg at T0, at the same time, establish a venous access and intravenously inject nitroglycerin 0.4mg; start contacting the receiving hospital with a catheter room at T0+5 minutes; prepare the transfer equipment at T0+10 minutes; start transfer at T0+15 minutes, continuously monitor vital signs during the transfer; arrive at the receiving hospital at T0+40 minutes, complete the patient handover; and perform coronary intervention treatment at T0+60 minutes.

[0156] In this embodiment, by introducing the time-dependent relationship in the initial action space and constructing the action transition graph, the sequence and mutual restriction between different treatment actions can be accurately described, and unreasonable action combinations in the scheme generation process can be avoided. The time sequence constraint function is calculated based on the action transition graph and combined with the state vector to obtain the predicted state transition result, which can realize the dynamic prediction of the treatment effect. Compared with the static rule derivation method, the method is more flexible and forward-looking. By calculating the state improvement degree of the predicted state transition result and combining the resource consumption to generate the immediate reward value, the search nodes are selected by using the Monte Carlo search and the confidence upper limit interval, the balance between exploration and utilization can be achieved, and the convergence efficiency and accuracy of the scheme optimization are improved. The predicted state transition result and the predicted expected return are introduced into the state-action value function and iteratively optimized by the time difference algorithm, which can continuously improve the quality of the scheme and realize the optimal balance.

[0157] In a second aspect, the embodiment of the present application provides a special disease emergency whole-process auxiliary decision system based on multi-modal perception and AI, which comprises:

[0158] The first unit is configured to acquire multi-modal data of an emergency scene and perform feature extraction to obtain an initial feature vector, calculate a time sequence change of the initial feature vector, and construct a feature dependency graph by combining a conditional probability, calculate a time sequence correlation degree between feature nodes in the feature dependency graph, determine a key feature sequence based on the time sequence correlation degree, and identify a specific disease type and a corresponding symptom feature vector.

[0159] The second unit is configured to perform partition indexing on pre-set specific disease diagnosis and treatment data based on the symptom feature vector, determine a to-be-indexed subset and calculate an initial similarity between the symptom feature vector in the to-be-indexed subset and a historical case, determine a candidate case and perform time sequence analysis on the candidate case to determine a symptom evolution rule, calculate a treatment matching degree between current physiological data and the symptom evolution rule, determine a reference case based on the treatment matching degree, and extract a treatment path corresponding to the reference case as a reference path.

[0160] The third unit is configured to generate an initial treatment scheme based on the reference path, construct an initial action space by combining a clinical standard, construct a state vector by combining the current physiological data, select a treatment action in the initial action space and calculate an immediate reward value, predict an expected return by combining a Monte Carlo search, and update the initial treatment scheme based on the predicted expected return to obtain an optimal treatment scheme.

[0161] In a third aspect, an electronic device is provided, including:

[0162] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0163] In a fourth aspect, a computer-readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0164] The present application can be a method, device, system or computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions stored therein, which are used to perform various aspects of the present application.

[0165] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for AI-based auxiliary decision-making of a whole process of emergency treatment of a specific disease based on multi-modal perception, characterized in that, The method comprises: obtaining multi-modal data of an emergency scene and performing feature extraction to obtain an initial feature vector, calculating the time sequence change of the initial feature vector and constructing a feature dependency graph combining conditional probability, calculating the time sequence correlation between feature nodes in the feature dependency graph, determining a key feature sequence based on the time sequence correlation and identifying a specific disease type and a corresponding symptom feature vector; based on the symptom feature vector, partition index is carried out on the pre-set specific disease diagnosis and treatment data, the initial similarity of the symptom feature vector in the to-be-indexed subset and the historical cases is calculated, the candidate cases are determined and the symptom evolution rule is determined through time sequence analysis, the treatment matching degree of the current physiological data and the symptom evolution rule is calculated, the reference case is determined based on the treatment matching degree, and the treatment path corresponding to the reference case is extracted as a reference path; based on the reference path, an initial treatment scheme is generated and an initial action space is constructed combining clinical norms, a state vector is constructed combining the current physiological data, a treatment action is selected in the initial action space and an immediate reward value is calculated, an expected return is predicted combining Monte Carlo search, and the initial treatment scheme is updated based on the predicted expected return to obtain an optimal treatment scheme.

