Cardiopulmonary sound annotation and medical education system and method based on interactive learning and dynamic weighted aggregation, and medium

Abstract

The invention relates to the field of medical health, and particularly provides an interactive mobile terminal system for cardiopulmonary sound annotation and medical education. The system integrates a two-stage feedback learning method, an auxiliary information labeling method and a labeling result aggregation method based on dynamic weighting. Through the first-stage standard teaching and the second-stage real-time feedback type practical training, the cognition and judgment ability of the user on heart and lung sound classification is gradually improved; through auxiliary information such as an audio spectrogram, clinical background information and a category feature quick check card, the memory load of a user is reduced, and the classification accuracy is enhanced; weighted optimization and high-quality aggregation of annotation data are realized through user capability dynamic modeling and an abnormal annotation filtering mechanism. The system adopts mobile terminal application deployment, supports data desensitization processing and annotation anomaly detection, and guarantees data security and privacy. The method can significantly improve the labeling accuracy of non-medical background users, improves the scale and quality of the data set, and is suitable for data set construction and training of an artificial intelligence auxiliary medical diagnosis system.

Description

Technical Field

[0001] The present invention relates to the interdisciplinary field of medical health and artificial intelligence, and in particular to a cardiopulmonary sound annotation and medical education system, method, and medium based on interactive learning and dynamic weighted aggregation. Background Art

[0002] With the rapid development of medical technology, especially the widespread application of digitalization and artificial intelligence in medicine, medical audio data is playing an increasingly important role in disease diagnosis, health monitoring, and medical research. Specifically, physiological audio data such as heart sounds, lung sounds, and respiratory sounds have become crucial tools for early disease detection and health monitoring. AI-driven diagnostic systems, especially those based on machine learning and deep learning, rely on high-quality, large-scale audio datasets to train and optimize their diagnostic capabilities.

[0003] However, the annotation of medical audio data still faces a series of technical challenges. The main audio data annotation methods include expert annotation and crowdsourcing annotation:

[0004] · Expert annotation involves the labeling of audio samples by experts with medical expertise. Although this method ensures high quality and accuracy, it suffers from high costs, slow annotation speed, and difficulty in scalability.

[0005] · Crowdsourced annotation, by leveraging the participation of non-professionals such as the general public and medical students, offers lower costs and greater scalability. However, due to the limited medical background of the annotators, annotation errors are prone to occur, resulting in reduced accuracy and reliability of the dataset. Crowdsourced annotation also typically requires extensive post-processing, further increasing costs and time.

[0006] Therefore, there is an urgent need for a new annotation method that can both ensure the quality of data annotation and expand the size of the data set to support the further application and development of artificial intelligence in the medical field. Summary of the Invention

[0007] In response to the above-mentioned shortcomings of the existing technology, the present invention provides a cardiopulmonary sound annotation and medical education system, method and medium based on interactive learning and dynamic weighted aggregation, aiming to solve the problems of low data annotation accuracy, low efficiency and uneven data distribution in traditional methods.

[0008] The technical solution of the present invention includes the following aspects:

[0009] Aspect 1: Two-stage feedback learning method

[0010] The present invention provides a two-stage feedback learning method, characterized in that:

[0011] 1. Adopt a two-stage progressive learning mechanism, including:

[0012] o Phase 1: Users learn physiological sound classification characteristics through standard teaching materials and audio samples, including the distinction between normal and abnormal sounds, auditory descriptions of common abnormal sounds, and spectral characteristics.

[0013] o Phase 2: Users practice real-time physiological audio classification. The system provides immediate feedback based on the user's annotation results. If the annotation is incorrect, the system returns the correct answer and analyzes the feature differences.

[0014] 2. In the second stage, questions are presented in increasing order of difficulty, including "differentiation between normal and abnormal sounds → identification of common abnormal sounds → classification of all categories of abnormal sounds" to reduce the user's learning cost.

[0015] 3. The system dynamically calculates the accuracy of users’ answers. When it detects that the user’s accuracy in a certain category is low, it automatically increases the frequency of questions in that category to promote targeted training.

[0016] Beneficial effects:

[0017] Through phased learning and instant feedback, it conforms to the user's cognitive laws and reduces learning costs; through progressive presentation, it simulates clinical thinking and improves the ability to recognize complex patterns; through real-time error correction, it enhances the user's self-correction ability, thereby indirectly improving the quality and consistency of crowdsourced annotation data.

