Medical data classification labeling method and system applied to large model mining

Abstract

The application relates to the technical field of medical big data processing, and relates to a medical data classification and labeling method and system applied to big model mining, which comprises a heterogeneous data phasor measurement unit, which samples multi-modal medical data and calculates a semantic deviation angle and a semantic angular velocity; a transient energy evaluation engine, which is configured with a Lyapunov function calculation core and is used for real-time calculation of semantic kinetic energy, semantic potential energy and instantaneous total energy of the sample; a stability discrimination and shunt controller, which calculates an energy margin and shunts the data to different channels; and an excitation feedback adjustment module, which executes an adaptive excitation control strategy and dynamically adjusts big model parameters in the medical field. The application quantifies data semantic reliability by using transient energy stability theory, accurately identifies critical high-value samples by calculating an energy margin, effectively suppresses model hallucinations by adaptive parameter adjustment while greatly reducing manual labeling costs, and significantly improves the clinical safety and robustness of the medical big model.

Description

Technical Field

[0001] This invention relates to the field of medical big data processing technology, and more specifically, to a method and system for classifying and labeling medical data for large model mining. Background Technology

[0002] With the rapid development of deep learning technology, large-scale medical language models have demonstrated enormous application potential in scenarios such as clinical auxiliary diagnosis, medical image report generation, and structured processing of electronic medical records. Medical data is highly multimodal and heterogeneous, encompassing unstructured electronic medical record text descriptions, high-dimensional DICOM image data, and continuously changing vital sign time-series waveforms. To improve the reasoning ability and accuracy of models in complex clinical scenarios, it is typically necessary to construct large-scale and high-quality instruction fine-tuning datasets to map the raw heterogeneous medical data to a standard medical knowledge space. Currently, academia and industry are working to extract samples with high medical cognitive value from massive amounts of routine hospital diagnostic and treatment data using automated data mining techniques, enabling continuous iteration and evolution of models. This process usually involves deep extraction of semantic features from multimodal data, mapping feature vectors to Riemannian manifold space, and alignment analysis with standard medical knowledge graphs, aiming to establish an efficient and intelligent data closed-loop flow system to support the deep learning and mastery of professional medical knowledge by large-scale medical models.

[0003] On the one hand, existing medical data screening and annotation methods are mostly based on static uncertainty sampling strategies such as minimum confidence or maximum entropy sampling. However, these methods only focus on the predicted probability distribution of the model at a single inference moment, ignoring the dynamic trajectory of semantic changes throughout the inference process. As a result, it is difficult to effectively distinguish between low-quality noisy data and critically difficult samples with high training value. On the other hand, traditional data cleaning methods often lack deep physical interpretability and cannot accurately quantify the specific degree to which samples deviate from standard clinical guidelines. Moreover, when faced with multi-source heterogeneous data from different medical institutions, it is difficult to eliminate systematic distribution biases caused by differences in equipment parameters or recording habits. Furthermore, relying solely on human experts to review all the screened long-tail samples is not only time-consuming, labor-intensive, and inefficient, but also fails to meet the urgent needs of the era of large models for real-time throughput and rapid iteration of massive amounts of data. This leads to logical collapse or factual errors when models face critical data at the knowledge boundary due to a lack of targeted training. Summary of the Invention

[0004] The purpose of this invention is to provide a medical data classification and labeling method and system for large model mining. It utilizes transient energy stability theory to quantify the semantic reliability of data, accurately identifies critical high-value samples by calculating energy margin, significantly reduces the cost of manual labeling, and effectively suppresses model illusion through adaptive parameter adjustment, thereby significantly improving the clinical safety and robustness of large medical models.

[0005] This invention is achieved through the following technical solution:

[0006] A medical data classification and annotation method applied to large model mining, the method includes the following steps:

[0007] Deploy multimodal data sensing probes in the medical data lake to capture raw heterogeneous data streams in real time, including electronic medical record text, image DICOM data, and vital sign time-series waveforms;

[0008] Based on the original heterogeneous data stream, feature vectors are extracted through a pre-set large model in the medical field, and the feature vectors are mapped to the Riemannian manifold space. By calculating the semantic deviation angle and semantic angular velocity relative to the standard medical knowledge graph, a semantic phase plane trajectory is constructed.

[0009] Based on the semantic phase plane trajectory, a Lyapunov energy function containing semantic kinetic energy and semantic potential energy is constructed, and the instantaneous total energy of each data sample causing perturbation of the prediction probability distribution of the medical big language model is calculated.

[0010] By combining the instantaneous total energy, the energy margin of the data sample is calculated using the equal area rule, and the energy margin is compared with the preset critical stability threshold to quantify the annotation value and cognitive risk of the data.

