A multi-dimensional adaptive optimization method for medical data

Abstract

The application relates to a multidimensional adaptive optimization method for medical data. The method comprises the following steps: acquiring a multidimensional feedback vector and a multidimensional scene feature vector; acquiring a multidimensional feedback vector and a multidimensional scene feature vector; constructing a dynamic weight network; training; splicing the multidimensional feedback vector and the multidimensional scene feature vector to form a joint input vector, and inputting the joint input vector into the trained dynamic weight network to obtain a dynamic weight coefficient vector alpha, and calculating a comprehensive reward value R by using the dynamic weight coefficient vector alpha; and optimizing the medical data by using the comprehensive reward value R. The application realizes multi-angle and structured measurement of the value of single human-computer interaction by using a fine feedback quantification model of the multidimensional feedback vector and the multidimensional scene feature vector which are deeply matched with the characteristics of the medical field. The application overcomes the extensive nature of traditional binary or shallow multidimensional feedback, and provides a much richer and more accurate guidance information source for engine evolution.

Description

Technical Field

[0001] This invention relates to the fields of medical information processing and artificial intelligence technology, and more specifically, to a multi-dimensional adaptive optimization method for medical data. Background Technology

[0002] In the process of medical informatization, massive amounts of clinical data are generated from different medical institutions, systems, and departments, resulting in significant heterogeneity in terminology. Accurately mapping unstructured clinical expressions (such as "chest pain") to standardized concepts in standard medical terminology systems (such as SNOMED CT and ICD) (such as "chest pain" or "angina"), i.e., semantic normalization, is a crucial foundation for achieving cross-institutional data interoperability, assisting clinical decision-making, and supporting medical research. Traditional rule-based or static dictionary-based methods struggle to address the diversity and dynamism of terminology usage.

[0003] In recent years, adaptive governance engines based on artificial intelligence, particularly human-machine collaboration and reinforcement learning frameworks, have become a research hotspot. The core idea is that the engine continuously learns through interaction with clinical users: the system provides initial terminology mapping suggestions, which users (such as doctors) confirm, select, or correct; the system then uses user feedback as evolutionary signals to optimize its internal mapping model (such as knowledge graphs or neural networks). However, existing technologies have significant shortcomings in the core link of this evolutionary loop—how ​​to quantify the value of a single human-machine interaction into a precise and reliable reward signal.

[0004] First, at the level of feedback quantification, most existing methods employ coarse-grained metrics. The most common is binary feedback, which simply judges whether the user's behavior is "accept" or "reject" the system's suggestion. This approach completely ignores the complexity of clinical decision-making. A single correction might stem from regional terminology, patient-specific factors (such as rare complications), or updates to the latest clinical guidelines. Simple negative penalties may cause the system to incorrectly weaken a reasonable mapping path within a specific context, or fail to learn from high-value corrections. A few studies have attempted to introduce multidimensional scoring, but the dimension design is often superficial (such as grammatical or semantic similarity), failing to deeply integrate key factors unique to the medical field, such as the decision-maker's professional authority, the logical consistency between the decision and the patient's complete condition, and the continuity of the decision-maker's personal historical behavioral patterns. Therefore, the signals provided by the feedback model itself are coarse and incomplete.

[0005] Secondly, regarding how to integrate feedback from multiple dimensions to form a single reward signal, existing technologies generally employ a static linear weighted summation method. This means that developers pre-set fixed weights for each dimension, using the same weight combination regardless of the decision-making scenario (e.g., emergency room vs. chronic disease management, tertiary hospital vs. community clinic). This static fusion strategy has fundamental limitations: First, medical scenarios are highly heterogeneous, with the credibility and importance of each feedback dimension varying drastically across different scenarios. For example, in an emergency room scenario, the match between the decision and the context may be crucial; while in a community general practice scenario, the experience (authority) of senior doctors may be more reliable. Static weights cannot adapt to this change, leading to a mismatch between the generated evolutionary signal and the scenario, thus guiding the model to make inefficient or even erroneous updates. Second, fixed weights rely on the prior experience of experts, lacking self-optimization capabilities and unable to learn better fusion strategies from historical interaction data. Third, the static weighting process is a "black box," making it difficult to explain why the final reward is a certain value, thus damaging the system's interpretability and clinical credibility. Summary of the Invention

[0006] Therefore, it is necessary to provide a multi-dimensional adaptive optimization method for medical data that can achieve better governance results in addressing the aforementioned technical issues.

[0007] One method for multi-dimensional adaptive optimization of medical data includes: S1: Obtain multi-dimensional feedback vectors and multi-dimensional scene feature vectors; S2: Construct a dynamic weighted network; S3: Train the dynamic weight network to obtain a trained dynamic weight network; S4: Concatenate the multi-dimensional feedback vector with the multi-dimensional scene feature vector to form a joint input vector, and input it into the trained dynamic weight network to obtain the dynamic weight coefficient vector α. Calculate the comprehensive reward value R using the dynamic weight coefficient vector α. S5: Optimize medical data using the overall reward value R.

[0008] Furthermore, in step S1, the multi-dimensional feedback vector includes: The decision correctness dimension is used to discretize and assign values ​​based on whether the user's action is to confirm the first choice, select a candidate, modify, or reject, thereby quantifying the degree of consistency between the user's decision and the system's recommendation. The context matching dimension calculates the logical matching probability between user-defined standard terminology text and patient clinical context description text through a pre-trained, fine-tuned, and probability-calibrated medical text sentence pair classification model. The user authority dimension is comprehensively quantified by a weighted formula that combines the user's static professional title level with a dynamic effectiveness score based on post-evaluation of historical feedback, where the weighting coefficients are trainable parameters. The decision history consistency dimension is quantified by calculating the average Jaccard similarity between the current decision action and the user's historical set of similar decision actions.

[0009] Furthermore, in step S1, the multi-dimensional scene feature vector includes: Data source institution type: based on medical institution level code; Specialty classification: coded according to the degree of specialization of the medical department; Task urgency: coded according to the type of medical visit; Historical consensus: Calculates the percentage of standard terms that appeared most frequently when the same department mapped the same original terminology in the past week; Scene volatility: Calculate the standard deviation of the group consensus over the past three days. The larger the standard deviation, the higher the volatility. Normalize to the [0,1] interval after the maximum and minimum values.

[0010] Furthermore, in step S2, the dynamic weight network is an attention network, with the following specific structure: The first layer is the feature fusion layer: H = tanh(W_h·X_norm+b_h), W_h∈R^{16×9}, b_h∈R^{16}; The second layer is the attention score layer: A = W_a·H+b_a, W_a∈R^{4×16}, b_a∈R^{4}; Dynamic weight generation: α = softmax(A); α is a four-dimensional vector, [α_correct, α_context, α_authority, α_consistency], and Σα_i = 1.

[0011] Furthermore, step S3 also includes constructing training data for training the dynamic weight network, specifically as follows: Obtain historical interaction logs and construct sample triples (F_base, S, R_ideal) offline from the historical interaction logs; F_base and S are recalculated based on the historical original data; the R_ideal label is obtained through a weakly supervised method based on delayed group consensus and behavioral effectiveness evaluation.