[0018] Second aspect: auxiliary information annotation method

[0019] The present invention provides an auxiliary information marking method, which is characterized by:

[0020] 1. Multimodal auxiliary information fusion mechanism, including but not limited to:

[0021] o Audio spectrum visualization: Dynamically generate audio time-frequency diagrams (such as STFT or MFCC) during annotation

[0022] Spectrogram), which intuitively displays the frequency distribution and time variation characteristics;

[0023] o Presentation of clinical background information: Provide basic patient information (such as age, gender, symptoms, and medical history) to assist in classification;

[0024] oCategory Feature Quick Reference Cards: These provide reference information such as auditory descriptions and spectral feature diagrams of typical features of each category.

[0025] 2. Auxiliary information is embedded in the annotation interface, such as displaying a summary of category features directly next to the options to reduce the user's cognitive load during the classification process.

[0026] Beneficial effects:

[0027] It reduces the user's memory load, improves perception and decision-making capabilities in complex scenarios, and enables non-professional users' labeling logic to approach the thinking of medical experts, thereby significantly improving labeling accuracy and learning efficiency.

[0028] The third aspect: annotation aggregation method based on dynamic weighting

[0029] The present invention provides a method for aggregating annotation results based on dynamic weighting of user capabilities, which is characterized by:

[0030] 1. Two-stage hierarchical labeling mechanism:

[0031] oThe first stage is to perform a rough classification of “normal / abnormal” of heart and lung sounds;

[0032] In the second stage, abnormal sound samples are classified in a fine-grained manner (such as wheezing and moist rales), and the classification is guided by the medical feature library.

[0033] 2. Abnormal annotation filtering mechanism: Use statistical outlier detection (such as Z-score analysis, cluster analysis, and overall accuracy analysis) to identify and remove abnormally labeled data.

[0034] 3. Dynamic capability weight model: Dynamically adjust the user's classification weight for unlabeled audio based on the user's historical annotation performance (accuracy, response time, consistency).

[0035] 4. Two-step crowdsourcing annotation aggregation algorithm, including:

[0036] For each audio sample, if the number of valid annotations is lower than the threshold, it will be removed;

[0037] If the number of valid annotations is reached, the major category labels are determined based on the normal / abnormal classification weights.

[0038] If it is an abnormal sound, the final classification is determined based on the accumulated user weight value in each abnormal category.

[0039] Beneficial effects:

[0040] By introducing anomaly detection and dynamic weighting mechanisms, we can fully utilize the judgment acumen of non-professional users and improve the accuracy and reliability of labeled datasets.

[0041] The fourth aspect: medical audio teaching, annotation and classification system

[0042] The present invention further provides an interactive mobile terminal system integrating the aforementioned method, characterized in that:

[0043] 1. Mobile integration: Integrate functions such as two-stage feedback learning, auxiliary information presentation, and dynamic weighted annotation aggregation into mobile applications, lowering the operation and learning thresholds and accelerating the learning and annotation process.

[0044] 2. Data security and privacy protection mechanisms, including:

[0045] oData desensitization: Remove patient identity information and retain only necessary desensitized labels (such as age, gender, and disease category).

[0046] Anomaly detection: Based on user behavior and consistency analysis, it can identify and filter malicious or low-quality annotations in real time.

[0047] Beneficial effects:

[0048] Ensure that data collection complies with medical ethics and data protection regulations, while effectively improving the quality of training data and supporting the performance optimization and promotion and application of AI diagnostic systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a schematic diagram of the overall structure of an interactive mobile terminal system for cardiopulmonary sound annotation and medical education provided by an embodiment of the present invention;

[0050] Figure 2 1 is a flow chart of a two-stage feedback learning method according to an embodiment of the present invention;

[0051] Figure 3 1 is a schematic structural diagram of an auxiliary information marking method according to an embodiment of the present invention;

[0052] Figure 4 This is a processing flow chart of a method for aggregating annotation results based on dynamic weighting in an embodiment of the present invention. DETAILED DESCRIPTION

[0053] To make the purpose, technical solutions and advantages of the present invention clearer, the following detailed description of the embodiments of the present invention is given in conjunction with the accompanying drawings and examples. It should be noted that the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, other embodiments obtained by those skilled in the art without creative work are all within the scope of protection of the present invention.

