Online diagnosis data intelligent analysis system based on big data

Abstract

The application belongs to the technical field of online consultation, and provides an online consultation data intelligent analysis system based on big data, which comprises: a data feature acquisition module, a strategy rule library construction and dynamic updating module, and a medical resource pool dynamic scheduling and management module, aiming at solving the problems of slow response in emergency scenes and resource waste in non-emergency scenes, the system contains three core modules: the data feature acquisition module, the strategy rule library construction and dynamic updating module, and the medical resource pool dynamic scheduling and management module, the data feature acquisition module is used for identifying emergency keywords in words and classifying, identifying emergency probability of physiological signals and classifying by using a light 1D-CNN, and fusing the three after processing periodic data to identify real-time flow / periodic static data, the strategy rule library construction and dynamic updating module is used for establishing an emergency degree-data type-resource response two-dimensional rule library based on the classification results and the data type, storing the rule library in a Redis cluster and setting an online updating mechanism, and the medical resource pool dynamic scheduling and management module is used for managing doctors / algorithm pools in layers and dynamically allocating resources according to the rule library, so that efficient response of emergency and precise utilization of medical resources are realized.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of online consultation, and in particular to an online consultation data intelligent analysis system based on big data. BACKGROUND

[0002] With the rapid development of Internet medical treatment, online consultation has become an important supplement to medical services due to its convenience and timeliness, and is widely used in the follow-up of chronic diseases, daily health consultation and preliminary screening of emergency situations. However, the current online consultation data intelligent analysis system still has many technical defects, which cannot meet the actual needs of clinical practice:

[0003] Multi-modal consultation data processing is not accurate, and the emergency degree is large: online consultation data covers multi-modal data such as text complaints (such as patient symptom description), time-series physiological signals (such as electrocardiogram, blood pressure waveform), periodic static data (such as weekly blood glucose records of patients with chronic diseases), etc. The existing system mainly processes single type of data (such as relying only on text keywords to determine the emergency degree), and does not realize the fusion of multi-modal data, which is easy to cause emergency degree misjudgment due to one-sided information - for example, only through the "chest tightness" text description, the ordinary symptom is misjudged, the ST segment elevation signal of electrocardiogram is missed in the emergency warning of myocardial infarction, or the periodic blood glucose fluctuation of patients with chronic diseases is misjudged as an emergency, causing invalid calling of medical resources.

[0004] Resource response strategy is rigid, and emergency delay and resource waste coexist: the existing system mainly adopts fixed resource allocation mode, and lacks dynamic adaptation mechanism of "scene-resource". On the one hand, emergency scenes (such as sudden chest pain) need to respond in seconds, but fixed flow processing link (such as redundant disk storage and full model reasoning) leads to delay of more than 30 seconds, which delays the treatment; on the other hand, non-emergency scenes (such as common cold consultation and chronic disease routine follow-up) excessively pursue real-time, occupy high computing power resources (such as GPU) and high seniority doctor time, causing idle rate of computing power more than 60%, mismatch of doctor resources, and increase of medical cost.

[0005] The rule base lacks dynamic updating ability and has poor adaptability: the resource scheduling rules of the existing system are mainly static configuration (such as fixed "emergency corresponding to chief physician"), which cannot adapt to the changes of clinical scenes in real time - for example, the new emergency type (such as new respiratory emergency) is not included in the keyword library, and the doctor scheduling adjustment leads to the offline of senior doctors, the system cannot update the rules in time, and still allocates resources according to the old strategy, further aggravating the delay of emergency response and waste of resources.

[0006] In summary, there is an urgent need for an online consultation data intelligent analysis system that can accurately process multi-modal data, dynamically adapt to scene requirements, and realize efficient scheduling of resources, to solve the above technical problems.

[0007] To this end, the application provides an online consultation data intelligent analysis system based on big data. SUMMARY

[0008] In order to make up for the deficiencies of the prior art, at least one technical problem raised in the background art is solved.

[0009] The technical scheme adopted by the application to solve its technical problems is:

[0010] One of the purposes of the application is to provide an online consultation data intelligent analysis system based on big data, comprising:

[0011] The data feature acquisition module: constructs a medical keyword library, identifies medical emergency keywords in online consultation text data, and performs hierarchical processing according to the identification results. A lightweight 1D-CNN model is designed to identify the probability of emergency in the time series physiological signals in the online consultation text. The data is classified according to the probability of emergency. The periodic data in the online consultation text is identified and classified, and the Chinese data and time series physiological signals are fused to identify real-time streaming data and periodic static data in two types of online consultation text data.

[0012] The strategy rule library construction and dynamic updating module: based on the hierarchical processing results and the online consultation text data type, an emergency-data type-resource response strategy rule library is constructed, and an online updating mechanism is set for the emergency-data type-resource response strategy rule library to cope with high concurrency scenarios.

[0013] The medical resource pool dynamic scheduling and management module: establishes a doctor resource pool management and dynamic allocation mechanism, and dynamically allocates doctor resources in combination with the emergency-data type-resource response strategy rule library. A hierarchical and containerized management mechanism for the computing resource pool is established to meet the computing resource demand in the emergency-data type-resource response strategy rule library.

[0014] As a further improvement, the specific process of identifying medical emergency keywords in online consultation text data and performing hierarchical processing according to the identification results is as follows:

[0015] According to the historical online consultation text data, the medical emergency keywords are extracted;

[0016] And a Trie tree data structure is used to store the medical emergency keywords, and all the medical emergency keywords in the medical keyword library are inserted into the Trie tree. Each node of the Trie tree stores one character, and the path from the root node to the leaf node constitutes a complete keyword, realizing linear time complexity matching of text medical emergency keywords and ensuring millisecond-level matching speed. The leaf node additionally records the emergency weight and associated medical emergency name of the medical emergency keyword.

[0017] First, the text data in the patient online consultation text is preprocessed, punctuation, stop words are removed, and core medical words are retained;

[0018] Using the Trie tree matching method, the preprocessed online consultation text core medical words are sequentially input into the Trie tree, and all online medical emergency keywords and emergency degree weights of medical emergency keywords are matched;

[0019] Sum all the emergency degree weights of the medical emergency keywords;

[0020] If the total medical emergency keyword is greater than or equal to 0.8, it is marked as the highest priority;

[0021] If the total medical emergency keyword is less than 0.8 and greater than or equal to 0.3, it is marked as medium priority;

[0022] If the total medical emergency keyword is less than 0.3, it is marked as low priority.

[0023] As a further improvement scheme, the specific process of identifying the probability of emergency in the time sequence physiological signal in the online consultation text is:

[0024] The time sequence physiological signal includes but is not limited to electrocardiogram, blood pressure waveform, and blood oxygen saturation curve;

[0025] The time sequence physiological signal is preprocessed, and a 30-second sliding time window is real-time intercepted, the waveform data in the sliding time window is normalized, the data is mapped to the interval [−1, 1], and the baseline drift and dimension difference of different patients and different devices are eliminated;

[0026] The normalized time sequence physiological signal is used as the input of the lightweight 1D-CNN model;

[0027] The lightweight 1D-CNN model is designed to contain 3 layers of convolution, 2 layers of maximum pooling, and the final output layer is a Sigmoid activation function, which outputs the probability of whether it is an emergency.

