Medical concept identification method based on three-level noise filtering and multi-modal fusion
By employing a three-level noise filtering and multimodal fusion approach, the problem of noise and dynamic changes in knowledge in unlabeled medical literature is solved, achieving high-precision medical concept recognition and concept evolution tracking. This approach is applicable to medical literature and can be transferred to other fields that require learning dynamic knowledge from weakly supervised or noisy data.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- THE FIRST AFFILIATED HOSPITAL OF ZHENGZHOU UNIV
- Filing Date
- 2026-01-12
- Publication Date
- 2026-04-10
AI Technical Summary
Existing technologies for identifying medical concepts in unlabeled medical academic literature suffer from problems such as high-cost manual annotation, severe noise in remote supervision, and static models that are difficult to adapt to dynamic changes in knowledge, resulting in insufficient accuracy and reliability of medical concept recognition.
A three-level noise filtering and multimodal fusion approach is adopted, including multi-knowledge source fusion, adversarial curriculum learning, and concept drift adaptation, to construct a high-quality training dataset. Medical concept recognition is performed through a heterogeneous multimodal deep learning model, combined with concept evolution tracking and interpretable output.
It achieves recognition accuracy close to that of supervised learning without human annotation, improves the robustness and adaptability of the model, and can automatically track concept changes and discover emerging concepts, meeting the safety and interpretability requirements of medical artificial intelligence.
Smart Images
Figure CN121835673A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical concept recognition technology, and more specifically to a medical concept recognition method based on three-level noise filtering and multimodal fusion. Background Technology
[0002] Medical knowledge graphs are a critical infrastructure supporting precision medicine, clinical decision support systems, and new drug development. Their construction heavily relies on the accurate and efficient extraction of standardized medical concepts from massive amounts of medical literature. Traditional methods face three main bottlenecks when processing unlabeled medical academic literature:
[0003] 1. High cost of manual annotation: High-quality annotated data heavily relies on medical experts, which is costly, time-consuming, and difficult to adapt to the needs of large-scale data processing.
[0004] 2. Severe noise in remote supervision: While remote supervision methods that utilize existing knowledge bases (such as UMLS) for automatic annotation avoid manual annotation, they are prone to introducing a large amount of false positives (incorrect annotations) and false negatives (missed annotations) due to the limitations of the knowledge base's coverage or the inherent biases of the annotation tools. This noise can seriously affect the accuracy and reliability of subsequent model training, thereby reducing the quality of medical concept recognition.
[0005] 3. Static models struggle to adapt to dynamic knowledge evolution: Medical knowledge is constantly updated and evolved as research progresses, and this "concept drift" phenomenon is prevalent. For example, disease definitions may be updated, and new treatments or drugs are constantly emerging. Traditional static models cannot effectively track the dynamic changes in concepts, nor can they discover emerging concepts not yet included in existing knowledge bases, leading to the lag in knowledge graphs.
[0006] Existing technologies also attempt to incorporate self-supervised learning or large language models to address the problem of medical concept recognition. For example, some studies have applied self-supervised learning to continuous learning of chest CT images, or to decouple fine-grained concepts from echocardiograms. Other frameworks utilize large language models to extract medical concepts from electronic health records.
[0007] However, none of these methods specifically address the two core challenges of annotation noise and dynamic concept evolution caused by remote supervision in the specific scenario of "unlabeled academic literature," and fail to propose a systematic end-to-end solution. Especially when faced with massive amounts of unlabeled text data, how to effectively extract real signals from noise and enable the recognition model to continuously adapt to the latest developments in knowledge is a critical problem that urgently needs to be solved in the field of medical concept recognition.
[0008] Therefore, it is necessary to propose a medical concept recognition method based on three-level noise filtering and multimodal fusion to solve the above problems. Summary of the Invention
[0009] The purpose of this invention is to solve the problems mentioned in the background art.