2. The method of claim 1, wherein, obtaining multi-modal data of an emergency scene and performing feature extraction to obtain an initial feature vector, calculating the time sequence change of the initial feature vector and constructing a feature dependency graph comprises: performing feature extraction on the multi-modal data of the emergency scene to obtain an initial feature vector, calculating the first-order difference value of the initial feature vector, and determining the inflection point based on the first-order difference value and a preset adaptive threshold value; According to the inflection point, the initial feature vector is divided into a plurality of feature segments, each feature segment is modeled and a segmented feature function is constructed, and a transition connection function of time-varying weight between adjacent feature segments is constructed based on the segmented feature function to obtain a time sequence change sequence; based on the time sequence change sequence, the conditional probability is calculated, when the conditional probability exceeds a preset dynamic threshold value, a directed edge is established between the corresponding feature nodes, and the directed edge is repeatedly constructed to obtain the feature dependency graph.

3. The method of claim 1, wherein, calculating the time sequence correlation between feature nodes in the feature dependency graph, determining a key feature sequence based on the time sequence correlation and identifying a specific disease type and a corresponding symptom feature vector comprises: for each feature node in the feature dependency graph, the PageRank centrality is calculated combining the neighbor nodes of the current feature node, the connection probability of the current feature node and the neighbor nodes is calculated to determine the information entropy, and the importance weight of the current feature node is calculated based on the PageRank centrality and the information entropy; the distance between different feature nodes is calculated and the inter-node similarity is determined, the initial correlation is calculated based on the inter-node similarity and the importance weight, and the time sequence correlation is obtained by time sequence adjustment based on the initial correlation and a preset time sequence decay factor; Extract a feature mapping relationship corresponding to each special disease type from a pre-set special disease type knowledge base, take a feature node with a time sequence correlation greater than a pre-set correlation threshold as a key node, analyze combination correlation between the key nodes, and construct a key feature sequence by combining the feature mapping relationship, calculate a matching degree of each feature in the key feature sequence with a special disease type to obtain a special disease type probability value, and determine a special disease type corresponding to the current feature, and extract a time sequence correlation corresponding to the special disease type to obtain a symptom feature vector.

4. The method of claim 1, wherein, Based on the symptom feature vector, pre-set special disease diagnosis and treatment data are indexed, a to-be-indexed subset is determined, and an initial similarity between the symptom feature vector in the to-be-indexed subset and a historical case is calculated, including: Determine the abnormality degree of each symptom by calculating the difference between each symptom feature value in the symptom feature vector and a pre-set normal reference mean value, and determine the danger level corresponding to each symptom based on the abnormality degree; Based on the danger level, the symptom feature vector is divided into multiple priority partitions, a first warning threshold is solved based on the danger level and a pre-set first balance parameter, for the symptom feature vector in each priority partition, a second warning threshold is solved by calculating a change trend and combining a pre-set second balance parameter, a combined warning value is obtained based on the first warning threshold and the second warning threshold, and an index structure is constructed based on the combined warning value and the symptom feature vector; Based on the index structure, a to-be-matched partition is located from the special disease diagnosis and treatment data, a partition matching degree between the symptom feature vector and each to-be-matched partition is calculated, and a to-be-indexed subset is obtained by combining three to-be-matched partitions with the highest partition matching degrees; Calculate the Euclidean distance between the symptom feature vector and each historical case in the to-be-indexed subset to determine the feature similarity, calculate the timeliness weight based on the time information of the historical cases, and calculate the initial similarity by combining the feature similarity.

