Artificial Intelligence-Based Method and System for Processing Feedback Data in Respiratory Disease Treatment

Abstract

This application provides an artificial intelligence-based method and system for processing respiratory disease treatment feedback data. By combining the high decision-making performance of a first treatment feedback matching model with the low decision-making error of a second treatment feedback matching model, the accuracy of evaluating the semantic correlation of feedback content between respiratory disease treatment feedback data is significantly improved. By introducing a training cost mechanism and dynamically adjusting the iteration number of each training sequence accordingly, the model's ability to distinguish complex feedback data is effectively enhanced, especially when dealing with counterexample data that differs significantly from the target respiratory disease treatment feedback data and is difficult to distinguish. Thus, not only is the model's convergence speed accelerated, but the iterative parameter interaction training strategy also integrates the advantages of both models. The final target treatment feedback matching model maintains high processing efficiency while significantly reducing decision error.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more specifically, to an artificial intelligence-based method and system for processing respiratory disease treatment feedback data. Background Technology

[0002] In the treatment of respiratory diseases, timely and accurate acquisition and analysis of patient feedback data are crucial for adjusting treatment plans and evaluating treatment effectiveness. However, traditional feedback data processing methods often rely on manual review and analysis, which is not only inefficient but also susceptible to subjective factors, making it difficult to quickly and accurately process large-scale and diverse feedback data.

[0003] With the rapid development of artificial intelligence technology, especially breakthroughs in natural language processing and deep learning, new approaches have been provided for the automated processing of respiratory disease treatment feedback data. Existing technologies have developed methods that utilize single models to perform semantic correlation analysis on feedback data. While these methods have improved processing efficiency to some extent, they still suffer from problems such as large model decision errors and insufficient generalization ability. Particularly when faced with complex and ever-changing respiratory disease treatment feedback data, a single model often struggles to balance efficiency and accuracy.

[0004] To address these issues, related technologies have begun exploring strategies that combine the strengths of multiple models. One effective approach is to introduce multiple treatment feedback matching models and leverage their complementarity to optimize overall performance. Specifically, the efficient decision-making performance of one model is used to quickly process large amounts of data, while the low decision-making error of another model ensures the accuracy of the evaluation results. However, how to effectively integrate these models so that they can work together to generate a more accurate target treatment feedback matching model remains a problem that needs to be solved.

[0005] Furthermore, dynamically adjusting training strategies based on feedback data of varying complexity during training to fully utilize limited computing resources and accelerate model convergence is a significant challenge currently facing the technology. Traditional methods of uniformly distributing training resources often result in insufficient model recognition capabilities for certain complex feedback data, thus impacting overall performance. Summary of the Invention

[0006] In view of the aforementioned problems, and in conjunction with the first aspect of this application, embodiments of this application provide an artificial intelligence-based method for processing respiratory disease treatment feedback data, the method comprising:

[0007] A first treatment feedback matching model and a second treatment feedback matching model are obtained. The first treatment feedback matching model and the second treatment feedback matching model are configured to determine the semantic correlation of feedback content between respiratory disease treatment feedback data. The model decision error of the first treatment feedback matching model is greater than the model decision error of the second treatment feedback matching model, and the model decision performance of the first treatment feedback matching model is greater than the model decision performance of the second treatment feedback matching model.

[0008] Multiple training sequences of respiratory disease treatment feedback data are acquired. Each training sequence of respiratory disease treatment feedback data includes target respiratory disease treatment feedback data and counterexample respiratory disease treatment feedback data corresponding to the target respiratory disease treatment feedback data. Different training costs are corresponding to the counterexample respiratory disease treatment feedback data in different training sequences of respiratory disease treatment feedback data. The training cost is used to represent the complexity index of distinguishing the target respiratory disease treatment feedback data and the corresponding counterexample respiratory disease treatment feedback data.

[0009] The number of training iterations for each respiratory disease treatment feedback data training sequence is determined, and the number of training iterations for any respiratory disease treatment feedback data training sequence is positively correlated with the training cost of the counterexample respiratory disease treatment feedback data in the corresponding respiratory disease treatment feedback data training sequence.

[0010] Based on the ascending order of training iterations, the first treatment feedback matching model is used to perform iterative parameter interaction training on the training sequence of the multiple respiratory disease treatment feedback data and the second treatment feedback matching model to generate the target treatment feedback matching model.

[0011] For example, in one possible implementation of the first aspect, the method further includes, prior to performing iterative parameter interaction training:

[0012] A first sample sequence is obtained, which includes at least one first combined sample set. Each first combined sample set includes: target respiratory disease treatment feedback data, a first positive example corresponding to the target respiratory disease treatment feedback data, and a first negative example corresponding to the target respiratory disease treatment feedback data; wherein the target respiratory disease treatment feedback data in different first combined sample sets are different.

[0013] Based on the first sample sequence, feature comparison learning is performed on the first treatment feedback matching model.

[0014] For example, in one possible implementation of the first aspect, the first treatment feedback matching model extracts semantic feature data from each of the two respiratory disease treatment feedback data sets, and uses the semantic feature matching degree between the two extracted semantic feature data sets as the semantic correlation of the feedback content between the two respiratory disease treatment feedback data sets; the second treatment feedback matching model fuses the two respiratory disease treatment feedback data sets to obtain fused respiratory disease treatment feedback data, estimates the semantic feature data of the fused respiratory disease treatment feedback data, and generates the semantic correlation of the feedback content between the two respiratory disease treatment feedback data sets based on the extracted semantic feature data.

[0015] In another aspect, embodiments of this application also provide an artificial intelligence-based respiratory disease treatment feedback data processing system, including a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the machine-readable storage medium to implement the above-described method.

[0016] Based on the above, this application embodiment significantly improves the evaluation accuracy of the semantic correlation between respiratory disease treatment feedback data by combining the efficient decision-making performance of the first treatment feedback matching model with the low decision-making error characteristics of the second treatment feedback matching model. By introducing a training cost mechanism and dynamically adjusting the iteration number of each training sequence accordingly, the model's ability to distinguish complex feedback data is effectively enhanced, especially when dealing with counterexample data that differs significantly from the target respiratory disease treatment feedback data and is difficult to distinguish. Thus, not only is the model's convergence speed accelerated, but the iterative parameter interaction training strategy also integrates the advantages of both models. The final target treatment feedback matching model maintains high processing efficiency while significantly reducing decision-making errors, providing more accurate and reliable feedback data support for adjusting respiratory disease treatment plans, thereby promoting the optimization of patient treatment outcomes and the realization of personalized medical services. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the execution flow of the artificial intelligence-based respiratory disease treatment feedback data processing method provided in the embodiments of this application.

[0018] Figure 2 This is a schematic diagram of the hardware architecture of the AI-based respiratory disease treatment feedback data processing system provided in this application embodiment. Detailed Implementation

[0019] The present application will now be described in detail with reference to the accompanying drawings. Figure 1This is a flowchart illustrating an artificial intelligence-based respiratory disease treatment feedback data processing method according to an embodiment of this application. The following is a detailed description of this artificial intelligence-based respiratory disease treatment feedback data processing method.

[0020] Step S110: Obtain a first treatment feedback matching model and a second treatment feedback matching model. The first and second treatment feedback matching models are configured to determine the semantic correlation of feedback content between respiratory disease treatment feedback data. The model decision error of the first treatment feedback matching model is greater than that of the second treatment feedback matching model, and the model decision performance of the first treatment feedback matching model is greater than that of the second treatment feedback matching model.

[0021] In this embodiment, in order to improve the accuracy of treatment feedback matching, two treatment feedback matching models are developed and optimized: a first treatment feedback matching model (model A) and a second treatment feedback matching model (model B).

[0022] Specifically, the server first connects to the data warehouse, retrieving pre-trained versions of Model A and Model B. Both Model A and Model B are designed to evaluate the semantic correlation of feedback content between respiratory disease treatment feedback data, but they differ in design and performance.

[0023] Model A (First Treatment Feedback Matching Model): Employs a deep learning architecture, using multi-layer neural networks to extract semantic features from textual data, and then calculates the correlation between respiratory disease treatment feedback data. Although Model A may produce significant errors in the decision-making process (e.g., inaccurate feedback assessment for some complex cases), it boasts high decision-making performance, enabling rapid response and processing of large amounts of data.

[0024] Model B (Second Treatment Feedback Matching Model): Employs an ensemble learning approach, combining the predictions of multiple base models (such as SVM, decision trees, etc.) and using a voting mechanism to arrive at the final correlation assessment. Model B has a relatively small decision error, but its efficiency is slightly lower than Model A when processing large-scale data. After the server acquires these two models, it loads them into memory, preparing them for subsequent iterative parameter interaction training.