[0010] To achieve the above objectives, the present invention specifically adopts the following technical solution:
[0011] A medical concept recognition method based on three-level noise filtering and multimodal fusion includes the following steps:
[0012] Step S1: Input unlabeled original medical academic literature, perform text cleaning, sentence and word segmentation, and extract the literature's metadata;
[0013] Step S2: Based on improved remote supervision technology, a three-level progressive noise filtering is performed on the preprocessed literature to construct a high-quality medical concept annotation training dataset. The three-level progressive noise filtering includes: the first level is multi-knowledge source fusion and initial confidence screening, which involves calling multiple medical concept annotation tools in parallel and fusing their results, and calculating the initial confidence by combining knowledge graph embedding similarity and concept co-occurrence frequency; the second level is noise identification based on adversarial course learning, which identifies and deweights potential noise samples by generating adversarial network frameworks; and the third level is dynamic data scheduling that is adaptive to concept drift, which monitors the performance changes of the model on time-series data and adjusts the data sampling weights when concept drift is detected.
[0014] Step S3: Using the training dataset constructed in step S2, train a heterogeneous multimodal deep learning model and use the model to identify medical conceptual entities from medical literature;
[0015] Step S4: Post-process the identified medical concept entities, including: linking them to a standard medical knowledge base for normalization and localization, and performing concept evolution tracking analysis to reveal changes in the connotation of the concepts and emerging concepts.
[0016] Furthermore, in step S2, the first-level multi-knowledge source fusion and initial confidence screening specifically includes:
[0017] At least two different automated medical concept annotation tools are invoked in parallel to process the same text; by comparing and fusing the annotation results of different tools, only annotations with high consistency are retained;
[0018] The association between candidate concepts and known concepts is calculated based on the similarity of knowledge graph embedding, and an initial confidence score is calculated for each candidate label based on the co-occurrence frequency of the concept in the corpus.
[0019] Furthermore, in step S2, the second-level noise identification based on adversarial course learning specifically includes:
[0020] Construct a generative adversarial network framework, in which a generator is used to synthesize simulated noisy data, and a discriminator is used to distinguish the cleaned high-quality annotations from the noise synthesized by the generator;
[0021] Adversarial training enables the discriminator to identify noise in real data, and a course learning strategy is used to learn noise patterns from easy to difficult, thereby denoising and reweighting the labels generated in the first level.
[0022] Furthermore, in step S2, the third-level concept drift adaptive dynamic data scheduling specifically includes:
[0023] The literature data was structured into a time-series data stream according to the publication year;
[0024] The performance metrics of the medical concept recognition model are continuously monitored on the latest batch of literature data. When a significant decline in performance is detected, concept drift is determined to have occurred, and the data sampling strategy in subsequent model training is automatically adjusted to increase the sampling weight of recent literature data in order to quickly adapt to the new knowledge distribution. Lightweight incremental learning of the model may also be initiated.
[0025] Furthermore, the architecture of the heterogeneous multimodal deep learning model in step S3 includes a pre-trained language model encoding layer, a convolutional neural network layer, and a conditional random field decoding layer; and introduces an adaptive multimodal fusion gating mechanism, which can dynamically integrate the fusion weights of text features with structured embedded features from medical charts, chemical formulas, and knowledge graphs.
[0026] Furthermore, in step S3, the loss function of the deep learning model adopts an improved focus-based multi-class cross-entropy loss to alleviate the imbalance problem of medical entity categories.
[0027] Furthermore, the concept evolution tracking analysis in step S4 specifically includes: for the same standardized medical concept, according to the time window of literature publication, statistically analyzing its frequency of occurrence, changes in the set of co-occurring concepts, and the evolution of the contextual semantic vector in different time periods; and visualizing it in the form of time series maps or heat maps to reveal the migration of research hotspots and the evolution of the definition boundary of the concept.
[0028] Furthermore, step S4 also includes: for entities that the model identifies with high confidence but cannot be linked to existing standard knowledge bases, marking them as emerging candidate concepts and recording the literature evidence and context in which they appear for further review by domain experts.