[0011] Based on the comparison between energy margin and critical stability threshold, the data stream is split through an adaptive excitation control strategy, and directed to an automatic annotation channel, a human expert review channel, or an adversarial sample cleaning channel, respectively, to generate a command fine-tuning dataset.

[0012] Optionally, the construction of the semantic phase plane trajectory specifically involves:

[0013] Define the standard diagnosis and treatment path in the medical knowledge graph as the synchronous rotating coordinate system of the system, and use the feature vector of the data sample i to be mined. The projection phase difference in this coordinate system is defined as the semantic deviation angle. ;

[0014] The information entropy of the predicted probability distribution of sample i in the medical big language model during inference is calculated as the first derivative with the number of inference steps t, and is defined as the semantic angular velocity. :

[0015]

[0016] in, Information entropy predicted by large models in the medical field. These are the normalization coefficients;

[0017] With semantic deviation angle x-axis, semantic angular velocity Using the vertical axis as the ordinate, construct the semantic phase plane trajectory.

[0018] Optionally, the Lyapunov energy function and its instantaneous total energy are calculated as follows:

[0019]

[0020] in, The sample information inertia is equal to the product of the number of medical entities in the sample and the modal mutual information, and is used to characterize the computational cost required to change the predicted distribution of the sample by a large medical model. Let i be the semantic angular velocity of sample i; The reference semantic deviation angle is set to 0, which represents a state that fully complies with the clinical guidelines. The model output power is defined as the confidence score of the top-1 diagnostic conclusions generated by the medical big language model for the samples. The graph constraint power is defined as the deviation angle of the medical knowledge graph. The logical error correction force applied to the samples; Let be the instantaneous total energy of sample i.

[0021] Optionally, the sample information inertia In multi-source data scenarios, the computation specifically employs the multi-machine system equivalent inertia center algorithm, and the calculation formula is as follows:

[0022] Treating different data source nodes as independent generator sets, calculate the global inertia center location. ;

[0023] Where N is the total number of data source nodes participating in data mining; j is the index of the data source node; The average information inertia of the dataset provided for the j-th node; Let be the average semantic deviation angle of the dataset at node j relative to the standard knowledge graph;

[0024] The semantic deviation angle Revised to To eliminate systematic biases.

[0025] Optionally, the spectrum constraint power The maximum amplitude in These are time-varying parameters, specifically:

[0026] The The value and the attention coefficient of current clinical guidelines Proportional relationship: ;

[0027] Where t is the current system time; The preset reference constraint power constant; This is the sensitivity adjustment coefficient for clinical guidelines, and its value is greater than 0. This represents the current level of attention given to clinical guidelines for a specific disease.

[0028] Optionally, the application of the equal area rule to calculate the energy margin of the data sample specifically involves:

[0029] The unstable equilibrium point of the computational system corresponds to the semantic deviation angle reaching a critical value. and satisfy the equation ;

[0030] Calculate the critical potential energy at this unstable equilibrium point. : , where the angular velocity is 0;

[0031] Calculate the energy margin of sample i The calculation formula is as follows:

[0032]

[0033] in, This represents the instantaneous total energy.

[0034] Optionally, the adaptive excitation control strategy is specifically as follows:

[0035] when If the data is determined to be an unstable sample, the sample will be routed to the human expert review channel.

[0036] when At that time, among them, If the data is determined to be a critical sample based on a preset safety threshold constant, the sample will be routed to the adversarial sample cleaning channel.

[0037] when When the data is determined to be a steady-state sample, the sample is routed to the automatic labeling channel.

[0038] Optionally, an adaptive excitation control strategy is implemented, specifically based on the energy margin. Dynamically adjusting the sampling temperature parameters of large medical models The calculation formula is:

[0039]

[0040] in, The fundamental temperature constant for large-scale model reasoning in the medical field; This is the excitation gain coefficient; This represents the absolute value of the energy margin.

[0041] Optionally, the generation instruction fine-tuning dataset is specifically based on energy margin. Assign training loss weights to each sample routed to the auto-labeling channel or the adversarial sample cleaning channel. The calculation formula is:

[0042]

[0043] in, To prevent smoothing constants with denominators of zero.

[0044] A medical data classification and annotation system applied to large model mining, comprising:

[0045] The heterogeneous data phasor measurement unit samples multimodal medical data and calculates semantic deviation angle and semantic angular velocity;

[0046] The transient energy assessment engine is equipped with a Lyapunov function calculation kernel to calculate the semantic kinetic energy, semantic potential energy and instantaneous total energy of the sample in real time.

[0047] Stability assessment and shunt controller calculate energy margin and shunt data to different channels;

[0048] The excitation feedback adjustment module executes an adaptive excitation control strategy and dynamically adjusts the parameters of a large medical model.