[0054] Example 1: System overall structure implementation

[0055] The present invention provides an interactive mobile terminal system for cardiopulmonary sound annotation and medical education, the overall structure of which is as follows: Figure 1 As shown. The system includes at least the following modules:

[0056] ·The mobile user interface module is used to present teaching content, audio annotation tasks, and auxiliary reference information to users, and receive interactive feedback and annotation operations from users;

[0057] · A real-time feedback processing module is used to generate timely feedback information based on the user's input in the audio classification task, including correct answer prompts, feature difference analysis, and subsequent learning guidance;

[0058] · Auxiliary information generation module, used to dynamically generate and display auxiliary information during the user's annotation process, including audio spectrograms, patient background information prompts, and reference content for each classification feature;

[0059] · The user capability assessment and management module is used to collect, count, and analyze user behavior data during the annotation process, establish a dynamic user capability profile, and provide a basis for subsequent data processing;

[0060] · Dynamic annotation aggregation and optimization module, which is used to screen, weight, filter, and fuse the collected annotation results based on user ability weights, and output high-quality and well-balanced audio tags;

[0061] · Data security and privacy protection module, which automatically performs desensitization, rights management, and anomaly detection during data collection and processing to ensure patient privacy and data compliance;

[0062] · The server-side data storage and processing module is used to store user interaction data, audio files and annotation results, and perform subsequent data analysis, training set construction and model optimization tasks.

[0063] In one embodiment of the present invention, data exchange between the mobile user interface module and the server-side module is achieved through an encrypted communication protocol (such as HTTPS) to ensure the integrity and security of data during transmission. These modules can be deployed on the same hardware platform (such as an integrated server) or distributed across different physical nodes. Data exchange between modules is performed through a standard application programming interface (API), ensuring good scalability and maintainability of the system.

[0064] Specifically:

[0065] · When the user logs into the system, the mobile user interface module pushes teaching materials and feedback teaching content to the user;

[0066] ·After the user completes each audio annotation, the real-time feedback processing module immediately analyzes the user input and provides feedback on the correct answer and key information to help the user make cognitive corrections;

[0067] · After the user completes the tutorial, the mobile user interface module pushes the user level test to the user, thereby calculating the user's classification weight in different audio categories. This weight is used in the dataset construction process;

[0068] · The system pushes annotation tasks to users, and the auxiliary information generation module simultaneously presents audio spectrograms and reference materials;

[0069] · Each user's operation record (including answers, response time, selection process, etc.) is uploaded to the user ability assessment and management module in real time;

[0070] · During the dataset construction phase, the dynamic annotation aggregation and optimization module conducts a comprehensive analysis of the annotated data based on the user's ability weight, removes abnormal annotations, and generates the final labels;

[0071] · Throughout the entire process, the data security and privacy protection module continuously monitors data access behavior to ensure that all collected, stored and processed data comply with medical data protection requirements.

[0072] Through the collaborative work of the above modules, the present invention realizes a complete closed loop of cardiopulmonary audio teaching and training, crowdsourcing annotation, annotation optimization and privacy protection, greatly improving the annotation accuracy of users with non-medical backgrounds, reducing the system's human review costs, and ensuring the security and availability of data.

[0073] The core mechanism of the present invention and its specific implementation process are described in detail below in conjunction with the system structure.

[0074] Example 2: Implementation steps of the two-stage feedback learning method

[0075] The present invention provides a two-stage feedback learning method for improving the cardiopulmonary audio classification ability among users with non-medical backgrounds, thereby significantly improving the accuracy and consistency of crowdsourced annotation data.

[0076] The two-stage feedback learning method comprises at least the following steps:

[0077] 1. Phase 1: Standard material learning and basic training

[0078] In this stage, the mobile user interface module pushes the system's preset standard teaching materials and audio samples to users. The specific process of this stage includes:

[0079] · Standard teaching materials should be presented: Demonstrate the basic characteristics of normal heart and lung sounds and common abnormal sounds (such as wheezing, moist rales, and rhonchi), including but not limited to:

[0080] o The time of day during the respiratory cycle when the sound occurs (e.g., end of inspiration, early expiration);

[0081] oSound frequency characteristics (such as high frequency, low frequency);

[0082] o Auditory description (e.g., continuous sound, discontinuous sound);

[0083] oExample of a typical spectrogram.