[0012] Furthermore, the sample triples are constructed offline from historical interaction logs as follows: For a historical interaction in the historical interaction log, a delayed observation window is set for a certain number of days after the interaction occurs. During this window, the mapping results of the same original terminology in all diagnosis and treatment contexts are collected from all physicians in the same department. A weighted voting mechanism is also used to calculate the delayed group consensus terminology. If the delayed group consensus terminology exists, that is, the weight of the term with the highest vote exceeds a certain proportion of the total weight, then the user decision terminology of this historical interaction is compared with the delayed group consensus terminology. If they are the same, R_ideal = +0.8 is assigned; if they are different, R_ideal = -0.5 is assigned. Key structured fields are extracted from the current patient's medical record, including: age, gender, BERT vector of the chief complaint text, and BERT vector of the first few characters of the present medical history text; these are concatenated to obtain a multi-dimensional vector; cosine similarity is used to calculate the similarity between the context feature vector of the current interaction and the context feature vector of the historical interaction records; the threshold is based on the quantile of the similarity between samples within the same class after performing K-means clustering on all historical context feature vectors to ensure intra-class similarity.

[0013] Furthermore, constructing the training data also includes: if a delayed consensus cannot be formed, an assignment strategy linked to the long-term performance trend is adopted: the proportion of all mapping decisions of a user that are supported by the subsequent delayed consensus within a certain period after the interaction occurs is statistically analyzed and denoted as P_support; then, scaling is performed based on the f_correct value of this interaction: if f_correct is +1.0, indicating confirmation of the first choice, then R_ideal = 0.5 + 0.3 * P_support; if f_correct is +0.5, indicating selection of a candidate, then R_ideal = 0.2 + 0.2 * P_support; if f_correct is -1.0, indicating correction, then R_ideal = -0.3 - 0.3 * (1 - P_support); if f_correct is 0.0, indicating rejection, then R_ideal = 0.0.

[0014] Furthermore, in step S3, specifically: Based on the training data, the parameters are jointly trained using the Adam optimizer with mean squared error as the loss function; The loss function is defined as: L(Θ,λ) = (1 / N)Σ_{i=1}^{N}(R_pred_i-R_ideal_i)^2 For the i-th sample, R_pred_i = α_i·F_base_i, and the calculation of α_i depends on the parameters Θ and input of AttnNet. During training, the gradient updates the parameters Θ and λ of AttnNet simultaneously through the backpropagation algorithm. The training set, validation set, and test set are divided in a 7:2:1 ratio. An early stopping strategy is used. Finally, the MSE of the predicted reward and R_ideal of the model on the test set is reduced to below 0.05.

[0015] Furthermore, the comprehensive reward value R is calculated using the dynamic weight coefficient vector α as follows: The comprehensive reward value R for scene perception in this human-computer interaction is calculated using the formula: R = α·F_base = α_correct * f_correct + α_context * f_context + α_authority * f_authority + α_consistency * f_consistency.

[0016] Further, in step S5, specifically: The proximal policy optimization algorithm is used as the core of the online evolver to drive the parameter updates of the underlying semantic mapping model: The problem is modeled as a Markov decision process: State s_t: The decision environment at time t, is a fused feature vector composed of three parts: Features of the original term e_t: The original term text is encoded using the BERT model, the 768-dimensional vector of the [CLS] tag is taken, and the dimension is reduced to 128 dimensions through a fully connected layer to obtain the vector v_term; Patient clinical context features c_t: compressed to 128 dimensions through a fully connected layer to obtain a vector v_context; Decision-making scenario feature vector S_t; Finally, s_t = Concat(v_term, v_context, S_t); Action a_t: The engine selects a term from a fixed subset of the standard terminology library UMLS as the final mapping output; the action space is discrete, and each action corresponds to the ID of a standard term; Policy π_θ(a_t|s_t): A probability function represented by a deep neural network with parameter θ; the policy network takes state s_t as input, passes through two hidden layers, and finally outputs the probability distribution over multiple actions through a softmax layer; Reward r_t: Scene-aware comprehensive reward R, obtained immediately after the engine executes action a_t and the user completes the interaction and makes a final decision; The online evolver continuously collects interaction data and triggers a PPO algorithm update every certain number of interactions, updating the policy network parameters θ; each update includes the following specific steps: Data collection: Using the current policy π_θ, interact with the environment to collect a batch of empirical data, which contains empirical tuples (s_t, a_t, r_t, s_{t+1}) for T=2048 time steps; Using generalized dominance estimation (GAE): A value function network V_φ(s_t) is used to estimate the state value; this network shares the input layer and the first hidden layer with the policy network, then separates it into a separate 128-dimensional hidden layer, and finally outputs a scalar state value estimate. For the entire collected trajectory, the advantage estimate Â_t is recursively calculated from back to front, where the discount factor γ=0.99 and the GAE parameter λ_gae=0.95; Policy updates are performed using the PPO-Clip objective function: The goal is to optimize the clipping objective function: L^{CLIP}(θ) = (1 / |B|) Σ_{t∈B} [ min(ratio_t * Â_t, clip(ratio_t, 1-ε, 1+ε) * Â_t ) ] Where B is a small batch (size = 64) randomly sampled from the batch, ratio_t = π_θ(a_t|s_t) / π_θ_old(a_t|s_t), π_θ_old is the old strategy when collecting data, and ε is the pruning hyperparameter, set to 0.2; Value function update: The value network parameters φ are updated by minimizing the mean squared error loss of the value function: L^{VF}(φ) = (1 / |B|) Σ_{t∈B} [ (V_φ(s_t) - V_t^{target})^2 ], where V_t^{target} = Â_t +V_φ_old(s_t); Joint optimization: The total loss function is: L^{total}(θ, φ) = E_t [ L^{CLIP}(θ) - c1 * L^{VF}(φ) +c2 * H(π_θ(·|s_t)) ], where c1=0.5, c2=0.01, and H is the policy entropy; Use the Adam optimizer to perform mini-batch iterative updates on the total loss; Evolutionary execution and interpretability record: After the PPO algorithm is updated, the new policy network π_θ is deployed; it will fully record the four-dimensional feedback vector F_base, the five-dimensional scene feature vector S, the dynamic weight vector α, the comprehensive reward R, and the difference in the output probability distribution of the current state s_t before and after the policy network update for each interaction.

[0017] This invention utilizes a refined feedback quantification model that deeply integrates multi-dimensional feedback vectors and multi-dimensional scene feature vectors with the characteristics of the medical field. This model enables a multi-faceted and structured measurement of the value of a single human-computer interaction. It overcomes the coarseness of traditional binary or shallow multi-dimensional feedback, providing a far richer and more precise source of guiding information for engine evolution. Attached Figure Description

[0018] Figure 1 Here is a flowchart of a multi-dimensional adaptive optimization method for medical data in one embodiment; Figure 2 This is a diagram of a multi-dimensional adaptive engine evolution system architecture in one embodiment; Figure 3 This is a diagram of the internal structure of the dynamic weight calculator (AttnNet) in one embodiment; Figure 4 This is a schematic diagram illustrating the dynamic weight comparison under different decision-making scenarios in one embodiment; Detailed Implementation To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] Example 1: This embodiment provides, as follows: Figure 1 The method shown is a multi-dimensional adaptive optimization method for medical data, including: S1: Obtain multi-dimensional feedback vectors and multi-dimensional scene feature vectors; S2: Construct a dynamic weighted network; S3: Train the dynamic weight network to obtain a trained dynamic weight network; S4: Concatenate the multi-dimensional feedback vector with the multi-dimensional scene feature vector to form a joint input vector, and input it into the trained dynamic weight network to obtain the dynamic weight coefficient vector α. Calculate the comprehensive reward value R using the dynamic weight coefficient vector α. S5: Optimize medical data using the overall reward value R.