[0049] The technical solution of the present invention has at least the following advantages and beneficial effects:

[0050] This invention can accurately calculate the energy margin of each data sample, thus predicting whether a sample will cause the model's inference to diverge or produce hallucinations that contradict medical common sense before model training. Furthermore, through an adaptive excitation control strategy, the system dynamically adjusts the model's sampling temperature and training loss weights based on the energy margin, enabling targeted reinforcement learning for high-value samples in a "critically stable" state, effectively solving the balance problem between hard example discovery and noise cleansing. In addition, the use of the multi-machine system equivalent center of inertia (COI) algorithm effectively eliminates systematic biases in multi-source data, significantly reducing reliance on expensive human expert annotation while significantly improving the robustness, safety, and diagnostic accuracy of large-scale medical models in complex clinical environments. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating the medical data classification and annotation method for large model mining provided by the present invention.

[0052] Figure 2 This is a schematic diagram illustrating the principle of the medical data classification and annotation system for large model mining provided by the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0054] like Figure 1 , Figure 2 As shown, this invention provides a method and system for medical data classification and annotation applied to large-scale model mining. The core idea of ​​this method is to analogize the dynamic reasoning process of a large medical language model processing data samples to the transient response process of a power system after being disturbed. By constructing an energy function to quantify the semantic stability of the model, it is possible to accurately identify and evaluate those critically stable data samples that are at the model's cognitive boundary and have the most training value, thereby achieving an efficient, intelligent, and safe data annotation and model iteration closed loop.

[0055] The medical data classification and labeling system for large-scale model mining, as described in this invention, can be deployed on cloud server clusters or local data centers with powerful computing capabilities. At the hardware level, the system typically includes a graphics processing unit (GPU) array for massively parallel computing; an InfiniBand or RoCE network for high-speed data exchange; and a distributed storage system for storing massive amounts of raw medical data and intermediate computation results.

[0056] At the software level, the system consists of multiple collaborative program modules that jointly implement the method described in this invention. The core modules of the system include: a heterogeneous data phasor measurement unit, a transient energy assessment engine, a stability discrimination and shunt controller, and an excitation feedback regulation module. The program instructions for all these modules are stored in non-volatile memory, such as a solid-state drive or network-attached storage, and are loaded and executed by the central processing unit and graphics processing unit.

[0057] The heterogeneous data phasor measurement unit is the central hub for data input and preprocessing in the entire system. This unit is responsible for interfacing with various information systems within the hospital. For example, it retrieves unstructured text data, such as chief complaints, present medical history, past medical history, physical examination records, and treatment plans, from electronic medical record systems or hospital information systems in real time via a HL7-compliant message protocol interface. Simultaneously, it accesses image archiving and communication systems via the DICOM protocol to obtain raw file streams of medical images such as CT, MRI, and X-rays. Furthermore, this unit can connect to vital sign monitoring equipment in the intensive care unit via IoT protocols such as MQTT to capture high-frequency time-series waveform data such as electrocardiograms, electroencephalograms, blood pressure, and blood oxygen saturation. Internally, this unit integrates sub-modules for data cleaning, format conversion, and anonymization to ensure that the data input to subsequent processes is clean and compliant. The core function of this unit is to execute the steps described in detail later: extracting data features and constructing semantic phase plane trajectories, outputting two key fundamental variables for energy assessment.

[0058] The transient energy assessment engine is the core computational unit of this invention. This engine deploys an optimized Lyapunov function computational kernel specifically designed to perform the semantic transient energy calculations proposed in this invention. It receives fundamental variables from the phasor measurement unit, and based on a set of physically meaningful computational logic, calculates the semantic kinetic energy and semantic potential energy of each data sample in real time, ultimately summing them to obtain the instantaneous total energy. This engine typically runs on a high-performance graphics processing unit (GPU) and utilizes CUDA technology for parallel acceleration to meet the performance requirements of real-time energy assessment of massive data streams.

[0059] The stability assessment and triage controller acts as the decision-making and scheduling center. It receives energy values ​​calculated by the transient energy assessment engine and performs two key tasks. First, based on the principle of equal area, it calculates the energy margin for each sample, which is the final quantitative indicator of sample stability. Second, it compares the magnitude of the energy margin with multiple pre-set thresholds to determine whether the data sample is steady-state, critical, or unstable. Based on this determination, the controller dynamically routes the data samples to different processing channels; for example, it sends steady-state samples to the automatic labeling process, pushes unstable samples to the human expert review interface, and imports critical samples into specific data augmentation or cleaning modules.