[0084] · Supporting audio sample playback: Real recorded audio corresponding to each category is provided simultaneously for users to listen to repeatedly and deepen their memory.

[0085] · Guided quizzes: Small-scale classification tests are interspersed throughout the learning process to test the user's understanding of the knowledge points.

[0086] Through learning with standard materials and samples, users can establish a basic cognitive framework for heart and lung sounds, laying the foundation for subsequent practical exercises.

[0087] 2. Second stage: real-time feedback classification practice

[0088] In this stage, users begin to practice classifying the cardiopulmonary audio samples pushed by the system and receive feedback from the system immediately after each answer. The specific process includes:

[0089] · Audio sample classification task: The user needs to make a classification judgment for each audio sample played (such as

[0090] “Normal”, “Abnormal - Moist Rales”, “Abnormal - Dry Rales”, etc.).

[0091] · Instant feedback generation:

[0092] When the user's classification is correct, the system prompts "correct" and optionally displays the corresponding correct feature summary to reinforce the correct identification experience;

[0093] oWhen the user makes a classification error, the system will immediately prompt "error" and automatically present the correct category and the typical auditory and spectral features of that category, and compare them with the category selected by the user.

[0094] · Error-guided learning mechanism: For categories in which users repeatedly make mistakes, the system records them and appropriately increases the frequency of samples in this category in subsequent tasks to promote targeted learning.

[0095] 3. Difficulty Progression Mechanism

[0096] To comply with the user's cognitive development rules, the present invention sets a gradually progressive difficulty level in the second stage, including:

[0097] · Primary stage: distinguishing normal sounds from abnormal sounds (two-classification task), helping users build basic abnormality recognition capabilities;

[0098] · Intermediate stage: Further subdivide common abnormal categories (such as moist rales and wheezing) among abnormal sounds to train users to identify common disease signs;

[0099] · Advanced stage: covers all preset exception categories and requires users to be able to perform fine-grained classification judgments in multiple categories.

[0100] As the training process progresses, the system dynamically adjusts the proportion and difficulty of samples of different categories in the task pool to gradually improve the user's comprehensive auscultation level.

[0101] In one embodiment of the system, a fixed number of questions are assigned to different difficulty levels, such as "10 questions for beginners, 20 questions for intermediate, and 20 questions for advanced"; in another independent embodiment, the user starts from the beginner level, and the system automatically increases the difficulty level when a certain accuracy rate is achieved.

[0102] 4. User performance tracking and capability profile generation

[0103] During the two-stage learning process, the system records the following user behavior data in real time through the user capability assessment and management module:

[0104] · Classification accuracy under each category;

[0105] · Response time distribution;

[0106] · Error frequency and error types;

[0107] · Difficulty level advancement.

[0108] The system dynamically constructs a user capability profile based on the above data, providing a basis for subsequent data annotation weight assignment (see Example 4).

[0109] 5.Technological advantages

[0110] By adopting a two-stage feedback learning method, the present invention is able to:

[0111] · Guide users to gradually master complex heart and lung sound classification tasks with low cognitive load;

[0112] · Facilitate self-correction and accelerate the learning curve through immediate feedback;

[0113] · The progressive difficulty design simulates the real clinical auscultation process and improves the user's pattern recognition ability;

[0114] · Through ability tracking and difficulty reinforcement, users' recognition level in weak categories is improved in a personalized manner.

[0115] This significantly improves the accuracy and consistency of audio annotation among users with non-medical backgrounds and optimizes the quality of crowdsourcing datasets.

[0116] Example 3: Implementation of the auxiliary information annotation method

[0117] The present invention provides an auxiliary information labeling method for reducing cognitive load, improving classification accuracy and learning efficiency by providing multimodal auxiliary information when users perform cardiopulmonary sound classification and labeling.

[0118] The auxiliary information annotation method at least includes the following contents:

[0119] 1. Audio spectrum visualization assistance

[0120] In a specific embodiment of the present invention, the auxiliary information generation module can generate a spectrogram based on the audio signal in real time and display it to the user simultaneously. The specific process includes:

[0121] · Audio signal processing: Perform short-time Fourier transform (STFT) or Mel-frequency cepstral coefficient (MFCC) extraction on the input audio samples to generate a two-dimensional time-frequency distribution map.