[0029] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0030] 1. This invention fundamentally solves the noise pollution problem of remote supervision through an original three-level noise filtering mechanism, achieving recognition accuracy close to that of supervised learning without manual annotation, and improving the model's robustness to noisy data.
[0031] 2. This invention integrates a concept drift detection mechanism and a time-series analysis method. It is not only a static identification tool, but also a dynamic knowledge monitoring mechanism that can automatically track changes in the connotation, extension and research hotspots of concepts, and promptly discover cutting-edge emerging concepts.
[0032] 3. This invention provides a confidence score, source tracing basis, and conceptual evolution context information for each identification result, meeting the stringent requirements of medical artificial intelligence applications for security and interpretability.
[0033] 4. This invention provides a complete technical framework encompassing data preprocessing, noise reduction, model training, post-processing, and knowledge management. This method is not only applicable to medical literature, but its core methodology can also be transferred to other fields requiring the learning of dynamic knowledge from weakly supervised or noisy data, demonstrating broad application prospects. Attached Figure Description
[0034] Figure 1 This is the overall flowchart of the present invention.
[0035] Figure 2 This is a flowchart illustrating the three-stage noise filtering process of the present invention. Detailed Implementation
[0036] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0037] Please see Figures 1 to 2 A medical concept recognition method based on three-level noise filtering and multimodal fusion includes the following steps:
[0038] Step S1: Data Preprocessing. This step receives unlabeled raw medical academic literature as input. First, text cleaning is performed, including removing irrelevant characters, HTML tags, and formatting errors. Then, the cleaned text is segmented into sentences and words, breaking down continuous text into processable linguistic units. Simultaneously, important metadata is extracted from the literature, such as publication year, author information, and journal source. The publication year is crucial for subsequent concept evolution tracing.
[0039] Step S2: Three-level progressive noise filtering and training data construction based on improved remote supervision. This step decouples the traditional single remote supervision process into three progressively advancing, collaborative subsystems, aiming to systematically reduce the noise introduced in remote supervision annotation:
[0040] S2.1: First Level, Multi-Knowledge Source Fusion and Initial Confidence Screening. This step involves parallelly calling multiple different automated medical concept annotation tools, such as MetaMap and cTAKES, to annotate the same text segment with concepts. Each tool identifies medical concepts in the text based on its built-in medical dictionary and rules, and attempts to link them to standard concept identifiers, such as UMLS's CUI. By comparing and fusing the annotation results of different tools—for example, retaining only those concepts annotated with the same CUI by at least two tools—cross-validation is performed to generate initial annotations that are more reliable than those from a single tool. Simultaneously, an initial confidence score is calculated for each candidate annotation. This score can be based on multiple factors, such as the consistency between annotation tools, the centrality of the concept in the existing knowledge graph, calculations using knowledge graph embedding similarity (e.g., calculating the average cosine similarity between the concept's embedding vector and its neighboring concept embedding vectors), and the co-occurrence frequency of the concept in the current corpus—that is, the number of times the concept appears alongside high-frequency medical concepts in its context.
[0041] Furthermore, a multi-dimensional confidence evaluation function is designed to quantify the credibility of each candidate label:
[0042]
[0043] in, This indicates the consistency score among annotation tools. This represents the cosine similarity between a candidate concept and a known concept in the knowledge graph embedding space. Representing the normalized values of co-occurrence frequencies, α, β, These are adjustable weighting coefficients.
[0044] Cosine similarity calculation:
[0045]
[0046] in Let be the embedding vector of the concept in the knowledge graph. This formula integrates tool consistency, semantic relevance, and contextual co-occurrence information, providing a reliable initial signal for subsequent noise identification.
[0047] S2.2: Second-level noise identification based on adversarial curriculum learning. This step constructs a Generative Adversarial Network (GAN) framework to identify and filter noise. The generator... Learning the distribution of simulated noise data, discriminator The algorithm then learns to distinguish between high-quality labeled samples and noisy samples. Its objective function is as follows:
[0048]
[0049] Furthermore, a course learning strategy is introduced, enabling the discriminator to learn noise patterns progressively from easy to difficult. The course scheduling function is as follows:
[0050]
[0051] in For training rounds, , Through progressive learning, the discriminator can gradually improve its ability to identify complex noise and perform denoising and reweighting on the first-stage output.