[0025] Step S120: Obtain multiple training sequences of respiratory disease treatment feedback data. Each training sequence includes target respiratory disease treatment feedback data and corresponding counterexample respiratory disease treatment feedback data. Different training costs are assigned to the counterexample respiratory disease treatment feedback data in different training sequences. The training cost represents the complexity index for distinguishing between the target respiratory disease treatment feedback data and the corresponding counterexample respiratory disease treatment feedback data.

[0026] In order to train and optimize the treatment feedback matching model, the server needs to extract actual treatment feedback data from the treatment feedback database as training samples.

[0027] Specifically, the server first connects to a treatment feedback database, which contains treatment feedback records for tens of thousands of respiratory disease patients over the past few years. Each treatment feedback record details the patient's diagnosis, treatment plan, and feedback on treatment effectiveness.

[0028] The server randomly selects a certain number of target respiratory disease treatment feedback data from the database, or according to a specific strategy (such as by disease type, treatment period, etc.). These target respiratory disease treatment feedback data will be used as positive examples in each training sequence for model learning and evaluation.

[0029] For each target respiratory disease treatment feedback data set, the server needs to find corresponding counterexample respiratory disease treatment feedback data, i.e., records that differ significantly from the target respiratory disease treatment feedback data in terms of treatment content or feedback semantics. To generate high-quality counterexample respiratory disease treatment feedback data, Model A (a model with relatively good initial performance) is used to perform a preliminary evaluation of each record in the treatment feedback database, calculating the semantic correlation between their feedback content and the target respiratory disease treatment feedback data. Based on the correlation score, records are sorted from high to low, and the top X records with the lowest scores are selected as counterexample respiratory disease treatment feedback data. Here, X is a preset positive integer representing the number of counterexamples corresponding to each target respiratory disease treatment feedback data set.

[0030] This involves assigning a training cost to each counterexample respiratory disease treatment feedback data point. The training cost is dynamically calculated based on factors such as the semantic differences between the counterexample and target respiratory disease treatment feedback data, and the complexity of the counterexample data. Greater differences and higher complexity result in higher training costs.

[0031] The server combines each target respiratory disease treatment feedback data and its corresponding counterexample respiratory disease treatment feedback data into a training sequence. Each training sequence contains a clear target (positive example) and a disturbance item (counterexample) that needs to be distinguished, as well as corresponding training cost information.

[0032] Step S130: Determine the number of training iterations for each respiratory disease treatment feedback data training sequence. The number of training iterations for any respiratory disease treatment feedback data training sequence is positively correlated with the training cost of the counterexample respiratory disease treatment feedback data in the corresponding respiratory disease treatment feedback data training sequence.

[0033] In this embodiment, during the iterative parameter interaction training process, different training sequences require different training iterations to achieve the best training effect due to the complexity of their counterexample respiratory disease treatment feedback data and the different training costs.

[0034] In detail, the server iterates through all training sequences, determining the number of training iterations based on the training cost of the counterexample respiratory disease treatment feedback data in each sequence. Specifically, for each training sequence, the server first calculates the average training cost of all its counterexample respiratory disease treatment feedback data. Based on the average training cost, the server assigns a different number of training iterations to each training sequence. Higher costs result in more iterations, ensuring the model can fully learn and distinguish complex counterexample respiratory disease treatment feedback data. Finally, the server sorts all training sequences in ascending order of iteration count, preparing them for subsequent iterative parameter interaction training.

[0035] Step S140: Based on the ascending order of training iterations, the first treatment feedback matching model is used to perform iterative parameter interaction training based on the training sequence of the multiple respiratory disease treatment feedback data and the second treatment feedback matching model to generate the target treatment feedback matching model.

[0036] In this embodiment, after determining the number of training iterations for each training sequence, the server begins to execute an iterative parameter interaction training process, aiming to combine the efficiency of model A and the accuracy of model B to generate a higher-performing target therapy feedback matching model.

[0037] In detail, the server first initializes an empty target treatment feedback matching model (model C), which will be gradually built based on the interactive training results of model A and model B.

[0038] Therefore, the server selects the training sequence with the fewest training iterations and uses Model A to perform a preliminary evaluation of the target respiratory disease treatment feedback data and counterexample respiratory disease treatment feedback data in this sequence, generating the first semantic association of the feedback content. Simultaneously, Model B performs the same evaluation to generate the second semantic association of the feedback content.

[0039] Next, the server compares the evaluation results of model A and model B, calculating the error between them. Based on the error information, the server optimizes the weights and bias parameters of model A to make it closer to the accurate evaluation result of model B. The optimized model A (or some of its parameters) will then be used to update the target model C.

[0040] The server processes the remaining training sequences in ascending order of training iteration count. During the processing of each sequence, the initial evaluation, error calculation, and parameter optimization steps described above are repeated. As the iterations progress, the target model C gradually integrates the efficient processing capabilities of model A and the accurate evaluation capabilities of model B.

[0041] After multiple rounds of iterative training, the server evaluates the performance of the target model C. If the model performance reaches the preset standard or the convergence trend is obvious, training stops; otherwise, the training strategy is adjusted based on the latest evaluation results, and iterative training continues until the requirements are met.

[0042] Once the target model C has been trained and validated, the server deploys it to the actual treatment feedback matching system. Doctors and patients can use this system to quickly and accurately obtain semantic relevance assessment results of treatment feedback, providing strong support for adjusting treatment plans for respiratory diseases.

[0043] Based on the above steps, by combining the advantages of the two initial treatment feedback matching models, a higher-performing targeted treatment feedback matching model was generated. The application of this targeted treatment feedback matching model will significantly improve the accuracy and efficiency of treatment feedback matching in medical institutions.

[0044] Based on the above steps, this embodiment of the application significantly improves the evaluation accuracy of the semantic correlation between respiratory disease treatment feedback data by combining the efficient decision-making performance of the first treatment feedback matching model with the low decision-making error characteristics of the second treatment feedback matching model. By introducing a training cost mechanism and dynamically adjusting the iteration number of each training sequence accordingly, the model's ability to distinguish complex feedback data is effectively enhanced, especially when dealing with counterexample data that differs significantly from the target respiratory disease treatment feedback data and is difficult to distinguish. Therefore, not only is the model's convergence speed accelerated, but the advantages of both models are also integrated through an iterative parameter interaction training strategy. The final target treatment feedback matching model maintains high processing efficiency while significantly reducing decision-making errors, providing more accurate and reliable feedback data support for adjusting respiratory disease treatment plans, thereby promoting the optimization of patient treatment outcomes and the realization of personalized medical services.

[0045] In one possible implementation, step S120 includes:

[0046] Step S121: Obtain treatment feedback data and a treatment feedback database for the target respiratory disease, wherein the treatment feedback database includes multiple treatment feedback samples.

[0047] In this embodiment, the server first connects to the medical institution's data warehouse to retrieve the latest treatment feedback data for the target respiratory disease. This treatment feedback data for the target respiratory disease is typically patient treatment feedback records collected recently, containing detailed diagnostic information, treatment plans, and feedback on treatment effects.

[0048] Simultaneously, the server is connected to a treatment feedback database, which stores treatment feedback samples from tens of thousands of respiratory disease patients over the past few years. These treatment feedback samples serve as a source of potential counterexample respiratory disease treatment feedback data, used to train and optimize the treatment feedback matching model.

[0049] Step S122: The target treatment feedback matching model is invoked to estimate the semantic correlation of the feedback content between the target respiratory disease treatment feedback data and each treatment feedback sample in the treatment feedback database, generating the target feedback content semantic correlation for each treatment feedback sample. The target feedback content semantic correlation for any treatment feedback sample is positively correlated with the training cost of the corresponding treatment feedback sample.

[0050] In this embodiment, the server selects an initial target treatment feedback matching model (which can be model A or model B; in this example, it is assumed that model A is used as the initial estimation model). Then, the server iterates through each treatment feedback sample in the treatment feedback database and uses model A to estimate the semantic association of the feedback content between each treatment feedback sample and the target respiratory disease treatment feedback data.

[0051] In detail, the server loads model A into memory. For each treatment feedback sample in the treatment feedback database, the server extracts its text content and passes it to model A for processing. Model A extracts the semantic features of the sample through a multi-layer neural network and compares them with the semantic features of the target respiratory disease treatment feedback data to calculate the semantic relevance score between the two. The server records the relevance score corresponding to each treatment feedback sample, which serves as the "semantic relevance of the target feedback content" for that sample.

[0052] Step S123: Based on the descending order of the semantic relevance of the target feedback content corresponding to each treatment feedback sample, extract X treatment feedback samples from the treatment feedback database as counterexample respiratory disease treatment feedback data corresponding to the target respiratory disease treatment feedback data, where X is a positive integer.