[0060] The excitation feedback adjustment module constitutes a sophisticated closed-loop feedback control system. It also receives the calculated energy margin, but its role is to adjust the behavior of the pre-trained medical large-scale language model in reverse. Based on the energy margin, this module generates control commands in real time, dynamically adjusting the model's internal parameters when processing specific samples. For example, it can adjust the model's sampling temperature parameter to control the creativity and determinism of the model's output. Furthermore, it is responsible for generating a dynamic training weight for the samples about to enter the training process. The magnitude of this weight is directly related to the criticality of the sample, ensuring that the model focuses on the most error-prone and most learning-intensive samples during subsequent fine-tuning phases.

[0061] like Figure 1 As shown, the present invention details how the above system architecture works together to achieve the method described in the present invention through the following complete processing flow.

[0062] Step 1: Perception and capture of multimodal heterogeneous data.

[0063] The process begins with the activation of the heterogeneous data phasor measurement unit. This unit first deploys a series of data-aware probes within the target medical data storage system. These probes are essentially lightweight software agents or services configured to monitor changes in table entries in specific databases, directory updates in file systems, or new messages in message queues.

[0064] When a new diagnostic event occurs, such as a patient completing a CT scan and generating an imaging report, the relevant data streams are triggered. The data-aware probe captures a series of related data: a set of CT image files stored in DICOM format, an imaging diagnostic report text written by a radiologist and recorded in the electronic medical record system, and time-series data of the patient's vital signs recorded synchronously during the examination. These raw, heterogeneous data streams are encrypted by the probe and transmitted in real time via a high-speed network to the main processing module of the heterogeneous data phasor measurement unit.

[0065] Step 2: Construction of semantic phase plane trajectories.

[0066] Once the data arrives, the core concept construction phase of this invention begins: transforming discrete, multimodal data points into phasors capable of describing their dynamic characteristics in mathematical space. This step is accomplished by a heterogeneous data phasor measurement unit, which internally calls a key component—a pre-trained medical large-scale language model. This model is a deep learning model that has been pre-trained on large-scale medical literature, textbooks, and anonymous medical records, such as a language model with hundreds of billions of parameters based on the Transformer architecture, which has a deep understanding of medical terminology and clinical logic.

[0067] The goal of this step is to calculate two core coordinates for each data sample: the semantic deviation angle and the semantic angular velocity. These two coordinates together define the sample's position and motion trend on the semantic phase plane.

[0068] The calculation process for the semantic deviation angle is as follows:

[0069] The system requires an absolute reference standard, provided by a standard medical knowledge graph. This knowledge graph is a vast semantic network where nodes represent medical concepts such as diseases, symptoms, drugs, and examinations, and edges represent relationships between them, such as treatment, manifestation, and discontinuation. From this graph, the system extracts one or more standard diagnostic and treatment pathways for specific clinical problems, such as the diagnosis of lung cancer. Mathematically, this pathway is represented as a high-dimensional reference feature vector, equivalent to the most typical and unambiguous ideal case described in textbooks.

[0070] The captured data samples to be mined, such as the aforementioned CT images and diagnostic report texts, are input into a pre-trained medical large language model. The model encodes this multimodal data, representing it uniformly as a high-dimensional feature vector of the same dimension. This vector contains the model's comprehensive understanding of all information in the current sample.

[0071] The angle or phase difference between the feature vector of the data sample to be mined and the reference feature vector of the standard diagnostic path in the Riemannian manifold space they both reside in is calculated. This calculated angle is defined as the semantic deviation angle of the data sample. The physical meaning of this angle is very intuitive: it quantifies the degree of deviation between the current real-world case and the most standard, ideal textbook case. A sample with a deviation angle of zero or close to zero means that it is a very typical case; while a sample with a large deviation angle may mean that it is a rare, atypical, or complex case.

[0072] The calculation process for semantic angular velocity focuses on dynamic changes:

[0073] When data samples are input into a pre-trained medical big data language model for inference analysis, the model's internal computation is not instantaneous, but rather undergoes a layered or time-step-by-time generation process. In each inference step, the model generates a probability distribution for the final conclusion, such as the diagnosis.

[0074] The first sub-step in calculating semantic angular velocity is to calculate the information entropy of the output probability distribution at each inference step of the model. Information entropy is a concept in information theory that measures the degree of uncertainty of a probability distribution. If the model is very certain about a certain conclusion, its probability distribution will be sharp, and the information entropy will be low; conversely, if the model is confused and hesitates between multiple possibilities, its probability distribution will be flat, and the information entropy will be high.

[0075] Therefore, as the reasoning proceeds, a sequence of information entropy changes with the number of reasoning steps is obtained.

[0076] The second sub-step involves calculating the first derivative of this information entropy sequence, which is the rate of change of information entropy. In discrete computation, this can be approximated as the difference in information entropy between two adjacent inference steps. This rate of change reflects the speed at which the uncertainty of the model changes during the inference process.