[0028] As a further improvement scheme, the specific process of grading according to the probability of emergency is:

[0029] The time sequence physiological signal in the historical online consultation text is used to train the lightweight 1D-CNN model, and the accuracy of the lightweight 1D-CNN model is verified through the test set, which meets the accuracy requirement;

[0030] Wherein, the parameter quantity of the lightweight 1D-CNN model is less than 100,000, and the single sample inference time is ensured within the preset time;

[0031] If the emergency probability output by the lightweight 1D-CNN model is greater than or equal to 0.8, it is marked as the highest priority;

[0032] If the emergency probability output by the lightweight 1D-CNN model is greater than or equal to 0.3 and less than 0.8, it is marked as medium priority; if the emergency probability output by the lightweight 1D-CNN model is less than 0.3, it is marked as low priority.

[0033] As a further improvement scheme, the specific process of identifying periodic data in online consultation text and performing hierarchical processing is:

[0034] Periodic data includes but is not limited to weekly blood glucose and blood pressure records of patients with chronic diseases; the upload timestamp of the periodic data in the online consultation text of the patient is extracted, and the year-month-day format is parsed;

[0035] The upload timestamps of the patient's last 10 times of the same type of data are obtained, and the time interval between adjacent two uploads is calculated;

[0036] If the time interval of the last 3 consecutive uploads meets the fixed periodicity rule, it is determined as periodic data; otherwise, it is determined as non-periodic data;

[0037] The last 3 times of the same type of periodic data of the patient are obtained;

[0038] The standard deviation of the 3 times of periodic data is calculated, and it is detected whether there is a critical value in the 3 times of periodic data: if any one of the periodic data is greater than or equal to the first critical value; or less than or equal to the second critical value, the critical value flag is triggered;

[0039] If the periodic data in the online consultation text is periodic data and has no critical value, and the standard deviation is less than the standard deviation threshold, it is marked as medium priority;

[0040] If the standard deviation of the periodic data in the online consultation text is greater than or equal to the standard deviation threshold or there is a critical value, it is marked as the highest priority;

[0041] If the periodic data in the online consultation text is non-periodic but has a critical value, it is marked as the highest priority.

[0042] As a further improvement scheme, the specific process of fusing Chinese character data and time series physiological signals for identifying real-time stream data and periodic static data of two types of online consultation text data is:

[0043] The feature vectors of text data and physiological signals are mapped to the same dimension through a linear transformation layer to obtain , ;

[0044] Calculate attention weights of the text data and the physiological signals to obtain the attention weights of the text data and the physiological signals 、 ;

[0045] Fuse the attention weights of the text data and the physiological signals into a feature vector: ;

[0046] The fused feature vector is input into a Softmax classification layer in a lightweight 1D-CNN model, and a type label of real-time streaming data or periodic static data is output.

[0047] As a further improvement scheme, the specific process of constructing the emergency-data type-resource response strategy rule library is as follows:

[0048] The emergency-data type-resource response strategy rule library is constructed, RedisCluster distributed cluster is used as the emergency-data type-resource response strategy rule library storage engine, and a refined Key-Value data model is designed:

[0049] The Key of the emergency-data type-resource response strategy rule library adopts the form of a combination of emergency priority and data type strings, so as to cover the scene division in the two dimensions of emergency and data type;

[0050] The highest priority-real-time streaming: corresponding to emergency diagnosis scene;

[0051] Medium priority-periodic static: corresponding to chronic disease routine follow-up scene;

[0052] Low priority-periodic static: corresponding to ordinary non-emergency scene;

[0053] The Value adopts JSON format and contains the following key fields:

[0054] Doctor resource priority: clearly needs to dispatch the doctor rank, ensures that the emergency scene allocates high seniority doctors first;

[0055] Response time requirement: define the maximum time length that the doctor needs to complete the response, match the time efficiency demand of different scenes;

[0056] Computing power resource demand: specify the required computing power type and resource pool, realize the precise scheduling of computing power;

[0057] Result presentation form: define the output mode of the diagnosis result, meet the information transmission demand of different scenes;

[0058] Doctor resource pool identifier: associate specific doctor grouping, provide precise target for the doctor allocation of the medical resource pool dynamic scheduling and management module;

[0059] Computing resource pool identifier: associated with a specific computing power group, providing direct basis for computing power scheduling of medical resource pool dynamic scheduling and management module;

[0060] Notification trigger mode: define the notification channel between doctors and patients to ensure timely response;

[0061] Adopt 3 master 3 slave RedisCluster cluster deployment, data sharding through consistent hashing, avoid single node hotspot and performance bottleneck; Start RDB plus AOF hybrid persistence, ensure that rules are not lost after service restart; Set the heartbeat detection interval between cluster nodes to 100ms to ensure the rapid discovery and switching of fault nodes.

[0062] As a further improvement scheme, the specific process of setting an online update mechanism for the urgency-data type-resource response strategy rule library is:

[0063] Develop a web-based management interface to support the full life cycle management of rules in the urgency-data type-resource response strategy rule library, including adding, modifying, deleting, and querying rules by operation and maintenance personnel or clinical experts:

[0064] Rule addition: users select the urgency priority and data type through the drop-down box, fill in all the Value fields of the doctor resource priority and response time requirements, and submit them to trigger the rule synchronization process;

[0065] Rule modification: after querying the target rule through Key combination, any Value field can be modified, and the corresponding Value in Redis is updated in real time after submission;

[0066] Rule deletion: after locating the target rule through Key combination, confirm the deletion operation, and the corresponding Key-Value in Redis is removed immediately;

[0067] Rule query: fuzzy / accurate query is supported according to urgency priority, data type, or Key combination, and all strategies in the current rule library are displayed.

[0068] As a further improvement scheme, the specific process of dynamically allocating doctor resources is:

[0069] Based on the doctor resource priority in the urgency-data type-resource response strategy rule library dynamic allocation, doctors are divided into three target resource pools:

[0070] Chief physician pool: includes doctors with high-level titles and emergency and critical care experience, covering cardiology and emergency departments;

[0071] Attending physician pool: includes doctors with intermediate titles and the ability to manage common diseases and chronic diseases, covering internal medicine and general practice;

[0072] Resident pool: contains junior doctors, responsible for routine consultation and follow-up; each resource pool is associated with the doctor resource priority field in the urgency-data type-resource response strategy rule base, ensuring that the strategy instruction can accurately point to the target pool;

[0073] WebSocket long connection and heartbeat detection technology is used to real-time perceive the online status and current load of doctors:

[0074] WebSocket long connection is established for each doctor's workstation, and the server sends a heartbeat packet to the client every 5 seconds; if no response is received from the client for 3 consecutive times, the doctor is marked as offline; if a response is received, the doctor is marked as online;

[0075] When a doctor sees a patient, the workstation sends a start signal to the server; when the patient finishes the consultation or the doctor marks the consultation as complete, an end signal is sent; the server real-time counts the number of patients each doctor is seeing as a load indicator;

[0076] Select an online doctor with the lowest current load from the target resource pool;

[0077] If all doctors in the target resource pool are in a high load state, trigger load preemption:

[0078] Select a doctor in the target resource pool with the highest load but with non-urgent patients, and push an urgent task preemption prompt to his workstation;

[0079] If the doctor confirms the transfer, the system automatically assigns the current non-urgent patient to a doctor with the second lowest load in the target resource pool, and immediately assigns the emergency patient to the doctor; if the doctor refuses to transfer, the system continues to select other doctors in the same pool; if there is no available doctor in the pool, trigger cross-target resource pool scheduling, and select a senior doctor with emergency capabilities from the attending physician pool for temporary assistance.