[0052] The generator learns to simulate the distribution of real-world noise data, for example, by randomly perturbing high-quality labeled data to generate noise samples, or by learning error patterns generated by remote monitoring tools in the real world. The discriminator learns to distinguish between cleaned high-quality labeled data and noise data synthesized by the generator. Through repeated adversarial training, it can effectively identify potential noise samples in the labels generated in stage S2.1 and downweight or remove them directly. This process employs a curriculum learning strategy, whereby the discriminator first learns from easily distinguishable noise patterns, such as entities that are completely inconsistent with the standard knowledge base, or labels with extremely low confidence and inconsistencies across multiple tools. Then, it gradually introduces more complex and subtle noise patterns, thereby improving its noise recognition capabilities.
[0053] S2.3: Level Three, Dynamic Data Scheduling Adaptive to Concept Drift. This step aims to address the dynamic evolution of medical knowledge. Literature data is sorted by publication year to construct a time-series data stream. The system continuously monitors performance changes of the medical concept recognition model on new batches of literature data. If a significant performance decline (i.e., concept drift) is detected, the system will automatically adjust the data sampling strategy. The performance change detection formula is:
[0054]
[0055] like ( If the threshold is set to a preset value, a drift alarm will be triggered. At this point, the system will dynamically increase the sampling weight of recent documents, with the weight update formula being:
[0056]
[0057] in The publication date of the document. This is the attenuation coefficient. Through this mechanism, the model can continuously adapt to changes in knowledge distribution and maintain the cutting edge of its recognition capabilities.
[0058] For example, the system calculates the model's F1 score on a small amount of new data or its matching degree with known high-confidence concepts. When a significant decline in model performance is detected, the system automatically triggers a "concept drift" alert. For drifting data, the system dynamically adjusts the data sampling weights, for example, increasing the sampling probability of recent literature data in subsequent model training. In some cases, the system may also initiate lightweight incremental learning of the model, such as fine-tuning the model on a small amount of new data, to ensure that the training data pool continuously represents the latest distribution of medical knowledge, thereby adapting to the evolution of concepts.
[0059] Step S3: Heterogeneous Multimodal Deep Learning Model Training and Concept Recognition. This step uses the high-quality, low-noise training data produced in stage S2 to train an end-to-end medical named entity recognition (NER) model. This model employs a hybrid architecture of "pre-trained language model (such as BERT) + convolutional neural network (CNN) + conditional random field (CRF)". Specifically, the pre-trained language model is responsible for extracting rich contextual semantic features from the text; the CNN layer is used to capture local feature patterns; and the CRF layer is used to model the dependencies between labels, thereby improving the accuracy of sequence labeling.
[0060] Specifically, this method proposes an adaptive multimodal fusion gating mechanism. Medical literature often contains multimodal information such as figures, diagrams, and chemical structural formulas. To fully utilize this information, this invention proposes an adaptive multimodal fusion gating mechanism. This mechanism can dynamically adjust the fusion weights of different modal features based on the context.
[0061] Specifically, the model receives text features Visual features and knowledge graph embedding The fusion weights are calculated using a learnable gating unit:
[0062]
[0063]
[0064] in: For the sigmoid function, , These are trainable parameters. This mechanism enables the model to establish semantic alignment between textual descriptions and images, as well as structured information, significantly improving its ability to recognize complex medical concepts such as anatomical structures and drug molecules.