[0053] After calculating the correlation score between each treatment feedback sample and the target respiratory disease treatment feedback data, the server needs to assign a training cost to each treatment feedback sample. The training cost represents the complexity of the model in distinguishing between target samples and negative samples.

[0054] The server initially assesses the complexity and differentiation difficulty of samples based on their relevance scores. Generally, samples with lower relevance scores (i.e., greater differences from the target sample) incur higher training costs. The server may also consider other factors, such as the text length of the sample, the number of technical terms included, and the degree of semantic ambiguity, to dynamically adjust the training cost. Ultimately, each treatment feedback sample is assigned a specific training cost for subsequent training sequence construction and iteration allocation.

[0055] The server extracts a certain number of samples from the treatment feedback database as counterexample respiratory disease treatment feedback data corresponding to the target respiratory disease treatment feedback data, based on the descending order of the semantic relevance of the target feedback content corresponding to each treatment feedback sample.

[0056] In detail, the server first sorts the samples in the treatment feedback database from lowest to highest according to their relevance scores. Then, the server selects the X lowest-scoring treatment feedback samples as negative examples of respiratory disease treatment feedback data. Here, X is a preset positive integer representing the number of negative examples corresponding to each target sample. If the number of samples in the treatment feedback database is much larger than X, the server only needs to select the top X treatment feedback samples; if the number of samples is less than X, all available samples are selected.

[0057] Step S124: Based on the target respiratory disease treatment feedback data and the extracted X counterexample respiratory disease treatment feedback data, configure multiple respiratory disease treatment feedback data training sequences.

[0058] In this embodiment, the server combines each target respiratory disease treatment feedback data and its corresponding X counterexample respiratory disease treatment feedback data into a training sequence. These training sequences will be used in subsequent iterative parameter interaction training processes.

[0059] Specifically, for each target respiratory disease treatment feedback data point, the server marks it as a positive example. The previously extracted X negative respiratory disease treatment feedback data points are combined with this positive example data point to form a training sequence. Each training sequence contains a clearly defined target (positive example), a disturbance term to be distinguished (negative example), and training cost information corresponding to each negative respiratory disease treatment feedback data point. The server stores all training sequences, ready for subsequent iterative training.

[0060] Through the steps described above, the server retrieves multiple training sequences from the treatment feedback database, each containing treatment feedback data for the target respiratory disease and corresponding counterexample respiratory disease treatment feedback data. These training sequences will be used in subsequent iterative parameter interaction training processes to optimize the performance of the treatment feedback matching model.

[0061] In one possible implementation, when the target treatment feedback matching model is the first treatment feedback matching model, step S124 includes:

[0062] Step S1241: Sort the X counterexample respiratory disease treatment feedback data according to the descending order of the semantic relevance of the target feedback content corresponding to the extracted X counterexample respiratory disease treatment feedback data to generate the target respiratory disease treatment feedback data distribution.

[0063] In this embodiment, within a medical institution, the server is using a first treatment feedback matching model (Model A) as the initial target treatment feedback matching model to construct and optimize a model for respiratory disease treatment feedback matching. This scenario details how, when the target model is Model A, the server configures multiple training sequences based on target respiratory disease treatment feedback data and extracted counterexample respiratory disease treatment feedback data.

[0064] Specifically, based on the initial assessment of Model A, the server has extracted X counterexample respiratory disease treatment feedback data from the treatment feedback database, which have the lowest correlation with the target respiratory disease treatment feedback data. Next, the server needs to sort these counterexample respiratory disease treatment feedback data according to the semantic relevance of the target feedback content (generated by Model A).

[0065] The server iterates through the X counterexamples of respiratory disease treatment feedback data, obtaining a semantic relevance score for the target feedback content for each counterexample. It then sorts these counterexamples from lowest to highest score (i.e., from weakest to strongest relevance). A lower score indicates a greater difference from the target data, making it more challenging for model training. After sorting, the server stores these counterexamples in a new order, generating a distribution of the target respiratory disease treatment feedback data. This distribution reflects the difficulty level of the counterexamples from the perspective of model A.

[0066] Step S1242: Estimate the semantic correlation of the feedback content between the target respiratory disease treatment feedback data and each counterexample respiratory disease treatment feedback data using the second treatment feedback matching model, and generate candidate semantic correlation of the feedback content corresponding to each counterexample respiratory disease treatment feedback data.

[0067] Step S1243: Sort the X counterexample respiratory disease treatment feedback data according to the descending order of the semantic relevance of the candidate feedback content corresponding to each counterexample respiratory disease treatment feedback data, and generate a candidate respiratory disease treatment feedback data distribution.

[0068] Step S1244: Perform multiple rounds of data sampling in both the target respiratory disease treatment feedback data distribution and the candidate respiratory disease treatment feedback data distribution, according to the reverse sampling direction. A respiratory disease treatment feedback data training sequence includes Y counterexample respiratory disease treatment feedback data, where Y is a positive integer. One round of data sampling is used to extract Z counterexample respiratory disease treatment feedback data from both the target and candidate respiratory disease treatment feedback data distributions, where Z ∈ [Y / 2, X].

[0069] Step S1245: After the m-th round of data sampling, based on the Z counterexample respiratory disease treatment feedback data extracted from the target respiratory disease treatment feedback data distribution and the candidate respiratory disease treatment feedback data distribution in this round, configure the counterexample respiratory disease treatment feedback data list corresponding to the m-th training iteration number.

[0070] Step S1246: Based on the target respiratory disease treatment feedback data and the list of counterexample respiratory disease treatment feedback data configured in this round, configure the respiratory disease treatment feedback data training sequence for the m-th training iteration, where m is a positive integer.

[0071] To introduce diversity and potentially uncover features that Model A failed to capture, the server then uses a second treatment feedback matching model (Model B) to reassess the semantic association of the feedback content between the target respiratory disease treatment feedback data and each counterexample respiratory disease treatment feedback data.

[0072] Specifically, the server loads Model B into memory. For each counterexample respiratory disease treatment feedback data in the sorted dataset, the server uses Model B to calculate a semantic relevance score between the counterexample and target respiratory disease treatment feedback data. Based on these scores, the server sorts the counterexample respiratory disease treatment feedback data again, this time based on the evaluation results of Model B. After sorting, a distribution of candidate respiratory disease treatment feedback data is generated.

[0073] Next, the server now has two sorted distributions of negative example respiratory disease treatment feedback data: one is the target distribution based on model A, and the other is the candidate distribution based on model B. To combine the advantages of both models, the server will alternately backsample from these two distributions to construct multiple training sequences.

[0074] Specifically, the server sets the number Y of counterexample respiratory disease treatment feedback data that each training sequence should include, and determines the number of samples Z per round, where Z ranges from [Y / 2, X).

[0075] In each round of sampling, the server first selects Z counterexamples of respiratory disease treatment feedback data from the target distribution, and then selects Z counterexamples of respiratory disease treatment feedback data from the candidate distribution. This inverse sampling method ensures that the training sequence includes both samples that model A considers most challenging and samples that model B may have different views on.

[0076] When Z is greater than Y / 2, the server will randomly select a subset of data from each of the two distributions to ensure that each training sequence ultimately contains Y counterexamples of respiratory disease treatment feedback data. For example, if Y=6 and Z=5, the server may select 3 from the target distribution, 3 from the candidate distribution, and then randomly discard one, or select according to other strategies.

[0077] After each round of sampling, the server configures a list of counterexample respiratory disease treatment feedback data corresponding to the number of training iterations based on the counterexample respiratory disease treatment feedback data extracted in that round. Then, the server combines this list with the target respiratory disease treatment feedback data to generate a complete training sequence.

[0078] Specifically, after the m-th round of sampling (where m is a positive integer), the server has Z counterexample respiratory disease treatment feedback data extracted from the two distributions. The server organizes these data into a list, labeled as the counterexample respiratory disease treatment feedback data list corresponding to the m-th training iteration.

[0079] Next, the server combines this list with the target respiratory disease treatment feedback data (as positive examples) to form a training sequence containing Y+1 samples (Y negative examples + 1 positive example).

[0080] The server stores this training sequence, ready for use in subsequent iterative parameter interaction training processes. As m increases, the server repeats the above sampling and configuration process until all the required training sequences are generated.

[0081] Through the above steps, the server is configured with multiple training sequences of respiratory disease treatment feedback data. These training sequences combine the evaluation results of model A and model B, aiming to optimize the final target treatment feedback matching model through iterative parameter interaction training.

[0082] In one possible implementation, when Z is greater than Y / 2, step S1245 includes:

[0083] Step S1245-1: In this round, from the Z counterexample respiratory disease treatment feedback data extracted from the target respiratory disease treatment feedback data distribution, Q counterexample respiratory disease treatment feedback data are randomly selected.