[0080] As a further improvement scheme, the specific process of establishing a hierarchical and containerized management mechanism for the algorithm resource pool is:

[0081] Based on the algorithm resource demand in the urgency-data type-resource response strategy rule base, the algorithm is divided into three algorithm resource pools:

[0082] Exclusive GPU resource pool: reserved for the highest priority-real-time streaming scenario, containing 4 NVIDIA T4 GPUs; each GPU is divided into 4 exclusive cores through Kubernetes container orchestration technology, and a Guaranteed QoS level Pod is created for each exclusive core, with a GPU limit of 0.25 cores;

[0083] Shared CPU resource pool: for medium priority-periodic static scenarios, contains 16-core Intel Xeon CPU; shared through container sharing mechanism, allows multiple regular tasks to time-share CPU threads;

[0084] Idle period CPU resource pool: for low priority-periodic static scenarios, uses the idle period of computing power from 1:00 to 3:00 in the morning, contains 8-core Intel Xeon CPU;

[0085] Containerized scheduling and resource isolation uses K8s for containerized management of computing power resources:

[0086] Create a Pod with Guaranteed QoS level for each exclusive core of the exclusive GPU resource pool, set CPU limit to 0.25 cores and GPU limit to 0.25 cores;

[0087] Create a Pod with Burstable QoS level for the shared CPU resource pool, set CPU request to 0.1 cores and CPU limit to 0.5 cores, allowing tasks to burst to higher computing power when idle;

[0088] Create a Pod with BestEffort QoS level for the idle period CPU resource pool, without setting a fixed computing power limit;

[0089] If it is an exclusive GPU core, the K8s scheduler selects an idle Guaranteed Pod from the exclusive GPU resource pool, and submits the lightweight 1D-CNN model inference task to the Pod, binding the GPU core;

[0090] If it is a shared CPU thread, the scheduler selects a Burstable Pod from the shared CPU resource pool, and submits the task to the Pod with the lowest load according to the current computing power load;

[0091] If it is an idle period CPU thread, the scheduler puts the task into a delayed task queue and sets the trigger time to 2:00 in the morning. Based on the computing power load during the idle period monitored by Prometheus, the scheduler dynamically selects the time point with the lowest load to trigger the task. The default fallback is 2:00 in the morning, and the task will be submitted to the BestEffort Pod in the idle resource pool at that time;

[0092] During task execution, the K8s probe monitors the Pod state; if the Pod is abnormal, the Pod is restarted and the task is rescheduled immediately;

[0093] After the task is completed, K8s automatically releases the Pod resources for subsequent tasks.

[0094] Two purposes of the present application are to provide an online consultation data intelligent analysis method based on big data, comprising:

[0095] S1: Construct a medical keyword library, identify the medical emergency keywords of the character data in the online consultation text, and perform hierarchical processing according to the identification result, identify the probability of emergency in the time sequence physiological signal in the online consultation text by designing a lightweight 1D-CNN model, and perform hierarchical processing according to the probability of emergency; identify the periodic data in the online consultation text and perform hierarchical processing, and fuse the Chinese character data and the time sequence physiological signal for identifying two types of online consultation text data types of real-time stream data and periodic static data;

[0096] S2: Based on the hierarchical processing result and the online consultation text data type, an emergency-data type-resource response strategy rule library is constructed, and an online updating mechanism is set for the emergency-data type-resource response strategy rule library to cope with high concurrency scenarios;

[0097] S3: Establish a doctor resource pool management and dynamic allocation mechanism, and dynamically allocate doctor resources in combination with the emergency-data type-resource response strategy rule library, establish a hierarchical and containerized management mechanism of the computing resource pool for meeting the computing resource demand in the emergency-data type-resource response strategy rule library.

[0098] The beneficial effects of the present application are as follows:

[0099] Improve the accuracy of disease severity determination and reduce the risk of misjudgment: through the data feature acquisition module, multi-modal fusion processing of character data, time sequence physiological signal and periodic data is realized, and real-time stream / periodic static data types are identified, avoiding the information one-sidedness caused by single data mode, making the emergency classification more comprehensive and accurate, and providing a reliable basis for subsequent resource scheduling.

[0100] Rely on the strategy rule library construction and dynamic updating module to construct the rule library in two dimensions of emergency-data type, and combine Redis Cluster distributed cluster to guarantee millisecond-level query under high concurrency, and the online updating mechanism adapts to the changes of clinical scenarios; the medical resource pool dynamic scheduling and management module allocates doctor and computing power resources in layers accordingly, which not only ensures ≤30 seconds response in emergency scenarios, but also reduces the idle rate of computing power in non-emergency scenarios and the mismatch of doctor resources, and reduces the waste of medical resources. BRIEF DESCRIPTION OF DRAWINGS

[0101] The present application will be further described below in conjunction with the drawings.

[0102] Figure 1 is the system module diagram of the online consultation data intelligent analysis system based on big data of the present application;

[0103] Figure 2 is a step flow chart of the online diagnosis data intelligent analysis method based on big data of the present application. DETAILED DESCRIPTION

[0104] In order to make the technical means, creative features, purposes and effects of the present application easy to understand, the present application will be further described below in combination with specific embodiments.