[0020] This embodiment employs a refined feedback quantification model that deeply aligns with the characteristics of the medical field through multi-dimensional feedback vectors and multi-dimensional scene feature vectors. This enables a multi-faceted and structured measurement of the value of a single human-computer interaction. It overcomes the limitations of traditional binary or shallow multi-dimensional feedback, providing a far richer and more precise source of guiding information for engine evolution.

[0021] Example 2: This embodiment further discloses information based on Embodiment 1: Furthermore, in step S1, the multi-dimensional feedback vector includes: The decision correctness dimension is used to discretize and assign values ​​based on whether the user's action is to confirm the first choice, select a candidate, modify, or reject, thereby quantifying the degree of consistency between the user's decision and the system's recommendation. The context matching dimension calculates the logical matching probability between user-defined standard terminology text and patient clinical context description text through a pre-trained, fine-tuned, and probability-calibrated medical text sentence pair classification model. The user authority dimension is comprehensively quantified by a weighted formula that combines the user's static professional title level with a dynamic effectiveness score based on post-evaluation of historical feedback, where the weighting coefficients are trainable parameters. The decision history consistency dimension is quantified by calculating the average Jaccard similarity between the current decision action and the user's historical set of similar decision actions.

[0022] Furthermore, in step S1, the multi-dimensional scene feature vector includes: Data source institution type: based on medical institution level code; Specialty classification: coded according to the degree of specialization of the medical department; Task urgency: coded according to the type of medical visit; Historical consensus: Calculates the percentage of standard terms that appeared most frequently when the same department mapped the same original terminology in the past week; Scene volatility: Calculate the standard deviation of the group consensus over the past three days. The larger the standard deviation, the higher the volatility. Normalize to the [0,1] interval after the maximum and minimum values.

[0023] Furthermore, in step S2, the dynamic weight network is an attention network, with the following specific structure: The first layer is the feature fusion layer: H = tanh(W_h·X_norm+b_h), W_h∈R^{16×9}, b_h∈R^{16}; The second layer is the attention score layer: A = W_a·H+b_a, W_a∈R^{4×16}, b_a∈R^{4}; Dynamic weight generation: α = softmax(A); α is a four-dimensional vector, [α_correct, α_context, α_authority, α_consistency], and Σα_i = 1.

[0024] Furthermore, step S3 also includes constructing training data for training the dynamic weight network, specifically as follows: Obtain historical interaction logs and construct sample triples (F_base, S, R_ideal) offline from the historical interaction logs; F_base and S are recalculated based on the historical original data; the R_ideal label is obtained through a weakly supervised method based on delayed group consensus and behavioral effectiveness evaluation.

[0025] Furthermore, the sample triples are constructed offline from historical interaction logs as follows: For a historical interaction in the historical interaction log, a delayed observation window is set for a certain number of days after the interaction occurs. During this window, the mapping results of the same original terminology in all diagnosis and treatment contexts are collected from all physicians in the same department. A weighted voting mechanism is also used to calculate the delayed group consensus terminology. If the delayed group consensus terminology exists, that is, the weight of the term with the highest vote exceeds a certain proportion of the total weight, then the user decision terminology of this historical interaction is compared with the delayed group consensus terminology. If they are the same, R_ideal = +0.8 is assigned; if they are different, R_ideal = -0.5 is assigned. Key structured fields are extracted from the current patient's medical record, including: age, gender, BERT vector of the chief complaint text, and BERT vector of the first few characters of the present medical history text; these are concatenated to obtain a multi-dimensional vector; cosine similarity is used to calculate the similarity between the context feature vector of the current interaction and the context feature vector of the historical interaction records; the threshold is based on the quantile of the similarity between samples within the same class after performing K-means clustering on all historical context feature vectors to ensure intra-class similarity.

[0026] Furthermore, constructing the training data also includes: if a delayed consensus cannot be formed, an assignment strategy linked to the long-term performance trend is adopted: the proportion of all mapping decisions of a user that are supported by the subsequent delayed consensus within a certain period after the interaction occurs is statistically analyzed and denoted as P_support; then, scaling is performed based on the f_correct value of this interaction: if f_correct is +1.0, indicating confirmation of the first choice, then R_ideal = 0.5 + 0.3 * P_support; if f_correct is +0.5, indicating selection of a candidate, then R_ideal = 0.2 + 0.2 * P_support; if f_correct is -1.0, indicating correction, then R_ideal = -0.3 - 0.3 * (1 - P_support); if f_correct is 0.0, indicating rejection, then R_ideal = 0.0.

[0027] Furthermore, in step S3, specifically: Based on the training data, the parameters are jointly trained using the Adam optimizer with mean squared error as the loss function; The loss function is defined as: L(Θ,λ) = (1 / N)Σ_{i=1}^{N}(R_pred_i-R_ideal_i)^2 For the i-th sample, R_pred_i = α_i·F_base_i, and the calculation of α_i depends on the parameters Θ and input of AttnNet. During training, the gradient updates the parameters Θ and λ of AttnNet simultaneously through the backpropagation algorithm. The training set, validation set, and test set are divided in a 7:2:1 ratio. An early stopping strategy is used. Finally, the MSE of the predicted reward and R_ideal of the model on the test set is reduced to below 0.05.

[0028] Furthermore, the comprehensive reward value R is calculated using the dynamic weight coefficient vector α as follows: The comprehensive reward value R for scene perception in this human-computer interaction is calculated using the formula: R = α·F_base = α_correct * f_correct + α_context * f_context + α_authority * f_authority + α_consistency * f_consistency.