[0084] Step S1245-2: From the Z counterexample respiratory disease treatment feedback data extracted from the candidate respiratory disease treatment feedback data distribution in this round, W counterexample respiratory disease treatment feedback data are randomly selected. Where Q and W are both positive integers, and Q + W = Y.

[0085] Step S1245-3: Based on the Q counterexample respiratory disease treatment feedback data and the W counterexample respiratory disease treatment feedback data, configure the counterexample respiratory disease treatment feedback data list corresponding to the m-th training iteration number.

[0086] During the construction of training sequences for respiratory disease treatment feedback data, the server needs to configure a list of counterexample data based on the results of each round of sampling. When the number of Z counterexample data extracted from the target respiratory disease treatment feedback data distribution and the candidate respiratory disease treatment feedback data distribution is greater than half the number of counterexample data required for the training sequence (i.e., Z>Y / 2), the server needs to adopt a specific strategy to select the counterexample data ultimately used for the training sequence.

[0087] In this embodiment, the server has completed the sampling from the target respiratory disease treatment feedback data distribution and extracted Z counterexample respiratory disease treatment feedback data. Since Z is greater than Y / 2, the server needs to randomly select a portion of these data to ensure that the total number of counterexample data in the final training sequence is Y.

[0088] In detail, the server first calculates the number of negative examples, Q, that need to be extracted from the target distribution. Since Q + W = Y, and given Z (the total number of negative examples extracted from the target distribution) and Y (the total number of negative examples required for the training sequence), the server can calculate Q = Y - W (although W is not yet determined at this point, the server can assume a reasonable starting value or range and adjust it later). However, in practice, since Z is already given and greater than Y / 2, a more direct approach is for the server to determine a Q value that is ≤ Z and as close as possible to Y / 2 (to maintain a balance between the contributions of the two distributions), while ensuring that Q + W = Y.

[0089] The server uses a random number generator or a similar algorithm to randomly select Q data points from Z negative examples extracted from the target distribution. These selected data points will be used for subsequent training sequences.

[0090] Similar to extracting data from the target distribution, the server also needs to randomly select a subset of counterexample data from the candidate respiratory disease treatment feedback data distribution. However, this selection is based on the previously determined Q-value and the total number of counterexample data Y required for the training sequence.

[0091] In detail, once the Q value is determined, the server can calculate the number of counterexamples W that need to be extracted from the candidate distribution, i.e., W = Y - Q.

[0092] The server then uses a random number generator or similar algorithm to randomly select W data points from the Z counterexamples extracted from the candidate distribution. These selected data points, along with Q data points extracted from the target distribution, constitute the final list of counterexamples used for training sequences.

[0093] After randomly selecting Q and W counterexamples from two distributions respectively, the server combines these data to form a list of counterexample respiratory disease treatment feedback data for the m-th training iteration. Specifically, the server merges the Q counterexamples from the target distribution and the W counterexamples from the candidate distributions into one list. This list now contains Y counterexamples, exactly the number required for the training sequence. The server configures this list, along with the target respiratory disease treatment feedback data (as positive examples), into the complete training sequence for the m-th training iteration. Once the training sequence is ready, the server can use it for subsequent iterative parameter interaction training processes.

[0094] Through the above steps, when Z is greater than Y / 2, the server configures a list of counterexample respiratory disease treatment feedback data corresponding to the m-th training iteration, ensuring the diversity and balance of counterexample data in the training sequence.

[0095] In one possible implementation, the target respiratory disease treatment feedback data is configured with positive respiratory disease treatment feedback data.

[0096] Step S123 includes:

[0097] Step S1231: Based on the positive respiratory disease treatment feedback data corresponding to the target respiratory disease treatment feedback data, optimize the treatment feedback database to generate an optimized treatment feedback database. The data optimization includes removing treatment feedback samples from the treatment feedback database that are identical to the positive respiratory disease treatment feedback data.

[0098] Step S1232: Based on the descending order of the semantic relevance of the target feedback content corresponding to each treatment feedback sample, extract X treatment feedback samples from the optimized treatment feedback database as counterexample respiratory disease treatment feedback data corresponding to the target respiratory disease treatment feedback data.

[0099] In constructing training sequences for respiratory disease treatment feedback data, the server needs to ensure that there are significant differences between the target respiratory disease treatment feedback data (including positive feedback) and its corresponding negative feedback data, so that the model can effectively learn how to distinguish them. To achieve this, the server first optimizes the treatment feedback database to remove samples that are duplicates or too similar to the target positive data, and then extracts negative feedback data from the optimized database.

[0100] Before extracting negative example data, the server has already identified the target respiratory disease treatment feedback data and configured corresponding positive example respiratory disease treatment feedback data for it. Positive example data typically provides positive feedback on the treatment effect of the target disease and is used as positive examples for model learning during training.

[0101] The server will then perform data optimization on the entire treatment feedback database in preparation for extracting negative example data. The main purpose of optimization is to remove samples that are identical to or highly similar to the target positive example data, as these samples are of no value in distinguishing between positive and negative examples.

[0102] In detail, the server first loads the target respiratory disease treatment feedback data and its corresponding positive examples into memory. Then, the server iterates through each sample in the treatment feedback database, using text similarity algorithms (such as cosine similarity, Jaccard similarity, etc.) to calculate a similarity score between each sample and the target positive examples. For samples with similarity scores exceeding a preset threshold, the server marks them as "duplicate" or "too similar" and removes them from the database. These removed samples are no longer involved in the subsequent negative example data extraction process.

[0103] After the above processing, the server generates an optimized treatment feedback database, in which the samples differ significantly from the target positive data.

[0104] After optimizing the treatment feedback database, the server can now extract counterexample data corresponding to the target respiratory disease treatment feedback data. Counterexample data typically represents negative feedback on treatment effectiveness or feedback related to other diseases, serving as negative examples for model learning during training. Specifically, the server uses the previously selected target treatment feedback matching model (e.g., Model A) to evaluate the semantic relevance of each sample in the optimized database to the target respiratory disease treatment feedback data. The server sorts the samples in the database according to their relevance scores in descending order. Samples with lower scores indicate greater differences from the target data and are more likely to be selected as counterexamples. The server extracts the top X samples from the sorted database as counterexample respiratory disease treatment feedback data. Here, X is a predefined positive integer representing the number of counterexamples corresponding to each target data point. These extracted counterexamples, along with the target respiratory disease treatment feedback data (including positive feedback), form part of the training sequence for subsequent iterative parameter interaction training.

[0105] Through the steps described above, the server is configured with positive and negative examples of treatment feedback data for the target respiratory disease. These negative examples will serve as a crucial component of the training sequence, helping the model learn how to accurately distinguish between good and bad treatment effects, thereby improving the accuracy of respiratory disease treatment feedback matching.

[0106] In one possible implementation, the training sequence of respiratory disease treatment feedback data with a training iteration number of m is represented as the m-th respiratory disease treatment feedback data training sequence, where m is a positive integer.

[0107] The method of using the first treatment feedback matching model to train the m-th respiratory disease treatment feedback data training sequence and the second treatment feedback matching model to perform the m-th round of parameter interaction training includes:

[0108] Step A110: Estimate the semantic correlation of feedback content between the target respiratory disease treatment feedback data and the counterexample respiratory disease treatment feedback data in the m-th respiratory disease treatment feedback data training sequence using the first treatment feedback matching model, and generate the first feedback content semantic correlation.

[0109] Step A120: Estimate the semantic correlation of feedback content between the target respiratory disease treatment feedback data and the counterexample respiratory disease treatment feedback data in the training sequence of the m-th respiratory disease treatment feedback data using the second treatment feedback matching model, and generate the second feedback content semantic correlation.

[0110] Step A130: Based on the error between the semantic relevance of the first feedback content and the semantic relevance of the second feedback content, optimize the weight information and bias information of the first treatment feedback matching model.

[0111] In this embodiment, during the construction and optimization of the respiratory disease treatment feedback matching model, the server gradually improves the model's performance by combining the advantages of the first treatment feedback matching model (Model A) and the second treatment feedback matching model (Model B) through iterative parameter interaction training. This scenario will detail how the server utilizes these two models for the m-th round of parameter interaction training.

[0112] Before starting the m-th round of parameter interaction training, the server has prepared a training sequence of respiratory disease treatment feedback data with m training iterations, denoted as the m-th training sequence. This sequence contains the target respiratory disease treatment feedback data (positive examples) and the corresponding negative respiratory disease treatment feedback data.

[0113] The server first loads the first treatment feedback matching model (Model A) into memory and prepares to use it to evaluate the semantic association of the feedback content between the target data and the counterexample data in the m-th training sequence.