[0077] Multiplying the calculated rate of change of information entropy by a preset normalization coefficient yields the semantic angular velocity of the sample. Its physical meaning is whether the model's internal processing of this sample tends towards stability or chaos. A positive angular velocity indicates that the model's uncertainty is rapidly increasing, potentially encountering contradictions or gaps in its knowledge system; a negative angular velocity indicates that the model is transitioning from uncertainty to certainty, finding a stable explanation; and an angular velocity close to zero indicates that the model's processing is very smooth.

[0078] Thus, each data sample is endowed with two dynamic attributes: semantic deviation angle and semantic angular velocity. Together, they constitute the trajectory of the sample on the semantic phase plane, laying the foundation for subsequent energy analysis.

[0079] Step 3: Calculation of instantaneous total energy.

[0080] After obtaining the coordinates on the semantic phase plane, the transient energy assessment engine begins operation. It introduces the Lyapunov energy function, a core tool in power system stability analysis, to construct a semantic transient energy function. The total energy value of this function consists of two parts: semantic kinetic energy and semantic potential energy.

[0081] The calculation process for semantic kinetic energy is as follows:

[0082] First, obtain the semantic angular velocity of the sample calculated in the previous step.

[0083] Second, we introduce a parameter called sample information inertia. This parameter is similar to mass in the physical world, representing the tendency of a data sample to maintain its current semantic state in the semantic space. A sample with a large amount of information and complex content, such as a medical record of a difficult and complicated disease containing decades of medical history and multiple surgical records, will be given a large information inertia, while a simple routine blood test report will have a smaller information inertia.

[0084] Third, square the semantic angular velocity, multiply it by the information inertia of the sample, and then multiply it by half. This calculation result is the semantic kinetic energy. It represents the degree of drastic change or momentum of the model's internal cognitive state when processing the sample. The larger the semantic angular velocity, the greater the kinetic energy, indicating that the model's reasoning process is more drastic and unstable.

[0085] When processing multi-source data from different hospitals or batches, an algorithm called the multi-machine system equivalent inertia center is used to calculate sample information inertia in order to eliminate systematic bias. The specific calculation process is as follows: First, data nodes from different data sources, such as different hospitals, are treated as independent generator units in a power grid. Then, for all samples within a batch, a weighted average of their semantic deviation angles is calculated, with the weight representing the information inertia of each sample. This weighted average angle is called the inertia center angle, which represents the overall semantic offset benchmark for this batch of data. Finally, the calculated inertia center angle is subtracted from the original semantic deviation angle of each sample to obtain a corrected relative deviation angle. All subsequent calculations will use this corrected deviation angle. This process effectively calibrates data distribution biases caused by differences in recording habits or equipment among different institutions, making energy assessment standards more uniform and fair.

[0086] The calculation of semantic potential is more complex; it measures the internal stress of the system caused by semantic deviation.

[0087] First, we need to define two antagonistic powers. The first is the model output power, which is defined as the confidence score of the pre-trained medical large language model in giving its final conclusion for the current sample. This score directly reflects the strength of the model's own tendency to output conclusions.

[0088] The second is the graph constraint power, which represents the objective constraint force from the standard medical knowledge graph. This power is calculated by taking the sine of the semantic deviation angle of the current sample (i.e., the corrected deviation angle) and multiplying it by a coefficient called the maximum amplitude. The characteristic of this constraint power is that it is zero when the deviation angle is zero; as the deviation angle increases, its constraint force first strengthens and then weakens, consistent with the restoring force characteristics of a physical system.

[0089] It's important to note that this maximum amplitude coefficient is not a fixed constant. It adjusts dynamically over time. The calculation involves using a baseline constant as a base, plus an increment proportional to the clinical guideline attention coefficient. This attention coefficient is maintained by a module in the system's backend that tracks the latest clinical guidelines published by major medical journals and health organizations. When there are significant updates to guidelines for a particular disease area, the corresponding attention coefficient is increased. This results in a corresponding increase in the maximum amplitude of the map constraint power related to that disease, meaning the system will exert a stronger constraint effect on data that may deviate from the latest guidelines.

[0090] Second, calculate the difference between the model output power and the power of the spectral constraints. This difference can be understood as the instantaneous power imbalance between the model output tendency and the objective law constraints at the current semantic deviation angle.

[0091] Third, this power imbalance is numerically integrated from the zero-degree deviation angle up to the semantic deviation angle of the current sample. The result of this integration is the semantic potential energy of the sample. It represents the total system tension or internal stress accumulated from a perfectly standard ideal state to the current state. The larger the deviation angle, or the greater the contradiction between the model output confidence and the spectral constraints, the higher the accumulated potential energy.