[0105] Embodiment 1

[0106] As shown in the figure, the online diagnosis data intelligent analysis system based on big data according to the embodiments of the present application comprises: Figure 1 The data feature acquisition module: construct a medical keyword library to identify medical emergency keywords in online diagnosis text word data, according to the identification result, perform hierarchical processing, identify the probability of emergency in the time sequence physiological signal in the online diagnosis text by designing a lightweight 1D-CNN model, and perform hierarchical processing according to the probability of emergency; identify periodic data in online diagnosis text and perform hierarchical processing, and fuse word data and time sequence physiological signal, which is used to identify two types of online diagnosis text data types of real-time stream data and periodic static data;

[0107] For the medical emergency keywords in the online diagnosis text word data, the construction of the medical keyword library is the fundamental way to quickly identify the medical emergency keywords, and the specific process of constructing the medical keyword library is as follows:

[0108] The medical keyword library is constructed according to the word data in the historical online diagnosis text, and the medical emergency keywords are extracted, for example, 500+ medical emergency keywords covering "chest pain", "dyspnea", "loss of consciousness", "massive hemorrhage", etc.;

[0109] And the Trie tree data structure is used to store the medical emergency keywords, and all the medical emergency keywords (such as "sudden chest pain", "ventricular fibrillation", "progressive dyspnea") in the medical keyword library are inserted into the Trie tree; each node of the Trie tree stores 1 character, and the path from the root node to the leaf node constitutes a complete keyword, realizing linear time complexity matching of text medical emergency keywords, ensuring millisecond-level matching speed; the leaf node additionally records the emergency degree weight (such as "ST segment elevation" weight 1.0, "chest tightness" weight 0.7, "fatigue" weight 0.2, the present application does not limit the emergency degree weight) and the associated medical emergency name (such as "sudden chest pain", "ventricular fibrillation", "progressive dyspnea") of the medical emergency keywords;

[0110]

[0111] ​After the construction of the medical keyword library, it is necessary to quickly identify the medical emergency keywords in the online consultation text for quick identification of patient symptoms, and according to the identification results, the specific hierarchical processing process is as follows:

[0112] Taking the patient's online consultation text as an example: "Male patient, 56 years old, sudden onset of crushing chest pain 10 minutes ago, accompanied by profuse sweating, and self-reported near-death feeling";

[0113] First, the patient's online consultation text is preprocessed, and punctuation, stop words (such as "patient" and "self-reported") are removed, and core medical words are retained: "sudden", "crushing chest pain", "profuse sweating", and "near-death feeling";

[0114] Using the Trie tree matching method, the core medical words in the preprocessed online consultation text are input into the Trie tree in turn, and all medical emergency keywords in the online consultation are matched, such as "crushing chest pain" and the emergency degree weight of the medical emergency keyword (emergency degree weight 1.0) "near-death feeling" (associated emergency: acute myocardial infarction, emergency degree weight 1.0);

[0115] Sum all the emergency degree weights of the medical emergency keywords;

[0116] If the total medical emergency keyword is greater than or equal to 0.8, it is marked as the highest priority;

[0117] If the total medical emergency keyword is less than 0.8 and greater than or equal to 0.3, it is marked as medium priority;

[0118] If the total medical emergency keyword is less than 0.3, it is marked as low priority;

[0119] For the time sequence physiological signal in the online consultation text, a lightweight model needs to be designed to quickly identify the time sequence physiological signal and identify the probability of emergency, and the probability of emergency identified by the lightweight model is classified, and the specific implementation process is as follows:

[0120] The time sequence physiological signal includes but is not limited to electrocardiogram, blood pressure waveform, and blood oxygen saturation curve, etc., which has strong real-time performance, and their common feature is that they are curves with waveform properties, which are difficult to identify. In some embodiments, millisecond-level emergency feature identification is required; taking electrocardiogram as an example, the implementation details are as follows:

[0121] First, the time sequence physiological signal is preprocessed, and a 30-second sliding time window (window step length 1 second, ensuring continuous monitoring without omission) is real-time intercepted, and the waveform data in the sliding time window is normalized to map the data to the interval [−1, 1], eliminating baseline drift and dimension differences of different patients and different devices;

[0122] Normalized temporal physiological signals are used as input to a lightweight 1D-CNN model;

[0123] Secondly, a lightweight 1D-CNN model is designed, consisting of 3 convolutional layers (with kernel sizes of 64, 32, and 16, and stride of 1 for each layer), 2 max pooling layers (with a pooling kernel size of 2), and the final output layer is a sigmoid activation function (outputting the probability of whether it is an emergency, such as the probability of myocardial infarction, ventricular fibrillation, etc.).

[0124] A lightweight 1D-CNN model was trained using temporal physiological signals from historical online medical consultation texts. The accuracy of the lightweight 1D-CNN model was verified using a test set, and it met the accuracy requirements (test set accuracy must be greater than 92%).

[0125] Among them, the lightweight 1D-CNN model has fewer than 100,000 parameters, ensuring that the inference time for a single sample is within a preset time (e.g., inference time is less than 20ms).

[0126] If the acute probability output by the lightweight 1D-CNN model is greater than or equal to 0.8, it is marked as the highest priority;

[0127] If the acute probability output by the lightweight 1D-CNN model is greater than or equal to 0.3 and less than 0.8, it is marked as medium priority; if the acute probability output by the lightweight 1D-CNN model is less than 0.3, it is marked as low priority.

[0128] For periodic data in online consultation texts, it is necessary to identify and classify them. The specific identification process is as follows:

[0129] Periodic data, including but not limited to: weekly blood glucose and blood pressure records for patients with chronic diseases; urgency needs to be analyzed by combining time patterns and historical fluctuations. The implementation steps are as follows:

[0130] Extract the upload timestamps of periodic data from patients' online consultation texts and parse them into year-month-day format;

[0131] Obtain the upload timestamps of the patient's last 10 similar data uploads, and calculate the time interval between two adjacent uploads;

[0132] If the time interval between three consecutive uploads follows a fixed periodic pattern, such as (the upload pattern for blood glucose is 7 days ± 1 day), then it is considered periodic data; otherwise, it is considered non-periodic data.

[0133] Obtain the patient's three most recent periodic data of the same type (such as fasting blood glucose values ​​over the past three weeks).

[0134] Calculate the standard deviation of the three periodic data, and detect whether there is a critical value in the three periodic data: for example, if the blood glucose is greater than or equal to the first critical value (high critical value, such as 16.7mmol / L); or less than or equal to the second critical value (low critical value, such as 3.9mmol / L), a critical value flag is triggered;

[0135] If the periodic data in the online consultation text is periodic data and there is no critical value, and the standard deviation is less than the standard deviation threshold (such as a blood glucose threshold of 2mmol / L), it is marked as a medium priority;

[0136] If the standard deviation of the periodic data in the online consultation text is greater than or equal to the standard deviation threshold or there is a critical value, it is marked as the highest priority;

[0137] If the periodic data in the online consultation text is non-periodic but there is a critical value, it is marked as the highest priority;

[0138] The first critical value, the second critical value, and the standard deviation threshold are all reference values set by industry technicians based on historical experience, and the present application does not limit them;

[0139] After completing the medical emergency keyword recognition and classification of the text data in the online consultation text and the emergency probability recognition and classification of the time-series physiological signals, in order to more comprehensively and accurately determine the emergency degree of the patient's condition, the classification results of the text data and the time-series physiological signal data need to be fused and processed;

[0140] The reason is that the text data can reflect the condition from the perspective of the patient's subjective description of symptoms, and the time-series physiological signal data can reflect the body condition from the perspective of objective physiological index waveform, and the combination of the two can reduce the misjudgment that may exist in a single data modality, thereby providing a more reliable basis for subsequent decisions such as reasonable allocation of doctor response speed, and the specific fusion process is as follows:

[0141] The feature vectors (dimensions are 192 and 64) of the text data and the physiological signals, wherein the feature vector of the text data is obtained by preprocessing through constructing a medical keyword library, medical field segmentation, medical pre-training word vector mapping, TextCNN model feature extraction, and global maximum pooling processing; the feature vector of the physiological signals (such as electrocardiogram) is obtained by 30-second sliding window interception, normalization processing, light-weight 1D-CNN model feature extraction, and full connection layer mapping;

[0142] Map them to the same dimension (such as 64) through a linear transformation layer to obtain 、 ;

[0143] Calculate attention weights of text data and physiological signals to obtain attention weights of text data and physiological signals 、 ;

[0144] Fusion feature vectors of attention weights of text data and physiological signals: ;

[0145] The fused feature vectors are input into the Softmax classification layer in the lightweight 1D-CNN model, and the type label of real-time streaming data or periodic static data is output. The classification task is trained using a cross-entropy loss function, and the classification accuracy on the test set is ≥95%.