[0114] Specifically, the server iterates through each negative example in the m-th training sequence. For each negative example, the server inputs it along with the target data into model A. Model A extracts semantic features from the text data using its multi-layer neural network structure and calculates the semantic relevance score between the target data and the negative example. The server records the first semantic relevance score for each negative example.

[0115] Next, the server loads the second treatment feedback matching model (Model B) into memory and uses it to re-evaluate the association between the target data and the counterexample data in the m-th training sequence to generate a second set of association scores. Specifically, similar to using Model A, the server iterates through each counterexample data in the m-th training sequence. For each counterexample data, the server inputs it along with the target data into Model B. Model B calculates the feedback content semantic association score between the target data and the counterexample data using its ensemble learning method (combining multiple base models). The server records the second feedback content semantic association score corresponding to each counterexample data.

[0116] After obtaining two sets of relevance scores (the first set and the second set), the server optimizes the weights and biases of the first treatment feedback matching model (Model A) based on the error between these two sets of scores. Specifically, the server calculates the error between the semantic relevance of the first feedback content and the semantic relevance of the second feedback content for each counterexample. The error can be the absolute difference, the squared difference, or other forms of difference measure between the two. For all counterexamples in the entire training sequence, the server summarizes these error values ​​and evaluates the performance of Model A based on the sum or average of the errors. Based on the error information, the server uses backpropagation or other optimization algorithms to adjust the weights and biases of Model A. The goal is to reduce the difference between the predictions of Model A (the first set of relevance scores) and the predictions of Model B (the second set of relevance scores). After multiple rounds of iterative adjustments, the server evaluates whether the performance of Model A has improved. If the improvement is significant, the next round of training continues; if the improvement is not significant or the preset training round limit has been reached, training stops and the optimized Model A is saved as the target treatment feedback matching model.

[0117] Through the above steps, the server completed the m-th round of parameter interaction training, further optimizing the performance of the first treatment feedback matching model (Model A). This process will be repeated iteratively until the model performance reaches a satisfactory level.

[0118] In one possible implementation, the counterexample respiratory disease treatment feedback data in the m-th respiratory disease treatment feedback data training sequence includes at least two, and each counterexample respiratory disease treatment feedback data in the m-th respiratory disease treatment feedback data training sequence corresponds to a first feedback content semantic association and a second feedback content semantic association.

[0119] Step A130 includes:

[0120] Step A131: Poll the counterexample respiratory disease treatment feedback data in the training sequence of the m-th respiratory disease treatment feedback data. Based on the training cost of the k-th counterexample respiratory disease treatment feedback data in this round of polling, assign a training importance factor to the corresponding counterexample respiratory disease treatment feedback data. The training importance factor is positively correlated with the training cost. k is a positive integer and is not greater than the number of counterexample respiratory disease treatment feedback data in the training sequence of the m-th respiratory disease treatment feedback data.

[0121] Step A132: Based on the error between the semantic correlation of the first feedback content and the semantic correlation of the second feedback content corresponding to the kth counterexample respiratory disease treatment feedback data, determine the parameter interaction training error parameter corresponding to the kth counterexample respiratory disease treatment feedback data.

[0122] Step A133: After all the counterexample respiratory disease treatment feedback data have been polled, the training error parameters of the first treatment feedback matching model are calculated based on the training importance factors of each counterexample respiratory disease treatment feedback data and the corresponding parameter interaction training error parameters.

[0123] Step A134: Based on the optimization objective of minimizing the training error parameters, optimize the weight information and bias information of the first treatment feedback matching model.

[0124] During the iterative parameter interaction training of the respiratory disease treatment feedback matching model, the server needs to meticulously process the counterexample data in each training sequence to more effectively optimize the model's weight and bias information. This scenario will detail how the server optimizes the first treatment feedback matching model (Model A) based on the error between the semantic relevance of the first and second feedback contents, as well as the training cost of the counterexample data.

[0125] In the m-th training sequence of respiratory disease treatment feedback data, there are at least two counterexample respiratory disease treatment feedback data. The server has used model A and model B to calculate the first and second semantic relevance scores between these counterexample data and the target data, respectively.

[0126] The server begins polling each negative example in the m-th training sequence. For each negative example (k-th), the server assigns a training importance factor based on its training cost.

[0127] In detail, the server maintains a record containing the training cost and an empty training importance factor field for each negative example. During polling, the server reads the training cost of the k-th negative example and calculates its training importance factor according to preset rules (such as linear mapping, logarithmic mapping, etc.). The training importance factor is positively correlated with the training cost; that is, the higher the training cost, the larger the training importance factor. The server stores the calculated training importance factor in the record corresponding to the negative example.

[0128] For each polled negative example (the k-th example), the server calculates the error between the semantic relevance scores of the first and second feedback content to determine the error contribution of that negative example during the parameter interaction training process. Specifically, the server reads the semantic relevance scores of the first and second feedback content corresponding to the k-th negative example. It calculates the error between the two scores, which can be in the form of absolute error, squared error, etc. The calculated error value is stored as the parameter interaction training error parameter corresponding to the k-th negative example.

[0129] After all negative examples have been polled and assigned training importance factors, and the parameter interaction training error parameters have been determined, the server calculates the overall training error parameters of the first treatment feedback matching model (Model A). Specifically, the server iterates through all negative example records, reading the training importance factor and parameter interaction training error parameters for each negative example. The overall training error parameters are calculated using a weighted summation method. Specifically, the parameter interaction training error parameter for each negative example is multiplied by its training importance factor, and all products are then summed. This sum is the overall training error parameter of Model A during the m-th round of parameter interaction training.

[0130] Finally, the server adjusts the weights and biases of model A based on the optimization objective of minimizing the overall training error parameters. Specifically, the server uses backpropagation or other optimization algorithms to optimize the parameters of model A. During backpropagation, the overall training error parameters are used as input to the loss function. The algorithm updates the weights and biases of model A using gradient descent or other optimization strategies to reduce the value of the overall training error parameters. After multiple rounds of iterative optimization, the server evaluates whether the performance of model A has improved. If the improvement is significant, the next round of training continues; if the improvement is not significant or the preset training conditions (such as maximum number of iterations, minimum error threshold, etc.) have been reached, training stops and the optimized model A is saved.

[0131] Through the steps described above, the server optimizes the weight and bias information of the first treatment feedback matching model (Model A) based on the error between the semantic relevance of the first and second feedback contents and the training cost of the counterexample data. This process is the core of iterative parameter interaction training and is crucial for improving the model's performance.

[0132] In one possible implementation, the method further includes, prior to performing iterative parameter interaction training:

[0133] Step B110: Obtain a first sample sequence, which includes at least one first combined sample set. Each first combined sample set includes: target respiratory disease treatment feedback data, a first positive example corresponding to the target respiratory disease treatment feedback data, and a first negative example corresponding to the target respiratory disease treatment feedback data. The target respiratory disease treatment feedback data are different in different first combined sample sets.

[0134] Step B120: Based on the first sample sequence, perform feature comparison learning on the first treatment feedback matching model.

[0135] In this embodiment, before iterative parameter interaction training, the server performs feature comparison learning on the first treatment feedback matching model (Model A) to improve its initial performance. This step aims to help the model better learn how to distinguish the semantic correlation features between different feedback contents by comparing the target respiratory disease treatment feedback data with its corresponding positive and negative examples.

[0136] The server first connects to a data warehouse or pre-prepared dataset to retrieve an initial sample sequence. This sequence contains a carefully selected set of combined samples, each built around a specific target of respiratory disease treatment feedback data.

[0137] Specifically, the server reads the first sample sequence file or database table and loads it into memory. Each first combined sample set contains at least three parts: a target respiratory disease treatment feedback data set, the first positive example feedback data corresponding to the target data set, and the first negative example feedback data corresponding to the target data set. The target respiratory disease treatment feedback data set in different combined sample sets is unique, ensuring data diversity and representativeness.

[0138] After acquiring the first sample sequence, the server begins using these samples to perform feature comparison learning on the first treatment feedback matching model (Model A). This process primarily focuses on how the model distinguishes the target data from its positive and negative examples based on semantic features. Specifically, the server first preprocesses each feedback data in the sample sequence, including text cleaning, word segmentation, and stop word removal, to prepare for subsequent feature extraction. Next, utilizing the existing structure of Model A (such as a multi-layer neural network), the server extracts features from each feedback data. These features are typically semantic representations of the text data, reflecting the inherent relationships between the data. For each first combined sample set, the server performs feature comparisons between the target data and the positive and negative examples. The comparisons may include Euclidean distance and cosine similarity of feature vectors, aiming to quantify the semantic differences between the data. Based on the results of the feature comparisons, the server calculates the loss value (e.g., cross-entropy loss) for the model in distinguishing between positive and negative examples. Then, the weights and bias parameters of Model A are adjusted using the backpropagation algorithm to minimize the loss value and improve the model's discriminative ability.