5. The method of claim 1, wherein, Determine a candidate case and perform time sequence analysis on the candidate case to determine a symptom evolution rule, calculate a treatment matching degree of current physiological data and the symptom evolution rule, determine a reference case based on the treatment matching degree, and extract a treatment path corresponding to the reference case as a reference path, including: Determine a candidate case in the to-be-indexed subset based on the initial similarity, perform time sequence analysis on the candidate case by a dynamic time warping method, calculate a time sequence distance corresponding to each candidate case, and determine a symptom evolution rule based on the time sequence distance; Acquire current physiological data, calculate a disease condition complexity matching value and a medical resource matching degree of the current physiological data and the symptom evolution rule, and sum them to obtain a treatment matching degree; Select the candidate case with the highest treatment matching degree as the reference case, construct a treatment optimization objective function according to a complication risk value, solve the treatment optimization objective function by a differential evolution algorithm to obtain an optimal solution, extract a feasibility score corresponding to the optimal solution, if the feasibility score is greater than a pre-set feasibility threshold, extract a treatment path and output it as a reference path.

6. The method of claim 1, wherein, generating an initial treatment scheme based on the reference path and constructing an initial action space combined with clinical norms, constructing a state vector combined with the current physiological data includes: temporally decomposing the reference path to obtain a plurality of treatment nodes, extracting treatment measures in each treatment node to construct an initial treatment scheme, constructing an initial action space combined with pre-acquired clinical norms, and optimizing constraint parameters in the initial action space through an adaptive fuzzy reasoning method to obtain optimized constraint parameters; constructing a fuzzy decision tree based on the optimized constraint parameters, evaluating the rationality of the initial action space based on the fuzzy decision tree to obtain a rationality score and initializing a feature mapping matrix based on the rationality score; mapping pre-acquired current physiological data into a feature space corresponding to the optimized constraint parameters based on the feature mapping matrix to obtain a state vector.

7. The method of claim 1, wherein, selecting a treatment action in the initial action space and calculating an immediate reward value, predicting an expected return based on the Monte Carlo search, and updating the initial treatment scheme based on the predicted expected return to obtain an optimal treatment scheme includes: collecting treatment actions in the initial action space and determining the temporal dependency relationship between different treatment actions, constructing an action transition graph based on the temporal dependency relationship and the treatment actions; determining an action vertex set and a dependency relationship edge set based on the action transition graph, calculating a temporal constraint function according to the dependency relationship edge set, and adding the state vector to the temporal constraint function to calculate a predicted state transition result; calculate the state improvement degree of the predicted state transition result relative to the state vector, calculate the immediate reward value combined with the treatment resource consumption, perform Monte Carlo search on the immediate reward value, calculate the confidence upper limit interval and select the search node to obtain the predicted expected return; add the predicted state transition result and the predicted expected return to the pre-set state-action value function and perform iterative optimization through the temporal difference algorithm, update the initial treatment scheme according to the optimization result to obtain the optimal treatment scheme.

8. A multi-modal perception and AI based auxiliary decision system for whole process of emergency treatment of a specific disease, for implementing the method of any one of the preceding claims 1-7, characterized in that, includes: The first unit is configured to acquire multi-modal data of the emergency scene and perform feature extraction to obtain an initial feature vector, calculate the temporal variation of the initial feature vector, and construct a feature dependency graph combined with conditional probability, calculate the temporal correlation between feature nodes in the feature dependency graph, determine a key feature sequence based on the temporal correlation, and identify a specific disease type and a corresponding symptom feature vector; The second unit is configured to partition and index pre-set specific disease diagnosis and treatment data based on the symptom feature vector, determine a to-be-indexed subset, calculate the initial similarity between the symptom feature vector in the to-be-indexed subset and historical cases, determine a candidate case, perform temporal analysis on the candidate case to determine a symptom evolution rule, calculate the treatment matching degree of the current physiological data and the symptom evolution rule, determine a reference case based on the treatment matching degree, and extract the treatment path corresponding to the reference case as a reference path. The third unit is configured to generate an initial treatment scheme based on the reference path, construct an initial action space in combination with clinical norms, construct a state vector in combination with the current physiological data, select a treatment action in the initial action space and calculate an immediate reward value, predict an expected return in combination with a Monte Carlo search, and update the initial treatment scheme based on the predicted expected return to obtain an optimal treatment scheme.

9. An electronic device, comprising: The computer program product comprises: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to execute the method of any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions, when executed by the processor, implement the method of any one of claims 1 to 7.

Images