[0122] · Spectrogram display: A spectrogram display area is embedded in the user annotation interface to display the time axis and frequency axis information of the audio signal in real time, allowing users to intuitively observe the frequency change pattern and energy distribution characteristics of the sound.

[0123] · Feature-assisted interpretation: Typical feature areas can be displayed synchronously next to the spectrogram, such as "high-frequency continuous components may correspond to wheezing sounds" and "mid- and low-frequency coarse scattered points may correspond to moist rales", etc., to help users interpret the spectrogram and make classification decisions.

[0124] By visually displaying abstract sound signal characteristics, users' subjective dependence on sound understanding is reduced, and the consistency and accuracy of annotations are improved.

[0125] 2. Clinical background information assistance

[0126] To further simulate the real clinical auscultation scenario, the present invention provides the patient's clinical background prompts associated with the audio sample in the annotation interface, including but not limited to:

[0127] · Patient age group (e.g., infants, young adults, middle-aged and elderly);

[0128] · Patient's gender;

[0129] · Description of relevant symptoms (such as cough, wheezing, fever);

[0130] During the labeling process, these background information are presented in a concise and clear manner near the audio playback area, and users can assist in audio classification judgment based on clinical inference habits.

[0131] In an optional embodiment, the background information can be set to be expandable / collapseable to avoid interfering with the user's independent judgment ability.

[0132] 3. Category feature quick check card assistance

[0133] In the annotation interface, a quick reference button is set next to each category option, and the user can instantly call up the corresponding category feature quick reference card. The card includes:

[0134] · Auditory description (e.g., "high-pitched, continuous, whistle-like");

[0135] · Common location of occurrence (e.g., "mainly in expiration");

[0136] · Description of typical spectrogram features;

[0137] · Key points for distinguishing easily confused categories.

[0138] In a specific embodiment, the category feature quick-check cards are modularly designed to allow for quick retrieval, and can dynamically prioritize the top display of several categories that are most relevant to the current audio prompt.

[0139] By reducing the information retrieval and recall burden on users during the annotation process, the annotation efficiency and accuracy can be further improved.

[0140] 4. Interactive Fusion of Auxiliary Information

[0141] In a preferred embodiment of the present invention, all auxiliary information is deeply integrated with the annotation interface, using the following design approach:

[0142] · Auxiliary information is presented by default, and users can freely choose whether to expand detailed information;

[0143] · During the user's answering process, auxiliary information can be intelligently recommended based on the current audio category. For example, high-frequency abnormal feature prompts are recommended for wheezing samples.

[0144] · Only applicable during the teaching phase, after answering incorrectly, auxiliary information will be replayed immediately to enhance learning effects.

[0145] Through the above-mentioned interactive auxiliary information presentation method, the annotation accuracy and learning experience are effectively improved without increasing the user's cognitive load.

[0146] 5.Technological advantages

[0147] By introducing the auxiliary information annotation method, the present invention can:

[0148] · Improve users' understanding of complex physiological sound characteristics through dual support of visualization and background prompts;

[0149] · By quickly checking category features, we can reduce reliance on short-term memory and alleviate cognitive stress during the labeling process.

[0150] · Through intelligent recommendation mechanism, targeted learning and strengthening of weaknesses can be achieved;

[0151] · It encourages users with non-medical backgrounds to make annotation decisions using logic close to that of professional doctors, thereby improving the accuracy and medical applicability of the dataset.

[0152] In summary, the auxiliary information annotation method of the present invention effectively improves the teaching effect, annotation efficiency and data quality of the system, and is suitable for the task of constructing medical audio datasets for large-scale crowdsourcing.

[0153] Example 4: Implementation of a method for aggregating annotation results based on dynamic weighting

[0154] The present invention provides a method for aggregating annotation results based on dynamic weighting, which is used to fully utilize the ability differences of users with non-medical backgrounds in the crowdsourcing annotation process of medical audio data to achieve high-quality and highly consistent dataset construction.

[0155] The method comprises the following steps:

[0156] 1. Phased labeling strategy

[0157] To adapt to the ability characteristics of non-professional users at different annotation difficulty levels, the present invention proposes a phased annotation mechanism, which specifically includes:

[0158] · Stage 1: Coarse Classification

[0159] Users label the heart and lung sound samples as "normal" or "abnormal".

[0160] In this stage, we take advantage of the higher sensitivity of non-professional users in anomaly detection to preliminarily screen out potential abnormal samples.