[0146] Strategy rule base construction and dynamic update module: based on the hierarchical processing results and online consultation text data types, an urgency-data type-resource response strategy rule base is constructed, and an online update mechanism is set for the urgency-data type-resource response strategy rule base to cope with high concurrency scenarios;

[0147] The urgency-data type-resource response strategy rule base is the core carrier of this step, which needs to meet the requirements of millisecond-level query, online non-stop update, and multi-dimensional strategy precise mapping. A distributed key-value storage and visual management interface and real-time synchronization mechanism technology architecture are adopted, and the specific implementation is as follows:

[0148] In order to guarantee the query efficiency and rule expansibility in high concurrency scenarios, an urgency-data type-resource response strategy rule base is constructed, RedisCluster distributed cluster is used as the storage engine of the urgency-data type-resource response strategy rule base, and a refined Key-Value data model is designed:

[0149] The Key of the urgency-data type-resource response strategy rule base adopts the combination of emergency priority and data type string, which covers the two-dimensional scenario division of emergency and data type, avoiding resource mismatch caused by single dimension;

[0150] Highest priority-real-time stream: corresponding to emergency diagnosis scenarios (such as real-time electrocardiogram and symptom text of patients with sudden chest pain);

[0151] Medium priority-periodic static: corresponding to chronic disease routine follow-up scenarios (such as weekly blood glucose monitoring data of diabetic patients);

[0152] Low priority-periodic static: corresponding to ordinary non-emergency scenarios (such as historical consultation records of common cold).

[0153] This two-dimensional combination ensures that high-emergency but non-real-time stream, low-emergency but timely attention, and other subdivided scenarios can all match the most suitable resource strategy.

[0154] Value adopts JSON format, encapsulates doctor resources, response time, computing power resources, result presentation, notification method core scheduling fields, and provides instructions for subsequent execution links;

[0155] Value adopts JSON format and contains the following key fields:

[0156] Doctor resource priority: Clearly define the doctor's job level that needs to be dispatched (such as "chief physician", "attending physician", "resident physician"), and ensure that high seniority doctors are preferentially allocated in emergency scenarios;

[0157] Response time requirement: Define the maximum time length that doctors need to complete the response (such as "≤30 seconds", "≤2 hours", "≤24 hours"), and match the time efficiency requirements of different scenarios;

[0158] Computing power resource demand: Specify the required computing power type and resource pool (such as "exclusive GPU core", "shared CPU thread", "idle period CPU thread"), and achieve precise scheduling of computing power;

[0159] Result presentation form: Define the output method of diagnostic results (such as "popup + red alert 'whether to send to hospital' conclusion", "web-based trend chart report", "daily summary email"), and meet the information transmission needs of different scenarios;

[0160] Doctor resource pool identifier: Associate specific doctor groups (such as "chief_physician_pool" corresponding to the chief physician pool), and provide precise targets for doctor allocation of the medical resource pool dynamic scheduling and management module;

[0161] Computing power resource pool identifier: Associate specific computing power groups (such as "exclusive_gpu_pool" corresponding to the exclusive GPU resource pool), and provide direct basis for computing power scheduling of the medical resource pool dynamic scheduling and management module;

[0162] Notification trigger method: Define the notification channel of doctors and patients (such as "WebSocket real-time popup", "email push", "SMS reminder"), and ensure the timeliness of the response.

[0163] RedisCluster cluster deployment with 3 master 3 slave is adopted, data sharding is performed through consistent hashing, single node hotspot and performance bottleneck are avoided; RDB (snapshot) and AOF (AppendOnlyFile) hybrid persistence are enabled to ensure that rules are not lost after service restart; The heartbeat detection interval between cluster nodes is set to 100ms to ensure the rapid discovery and switching of faulty nodes, achieve "99.99% availability", and meet the high reliability requirements of medical scenarios;

[0164] On the basis of completing the storage architecture construction and core data model design of the emergency degree-data type-resource response strategy rule base, ensuring that the rules have high reliability and high concurrent query capability, in order to cope with the dynamic changes such as the addition of emergency types, the adjustment of doctor scheduling, and the optimization of computing power allocation in clinical scenarios, an efficient online update mechanism needs to be designed for the rule base to realize online modification, real-time effectiveness, and no need to restart the system. The following is a specific description of its implementation:

[0165] Develop a web-based management interface to support operations personnel or clinical experts to add, modify, delete, and query the rules in the emergency degree-data type-resource response strategy rule base throughout their life cycle:

[0166] Rule addition: After the user selects the emergency priority and data type through the drop-down box and fills in all the Value fields of the doctor resource priority and response time requirements, the rule synchronization process is triggered (synchronized to all nodes of RedisCluster to ensure consistency of distributed cluster rules, and if synchronization times out, it will retry for 3 times, and if it fails, it will push an alarm to the operation and maintenance platform);

[0167] Rule modification: After querying the target rule through the Key combination, any Value field can be modified (such as compressing the response time of "highest priority-real-time stream" from "<=30 seconds" to "<=25 seconds"), and after submission, the corresponding Value in Redis is updated in real time;

[0168] Rule deletion: After locating the target rule through the Key combination, confirm the deletion operation, and the corresponding Key-Value in Redis is immediately removed;

[0169] Rule query: Supports fuzzy / accurate query by emergency priority, data type, or Key combination, and displays all strategies in the current rule base for manual verification and adjustment.

[0170] Medical resource pool dynamic scheduling and management module: Establish a doctor resource pool management and dynamic allocation mechanism, and dynamically allocate doctor resources based on the emergency degree-data type-resource response strategy rule base, establish a hierarchical and containerized management mechanism for the computing resource pool to meet the computing resource needs in the emergency degree-data type-resource response strategy rule base;

[0171] Based on the doctor resource priority in the dynamic allocation of the emergency degree-data type-resource response strategy rule base, doctors are divided into three target resource pools:

[0172] Chief physician pool: Includes doctors with high-level titles and experience in emergency and critical care, covering cardiology and emergency departments;

[0173] Attending physician pool: contains doctors with intermediate titles and the ability to manage common diseases and chronic diseases, covering internal medicine and general practice;

[0174] Resident physician pool: contains doctors with primary titles who are responsible for routine consultations and follow-ups;

[0175] WebSocket long connection and heartbeat detection technology is used to real-time perceive the online status and current load of doctors:

[0176] Online status detection: WebSocket long connection is established for each doctor's workstation (PC end / mobile end), and the server sends a heartbeat packet to the client every 5 seconds. If no response is received from the client for 3 consecutive times, the doctor is marked as offline. If a response is received, the doctor is marked as online.