[0029] Further, in step S5, specifically: The proximal policy optimization algorithm is used as the core of the online evolver to drive the parameter updates of the underlying semantic mapping model: The problem is modeled as a Markov decision process: State s_t: The decision environment at time t, is a fused feature vector composed of three parts: Features of the original term e_t: The original term text is encoded using the BERT model, the 768-dimensional vector of the [CLS] tag is taken, and the dimension is reduced to 128 dimensions through a fully connected layer to obtain the vector v_term; Patient clinical context features c_t: compressed to 128 dimensions through a fully connected layer to obtain a vector v_context; Decision-making scenario feature vector S_t; Finally, s_t = Concat(v_term, v_context, S_t); Action a_t: The engine selects a term from a fixed subset of the standard terminology library UMLS as the final mapping output; the action space is discrete, and each action corresponds to the ID of a standard term; Policy π_θ(a_t|s_t): A probability function represented by a deep neural network with parameter θ; the policy network takes state s_t as input, passes through two hidden layers, and finally outputs the probability distribution over multiple actions through a softmax layer; Reward r_t: Scene-aware comprehensive reward R, obtained immediately after the engine executes action a_t and the user completes the interaction and makes a final decision; The online evolver continuously collects interaction data and triggers a PPO algorithm update every certain number of interactions, updating the policy network parameters θ; each update includes the following specific steps: Data collection: Using the current policy π_θ, interact with the environment to collect a batch of empirical data, which contains empirical tuples (s_t, a_t, r_t, s_{t+1}) for T=2048 time steps; Using generalized dominance estimation (GAE): A value function network V_φ(s_t) is used to estimate the state value; this network shares the input layer and the first hidden layer with the policy network, then separates it into a separate 128-dimensional hidden layer, and finally outputs a scalar state value estimate. For the entire collected trajectory, the advantage estimate Â_t is recursively calculated from back to front, where the discount factor γ=0.99 and the GAE parameter λ_gae=0.95; Policy updates are performed using the PPO-Clip objective function: The goal is to optimize the clipping objective function: L^{CLIP}(θ) = (1 / |B|) Σ_{t∈B} [ min(ratio_t * Â_t, clip(ratio_t, 1-ε, 1+ε) * Â_t ) ] Where B is a small batch (size = 64) randomly sampled from the batch, ratio_t = π_θ(a_t|s_t) / π_θ_old(a_t|s_t), π_θ_old is the old strategy when collecting data, and ε is the pruning hyperparameter, set to 0.2; Value function update: The value network parameters φ are updated by minimizing the mean squared error loss of the value function: L^{VF}(φ) = (1 / |B|) Σ_{t∈B} [ (V_φ(s_t) - V_t^{target})^2 ], where V_t^{target} = Â_t +V_φ_old(s_t); Joint optimization: The total loss function is: L^{total}(θ, φ) = E_t [ L^{CLIP}(θ) - c1 * L^{VF}(φ) +c2 * H(π_θ(·|s_t)) ], where c1=0.5, c2=0.01, and H is the policy entropy; Use the Adam optimizer to perform mini-batch iterative updates on the total loss; Evolutionary execution and interpretability record: After the PPO algorithm is updated, the new policy network π_θ is deployed; it will fully record the four-dimensional feedback vector F_base, the five-dimensional scene feature vector S, the dynamic weight vector α, the comprehensive reward R, and the difference in the output probability distribution of the current state s_t before and after the policy network update for each interaction.

[0030] This invention utilizes a refined feedback quantification model that deeply integrates multi-dimensional feedback vectors and multi-dimensional scene feature vectors with the characteristics of the medical field. This model enables a multi-faceted and structured measurement of the value of a single human-computer interaction. It overcomes the coarseness of traditional binary or shallow multi-dimensional feedback, providing a far richer and more precise source of guiding information for engine evolution.

[0031] This embodiment employs a refined feedback quantification model that deeply aligns with the characteristics of the medical field through multi-dimensional feedback vectors and multi-dimensional scene feature vectors. This enables a multi-faceted and structured measurement of the value of a single human-computer interaction. It overcomes the limitations of traditional binary or shallow multi-dimensional feedback, providing a far richer and more precise source of guiding information for engine evolution.

[0032] Example 3: This embodiment provides, as follows: Figure 2 This paper presents a multi-dimensional adaptive optimization method for medical data, including: Step 1: Refined Construction of Multi-Dimensional Feedback Vectors and Scene Feature Vectors The goal of this step is to extract and calculate a standardized four-dimensional feedback vector from a single human-computer interaction instance, and at the same time construct a quantized five-dimensional scene feature vector as input for the subsequent dynamic weight calculator.

[0033] 1.1 Calculation of the multi-dimensional feedback vector F_base: Define the feedback vector F_base as a four-dimensional vector: F_base = [f_correct, f_context, f_authority, f_consistency]. The calculation methods for each dimension are as follows: (1) Decision correctness f_correct: Used to quantify the degree of consistency between the user's final decision and the system's initial recommendation. A discrete assignment strategy based on historical behavior pattern analysis and expert experience calibration is adopted. The specific assignment rules and basis are as follows: If the user confirms the system's recommended preferred standard terminology, a value of +1.0 is assigned, indicating high approval.

[0034] If a user does not confirm their preferred term but instead selects another term from the system's candidate list (top 5 recommendations), a value of +0.5 is assigned. This value reflects partial acceptance.

[0035] If a user completely deviates from any of the system's provided candidates and manually enters or selects a completely new standard term, the value is -1.0. This assignment indicates a fundamental correction, usually stemming from a blind spot in the system's knowledge or a serious misjudgment of context, and requires strong negative incentives.

[0036] If the user believes that the original term does not require normalization mapping, or if it is a meaningless input error, and rejects all options, assign a value of 0.0. This assignment indicates an invalid task with no learning value.

[0037] The above values ​​are the system default configurations. In actual deployment, they can be used as adjustable parameters, allowing administrators to fine-tune them within the range of [-1.5, +1.5] according to specific scenarios.

[0038] (2) Context matching degree f_context: used to quantify the logical consistency between the terminology finally determined by the user and the specific clinical context of the current patient. This dimension is quantified by the classification model through a pre-trained and probability-calibrated sentence: Model pre-training: a) Corpus Preparation: Based on the data elements and dataset standards defined in the "Guangdong Provincial Public Health Data Collection Catalog (2022 Edition)" and the "Shenzhen Public Health Information Platform Data Collection Standard Specification (Trial) V3.0", structured extraction and standardized cleaning were performed on the anonymized medical records obtained from cooperating tertiary hospitals. The constructed domain corpus contains over 800,000 outpatient medical records, inpatient medical record front pages, diagnostic records, surgical operation records, etc., that conform to the above standards and specifications.

[0039] b) Model Architecture and Training: The BERT-base-Chinese model architecture was adopted. Initial pre-training was performed on a large-scale general Chinese corpus, followed by domain-adaptive pre-training on the aforementioned medical corpus. The domain pre-training task was a masked language model (MLM). Training used the AdamW optimizer (learning rate 3e-5), with a batch size of 64, and was conducted on 8 NVIDIA A10 GPUs (24GB of VRAM each) for approximately 15 epochs, totaling approximately 300,000 training steps, until the loss function value on the validation set converged.

[0040] Model fine-tuning: a) Task and Data: The pre-trained model will be fine-tuned on a medical natural language reasoning task. The fine-tuning dataset will be a combination of publicly available and authoritative medical NLI datasets, including NLI format data such as MIMIC-III, i2b2 2010, and n2c2 2018.

[0041] b) Fine-tuning method: The model input format is "[CLS] terminology text [SEP] context text [SEP]". The final hidden state of the [CLS] tag is taken, and a classification layer is used to output three classification probabilities (implication, neutral, contradictory). A hierarchical learning rate strategy is adopted: the learning rate of the top-level classifier is 1e-4, the learning rate of the last 6 layers of the BERT model is 5e-5, and the learning rate of the first 6 layers is 1e-5. Fine-tuning is performed until the validation set accuracy reaches or exceeds 85%.