[0161] · Phase 2: Sub-classification

[0162] For audio samples marked as "abnormal", the user continues to complete fine-grained classification of abnormal types (such as wheezing, moist rales, dry rales, etc.).

[0163] In a preferred embodiment of the present invention, when the system enters the sub-classification stage, it combines auxiliary information (such as spectrograms, clinical background prompts) to improve the accuracy of sub-classification judgment.

[0164] Through the above two-stage process, the advantages of non-professional users in labeling tasks at different levels are fully utilized, reducing the risk of error propagation in the fine-grained classification process.

[0165] 2. Abnormal labeling detection and filtering mechanism

[0166] To eliminate low-quality or maliciously labeled data, the present invention designs an anomaly detection and automatic filtering mechanism, which specifically includes:

[0167] · Consistency detection: Count the degree of self-consistency of a single user in the same audio annotation. If the consistency is lower than the set threshold, it is judged as an abnormal annotation;

[0168] · Outlier detection: Statistical methods such as Z-score analysis and clustering algorithms are used to detect annotation records that significantly deviate from the majority opinion among all annotations;

[0169] · Annotation response time analysis: If the user response time is abnormally short (such as less than the preset reasonable minimum time), the corresponding annotation record is considered a potential anomaly.

[0170] In one embodiment of the present invention, when an abnormal annotation is detected, the system automatically removes the annotated data or reduces its influence weight on the final label.

[0171] In one embodiment of the present invention, the system mixes several repeated audios into the labeling task. If the user labels the same audio multiple times with different annotations, and the number of times exceeds a threshold, the system automatically removes the user's annotation data or reduces its influence on the final label.

[0172] 3. Dynamic Capability Weight Model

[0173] In order to rationally utilize the annotation results of different users, the present invention introduces a dynamic capability weight model based on historical performance. The specific implementation is as follows:

[0174] · Record each user's annotation accuracy, response time, and error type distribution in different categories during the teaching phase;

[0175] · Dynamically assign each user an annotation weight on different audio samples based on their historical performance;

[0176] · The weight update mechanism is based on a sliding window strategy to ensure that the weights can reflect recent changes in capabilities, balancing stability and flexibility.

[0177] In a preferred embodiment, the higher the user's audio annotation accuracy in a certain category, the higher the weight of the user in the annotation data of that category; conversely, if a certain category performs poorly, the weight is reduced accordingly to reduce its negative impact on the quality of the dataset.

[0178] In one embodiment, the system mixes a portion of the tagged audio into the audio to be tagged, so as to dynamically monitor the user's ability and thereby change the user's tagging weight in real time.

[0179] 4. Two-step crowdsourcing annotation aggregation algorithm

[0180] This paper adopts a two-step aggregation mechanism to further improve the accuracy and stability of the data set. The specific steps are as follows:

[0181] · Step 1: Effective annotation screening

[0182] For each audio sample, count the number of valid annotations after anomaly detection and filtering.

[0183] If the number of valid annotations is lower than the preset threshold (e.g., 5 annotations), the sample is discarded to avoid label deviation caused by a small number of opinions;

[0184] If the valid markup number is reached, proceed to the next step.

[0185] · Step 2: Weighted labeling decisions

[0186] For each valid audio sample, first calculate the total weight of the "normal" and "abnormal" labels;

[0187] If the sum of the normal label weights is greater than the sum of the abnormal label weights, the sample is ultimately labeled as “normal”;

[0188] If the sum of the weights of abnormal labels is greater than the sum of the weights of normal labels, the sample is marked as "abnormal".

[0189] For samples marked as abnormal, the weighted total scores of each category are further calculated in each abnormal category, and the abnormal category with the highest weighted score is selected as the final sub-classification result.

[0190] Through the above two-step aggregation, not only the influence of noise data on the final label is avoided, but also the judgment results of high-ability users are fully utilized at the fine-grained level.

[0191] 5.Technological advantages

[0192] By introducing a dynamic weighted annotation result aggregation method, the present invention can:

[0193] · Leverage the sensitivity of non-professional users in coarse classification (normal / abnormal) tasks to improve initial screening efficiency;

[0194] · Through the sub-classification stage and auxiliary information guidance, the probability of misjudgment of abnormal types is reduced;

[0195] · Through anomaly detection and dynamic weighting mechanism, noisy labeled data is significantly eliminated, improving the consistency and credibility of the data set;

[0196] · While ensuring the ability to expand the scale of the data set, it ensures that the data quality reaches a level close to that of medical experts, providing a reliable foundation for the training of AI-assisted diagnosis models.