[0177] Current load statistics: When a doctor sees a patient, the workstation sends a start signal to the server. When the patient finishes the consultation or the doctor marks the consultation as complete, an end signal is sent. The server real-time counts the number of patients each doctor is currently seeing as a load indicator (e.g., "Director A is seeing 2 people").

[0178] Select the doctor with the lowest current load from the target resource pool (e.g., the director physician pool). For example, there are 5 online doctors in the director physician pool with loads of 0, 1, 1, 2, and 2, respectively. The doctor with a load of 0 is selected first.

[0179] If all doctors in the target resource pool are in a high load state (all online doctors in the director physician pool have a load ≥ 2), load preemption is triggered.

[0180] Select the doctor with the highest load but seeing patients in non-urgent scenarios from the target resource pool, and push an urgent task preemption prompt (e.g., a pop-up window showing "There is an emergency patient who needs to be seen first. Do you want to transfer the current non-urgent patient to another doctor?") to the doctor's workstation.

[0181] If the doctor confirms the transfer, the system automatically assigns the current non-urgent patient to another doctor with the second lowest load in the same target resource pool, and immediately assigns the emergency patient to the doctor. If the doctor refuses to transfer, the system continues to select other doctors in the same pool. If there are no available doctors in the pool, cross-target resource pool scheduling is triggered, and a senior doctor with emergency capabilities is selected from the attending physician pool to provide temporary assistance ('Emergency Qualification Tag' is 'Yes'). This tag is added to eligible doctors during the initialization of the doctor resource pool.

[0182] After the doctor assignment is completed, a doctor assignment voucher is generated, containing patient ID, doctor ID, assignment time, urgency level, policy source information, and stored in a distributed database (such as MySQL Cluster) for subsequent feedback and audit;

[0183] To ensure that doctors see patients according to response time requirements, an emergency pop-up window is pushed to the target doctor's workstation, containing patient desensitization basic information, core symptoms (such as sudden chest pain for 20 minutes), and a lightweight 1D-CNN model for preliminary judgment (such as suspected heart attack), and showing a countdown of 30 seconds for admission;

[0184] If the doctor's workstation does not respond (such as 10 seconds after the pop-up window appears without operation), send an SMS message to the doctor's mobile phone (remaining response time = required response time in the rule library - elapsed time, such as 30 seconds total response time has elapsed 10 seconds, remaining 20 seconds), content: online consultation platform has emergency patients waiting for admission, patient ID: XXX, core symptoms: sudden chest pain, please log in to the workstation immediately;

[0185] If the doctor clicks the pop-up window admission button within the required response time (such as 30 seconds), it is marked as on-time response;

[0186] If it is not operated within the time limit (such as 30 seconds), it is marked as overtime response, triggering load preemption (reassigning the patient to other doctors in the same pool);

[0187] A hierarchical and containerized management mechanism for computing resource pools is established to meet the computing resource requirements in the urgency-data type-resource response strategy rule library, and the specific implementation process is as follows:

[0188] Based on the computing resource requirements in the urgency-data type-resource response strategy rule library, the computing power is divided into three computing resource pools:

[0189] Dedicated GPU resource pool: reserved for highest priority-real-time flow scenario, containing 4 NVIDIA T4 GPUs; each GPU is divided into 4 dedicated cores through Kubernetes (K8s) container orchestration technology, and a Pod (container group) with Guaranteed QoS level is created for each dedicated core, with GPU limit set to 0.25 cores to ensure exclusive computing power for emergency tasks;

[0190] Shared CPU resource pool: for medium priority-periodic static scenario, containing 16-core Intel Xeon CPU; through container sharing mechanism, multiple regular tasks are allowed to time-share CPU threads;

[0191] Idle Period CPU Resource Pool: For low-priority - periodic static scenario peak shaving scheduling, use the idle period of computing power from 1:00 to 3:00 in the morning, including 8-core Intel Xeon CPU (physically isolated from the shared pool to avoid affecting regular tasks);

[0192] Containerized scheduling and resource isolation use K8s for containerized management of computing power resources:

[0193] Create a Pod (container group) with Guaranteed QoS level for each dedicated core of the dedicated GPU resource pool, set CPU limit to 0.25 cores and GPU limit to 0.25 cores, to ensure that the computing power of emergency tasks is not disturbed by other tasks;

[0194] Create a Burstable QoS level Pod for the shared CPU resource pool, set CPU request to 0.1 cores and CPU limit to 0.5 cores, allow the task to burst to higher computing power when idle, and ensure that it is limited to a reasonable range when busy;

[0195] Create a BestEffort QoS level Pod for the idle period CPU resource pool, without setting a fixed computing power limit, fully utilizing idle resources;

[0196] If it is a dedicated GPU core, the K8s scheduler selects an idle Guaranteed Pod from the dedicated GPU resource pool, submits the lightweight 1D-CNN model inference task to the Pod, binds the GPU core, and ensures that the task exclusively occupies the computing power;

[0197] If it is a shared CPU thread, the scheduler selects a Burstable Pod from the shared CPU resource pool, and submits the task to the Pod with the lowest load based on the current computing power load (monitored by Prometheus);

[0198] If it is an idle period CPU thread, the scheduler puts the task into a delayed task queue, sets the trigger time to 2:00 in the morning, and dynamically selects the time point with the lowest load (e.g. 1:30) based on the idle period (1:00-3:00) computing power load monitored by Prometheus. The default fallback is 2:00 in the morning, and the task will be submitted to the BestEffort Pod in the idle resource pool at that time;

[0199] During task execution, the Pod state is monitored through K8s probes (LivenessProbe, ReadinessProbe); if the Pod is abnormal (such as GPU failure), restart the Pod and reschedule the task immediately;

[0200] After the task is completed (such as generating the "whether to send to the hospital" conclusion), K8s automatically releases the Pod resource for subsequent tasks, ensuring the recycling of computing power resources;

[0201] Embodiment 2

[0202] Based on the embodiment 1, the application provides an online consultation data intelligent analysis method based on big data, comprising:

[0203] S1: Construct a medical keyword library, identify the medical emergency keywords of the Chinese data in the online consultation text, and perform hierarchical processing according to the identification result; identify the probability of emergency in the time sequence physiological signal in the online consultation text by designing a lightweight 1D-CNN model, and perform hierarchical processing according to the probability of emergency; identify the periodic data in the online consultation text and perform hierarchical processing, and fuse the Chinese data and the time sequence physiological signal for identifying the real-time flow data and the periodic static data of two online consultation text data types;

[0204] S2: Based on the hierarchical processing result and the online consultation text data type, construct an emergency degree-data type-resource response strategy rule library, and set an online updating mechanism for the emergency degree-data type-resource response strategy rule library to cope with the high concurrency scene;

[0205] S3: Establish a doctor resource pool management and dynamic allocation mechanism, and dynamically allocate doctor resources in combination with the emergency degree-data type-resource response strategy rule library, establish a hierarchical and containerized management mechanism of the computing power resource pool, and use it to meet the computing power resource demand in the emergency degree-data type-resource response strategy rule library.