[0042] Model probability calibration: a) Calibration Method: After fine-tuning, the model's output probabilities are calibrated using temperature scaling. On the calibration set (split from the fine-tuning validation set or collected from other relevant data), a single temperature parameter T (T>0) is optimized to minimize the negative log-likelihood loss. The calibrated "implied" class probability is calculated as: P_calibrated(implied|T_term, T_context) = softmax(logits / T)_implied, where logits is the model's original output score for the "implied" class. The optimized temperature parameter T is saved as a fixed model parameter.

[0043] Online calculation: When providing online services, for the term-context pair (T_term, T_context) to be evaluated, the calibrated model is input, and the calibrated probability value of the "implication" category is taken as the context matching score, i.e., f_context = P_calibrated(implication | T_term, T_context). The calibrated probability has better probabilistic meaning and cross-sample comparability.

[0044] (3) User Authority f_authority: Used to comprehensively quantify the professional credibility of feedback providers, combining their static identity and short-term behavioral performance. The calculation formula is: f_authority = λ_static * (Level / L_max) + λ_behavior * C_recent in: Level: The user's professional title level code value, which is mapped according to the "Standards for the Professional Title Levels of Health Technicians" issued by the National Health Commission. For example, resident physician = 1, attending physician = 2, associate chief physician = 3, chief physician = 4.

[0045] L_max: The highest professional title level defined in the system, set to 4.

[0046] C_recent: This is the user's recent feedback consensus score. It calculates the proportion of feedback from this user over the past month where their final decision term T_user aligns with the "system's final adopted group consensus term" T_consensus. The "group consensus term" T_consensus is calculated as follows: within the same department (dept), for the same original term S_original, the mapping results of all physicians (excluding the current user) over the past week are collected, and a weighted vote is performed. The weight is the physician's static professional title level coefficient (Level / L_max). The term with the highest number of votes exceeding 40% of the total weight is designated as T_consensus. If no consensus can be reached, T_consensus is considered empty, and this feedback is not included in the C_recent calculation. If the user has fewer than 5 valid feedback responses in the past, C_recent takes the default value of 0.5.

[0047] λ_static and λ_behavior are dynamic harmonic coefficients, which are trainable parameters, initially set to λ_static = 0.7 and λ_behavior = 0.3. These two coefficients will be optimized together with the parameters of the AttnNet attention network in step two in the same end-to-end training process.

[0048] (4) Decision history consistency f_consistency: used to measure the similarity between the current decision and the user's own historical behavior patterns. The calculation method is based on Jaccard similarity.

[0049] Jaccard similarity is defined as follows: It measures the similarity between two sets, and is expressed as the size of their intersection divided by the size of their union. The formula is: J(A,B) = |A ∩ B| / |A ∪ B|.

[0050] Calculation process: a) Let the current decision action be A, denoted as (original term S_original, mapped standard term T_std).

[0051] b) Construct the concept set Set_A for the current action: containing T_std and its direct parent and direct child concepts in a standard terminology system (such as UMLS) (depth of 1).

[0052] c) Retrieve historical actions: Extract all decision actions containing the same original term S_original from the user's historical interaction logs, forming a historical action set {A_hist}. If the number is insufficient (e.g., less than 5 times), expand to decision actions corresponding to other original terms with high semantic similarity to S_original (based on word vector cosine similarity, threshold > 0.7).

[0053] d) Calculate similarity: For each historical action A_i, construct its concept set Set_i using the same method. Calculate the Jaccard similarity J(Set_A, Set_i) between Set_A and each Set_i.

[0054] e) Calculate f_consistency: Take the arithmetic mean of all J(Set_A, Set_i). If no historical action is available, assign the default value of 0.5.

[0055] 1.2 Construction of the feature vector S for the decision-making scenario: Define the scene feature vector S as a five-dimensional vector: S = [s_source, s_dept, s_urgency, s_consensus, s_volatility]. The definitions and quantization rules for each dimension are as follows: s_source (data source institution type): based on the medical institution level code, such as community service center = 0.2, secondary hospital = 0.5, tertiary-level A hospital = 1.0.

[0056] s_dept (Specialty Classification): Coded according to the specialization of the medical department, such as general practice = 0.3, internal medicine = 0.6, and specialties such as cardiology = 1.0.

[0057] s_urgency (task urgency): coded according to the type of visit, such as regular outpatient = 0.2, scheduled hospitalization = 0.5, emergency = 1.0.

[0058] s_consensus (Historical Group Consensus): Calculates the percentage of standard terms that appear most frequently when mapping the same original term S_original within the same department over the past week. The higher the percentage, the higher the consensus. The value ranges from [0,1].

[0059] s_volatility (scenario volatility): Calculates the standard deviation of the above group consensus over the past three days. The larger the standard deviation, the higher the volatility. Normalize to the [0,1] interval after the maximum and minimum values.

[0060] Step 2: Calculation and training of dynamic weight coefficients based on attention network The core of this step is to design a dynamic weight calculator, which is implemented by a specifically defined, lightweight feedforward attention neural network, responsible for intelligently integrating the four feedback dimensions according to the scenario.

[0061] 2.1 Network Input and Preprocessing: The F_base (4-dimensional) and S (5-dimensional) vectors generated in step one are concatenated to obtain the joint input vector X = Concat(F_base, S), which has a dimension of 9. Layer normalization is then performed: X_norm = LayerNorm(X).

[0062] 2.2 Attention Network (AttnNet) Structure and Forward Propagation: AttnNet is a two-layer fully connected neural network specifically designed for dynamic weight allocation in this system, and its structure is as follows: First layer (feature fusion layer): H = tanh(W_h · X_norm + b_h), W_h ∈ R^{16×9}, b_h ∈ R^{16}.

[0063] Second layer (attention score layer): A = W_a · H + b_a, W_a ∈ R^{4×16}, b_a ∈ R^{4}.

[0064] Dynamic weight generation, such as Figure 3As shown: α = softmax(A). α is a four-dimensional vector, [α_correct, α_context, α_authority, α_consistency], and Σα_i = 1.

[0065] The results of dynamic weight generation are as follows Figure 4 As shown, through side-by-side comparison, it is clearly evident that the context matching weight is highest (0.55) in the "Tertiary Grade A Hospital - Cardiology Department - Emergency Department" scenario, while the user authority weight is highest (0.45) in the "Community Health Service Center - General Practice Department - Routine Outpatient Clinic" scenario. This comparison intuitively demonstrates the core feature of this invention: the weights dynamically adapt to the scenario. The system can intelligently adjust the fusion weights of each feedback dimension according to scenario characteristics (such as institution level, departmental specialization, task urgency, etc.).