[0197] Therefore, the annotation aggregation mechanism based on the two-step crowdsourcing annotation aggregation algorithm is an important innovative link in the present invention to achieve efficient, accurate and low-cost construction of medical audio datasets.

Claims

1. A cardiopulmonary sound annotation and medical education system based on interactive learning and dynamic weighted aggregation, characterized by: include: The mobile user interface module is used to present teaching content, audio annotation tasks and auxiliary reference information to users, and receive user interactions and annotation results; A real-time feedback processing module is used to generate feedback information based on the user's input in the audio classification task, including correct answer prompts, feature difference analysis, and subsequent learning guidance; Auxiliary information generation module, used to dynamically generate and display auxiliary information during the user annotation process, the auxiliary information includes audio spectrogram, patient background information and reference information of each classification feature; User capability assessment and management module, used to collect, count and analyze user behavior data during the annotation process, build user capability profiles and dynamically adjust annotation weights; Dynamic annotation aggregation and optimization module, which is used to screen, weight, filter and fuse annotation results based on user ability weights and output the final audio tags; Data security and privacy protection module, used for data desensitization, rights management, and anomaly detection; The server-side data storage and processing module is used to store user interaction data, audio samples and annotation results, and perform subsequent data analysis and training set construction tasks.

2. The cardiopulmonary sound annotation and medical education system based on interactive learning and dynamic weighted aggregation according to claim 1 is characterized in that: The real-time feedback processing module compares the answers immediately after the user submits the annotations, displays the correct categories and feature comparisons when they are wrong, and dynamically adjusts subsequent training samples.

3. The cardiopulmonary sound annotation and medical education system based on interactive learning and dynamic weighted aggregation according to claim 1 is characterized in that: The auxiliary information generation module generates an audio spectrogram through short-time Fourier transform (STFT) or Mel-frequency cepstral coefficients (MFCC), and provides clinical background prompts and category feature quick-check cards.

4. The cardiopulmonary sound annotation and medical education system based on interactive learning and dynamic weighted aggregation according to claim 1 is characterized in that: The dynamic annotation aggregation and optimization module includes: Anomaly detection unit, used to remove outlier annotations through Z-score analysis or clustering algorithm; The two-step aggregation unit first performs coarse normal / abnormal classification, then performs fine-grained classification of abnormal samples, and calculates the final label based on user weights.

5. A two-stage feedback learning method, applied to the system according to any one of claims 1 to 4, characterized in that: include: In the first stage, users establish physiological sound classification features through standard teaching materials and audio samples; In the second stage, users practice real-time audio classification, and the system generates feedback information immediately after the user answers; Among them, the classification tasks include, in increasing order of difficulty, distinguishing between normal and abnormal sounds, identifying common abnormal sounds, and classifying all categories of abnormal sounds.

6. The two-stage feedback learning method according to claim 5, characterized in that: The second stage adopts a difficulty progression mechanism, sequentially performing normal / abnormal classification, common anomaly identification, and all-category fine-grained classification.

7. An auxiliary information annotation method, applied to the system according to any one of claims 1 to 4, characterized in that: include: Dynamically generate and display audio spectrograms; Provide clinical background information of the patient; Embed category feature quick-check cards in the annotation interface.

8. A method for aggregating annotation results based on dynamic weighting, applied to the system according to any one of claims 1 to 4, characterized in that: include: Staged labeling: first perform coarse normal / abnormal classification, then perform fine abnormal classification; Abnormal annotation filtering: Eliminate low-quality data through consistency detection, group outlier analysis, and response time analysis; Dynamic weighted aggregation: Calculates the final label based on user capability weights.

9. The method according to claim 8, characterized in that The dynamic weighted aggregation specifically includes: If the number of valid annotations is lower than the threshold, the sample is discarded; For valid samples, first calculate the sum of the weights of normal / abnormal labels to determine the major categories; For abnormal samples, the subcategory with the highest weighted score is selected as the final label.

10. A computer-readable storage medium storing a computer program, characterized in that: When the program is executed by a processor, the method according to any one of claims 5 to 9 is implemented.

Images