[0065] When the input text contains references to non-textual information such as diagrams, chemical structural formulas, etc., for example, "as shown in Table 1", "the structural formula is..." Figure 2 This mechanism dynamically adjusts the fusion weights of textual features and potential multimodal embedding features, such as visual features extracted from images, molecular graph embeddings extracted from chemical structures, or entity embeddings obtained from knowledge graphs. Through a learnable gating unit, it determines when and how to more effectively utilize multimodal information, thereby improving the ability to recognize medical concepts in complex descriptions. The model's loss function employs an improved Focused Multiclass Cross-Entropy Loss (FCCEL), which introduces a moderating factor to assign higher weights to inaccurately classified samples while reducing the weights of easily classified samples, effectively mitigating the imbalance problem in medical entity categories and improving the model's ability to recognize rare medical concepts.
[0066] When medical literature contains figures, diagrams, chemical formulas, or images, the system invokes the corresponding visual model (such as a CNN or graph neural network) to extract visual features and converts them into embedding vectors aligned with the text features. An adaptive gating mechanism dynamically adjusts the weights of text and visual features based on the context; for example, in paragraphs densely packed with figure captions or chemical structure descriptions, the weight of visual features is increased accordingly.
[0067] The model's loss function employs an improved focus-based multi-class cross-entropy loss to mitigate the class imbalance problem in medical entities, and its formula is as follows:
[0068]
[0069] in For the model to class The predicted probability, For category weights, To focus on parameters, this loss function makes the model pay more attention to hard-to-classify samples and rare categories, improving overall recognition robustness.
[0070] Step S4: Concept post-processing, evolution tracing, and interpretability output.
[0071] S4.1: Standardization and Localization. This step links the medical entities identified by the model to standard medical knowledge bases, such as UMLS (Unified Medical Language System), SNOMED CT (System of Clinical Terminology), or CMeKG (Chinese Medical Knowledge Graph), to standardize the concepts. For concepts identified in English literature, the system will further perform Chinese mapping, providing corresponding Chinese names or descriptions.
[0072] S4.2: Concept Evolution Tracking. The system maintains a dynamic concept evolution map. For the same standardized medical concept, the system slices it according to the publication time window, statistically analyzing its frequency of occurrence in different time periods, changes in the set of other concepts co-occurring with the concept, and the evolution of its contextual semantic vector. For example, the system quantifies the degree of semantic drift by calculating the average cosine distance of the concept's contextual word embeddings in different time periods. By analyzing these temporal characteristics, the system can automatically generate evolution reports on the definition, connotation, extension, research hotspots, or changes in related concepts of a specific concept, and visualize its research trends in the form of heatmaps, trend charts, etc.
[0073] S4.3: Emerging Concept Discovery and Interpretability. For entities identified by the model with high confidence, if they cannot be linked to any existing standard medical knowledge base, the system will mark them as "emerging candidate concepts." Simultaneously, the system will record the first literature evidence, contextual information, and frequency changes of these emerging concepts for further review and evaluation by domain experts. Furthermore, the model combines attention mechanisms and feature attribution techniques, such as LIME or SHAP, to generate explanations for each identification result, identifying key evidence fragments or the most contributing words in the text, enhancing the model's credibility and interpretability.
[0074] (1) Semantic drift measurement (reflecting changes in the connotation of a concept):
[0075]
[0076] (2) Similarity of co-occurring concept sets (reflecting changes in the extension of concepts):
[0077]
[0078] (3) Emerging Concept Discovery Scoring (Identifying cutting-edge concepts not yet included):
[0079]
[0080] in To identify confidence levels for the model, and These represent recent and historical frequencies, respectively. High-scoring candidate concepts will be flagged for expert review.
[0081] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. The scope of patent protection of the present invention shall be determined by the claims. Similarly, any equivalent structural changes made based on the content of the present invention's specification shall also be included within the scope of protection of the present invention.