[0206] The above shows and describes the basic principles, main features and advantages of the application. It should be understood by those skilled in the art that the application is not limited by the above examples, and the above examples and descriptions in the specification are only to illustrate the principles of the application. Without departing from the spirit and scope of the application, various changes and improvements can be made to the application, and these changes and improvements all fall within the scope of the claimed application. The scope of protection of the application is defined by the appended claims and their equivalents.

Claims

1. An online diagnosis data intelligent analysis system based on big data, characterized in that: Comprise: Data feature acquisition module: construct a medical keyword library, identify medical emergency keywords in online consultation text, according to the identification result, hierarchical processing, through the design of lightweight 1D-CNN model, identify the probability of emergency in the time sequence of physiological signals in online consultation text, according to the probability of emergency, hierarchical processing; Identify the periodic data in online consultation text and carry out hierarchical processing, and fuse the text data and time sequence physiological signals to identify real-time flow data and periodic static data of two online consultation text data types; Strategy rule base construction and dynamic updating module: based on the hierarchical processing result and online consultation text data type, construct emergency-data type-resource response strategy rule base, and set online updating mechanism for emergency-data type-resource response strategy rule base to cope with high concurrency scene; Medical resource pool dynamic scheduling and management module: establish doctor resource pool management and dynamic allocation mechanism, and combine emergency-data type-resource response strategy rule base to dynamically allocate doctor resources, establish hierarchical and containerized management mechanism of computing resource pool to meet the computing resource demand in emergency-data type-resource response strategy rule base; The specific process of constructing emergency-data type-resource response strategy rule base is: Construct emergency-data type-resource response strategy rule base, use RedisCluster distributed cluster as emergency-data type-resource response strategy rule base storage engine, and design fine Key-Value data model: The Key of emergency-data type-resource response strategy rule base adopts the string combination form of emergency priority plus data type, which covers the scene division of emergency and data type double dimensions; The highest priority-real-time flow: corresponding to emergency scene; Medium priority-periodic static: corresponding to chronic disease routine follow-up scene; Low priority-periodic static: corresponding to ordinary non-emergency scene; Value adopts JSON format and contains the following key fields: Doctor resource priority: clear the doctor rank that needs to be dispatched, ensure that high seniority doctors are allocated in emergency scene first; Response time requirement: define the maximum time length that doctors need to complete the response, match the time limit requirement of different scenes; Computing resource demand: specify the required computing type and resource pool to realize the precise scheduling of computing power; Result presentation form: define the output mode of diagnosis result to meet the information transmission demand of different scenes; Doctor resource pool identification: associate specific doctor grouping to provide precise target for doctor allocation of medical resource pool dynamic scheduling and management module; Computing resource pool identification: associate specific computing power grouping to provide direct basis for computing power scheduling of medical resource pool dynamic scheduling and management module; Notification triggering mode: define the notification channel of doctors and patients to ensure the timeliness of response. 2.The big data based online consultation data intelligent analysis system according to claim 1, characterized in that: The specific process of identifying medical emergency keywords in online consultation text according to the identification result is: According to the text data in the historical online consultation text, extract medical emergency keywords; And adopt the Trie tree data structure to store the medical emergency keywords, insert all the medical emergency keywords in the medical keyword library into the Trie tree; each node of the Trie tree stores 1 character, the path from the root node to the leaf node constitutes a complete keyword, realizes the linear time complexity matching of the text medical emergency keyword, ensures the millisecond level matching speed; the leaf node additionally records the emergency degree weight and the associated medical emergency name of the medical emergency keyword; First, pre-process the character data in the patient's online consultation text, remove punctuation, stop words, and retain core medical words; Using the Trie tree matching method, the core medical words in the pre-processed online consultation text are input into the Trie tree in turn, and all the medical emergency keywords and the emergency degree weight of the medical emergency keywords in the online consultation are matched; Sum all the emergency degree weights of the medical emergency keywords; If the total emergency degree weight of the medical emergency keywords is greater than or equal to 0.8, it is marked as the highest priority; If the total emergency degree weight of the medical emergency keywords is less than 0.8 and greater than or equal to 0.3, it is marked as medium priority; If the total emergency degree weight of the medical emergency keywords is less than 0.3, it is marked as low priority. 3.The big data based online consultation data intelligent analysis system according to claim 1, characterized in that: The specific process of identifying the probability of emergency in the time sequence physiological signal in the online consultation text is: The time sequence physiological signal includes but is not limited to electrocardiogram, blood pressure waveform, and blood oxygen saturation curve; Pre-process the time sequence physiological signal, real-time intercept 30 seconds long sliding time window, normalize the waveform data in the sliding time window, map the data to the interval [−1, 1], and eliminate the baseline drift and dimension difference of different patients and different devices; The normalized time sequence physiological signal is used as the input of the lightweight 1D-CNN model; Design a lightweight 1D-CNN model containing 3 convolutional layers and 2 max pooling layers, and the final output layer is a Sigmoid activation function, which outputs the probability of whether it is an emergency. 4.The big data based online consultation data intelligent analysis system according to claim 1, characterized in that: The specific process of grading according to the probability of emergency is: Train the lightweight 1D-CNN model through the time sequence physiological signal in the historical online consultation text, verify the accuracy of the lightweight 1D-CNN model through the test set, and meet the accuracy requirement; Wherein, the parameter quantity of the lightweight 1D-CNN model is less than 100,000, and the single sample inference time is ensured within the preset time; If the emergency probability output by the lightweight 1D-CNN model is greater than or equal to 0.8, it is marked as the highest priority; If the emergency probability output by the lightweight 1D-CNN model is greater than or equal to 0.3 and less than 0.8, it is marked as medium priority; if the emergency probability output by the lightweight 1D-CNN model is less than 0.3, it is marked as low priority. 5.The big data based online consultation data intelligent analysis system according to claim 1, characterized in that: The specific process of identifying periodic data in the online consultation text and grading is: Periodic data, including but not limited to: weekly blood glucose and blood pressure records of patients with chronic diseases; extract the upload timestamp of the periodic data in the patient's online consultation text, and parse it into year-month-day format; Get the upload timestamps of the patient's last 10 times of the same type of data, and calculate the time interval between the adjacent two uploads; If the time interval of the three consecutive uploads meets the fixed periodicity rule, it is determined as periodic data; otherwise, it is determined as non-periodic data; Obtain the last three periodic data of the same type of the patient; Calculate the standard deviation of the three periodic data, and detect whether there is a critical value in the three periodic data: if any periodic data is greater than or equal to the first critical value; Or less than or equal to the second critical value, a critical value flag is triggered; If the periodic data in the online consultation text is periodic data and has no critical value, and the standard deviation is less than the standard deviation threshold, it is marked as medium priority; If the standard deviation of the periodic data in the online consultation text is greater than or equal to the standard deviation threshold or there is a critical value, it is marked as the highest priority; If the periodic data in the online consultation text is non-periodic but has a critical value, it is marked as the highest priority. 