[0139] The above process will be repeated iteratively until the model's performance on the validation set stabilizes or the preset stopping condition is met. Each iteration updates the model parameters based on the new loss value, gradually optimizing the model's performance.

[0140] After completing the comparative feature learning, the server needs to evaluate whether the performance of model A has improved. This is typically done by testing the model on an independent validation set. Specifically, the server prepares a validation set containing feedback data on the treatment of the target respiratory disease and its positive and negative examples. Model A is used to predict the samples in the validation set, and metrics such as accuracy, recall, and F1 score are calculated to distinguish between positive and negative examples. These metrics are then compared with the model's performance before the comparative feature learning to evaluate the learning outcome.

[0141] In one possible implementation, the first treatment feedback matching model extracts semantic feature data from each of the two respiratory disease treatment feedback data sets, and uses the semantic feature matching degree between the two extracted semantic feature data sets as the semantic correlation of the feedback content between the two respiratory disease treatment feedback data sets. The second treatment feedback matching model fuses the two respiratory disease treatment feedback data sets to obtain fused respiratory disease treatment feedback data, estimates the semantic feature data of the fused respiratory disease treatment feedback data, and generates the semantic correlation of the feedback content between the two respiratory disease treatment feedback data sets based on the extracted semantic feature data.

[0142] In this embodiment, in the task of matching respiratory disease treatment feedback, the server uses two different types of models to evaluate the semantic correlation between feedback data: a first treatment feedback matching model (Model A) and a second treatment feedback matching model (Model B). These two models employ different strategies when processing feedback data, but their common goal is to quantify the similarity or correlation between feedback content.

[0143] The first treatment feedback matching model (Model A) evaluates the semantic relevance of the feedback content between two sets of respiratory disease treatment feedback data by extracting semantic feature data from each set and calculating the matching degree between these feature data. Specifically, the server feeds two sets of respiratory disease treatment feedback data (denoted as Data A and Data B) as input to Model A. Model A's internal multi-layer neural network structure processes the input data, extracting semantic features layer by layer through components such as convolutional layers, pooling layers, and fully connected layers. These features are high-dimensional vectors that capture key information in the text data. After extracting the semantic feature vectors of Data A and Data B, Model A calculates the similarity or matching degree between these two vectors. This is typically achieved using methods such as cosine similarity and Euclidean distance, aiming to quantify the semantic proximity of the two feedback data sets. Based on the calculated feature matching degree, Model A outputs a relevance score, representing the strength of the semantic relevance between Data A and Data B.

[0144] Unlike Model A, the second treatment feedback matching model (Model B) employs a fusion strategy to evaluate the semantic relevance of feedback content between two respiratory disease treatment feedback data sets. Similarly, the server inputs two respiratory disease treatment feedback data sets (Data A and Data B) into Model B. Model B first fuses Data A and Data B. The fusion can be achieved through simple concatenation, weighted averaging, or more complex attention mechanisms or recurrent neural network structures. The goal of the fusion is to generate a fused respiratory disease treatment feedback data set that incorporates comprehensive information from both sets. Next, Model B extracts semantic features from the fused data. This process is similar to Model A, using a multi-layered neural network structure to extract high-dimensional semantic feature vectors layer by layer. Since the fused data already contains comprehensive information from Data A and Data B, Model B can directly evaluate the semantic relevance of feedback content between the original two sets of data based on the semantic features of this fused data. This is typically achieved by calculating the similarity or difference between the fused data feature vector and a baseline vector (such as an all-zero vector or a prior knowledge vector specific to the task). However, in practical applications, Model B might directly output a score representing the strength of the association, without explicitly calculating the similarity to the baseline vector. Ultimately, Model B outputs an association score that reflects the semantic relevance of the feedback content between data A and data B from a fusion perspective.

[0145] In one possible implementation, the method further includes:

[0146] Step S150: Obtain the query sequence retrieved for the respiratory disease treatment feedback matching task, wherein the query sequence includes multiple reference treatment feedback data.

[0147] Step S160: Invoke the target treatment feedback matching model to extract each reference treatment feedback data in the query sequence and generate semantic feature data of each reference treatment feedback data.

[0148] Step S170: When the respiratory disease treatment feedback data to be matched is obtained, the respiratory disease treatment feedback data to be matched is extracted through the target treatment feedback matching model to generate semantic feature data of the respiratory disease treatment feedback data to be matched.

[0149] Step S180: Calculate the semantic feature matching degree between the semantic feature data of each reference treatment feedback data and the semantic feature data of the respiratory disease treatment feedback data to be matched, and use it as the semantic correlation of the feedback content between the corresponding reference treatment feedback data and the respiratory disease treatment feedback data to be matched.

[0150] Step S190: Based on the descending order of the semantic relevance of the feedback content, and based on the semantic relevance of the feedback content between each reference treatment feedback data and the respiratory disease treatment feedback data to be matched, at least one reference treatment feedback data is extracted from the query sequence as reference treatment feedback data. Based on the extracted at least one reference treatment feedback data, a matching result for the respiratory disease treatment feedback data to be matched is generated.

[0151] In this embodiment, within a medical institution, the server is responsible for performing respiratory disease treatment feedback matching tasks to assist doctors in quickly finding historical treatment feedback data that is similar to or related to the feedback data to be matched. This process involves multiple steps, including obtaining the query sequence, extracting semantic feature data, calculating the semantic feature matching degree, and generating the matching result.

[0152] First, the server receives a request for a respiratory disease treatment feedback matching task. To respond to this request, the server first needs to retrieve a query sequence from the database. This query sequence contains multiple reference treatment feedback records, which are patient treatment feedback records collected over a past period.

[0153] In detail, the server retrieves relevant records from the treatment feedback database based on parameters in the request (such as disease type, treatment period, etc.). The retrieved records are then organized into a query sequence, which contains multiple reference treatment feedback data.

[0154] After obtaining the query sequence, the server invokes the target treatment feedback matching model (a model optimized through iterative parameter interaction training) to process each reference treatment feedback data in the sequence to extract its semantic feature data. Specifically, the server loads the target treatment feedback matching model into memory. It iterates through each reference treatment feedback data in the query sequence, inputting its text content into the target treatment feedback matching model. The target treatment feedback matching model processes the input data through its internal multi-layer neural network structure, extracting high-dimensional semantic feature vectors. The server stores the semantic feature vector corresponding to each reference treatment feedback data for later use.

[0155] At some point, the server retrieved new respiratory disease treatment feedback data to be matched. This data may be from the current patient's latest treatment feedback record.

[0156] Specifically, the server receives and stores the respiratory disease treatment feedback data to be matched. The same target treatment feedback matching model is used to process the respiratory disease treatment feedback data to be matched, extracting its semantic feature data (semantic feature vector).

[0157] After extracting the semantic features of all reference treatment feedback data and the respiratory disease treatment feedback data to be matched in the query sequence, the server begins to calculate the semantic feature matching degree between them to evaluate the semantic relevance between the feedback content. Specifically, the server iterates through each reference treatment feedback data in the query sequence. For each reference treatment feedback data, it calculates the similarity or matching degree between its semantic feature vector and the semantic feature vector of the respiratory disease treatment feedback data to be matched. This can be achieved using methods such as cosine similarity and Euclidean distance. The calculated matching degree is stored as a semantic relevance score between the corresponding reference treatment feedback data and the respiratory disease treatment feedback data to be matched.

[0158] After evaluating the semantic relevance of the feedback content between all reference treatment feedback data and the respiratory disease treatment feedback data to be matched, the server extracts at least one reference treatment feedback data from the query sequence in descending order of relevance strength as the final reference treatment feedback data and generates the matching result.

[0159] In detail, the server sorts the reference treatment feedback data in the query sequence based on the semantic relevance score of the feedback content. Reference treatment feedback data is extracted from the sorted sequence according to preset rules (such as selecting the top N data with the highest scores). The extracted reference treatment feedback data is compiled into a matching result report, which may include detailed information about the reference treatment feedback data and the strength of its relevance to the respiratory disease treatment feedback data to be matched. The server sends the matching result report to the requesting party (such as a doctor's workstation) for further analysis and use.

[0160] In one possible implementation, step S190 includes:

[0161] Step S191: The extracted reference treatment feedback data are fused with the respiratory disease treatment feedback data to be matched to generate fused data corresponding to the corresponding reference treatment feedback data.

[0162] Step S192: The second treatment feedback matching model is invoked to estimate the semantic correlation of the feedback content between the corresponding reference treatment feedback data and the respiratory disease treatment feedback data to be matched based on the fusion data corresponding to each reference treatment feedback data.