[0066] 2.3 Network training and cloud-edge collaborative update mechanism: (1) Training data construction: Sample triples (F_base, S, R_ideal) are constructed offline from historical interaction logs. F_base and S are recalculated based on the original historical data. The R_ideal label is obtained through a weakly supervised method based on delayed group consensus and behavior effectiveness evaluation, with the following specific rules: a) Delayed Consensus Assessment (Main Method): For a historical interaction, a delayed observation window is set to 30 days after the interaction. During this window, the mapping results of the same original term S_original in all clinical contexts are collected from all physicians in the same department (same s_dept). A weighted voting mechanism is used (weight is the physician's static professional title coefficient Level / L_max) to calculate the delayed consensus term T_delayed_consensus. If T_delayed_consensus exists (i.e., the highest-voted term's weight exceeds 35% of the total weight), the user decision term T_user for this historical interaction is compared with T_delayed_consensus. If they are the same, R_ideal = +0.8 is assigned; if they are different, R_ideal = -0.5 is assigned. This method is based on the assumption that consensus formed over a longer period and with broader practice is more stable than immediate feedback or short-term consensus and can be used as an approximate truth value.

[0067] b) Definition and Similarity Calculation of Context Feature Vector c_t: To support the above evaluation and subsequent steps, the composition of the context feature vector c_t is explicitly defined here. c_t extracts key structured fields from the current patient's medical record, including: age (normalized to [0,1]), gender (one-hot encoded, 2D), BERT vector of the chief complaint text (reduced to 64D), and BERT vector of the first 200 characters of the present medical history text (reduced to 64D). These features are concatenated to obtain a vector of approximately 131 dimensions. When evaluating the "similar context" of historical interactions, cosine similarity is used to calculate the similarity between c_t of the current interaction and c_t' of historical interaction records. A threshold of 0.85 is determined based on the 95th percentile of the similarity between samples within the same cluster after performing K-means clustering on all historical c_t vectors (e.g., setting k=50) to ensure intra-cluster similarity.

[0068] c) Baseline Assignment of Feedback Behavior Validity (Auxiliary Method): If delayed consensus cannot be formed (i.e., no T_delayed_consensus), an assignment strategy linked to the long-term trend of engine performance is adopted. The proportion of all mapping decisions made by the user within one week after this interaction that are supported by subsequent (within 30 days) delayed consensus is calculated and denoted as P_support. Then, scaling is applied based on the f_correct value of this interaction: if f_correct is +1.0 (confirm preferred), then R_ideal = 0.5 + 0.3 * P_support; if f_correct is +0.5 (select candidate), then R_ideal = 0.2 + 0.2 * P_support; if f_correct is -1.0 (correction), then R_ideal = -0.3 - 0.3 * (1 - P_support); if f_correct is 0.0 (rejection), then R_ideal = 0.0. This method correlates a single action with the user's short-term validity, providing a noisy but directional monitoring signal.

[0069] (2) Model Training: In the cloud, using the aggregated historical dataset D, with mean squared error (MSE) as the loss function, the Adam optimizer (learning rate 1e-3) is used to jointly train the parameters Θ (including W_h, b_h, W_a, b_a) of AttnNet and the trainable parameters λ_static and λ_behavior in the f_authority formula. The loss function is defined as: L(Θ, λ) = (1 / N) Σ_{i=1}^{N} (R_pred_i - R_ideal_i)^2 For the i-th sample, R_pred_i = α_i · F_base_i, where the calculation of α_i depends on the AttnNet parameters Θ and the input (including f_authority_i calculated from the current λ value). During training, the gradient updates both the AttnNet parameters Θ and λ simultaneously via backpropagation. The training, validation, and test sets are divided in a 7:2:1 ratio. An early stopping strategy is used (patience value = 10). The final model's predicted reward on the test set with an MSE of R_ideal below 0.05.

[0070] (3) Cloud-Edge Collaborative Update: A "performance monitoring trigger + rolling update" mechanism is adopted. The cloud monitors the overall mapping accuracy of each edge engine instance over the past week (based on samples that can be determined by latency consensus). When the accuracy of a certain engine instance is consistently lower than the baseline by 2 percentage points for more than 3 days, or when a new version of AttnNet is trained in the cloud and its validation set performance improves by more than 5%, the update process is triggered. The update package includes model parameters Θ and λ, version number, and hash checksum. After the edge node downloads the update, it first runs in shadow mode to compare the differences between the output of the new and old models. If there are no major anomalies, it hot-swaps to the new model. At the same time, the first two versions are retained to support fast rollback. The update cycle is not fixed and can be once every 2-3 weeks.

[0071] Step 3: Scene-Aware Reward Generation and Online Engine Evolution This step utilizes the output of the first two steps to generate the core signal that drives the engine's evolution, and details the online evolution process based on the near-end policy optimization algorithm.

[0072] 3.1 Calculation of Comprehensive Rewards for Scene Awareness: The comprehensive reward value R for scene perception in this human-computer interaction is calculated using the formula: R = α · F_base = α_correct * f_correct + α_context * f_context + α_authority * f_authority + α_consistency * f_consistency. This R value is a scalar that deeply integrates information from all feedback dimensions and dynamically adjusts the weight of each dimension's contribution based on the current scene, serving as the final quantified value of this interaction.

[0073] 3.2 PPO-based online engine evolution drive: The R-value of 3.1 serves as the core immediate reward signal for the reinforcement learning agent and is input into the engine's online evolver. This invention employs a proximal policy optimization algorithm as the core of the online evolver, driving the parameter updates of the underlying semantic mapping model.

[0074] (1) The problem is modeled as a Markov decision process: State s_t: The decision environment at time t, is a fused feature vector composed of three parts: Features of the original term e_t: The original term text is encoded using the BERT model that has been fine-tuned and calibrated in step 1.1. The 768-dimensional vector of the [CLS] tag is taken and reduced to 128 dimensions through a fully connected layer to obtain the vector v_term.

[0075] Patient clinical context feature c_t: that is, the approximately 131-dimensional vector defined in step 2.3(1)b, which is then compressed to 128 dimensions through a fully connected layer to obtain the vector v_context.

[0076] The feature vector S_t of the decision-making scenario is the 5-dimensional vector calculated in step 1.2.

[0077] Finally, s_t = Concat(v_term, v_context, S_t) is a 261-dimensional vector.

[0078] Action a_t: The engine selects a term from a fixed subset of the standard terminology library UMLS (e.g., the 10,000 terms most relevant to the current department) as the final mapping output. The action space is discrete, with each action corresponding to the ID of a standard term.

[0079] Policy π_θ(a_t|s_t): A probability function represented by a deep neural network with parameter θ. This policy network takes state s_t as input, passes through two hidden layers (dimensions 256 and 128, using ReLU activation), and finally outputs a probability distribution over 10,000 actions through a softmax layer.

[0080] Reward r_t: This is the scene perception comprehensive reward R calculated by the present invention, which is obtained immediately after the engine executes action a_t (i.e. outputs mapping suggestions) and the user completes the interaction and makes a final decision.

[0081] (2) PPO algorithm update process: The online evolver continuously collects interaction data and triggers a PPO algorithm update every 1000 interactions, updating the policy network parameters θ. Each update includes the following specific steps: a) Data collection: Using the current policy π_θ, interact with the environment to collect a batch of empirical data, which contains empirical tuples (s_t, a_t, r_t, s_{t+1}) for T = 2048 time steps.

[0082] b) Dominance estimation (using generalized dominance estimation, GAE): A value function network V_φ(s_t) (with parameter φ) is used to estimate the state value. This network shares the input layer and the first hidden layer (256 dimensions) with the policy network, then separates it into a separate 128-dimensional hidden layer, and finally outputs a scalar state value estimate.