Claims
1. A medical concept recognition method based on three-level noise filtering and multimodal fusion, characterized in that, Includes the following steps: Step S1: Input unlabeled original medical academic literature, perform text cleaning, sentence and word segmentation, and extract the literature's metadata; Step S2: Based on the improved remote supervision technology, a three-level progressive noise filtering is performed on the preprocessed documents to construct a high-quality medical concept annotation training dataset; The three-stage progressive noise filtering includes: The first level of multi-knowledge source fusion and initial confidence screening involves calling multiple medical concept annotation tools in parallel and fusing their results, and then calculating the initial confidence by combining knowledge graph embedding similarity and concept co-occurrence frequency. The second level of noise identification is based on adversarial curriculum learning, which identifies and deweights potential noise samples through a generative adversarial network framework. The third-level concept drift adaptive dynamic data scheduling monitors the model's performance changes on time-series data and adjusts the data sampling weights when concept drift is detected. Step S3: Using the training dataset constructed in step S2, train a heterogeneous multimodal deep learning model and use the model to identify medical conceptual entities from medical literature; Step S4: Post-process the identified medical concept entities, including: linking them to a standard medical knowledge base for normalization and localization, and performing concept evolution tracking analysis to reveal changes in the connotation of the concepts and emerging concepts.
2. The medical concept recognition method based on three-level noise filtering and multimodal fusion according to claim 1, characterized in that, In step S2, the first-level multi-knowledge source fusion and confidence screening specifically involves: At least two different automated medical concept annotation tools are invoked in parallel to process the same text; by comparing and fusing the annotation results of different tools, only annotations with high consistency are retained; The association between candidate concepts and known concepts is calculated based on the similarity of knowledge graph embedding, and an initial confidence score is calculated for each candidate label based on the co-occurrence frequency of the concept in the corpus.
3. The medical concept recognition method based on three-level noise filtering and multimodal fusion according to claim 1, characterized in that, In step S2, the second-level noise identification based on adversarial course learning specifically involves: Construct a generative adversarial network framework, in which a generator is used to synthesize simulated noisy data, and a discriminator is used to distinguish the cleaned high-quality annotations from the noise synthesized by the generator; Adversarial training enables the discriminator to identify noise in real data, and a course learning strategy is used to learn noise patterns from easy to difficult, thereby denoising and reweighting the labels generated in the first level.
4. The medical concept recognition method based on three-level noise filtering and multimodal fusion according to claim 1, characterized in that, In step S2, the third-level concept drift adaptive dynamic data scheduling specifically includes: The literature data was structured into a time-series data stream according to the publication year; The performance metrics of the medical concept recognition model are continuously monitored on the latest batch of literature data. When a significant decline in performance is detected, concept drift is determined to have occurred, and the data sampling strategy in subsequent model training is automatically adjusted to increase the sampling weight of recent literature data in order to quickly adapt to the new knowledge distribution. Lightweight incremental learning of the model may also be initiated.
5. The medical concept recognition method based on three-level noise filtering and multimodal fusion according to claim 1, characterized in that, The architecture of the heterogeneous multimodal deep learning model in step S3 includes a pre-trained language model encoding layer, a convolutional neural network layer, and a conditional random field decoding layer. An adaptive multimodal fusion gating mechanism is introduced, which can dynamically integrate the fusion weights of text features with structured embedded features from medical charts, chemical formulas, and knowledge graphs.
6. The medical concept recognition method based on three-level noise filtering and multimodal fusion according to claim 1, characterized in that, In step S3, the loss function of the deep learning model adopts an improved focal multi-class cross-entropy loss to alleviate the imbalance problem of medical entity categories.
7. The medical concept recognition method based on three-level noise filtering and multimodal fusion according to claim 1, characterized in that, The concept evolution tracking analysis in step S4 specifically includes: for the same standardized medical concept, according to the time window of literature publication, statistically analyzing its frequency of occurrence, changes in the set of co-occurring concepts, and the evolution of the contextual semantic vector in different time periods; and visualizing it in the form of time series maps or heat maps to reveal the migration of research hotspots and the evolution of the definition boundary of the concept.
8. The medical concept recognition method based on three-level noise filtering and multimodal fusion according to claim 1, characterized in that, Step S4 further includes: for entities that the model identifies with high confidence but cannot be linked to existing standard knowledge bases, marking them as emerging candidate concepts and recording the literature evidence and context in which they appear for further review by domain experts.