[0092] Adding the semantic kinetic energy and semantic potential energy calculated earlier yields the instantaneous total energy of the sample at that moment. This total energy value combines the intensity of the model's processing, i.e., kinetic energy, and the risk of its conclusion deviating from objective standards, i.e., potential energy, and forms the basis for all subsequent decisions.

[0093] Step 4: Calculation of energy margin and quantification of data value.

[0094] After obtaining the instantaneous total energy, the stability assessment and shunt controller begin to perform its core function: applying the equal area rule to calculate the energy margin of the data sample. The equal area rule is a classic criterion in power system analysis for determining whether a system can resume synchronous operation after a large disturbance; here, it is creatively applied to semantic stability analysis.

[0095] The specific calculation logic is as follows:

[0096] The system needs to pre-calculate a critical energy value. This value corresponds to the maximum potential energy the system can withstand. In semantic space, it corresponds to an unstable equilibrium point, i.e., a semantic deviation angle. Once this angle is exceeded, the system's own recovery ability, i.e., the graph constraint power, will be unable to pull its reasoning process back to the stable region, and its reasoning process will inevitably lead to logical divergence or severe hallucinations. The potential energy corresponding to this critical point is the system's critical energy.

[0097] Subtract the instantaneous total energy of the current data sample from this pre-calculated critical energy value; this is the energy value calculated in step three. This difference is defined as the energy margin of the sample.

[0098] Energy margin is a highly valuable quantitative metric. A large positive margin means that the energy of the current sample is far below the system's stability limit, indicating a safe and stable sample. A positive margin close to zero means that the sample's energy is very close to the critical limit, the system is in a critically stable state, and this sample is a critical sample with significant mining value. Conversely, a negative margin means that the sample's energy has exceeded the critical value, the system is semantically unstable, and this sample is likely an unstable sample containing erroneous information, harmful information, or information that the model cannot understand at all. In this way, the previously ambiguous value of data annotation is precisely quantified into a concrete physical quantity.

[0099] Step 5: Adaptive excitation control and data splitting.

[0100] Based on the calculated energy margin, the stability assessment and the shunt controller and excitation feedback adjustment module work together to execute the final decision and control actions.

[0101] The logic for determining data splitting is as follows:

[0102] The controller has at least one configurable stability threshold. Once the energy margin of a data sample is calculated, the controller performs a three-branch decision:

[0103] In the first scenario, if the energy margin is less than zero, the sample is considered unstable. This typically means that the information in the sample severely conflicts with the knowledge base of the pre-trained medical large language model, or that the data itself has quality issues. The controller will package the sample and its related information, such as its energy trajectory and associated original medical records, and send it to a human expert review channel. This channel is usually a web application interface where human medical experts can view the high-risk cases screened by the system and review, correct, and annotate them.

[0104] In the second scenario, if the energy margin is greater than or equal to zero but less than a preset stability threshold, the sample is considered a critical sample. These samples are the core targets of this invention; they have no obvious errors and are just on the edge of the model's cognitive ability, making them the most valuable for the model's growth. The controller routes these samples to an adversarial sample cleaning channel. In this channel, the system may make minor perturbations to the sample to test the model's robustness, or use the sample as a seed to generate a series of similar but more challenging adversarial samples for subsequent model reinforcement training.

[0105] In the third scenario, if the energy margin is greater than or equal to a preset stability threshold, the sample is classified as a steady-state sample. This indicates that the model can handle the sample with great confidence and accuracy, and its internal logic is highly consistent with the knowledge graph. The controller will route such samples to the automatic labeling channel. In this channel, the output of the pre-trained medical large language model for the sample, such as a diagnostic conclusion or report summary, will be adopted by the system as a high-quality baseline label and directly stored in the instruction fine-tuning dataset without manual intervention.

[0106] Based on the above-mentioned data diversion, the system greatly improves the efficiency of data annotation, and accurately uses valuable human expert resources where they are most needed.

[0107] Simultaneously, an adaptive excitation control strategy is executed, which is a process of real-time, fine-tuning of the pre-trained medical large language model. This strategy is implemented by the excitation feedback adjustment module and includes two aspects of adjustment:

[0108] The first aspect is the dynamic adjustment of the sampling temperature parameter of the model. Sampling temperature is a key parameter when a language model generates text; the higher the temperature, the more random and creative the model's output; the lower the temperature, the more deterministic the output, and the more inclined it is to select the words with the highest probability. The adjustment logic of this invention is:

[0109] Obtain the absolute value of the energy margin for the current sample. Multiply this absolute value by a negative excitation gain coefficient. Calculate the natural exponential function value, with the result of the previous step as the exponent. Multiply this exponential function result by a base temperature constant. The final value is the dynamically set new sampling temperature for processing the current sample.