[0083] For the entire collected trajectory (length T), the advantage estimate Â_t is recursively calculated from back to front, where the discount factor γ=0.99 and the GAE parameter λ_gae=0.95.

[0084] c) Policy update (using the PPO-Clip objective function): The goal is to optimize the clipping objective function: L^{CLIP}(θ) = (1 / |B|) Σ_{t∈B} [ min(ratio_t * Â_t, clip(ratio_t, 1-ε, 1+ε) * Â_t ) ] Where B is a small batch (size = 64) randomly sampled from the batch, ratio_t = π_θ(a_t|s_t) / π_θ_old(a_t|s_t), π_θ_old is the old strategy when collecting data, and ε is the pruning hyperparameter, set to 0.2.

[0085] d) Value function update: The value network parameters φ are updated by minimizing the mean squared error loss of the value function: L^{VF}(φ) = (1 / |B|) Σ_{t∈B} [ (V_φ(s_t) - V_t^{target})^2 ], where V_t^{target} = Â_t +V_φ_old(s_t).

[0086] e) Joint optimization: The total loss function is: L^{total}(θ, φ) = E_t [ L^{CLIP}(θ) - c1 * L^{VF}(φ) +c2 * H(π_θ(·|s_t)) ], where c1=0.5, c2=0.01, and H is the policy entropy.

[0087] The total loss was updated using the Adam optimizer (policy network learning rate 3e-4, value network learning rate 1e-3) for approximately 10 mini-batch iterations.

[0088] (3) Evolutionary execution and interpretability record: After the PPO algorithm is updated, the new policy network π_θ is deployed. The system will fully record the four-dimensional feedback vector F_base, the five-dimensional scene feature vector S, the dynamic weight vector α, the comprehensive reward R, and the difference in the output probability distribution of the current state s_t before and after the policy network update for each interaction. These structured logs provide a clear attribution chain for each model evolution.

[0089] Through the closed-loop execution of the above three steps, this embodiment achieves efficient, robust, and interpretable continuous self-evolution of the medical data governance engine driven by precise quantification of multi-dimensional feedback and scenario-adaptive fusion. It employs a more robust and automatically obtainable delayed consensus and behavior effectiveness evaluation mechanism, ensuring that the system can operate normally and continuously optimize from the initial deployment stage.

[0090] This embodiment provides a refined feedback quantification model that is deeply aligned with the characteristics of the medical field: through four innovative dimensions (decision correctness, context matching, user authority, and decision history consistency), it achieves for the first time a multi-faceted and structured measurement of the value of a single human-computer interaction. This overcomes the crudeness of traditional binary or shallow multi-dimensional feedback, providing a far richer and more precise source of guiding information for engine evolution.

[0091] The system achieves contextualized intelligent fusion and adaptation of evolutionary stimulus signals: by introducing a dynamic weight calculator based on an attention neural network, the system can automatically perceive and understand the characteristics of the current decision-making scenario (such as institution type, specialty, urgency, etc.) and dynamically adjust the fusion weights of each feedback dimension. This ensures that, in a changing medical environment, the evolutionary signal can always focus on the most credible and critical feedback dimension in the current scenario, fundamentally solving the problems of scenario mismatch and signal distortion caused by static weighted fusion.

[0092] This endows the system with "meta-learning" capabilities, enhancing its autonomy: the attention neural network itself is trainable, enabling the system to learn not only "how to map terms more accurately" from feedback, but also "how to optimally balance feedback of different natures" from the overall utility of historical interactions. This iterative optimization of the learning strategy itself significantly reduces the reliance on manually setting and adjusting fusion weights, thus reducing long-term operational costs.

[0093] A complete interpretability and audit trail framework was constructed: by explicitly recording the dynamic weight vector α, the original feedback vector F_base, and the scenario vector S for each decision, the system can provide a clear attribution explanation for each update of the model parameters. This meets the stringent compliance and security requirements of auditable and traceable decision-making processes in medical AI applications, and significantly enhances clinical professionals' understanding and trust in the system's evolutionary behavior.

[0094] It significantly improves the evolutionary robustness and stability in complex conflict scenarios: When faced with seemingly contradictory feedback (e.g., expert corrections contradicting system recommendations, but the reasons for the corrections are highly consistent with the context and the experts have high authority), the dynamic fusion mechanism can rationally arbitrate based on the scenario, avoiding hasty evolutionary decisions due to extreme values ​​or fixed rules in a single dimension. This effectively prevents oscillations or degradation in model performance, ensuring the stability of the evolutionary process and the overall reliability of the final model.

Claims

1. A multi-dimensional adaptive optimization method for medical data, characterized in that, include: S1: Obtain multi-dimensional feedback vectors and multi-dimensional scene feature vectors; S2: Construct a dynamic weighted network; S3: Train the dynamic weight network to obtain a trained dynamic weight network; S4: Concatenate the multi-dimensional feedback vector with the multi-dimensional scene feature vector to form a joint input vector, and input it into the trained dynamic weight network to obtain the dynamic weight coefficient vector α. Calculate the comprehensive reward value R using the dynamic weight coefficient vector α. S5: Optimize medical data using the overall reward value R.

2. The multi-dimensional adaptive optimization method for medical data according to claim 1, characterized in that, In step S1, the multi-dimensional feedback vector includes: The decision correctness dimension is used to discretize and assign values ​​based on whether the user's action is to confirm the first choice, select a candidate, modify, or reject, thereby quantifying the degree of consistency between the user's decision and the system's recommendation. The context matching dimension calculates the logical matching probability between user-defined standard terminology text and patient clinical context description text through a pre-trained, fine-tuned, and probability-calibrated medical text sentence pair classification model. The user authority dimension is comprehensively quantified by a weighted formula that combines the user's static professional title level with a dynamic effectiveness score based on post-evaluation of historical feedback, where the weighting coefficients are trainable parameters. The decision history consistency dimension is quantified by calculating the average Jaccard similarity between the current decision action and the user's historical set of similar decision actions.

3. The multi-dimensional adaptive optimization method for medical data according to claim 1, characterized in that, In step S1, the multi-dimensional scene feature vector includes: Data source institution type: based on medical institution level code; Specialty classification: coded according to the degree of specialization of the medical department; Task urgency: coded according to the type of medical visit; Historical consensus: Calculates the percentage of standard terms that appeared most frequently when the same department mapped the same original terminology in the past week; Scene volatility: Calculate the standard deviation of the group consensus over the past three days. The larger the standard deviation, the higher the volatility. Normalize to the [0,1] interval after the maximum and minimum values.

4. The multi-dimensional adaptive optimization method for medical data according to claim 2, characterized in that, In step S2, the dynamic weight network is an attention network, with the following specific structure: The first layer is the feature fusion layer: H = tanh(W_h·X_norm+b_h), W_h∈R^{16×9}, b_h∈R^{16}; The second layer is the attention score layer: A = W_a·H+b_a, W_a∈R^{4×16}, b_a∈R^{4}; Dynamic weight generation: α = softmax(A); α is a four-dimensional vector, [α_correct, α_context, α_authority, α_consistency], and Σα_i = 1.