6.The big data based online consultation data intelligent analysis system according to claim 1, characterized in that: The specific process of fusing Chinese character data and time-series physiological signals for identifying real-time stream data and periodic static data two types of online consultation text data types is: The feature vectors of the text data and the physiological signals are mapped to the same dimension through a linear transformation layer to obtain , ; Calculate attention weights of the text data and the physiological signals to obtain the attention weights of the text data and the physiological signals , ; Fusing attention weights of text data and physiological signals into a feature vector: ; The fused feature vector is input into the Softmax classification layer in the lightweight 1D-CNN model, and the type label of real-time stream data or periodic static data is output. 7.The big data based online consultation data intelligent analysis system according to claim 1, characterized in that: The specific process of constructing the emergency-data type-resource response strategy rule base further includes: A 3-master 3-slave RedisCluster cluster deployment is adopted, data sharding is performed through consistent hashing, single node hot spot and performance bottleneck are avoided, RDB plus AOF hybrid persistence is started to ensure that the rules are not lost after service restart, and the heartbeat detection interval between cluster nodes is set to 100ms to ensure the rapid discovery and switching of fault nodes. 8.The big data based online consultation data intelligent analysis system according to claim 1, characterized in that: The specific process of setting an online update mechanism for the emergency-data type-resource response strategy rule base is: A web-based management interface is developed to support the full life cycle management of rules in the emergency-data type-resource response strategy rule base, including adding, modifying, deleting, and querying rules by operation and maintenance personnel or clinical experts: Rule addition: the user selects the emergency priority and data type through the drop-down box, fills in all the Value fields of the doctor resource priority and response time requirement, and submits, triggering the rule synchronization process; Rule modification: after querying the target rule through the Key combination, any Value field can be modified, and after submission, the corresponding Value in Redis is updated in real time; Rule deletion: after locating the target rule through the Key combination, confirm the deletion operation, and the corresponding Key-Value in Redis is removed immediately; Rule query: fuzzy / accurate query can be performed according to the emergency priority, data type, or Key combination, and all strategies in the current rule base are displayed. 9.The big data based online consultation data intelligent analysis system according to claim 1, characterized in that: The specific process of dynamically allocating doctor resources is: Based on the doctor resource priority in the emergency-data type-resource response strategy rule base dynamic allocation, doctors are divided into three target resource pools: Chief physician pool: including doctors with high-level titles and emergency and critical treatment experience, covering cardiology and emergency departments; Attending physician pool: including doctors with intermediate titles and common disease and chronic disease management capabilities, covering internal medicine and general practice; Resident physician pool: contains junior doctors responsible for routine consultation and follow-up; each pool is associated with the priority field of the doctor resource in the emergency-data type-resource response strategy rule base, ensuring that the strategy instruction can accurately point to the target pool; WebSocket long connection and heartbeat detection technology are used to real-time perceive the online status and current load of doctors: WebSocket long connection is established for each doctor's workstation, and the server sends a heartbeat packet to the client every 5 seconds; If the server does not receive a response from the client for 3 consecutive times, the doctor is marked as offline; if a response is received, the doctor is marked as online; When a doctor receives a patient, the workstation sends a start signal to the server; when the patient finishes the consultation or the doctor marks the consultation as complete, an end signal is sent; the server real-time counts the number of patients each doctor is currently receiving as a load indicator; Select an online doctor with the lowest current load from the target resource pool; If all doctors in the target resource pool are in a high-load state, trigger load preemption: Select a doctor in the target resource pool with the highest load but receiving non-urgent patients, and push a preemption prompt to their workstation; If the doctor confirms the transfer, the system automatically assigns the current non-urgent patient to a doctor with the second-lowest load in the target resource pool, and immediately assigns the emergency patient to the doctor with the highest load but receiving non-urgent patients; If the doctor refuses to transfer, the system continues to select other doctors in the pool; if there are no available doctors in the pool, trigger cross-target resource pool scheduling and select a senior doctor with emergency capabilities from the attending physician pool for temporary assistance. 10.The big data based online consultation data intelligent analysis system according to claim 1, characterized in that: The specific process of establishing a hierarchical and containerized management mechanism for the computing resource pool is as follows: Based on the computing resource demand in the emergency-data type-resource response strategy rule base, the computing power is divided into three computing resource pools: Exclusive GPU resource pool: reserved for the highest priority-real-time streaming scenario, containing 4 NVIDIA T4 GPUs; each GPU is divided into 4 exclusive cores through Kubernetes container orchestration technology, and a Guaranteed QoS level Pod is created for each exclusive core with a GPU limit of 0.25 cores; Shared CPU resource pool: for medium priority-periodic static scenario, containing 16-core Intel Xeon CPU; through container sharing mechanism, multiple regular tasks are allowed to time-share CPU threads; Idle period CPU resource pool: for low priority-periodic static scenario peak shifting, taking advantage of the idle period of computing power from 1:00 to 3:00, containing 8-core Intel Xeon CPU; Containerized scheduling and resource isolation use K8s for containerized management of computing resources: For each exclusive core of the exclusive GPU resource pool, create a Guaranteed QoS level Pod with a CPU limit of 0.25 cores and a GPU limit of 0.25 cores; For the shared CPU resource pool, create a Burstable QoS level Pod with a CPU request of 0.1 cores and a CPU limit of 0.5 cores, allowing tasks to burst to higher power when idle. Create a BestEffort QoS level Pod for the idle period CPU resource pool, without setting a fixed computing power limit; If it is a dedicated GPU core, the K8s scheduler selects an idle Guaranteed Pod from the dedicated GPU resource pool and submits the lightweight 1D-CNN model inference task to the Pod, binding the GPU core; If it is a shared CPU thread, the scheduler selects a Burstable Pod from the shared CPU resource pool and submits the task to the Pod with the lowest load according to the current computing power load; If it is an idle period CPU thread, the scheduler places the task in the delayed task queue and sets the trigger time to 2:00 AM. Based on the computing power load during the idle period monitored by Prometheus, the task is triggered at the time point with the lowest load. The default fallback is 2:00 AM, and the task is submitted to the BestEffort Pod in the idle resource pool at that time; During task execution, the Pod state is monitored through K8s probes; if the Pod is abnormal, the Pod is restarted and the task is rescheduled; After the task is completed, K8s automatically releases the Pod resources for subsequent tasks.

Images