[0110] According to this calculation process, when the energy margin of a sample is very close to zero, i.e., at the critical state, its absolute value is very small, the exponential term is close to zero, and the calculated exponential function value is close to one. Therefore, the final sampling temperature will be close to the base temperature constant, maintaining a relatively high level. The intention of this design is to encourage the model to explore further and try different inference paths when faced with a perplexing critical sample, rather than prematurely converging to a potentially unstable conclusion. Conversely, for stable or unstable samples with a large energy margin, the absolute value is large, the exponential term is a large negative number, and the exponential function value becomes very small, thus significantly reducing the sampling temperature and making the model more deterministic and conservative when dealing with these samples.

[0111] The second aspect is assigning training loss weights to the samples. This step prepares for subsequent model fine-tuning. The calculation process is as follows:

[0112] Obtain the energy margin of the current sample. Add a very small positive number to this energy margin; this value is called the smoothing constant, and its sole purpose is to prevent the mathematical error of dividing by zero when the energy margin is exactly zero. Calculate the reciprocal of the result from the previous step. This reciprocal value is the training loss weight for this sample.

[0113] The significance of this weighting lies in the fact that samples with smaller energy margins, i.e., more critical samples, have larger reciprocals and thus receive higher training weights. When these data samples, assigned different weights, collectively form a fine-tuning dataset and are used for the next round of model training, the model will pay more attention to the errors generated by samples with higher weights when calculating the total loss function. This forces the model to invest more learning resources in tackling difficult samples that were previously poorly processed, thereby achieving the most efficient knowledge absorption and capability improvement.

[0114] In summary, this embodiment successfully transfers the concept of transient stability analysis of power systems to the field of data mining for medical large-scale language models by constructing a complete closed loop from data perception, feature extraction, energy assessment to decision control. It not only provides a novel, physically interpretable perspective for evaluating data value but also establishes an efficient, safe, and automated mechanism for continuous model evolution through intelligent diversion and adaptive feedback control. This method can significantly improve the production efficiency of high-quality training data, reduce reliance on manual annotation, and ultimately enhance the accuracy, robustness, and safety of medical large-scale language models in real, complex clinical environments.

[0115] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A medical data classification and annotation method applied to large model mining, characterized in that, The steps of this method include: Deploy multimodal data sensing probes in the medical data lake to capture raw heterogeneous data streams in real time, including electronic medical record text, image DICOM data, and vital sign time-series waveforms; Based on the original heterogeneous data stream, feature vectors are extracted through a pre-set large model in the medical field, and the feature vectors are mapped to the Riemannian manifold space. By calculating the semantic deviation angle and semantic angular velocity relative to the standard medical knowledge graph, a semantic phase plane trajectory is constructed. Based on the semantic phase plane trajectory, a Lyapunov energy function containing semantic kinetic energy and semantic potential energy is constructed, and the instantaneous total energy of each data sample causing perturbation of the prediction probability distribution of the medical big language model is calculated. By combining the instantaneous total energy, the energy margin of the data sample is calculated using the equal area rule, and the energy margin is compared with the preset critical stability threshold to quantify the annotation value and cognitive risk of the data. Based on the comparison results of energy margin and critical stability threshold, the data stream is split through an adaptive excitation control strategy and directed to the automatic annotation channel, the manual expert review channel, or the adversarial sample cleaning channel, respectively, to generate a command fine-tuning dataset. The construction of the semantic phase plane trajectory specifically involves: Define the standard diagnosis and treatment path in the medical knowledge graph as the synchronous rotating coordinate system of the system, and use the feature vector of the data sample i to be mined. The projection phase difference in this coordinate system is defined as the semantic deviation angle. ; The information entropy of the predicted probability distribution of sample i in the medical big language model during inference is calculated as the first derivative with the number of inference steps t, and is defined as the semantic angular velocity. : in, Information entropy predicted by large models in the medical field. These are the normalization coefficients; With semantic deviation angle x-axis, semantic angular velocity Using the vertical axis as the ordinate, construct the semantic phase plane trajectory.

2. The medical data classification and labeling method applied to large model mining according to claim 1, characterized in that, The Lyapunov energy function and its instantaneous total energy calculation formula are as follows: in, The sample information inertia is equal to the product of the number of medical entities in the sample and the modal mutual information, and is used to characterize the computational cost required to change the predicted distribution of the sample by a large medical model. Let i be the semantic angular velocity of sample i; The reference semantic deviation angle is set to 0, which represents a state that fully complies with the clinical guidelines. The model output power is defined as the confidence score of the top-1 diagnostic conclusions generated by the medical big language model for the samples. The graph constraint power is defined as the deviation angle of the medical knowledge graph. The logical error correction force applied to the samples; Let be the instantaneous total energy of sample i.