5. The multi-dimensional adaptive optimization method for medical data according to claim 4, characterized in that, Step S3 also includes constructing training data for training the dynamic weight network, specifically as follows: Obtain historical interaction logs and construct sample triples (F_base, S, R_ideal) offline from the historical interaction logs; F_base and S are recalculated based on the historical original data; the R_ideal label is obtained through a weakly supervised method based on delayed group consensus and behavioral effectiveness evaluation.

6. The multi-dimensional adaptive optimization method for medical data according to claim 5, characterized in that, The offline construction of sample triples from historical interaction logs is as follows: For a historical interaction in the historical interaction log, a delayed observation window is set to a certain number of days after the interaction occurs. During this window, the mapping results of the same original terminology in all diagnosis and treatment contexts are collected from all physicians in the same department. A weighted voting mechanism is also used to calculate the delayed group consensus terminology. If the delayed group consensus terminology exists, that is, the weight of the term with the highest vote exceeds a certain proportion of the total weight, then the user decision terminology of this historical interaction is compared with the delayed group consensus terminology. If they are the same, then R_ideal = +0.8 is assigned. If they are different, assign the value R_ideal = -0.5; Key structured fields are extracted from the current patient's medical record, including: age, gender, BERT vector of the chief complaint text, and BERT vector of the first few characters of the present medical history text; these are concatenated to obtain a multi-dimensional vector; cosine similarity is used to calculate the similarity between the context feature vector of the current interaction and the context feature vector of the historical interaction records; the threshold is based on the quantile of the similarity between samples within the same class after performing K-means clustering on all historical context feature vectors to ensure intra-class similarity.

7. The multi-dimensional adaptive optimization method for medical data according to claim 6, characterized in that, The construction of training data also includes: if a consensus among the delayed groups cannot be formed, an assignment strategy linked to the long-term performance trend is adopted: the proportion of all mapping decisions of a user that are supported by the subsequent consensus among the delayed groups within a certain period after the interaction occurs is counted and denoted as P_support; then, scaling is performed according to the f_correct value of this interaction: if f_correct is +1.0, i.e., confirming the first choice, then R_ideal = 0.5 + 0.3 * P_support; if f_correct is +0.5, i.e., selecting a candidate, then R_ideal = 0.2 + 0.2 * P_support; if f_correct is -1.0, i.e., correction, then R_ideal = -0.3 - 0.3 * (1 - P_support); if f_correct is 0.0, i.e., rejection, then R_ideal = 0.

0.

8. The multi-dimensional adaptive optimization method for medical data according to claim 7, characterized in that, In step S3, specifically: Based on the training data, the parameters are jointly trained using the Adam optimizer with mean squared error as the loss function; The loss function is defined as: L(Θ,λ) = (1 / N)Σ_{i=1}^{N}(R_pred_i-R_ideal_i)^2 For the i-th sample, R_pred_i = α_i·F_base_i, and the calculation of α_i depends on the parameters Θ and input of AttnNet. During training, the gradient updates the parameters Θ and λ of AttnNet simultaneously through the backpropagation algorithm. The training set, validation set, and test set are divided in a 7:2:1 ratio. An early stopping strategy is used. Finally, the MSE of the predicted reward and R_ideal of the model on the test set is reduced to below 0.

05.

9. A multi-dimensional adaptive optimization method for medical data according to claim 8, characterized in that, The comprehensive reward value R is calculated using the dynamic weight coefficient vector α as follows: The comprehensive reward value R for scene perception in this human-computer interaction is calculated using the formula: R = α·F_base = α_correct * f_correct + α_context * f_context + α_authority * f_authority + α_consistency * f_consistency.

10. A multi-dimensional adaptive optimization method for medical data according to claim 9, characterized in that, In step S5, specifically: The proximal policy optimization algorithm is used as the core of the online evolver to drive the parameter updates of the underlying semantic mapping model: The problem is modeled as a Markov decision process: State s_t: The decision environment at time t, is a fused feature vector composed of three parts: Features of the original term e_t: The original term text is encoded using the BERT model, the 768-dimensional vector of the [CLS] tag is taken, and the dimension is reduced to 128 dimensions through a fully connected layer to obtain the vector v_term; Patient clinical context features c_t: compressed to 128 dimensions through a fully connected layer to obtain the vector v_context; Decision-making scenario feature vector S_t; Finally, s_t = Concat(v_term, v_context, S_t); Action a_t: The engine selects a term from a fixed subset of the standard terminology library UMLS as the final mapping output; The action space is discrete, and each action corresponds to a standard term ID; Policy π_θ(a_t|s_t): A probability function represented by a deep neural network with parameter θ; the policy network takes state s_t as input, passes through two hidden layers, and finally outputs the probability distribution over multiple actions through a softmax layer; Reward r_t: Scene-aware comprehensive reward R, obtained immediately after the engine executes action a_t and the user completes the interaction and makes a final decision; The online evolver continuously collects interaction data and triggers a PPO algorithm update every certain number of interactions, updating the policy network parameters θ; each update includes the following specific steps: Data collection: Using the current policy π_θ, interact with the environment to collect a batch of empirical data, which contains empirical tuples (s_t, a_t, r_t, s_{t+1}) for T=2048 time steps; Using generalized dominance estimation (GAE): A value function network V_φ(s_t) is used to estimate the state value; this network shares the input layer and the first hidden layer with the policy network, then separates it into a separate 128-dimensional hidden layer, and finally outputs a scalar state value estimate. For the entire collected trajectory, the advantage estimate Â_t is recursively calculated from back to front, where the discount factor γ=0.99 and the GAE parameter λ_gae=0.95; Policy updates are performed using the PPO-Clip objective function: The goal is to optimize the clipping objective function: L^{CLIP}(θ) = (1 / |B|) Σ_{t∈B} [ min( ratio_t* Â_t, clip(ratio_t, 1-ε, 1+ε) * Â_t ) ] Where B is a small batch (size = 64) randomly sampled from the batch, ratio_t = π_θ(a_t|s_t) / π_θ_old(a_t|s_t), π_θ_old is the old strategy when collecting data, and ε is the pruning hyperparameter, set to 0.2; Value function update: The value network parameters φ are updated by minimizing the mean squared error loss of the value function: L^{VF}(φ) = (1 / |B|) Σ_{t∈B} [ (V_φ(s_t) - V_t^{target})^2 ], where V_t^{target} = Â_t + V_φ_old(s_t); Joint optimization: The total loss function is: L^{total}(θ, φ) = E_t [ L^{CLIP}(θ) - c1 * L^{VF}(φ) + c2* H(π_θ(·|s_t)) ], where c1=0.5, c2=0.01, and H is the policy entropy; Use the Adam optimizer to perform mini-batch iterative updates on the total loss; Evolutionary execution and interpretability record: After the PPO algorithm is updated, the new policy network π_θ is deployed; it will fully record the four-dimensional feedback vector F_base, the five-dimensional scene feature vector S, the dynamic weight vector α, the comprehensive reward R, and the difference in the output probability distribution of the current state s_t before and after the policy network update for each interaction.

Images