[0163] Step S193: Based on the descending order of the semantic relevance of the feedback content, and based on the semantic relevance of each feedback content extracted by the second treatment feedback matching model, extract the first T reference treatment feedback data from at least one extracted reference treatment feedback data as target treatment feedback data, where T is a positive integer.

[0164] Step S194: Based on the extracted T target treatment feedback data, generate the matching result of the respiratory disease treatment feedback data to be matched.

[0165] In this embodiment, after extracting and initially evaluating the reference treatment feedback data, the server needs to further process this data to generate an accurate matching result for the respiratory disease treatment feedback data to be matched. This process involves steps such as data fusion, re-evaluating semantic relevance using a second treatment feedback matching model, and extracting target treatment feedback data based on the strength of relevance.

[0166] First, the server merges each extracted reference treatment feedback data with the respiratory disease treatment feedback data to be matched to generate fused data containing comprehensive information from both.

[0167] In detail, for each selected reference treatment feedback data in the query sequence, the server merges its text content with the text content of the respiratory disease treatment feedback data to be matched. The merging can be a simple concatenation or an intelligent fusion based on certain rules or algorithms, such as using an attention mechanism to highlight important information. The fused data not only contains the features of the reference treatment feedback data but also incorporates the characteristics of the respiratory disease treatment feedback data to be matched, providing a more comprehensive perspective for subsequent correlation assessment.

[0168] Next, the server invokes the second treatment feedback matching model (Model B) to process the fused data in order to reassess the semantic correlation of the feedback content between each reference treatment feedback data and the respiratory disease treatment feedback data to be matched.

[0169] In detail, the server loads Model B into memory. It iterates through all the fused data (each fused data point corresponds to a combination of reference treatment feedback data and the respiratory disease treatment feedback data to be matched). Each fused data point is input into Model B, which, through its unique fusion strategy and semantic feature extraction capabilities, generates a semantic feature vector for the fused data. Based on these semantic feature vectors, Model B calculates the semantic relevance score between the fused data and the respiratory disease treatment feedback data to be matched (from the fusion perspective).

[0170] After obtaining the semantic relevance scores of all fused data, the server extracts the top T reference treatment feedback data as target treatment feedback data based on the descending order of the scores.

[0171] In detail, the server sorts the semantic relevance scores of all fused data. Based on a preset T value (a positive integer), it extracts the reference treatment feedback data corresponding to the top T highest-scoring fused data from the sorted list. These extracted data are considered to be the target treatment feedback data that are semantically closest to or most relevant to the respiratory disease treatment feedback data to be matched.

[0172] Finally, the server generates matching results for the T extracted target treatment feedback data based on the respiratory disease treatment feedback data to be matched. In detail, the server organizes information from these T target treatment feedback data, including their original text content and correlation scores with the respiratory disease treatment feedback data to be matched. This information is then organized into a matching result report, which may also include a brief analysis or recommendations for the target data. The server sends the matching result report to the requesting party (such as a doctor's workstation or patient terminal) for further analysis, decision-making, or feedback.

[0173] Figure 2The illustration shows a schematic diagram of the hardware structure of an AI-based respiratory disease treatment feedback data processing system 100, provided in an embodiment of this application, for implementing the aforementioned AI-based respiratory disease treatment feedback data processing method. Figure 2 As shown, the AI-based respiratory disease treatment feedback data processing system 100 may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0174] In one possible design, the AI-based respiratory disease treatment feedback data processing system 100 can be a single server or a group of servers. The server group can be centralized or distributed (e.g., the AI-based respiratory disease treatment feedback data processing system 100 can be a distributed system). In some embodiments, the AI-based respiratory disease treatment feedback data processing system 100 can be local or remote. For example, the AI-based respiratory disease treatment feedback data processing system 100 can access information and / or data stored in machine-readable storage medium 120 via a network. As another example, the AI-based respiratory disease treatment feedback data processing system 100 can be directly connected to machine-readable storage medium 120 to access stored information and / or data. In some embodiments, the AI-based respiratory disease treatment feedback data processing system 100 can be implemented on an AI-based respiratory disease treatment feedback data processing system. By way of example only, the AI-based respiratory disease treatment feedback data processing system can include a private cloud, public cloud, hybrid cloud, community cloud, distributed cloud, internal cloud, multi-layered cloud, etc., or any combination thereof.

[0175] Machine-readable storage medium 120 may store data and / or instructions. In some embodiments, machine-readable storage medium 120 may store data acquired from an external terminal. In some embodiments, machine-readable storage medium 120 may store data and / or instructions used by the AI-based respiratory disease treatment feedback data processing system 100 to perform or use in order to accomplish the exemplary methods described in this application.

[0176] In a specific implementation, one or more processors 110 execute computer-executable instructions stored in the machine-readable storage medium 120, enabling the processor 110 to execute the AI-based respiratory disease treatment feedback data processing method as described in the above method embodiment. The processor 110, the machine-readable storage medium 120, and the communication unit 140 are connected via a bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0177] The specific implementation process of processor 110 can be found in the various method embodiments executed by the above-mentioned artificial intelligence-based respiratory disease treatment feedback data processing system 100. The implementation principle and technical effect are similar, and will not be repeated here.

[0178] Furthermore, this application embodiment also provides a readable storage medium, wherein computer-executable instructions are preset in the readable storage medium, and when the processor executes the computer-executable instructions, the above-mentioned artificial intelligence-based respiratory disease treatment feedback data processing method is implemented.

[0179] It should be noted that, in order to simplify the description disclosed in this application and thus help to understand one or more embodiments of the invention, the foregoing description of the embodiments of this application may sometimes combine multiple features into one embodiment, drawing or description thereof.

Claims

1. A method for processing respiratory disease treatment feedback data based on artificial intelligence, characterized in that, The method comprises: A first treatment feedback matching model and a second treatment feedback matching model are obtained. The first treatment feedback matching model and the second treatment feedback matching model are configured to determine the semantic correlation of feedback content between respiratory disease treatment feedback data. The model decision error of the first treatment feedback matching model is greater than the model decision error of the second treatment feedback matching model, and the model decision performance of the first treatment feedback matching model is greater than the model decision performance of the second treatment feedback matching model. The model decision performance is used to represent the performance of rapidly processing large amounts of data. Multiple training sequences of respiratory disease treatment feedback data are acquired. Each training sequence of respiratory disease treatment feedback data includes target respiratory disease treatment feedback data and counterexample respiratory disease treatment feedback data corresponding to the target respiratory disease treatment feedback data. Different training costs are corresponding to the counterexample respiratory disease treatment feedback data in different training sequences of respiratory disease treatment feedback data. The training cost is used to represent the complexity index of distinguishing the target respiratory disease treatment feedback data and the corresponding counterexample respiratory disease treatment feedback data. The number of training iterations for each respiratory disease treatment feedback data training sequence is determined, and the number of training iterations for any respiratory disease treatment feedback data training sequence is positively correlated with the training cost of the counterexample respiratory disease treatment feedback data in the corresponding respiratory disease treatment feedback data training sequence. Based on the ascending order of training iterations, the training sequence of respiratory disease treatment feedback data with m training iterations is represented as the m-th respiratory disease treatment feedback data training sequence, where m is a positive integer. The first treatment feedback matching model is used to perform the m-th round of parameter interaction training based on the m-th respiratory disease treatment feedback data training sequence and the second treatment feedback matching model. This includes: estimating the semantic correlation of feedback content between the target respiratory disease treatment feedback data and the counterexample respiratory disease treatment feedback data in the m-th respiratory disease treatment feedback data training sequence using the first treatment feedback matching model, generating a first semantic correlation of feedback content; estimating the semantic correlation of feedback content between the target respiratory disease treatment feedback data and the counterexample respiratory disease treatment feedback data in the m-th respiratory disease treatment feedback data training sequence using the second treatment feedback matching model, generating a second semantic correlation of feedback content; and optimizing the weight and bias information of the first treatment feedback matching model based on the error between the first and second semantic correlations of feedback content, generating a target treatment feedback matching model.

2. The method for processing respiratory disease treatment feedback data based on artificial intelligence according to claim 1, characterized in that, The acquisition of multiple respiratory disease treatment feedback data training sequences includes: Acquire treatment feedback data and a treatment feedback database for the target respiratory disease, wherein the treatment feedback database includes multiple treatment feedback samples; An initial treatment feedback matching model is invoked to estimate the semantic correlation of the feedback content between the target respiratory disease treatment feedback data and each treatment feedback sample in the treatment feedback database, thereby generating the semantic correlation of the target feedback content corresponding to each treatment feedback sample; wherein, the semantic correlation of the target feedback content corresponding to any treatment feedback sample is positively correlated with the training cost of the corresponding treatment feedback sample; Based on the descending order of the semantic relevance of the target feedback content corresponding to each treatment feedback sample, X treatment feedback samples are extracted from the treatment feedback database as counterexample respiratory disease treatment feedback data corresponding to the target respiratory disease treatment feedback data, where X is a positive integer. Based on the target respiratory disease treatment feedback data and the extracted X counterexample respiratory disease treatment feedback data, multiple respiratory disease treatment feedback data training sequences are configured.