3. The medical data classification and annotation method applied to large model mining according to claim 2, characterized in that, The sample information inertia In multi-source data scenarios, the computation specifically employs the multi-machine system equivalent inertia center algorithm, and the calculation formula is as follows: Treating different data source nodes as independent generator sets, calculate the global inertia center location. ; Where N is the total number of data source nodes participating in data mining; j is the index of the data source node; The average information inertia of the dataset provided for the j-th node; Let be the average semantic deviation angle of the dataset at node j relative to the standard knowledge graph; The semantic deviation angle Revised to To eliminate systematic biases.

4. The medical data classification and annotation method applied to large model mining according to claim 3, characterized in that, The spectrum constraint power The maximum amplitude in These are time-varying parameters, specifically: The The value and the attention coefficient of current clinical guidelines Proportional relationship: ; Where t is the current system time; The preset reference constraint power constant; This is the sensitivity adjustment coefficient for clinical guidelines, and its value is greater than 0. This represents the current level of attention given to clinical guidelines for a specific disease.

5. The medical data classification and annotation method applied to large model mining according to claim 4, characterized in that, The application of the equal area rule to calculate the energy margin of the data sample is specifically as follows: The unstable equilibrium point of the computational system corresponds to the semantic deviation angle reaching a critical value. and satisfy the equation ; Calculate the critical potential energy at this unstable equilibrium point. : , where the angular velocity is 0; Calculate the energy margin of sample i The calculation formula is as follows: in, This represents the instantaneous total energy.

6. The medical data classification and annotation method applied to large model mining according to claim 5, characterized in that, The adaptive excitation control strategy is as follows: when If the data is determined to be an unstable sample, the sample will be routed to the human expert review channel. when At that time, among them, If the data is determined to be a critical sample based on a preset safety threshold constant, the sample will be routed to the adversarial sample cleaning channel. when When the data is determined to be a steady-state sample, the sample is routed to the automatic labeling channel.

7. The medical data classification and annotation method applied to large model mining according to claim 6, characterized in that, The adaptive excitation control strategy is implemented, specifically based on the energy margin. Dynamically adjusting the sampling temperature parameters of large medical models The calculation formula is: in, The fundamental temperature constant for large-scale model reasoning in the medical field; This is the excitation gain coefficient; This represents the absolute value of the energy margin.

8. The medical data classification and annotation method applied to large model mining according to claim 6, characterized in that, The aforementioned instruction fine-tuning dataset is specifically based on energy margin. Assign training loss weights to each sample routed to the auto-labeling channel or the adversarial sample cleaning channel. The calculation formula is: in, To prevent smoothing constants with denominators of zero.

9. A medical data classification and annotation system applied to large model mining, characterized in that, include: The heterogeneous data phasor measurement unit deploys multimodal data sensing probes in the medical data lake to capture raw heterogeneous data streams in real time, including electronic medical record text, image DICOM data, and vital sign time-series waveforms. Based on the raw heterogeneous data streams, feature vectors are extracted through a pre-set large model in the medical field, and the feature vectors are mapped to the Riemannian manifold space. By calculating the semantic deviation angle and semantic angular velocity relative to the standard medical knowledge graph, a semantic phase plane trajectory is constructed. The transient energy assessment engine, based on semantic phase plane trajectories, constructs a Lyapunov energy function that includes semantic kinetic energy and semantic potential energy, and calculates the instantaneous total energy of each data sample that causes a perturbation in the prediction probability distribution of the medical big language model. The stability discrimination and diversion controller, combined with the instantaneous total energy, calculates the energy margin of the data sample using the equal area rule, and compares the energy margin with the preset critical stability threshold to quantify the annotation value and cognitive risk of the data. Based on the comparison results of the energy margin and the critical stability threshold, the data stream is diverted through an adaptive excitation control strategy, and directed to the automatic annotation channel, the human expert review channel, or the adversarial sample cleaning channel, respectively, to generate a command fine-tuning dataset. The excitation feedback adjustment module executes an adaptive excitation control strategy and dynamically adjusts the parameters of a large medical model. The construction of the semantic phase plane trajectory specifically involves: Define the standard diagnosis and treatment path in the medical knowledge graph as the synchronous rotating coordinate system of the system, and use the feature vector of the data sample i to be mined. The projection phase difference in this coordinate system is defined as the semantic deviation angle. ; The information entropy of the predicted probability distribution of sample i in the medical big language model during inference is calculated as the first derivative with the number of inference steps t, and is defined as the semantic angular velocity. : in, Information entropy predicted by large models in the medical field. These are the normalization coefficients; With semantic deviation angle x-axis, semantic angular velocity Using the vertical axis as the ordinate, construct the semantic phase plane trajectory.

Images