3. The method for processing respiratory disease treatment feedback data based on artificial intelligence according to claim 2, characterized in that, When the initial treatment feedback matching model is the first treatment feedback matching model, the step of configuring multiple respiratory disease treatment feedback data training sequences based on the target respiratory disease treatment feedback data and the extracted X counterexample respiratory disease treatment feedback data includes: Based on the descending order of the semantic relevance of the target feedback content corresponding to the X counterexample respiratory disease treatment feedback data, the X counterexample respiratory disease treatment feedback data are sorted to generate the target respiratory disease treatment feedback data distribution. The second treatment feedback matching model is used to estimate the semantic correlation of the feedback content between the target respiratory disease treatment feedback data and each counterexample respiratory disease treatment feedback data, and to generate candidate feedback content semantic correlations corresponding to each counterexample respiratory disease treatment feedback data. Based on the descending order of the semantic relevance of the candidate feedback content corresponding to each counterexample respiratory disease treatment feedback data, the X counterexample respiratory disease treatment feedback data are sorted to generate a candidate respiratory disease treatment feedback data distribution; Multiple rounds of data sampling are performed in both the target respiratory disease treatment feedback data distribution and the candidate respiratory disease treatment feedback data distribution, according to the reverse sampling direction. A respiratory disease treatment feedback data training sequence includes Y counterexample respiratory disease treatment feedback data, where Y is a positive integer. One round of data sampling is used to extract Z counterexample respiratory disease treatment feedback data from both the target and candidate respiratory disease treatment feedback data distributions, where Z ∈ [Y / 2, X). After the m-th round of data sampling, based on the Z counterexample respiratory disease treatment feedback data extracted from the target respiratory disease treatment feedback data distribution and the candidate respiratory disease treatment feedback data distribution in this round, a counterexample respiratory disease treatment feedback data list corresponding to the m-th training iteration number is configured. Based on the target respiratory disease treatment feedback data and the list of counterexample respiratory disease treatment feedback data configured in this round, configure the respiratory disease treatment feedback data training sequence for the m-th training iteration, where m is a positive integer.

4. The method for processing respiratory disease treatment feedback data based on artificial intelligence according to claim 3, characterized in that, When Z is greater than Y / 2, the step of configuring the list of counterexample respiratory disease treatment feedback data corresponding to the m-th training iteration number, based on the Z counterexample respiratory disease treatment feedback data extracted from the target respiratory disease treatment feedback data distribution and the candidate respiratory disease treatment feedback data distribution in this round, includes: In this round, from the Z counterexample respiratory disease treatment feedback data extracted from the target respiratory disease treatment feedback data distribution, Q counterexample respiratory disease treatment feedback data are randomly selected. In this round, from the Z counterexample respiratory disease treatment feedback data extracted from the candidate respiratory disease treatment feedback data distribution, W counterexample respiratory disease treatment feedback data are randomly selected; where Q and W are both positive integers, and Q+W=Y; Based on the Q counterexample respiratory disease treatment feedback data and the W counterexample respiratory disease treatment feedback data, configure the counterexample respiratory disease treatment feedback data list corresponding to the m-th training iteration number.

5. The method for processing respiratory disease treatment feedback data based on artificial intelligence according to any one of claims 2-4, characterized in that, The target respiratory disease treatment feedback data is configured with positive respiratory disease treatment feedback data; The step of extracting X treatment feedback samples from the treatment feedback database as counterexample respiratory disease treatment feedback data corresponding to the target respiratory disease treatment feedback data, based on the descending order of the semantic relevance of the target feedback content corresponding to each treatment feedback sample, includes: Based on the positive respiratory disease treatment feedback data corresponding to the target respiratory disease treatment feedback data, the treatment feedback database is optimized to generate an optimized treatment feedback database; wherein, the data optimization includes: removing treatment feedback samples in the treatment feedback database that are identical to the positive respiratory disease treatment feedback data; Based on the descending order of the semantic relevance of the target feedback content corresponding to each treatment feedback sample, X treatment feedback samples are extracted from the optimized treatment feedback database as counterexample respiratory disease treatment feedback data corresponding to the target respiratory disease treatment feedback data.

6. The method for processing respiratory disease treatment feedback data based on artificial intelligence according to claim 1, characterized in that, The counterexample respiratory disease treatment feedback data in the m-th respiratory disease treatment feedback data training sequence includes at least two, and each counterexample respiratory disease treatment feedback data in the m-th respiratory disease treatment feedback data training sequence corresponds to a first feedback content semantic association and a second feedback content semantic association. The step of optimizing the weight and bias information of the first treatment feedback matching model based on the error between the semantic relevance of the first feedback content and the semantic relevance of the second feedback content includes: The training sequence of the m-th respiratory disease treatment feedback data is polled for each counterexample respiratory disease treatment feedback data. Based on the training cost of the k-th counterexample respiratory disease treatment feedback data in this round of polling, a training importance factor is assigned to the corresponding counterexample respiratory disease treatment feedback data. The training importance factor is positively correlated with the training cost. k is a positive integer and is not greater than the number of counterexample respiratory disease treatment feedback data in the training sequence of the m-th respiratory disease treatment feedback data. Based on the error between the semantic correlation of the first feedback content and the semantic correlation of the second feedback content corresponding to the kth counterexample respiratory disease treatment feedback data, the parameter interaction training error parameter corresponding to the kth counterexample respiratory disease treatment feedback data is determined. After all the counterexample respiratory disease treatment feedback data have been polled, the training error parameters of the first treatment feedback matching model are calculated based on the training importance factor of each counterexample respiratory disease treatment feedback data and the corresponding parameter interaction training error parameters. Based on the optimization objective of minimizing the training error parameters, the weight information and bias information of the first treatment feedback matching model are optimized.

7. The method for processing respiratory disease treatment feedback data based on artificial intelligence according to claim 1, characterized in that, The method further includes: Obtain the query sequence retrieved for the respiratory disease treatment feedback matching task, wherein the query sequence includes multiple reference treatment feedback data; The target treatment feedback matching model is invoked to extract each reference treatment feedback data in the query sequence, and semantic feature data of each reference treatment feedback data is generated. When the respiratory disease treatment feedback data to be matched is obtained, the target treatment feedback matching model is used to extract the respiratory disease treatment feedback data to be matched and generate semantic feature data of the respiratory disease treatment feedback data to be matched. Calculate the semantic feature matching degree between the semantic feature data of each reference treatment feedback data and the semantic feature data of the respiratory disease treatment feedback data to be matched, and use it as the semantic correlation of the feedback content between the corresponding reference treatment feedback data and the respiratory disease treatment feedback data to be matched. Based on the descending order of the semantic relevance of the feedback content, and based on the semantic relevance of the feedback content between each reference treatment feedback data and the respiratory disease treatment feedback data to be matched, at least one reference treatment feedback data is extracted from the query sequence as reference treatment feedback data. Based on at least one extracted reference treatment feedback data, a matching result is generated for the respiratory disease treatment feedback data to be matched.

8. The method for processing respiratory disease treatment feedback data based on artificial intelligence according to claim 7, characterized in that, The process of generating a matching result for the respiratory disease treatment feedback data to be matched, based on at least one extracted reference treatment feedback data, includes: The extracted reference treatment feedback data are fused with the respiratory disease treatment feedback data to be matched to generate fused data corresponding to the corresponding reference treatment feedback data. The second treatment feedback matching model is invoked to estimate the semantic correlation of the feedback content between the corresponding reference treatment feedback data and the respiratory disease treatment feedback data to be matched based on the fusion data corresponding to each reference treatment feedback data. Based on the descending order of the semantic relevance of the feedback content, and based on the semantic relevance of each feedback content extracted by the second treatment feedback matching model, the first T reference treatment feedback data are extracted from at least one reference treatment feedback data as target treatment feedback data, where T is a positive integer; Based on the extracted T target treatment feedback data, a matching result is generated for the respiratory disease treatment feedback data to be matched.

9. A respiratory disease treatment feedback data processing system based on artificial intelligence, characterized in that, The AI-based respiratory disease treatment feedback data processing system includes a processor and a memory, the memory and the processor being connected. The memory is used to store programs, instructions or code, and the processor is used to execute the programs, instructions or code in the memory to implement the AI-based respiratory disease treatment feedback data processing method according to any one of claims 1-8.

Images