Adverse reaction report analysis method and system based on deep learning model

Abstract

The application discloses an adverse reaction report analysis method and system based on a deep learning model, acquires a text semantic feature vector, adopts a bidirectional long short-term memory network-condition random field model to identify key entities in the report, extracts the relationship between the entities through a graph convolution network based on an attention mechanism, and constructs a structured heterogeneous graph data; a heterogeneous graph neural network model containing multiple types of nodes and multiple types of edges is constructed, representation learning and a multi-task learning framework based on a meta path are adopted for joint training, and a contrast learning strategy is introduced to optimize the feature discrimination ability; finally, the report to be analyzed is input into the trained model to output a risk prediction result, and intelligent and accurate analysis of the adverse reaction report is realized.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a method and system for analyzing adverse reaction reports based on deep learning models. Background Technology

[0002] Adverse drug reactions (ADDRs) refer to harmful reactions unrelated to the intended use of a qualified drug at normal dosage and administration, and are a significant factor affecting patient medication safety. With the continuous increase in the variety and quantity of drugs, and the growing complexity of combination drug regimens, ADDR monitoring and early warning have become core components of pharmacovigilance. Traditional ADDR monitoring relies primarily on proactive reporting by healthcare professionals, periodic summaries and analyses by marketing authorization holders, and passive collection and analysis by regulatory authorities. With the development of information technology, the pharmacovigilance field has accumulated a large amount of unstructured ADDR reporting data, including individual pharmacovigilance safety reports, free-text descriptions in ICSRs, adverse reaction descriptions in electronic health records, case reports in medical literature, and drug experience sharing on social media. However, current technologies still fail to fully integrate multi-source heterogeneous information at the drug molecular level, clinical semantic level, and individual patient level, resulting in low accuracy and limited comprehensiveness in predictions; problems include high underreporting rates, delayed signal detection, and low analytical efficiency. Summary of the Invention

[0003] The purpose of this invention is to solve the above-mentioned problems by designing an adverse reaction report analysis method and system based on a deep learning model.

[0004] To achieve the above objectives, the technical solution of the present invention further includes the following steps in the above-mentioned adverse reaction report analysis method based on a deep learning model: Obtain multi-source adverse reaction reports, preprocess the multi-source adverse reaction reports, and then use a medical pre-trained language model to perform semantic encoding on the report text and extract the text semantic feature vector; Based on text semantic feature vectors, the BiLSTM-CRF model is used to identify key entities in the report, and then the relationships between entities are extracted through attention graph convolutional networks to construct structured heterogeneous graph data. Drug molecular structure features, adverse reaction semantic features, and patient individualization features are extracted from heterogeneous graph data. The three types of features are adaptively weighted and fused through a cross-modal attention mechanism to generate a multimodal fusion feature vector. A heterogeneous graph with multiple types of nodes and multiple types of edges is constructed. The multimodal fusion feature vector is input into the heterogeneous graph and representation learning is performed using a heterogeneous graph neural network based on meta-paths to obtain a risk analysis model. The report to be analyzed is input into the risk analysis model, which outputs risk prediction results. Attention mechanisms and gradient attribution methods are used to trace and locate key features, generating an interpretable analysis report.

[0005] Furthermore, in the aforementioned adverse reaction report analysis method based on a deep learning model, the step of obtaining multi-source adverse reaction reports, preprocessing the multi-source adverse reaction reports, and then using a medical pre-trained language model to semantically encode the report text and extract text semantic feature vectors includes: Acquire multi-source adverse reaction report data, which includes pharmacovigilance case safety reports, adverse reaction records in electronic health records, adverse reaction descriptions in medical literature, and self-reported adverse drug reactions in social media. The adverse reaction report data is cleaned and standardized, including removing irrelevant characters and noise information, standardizing medical terminology, and performing Chinese word segmentation and part-of-speech tagging to obtain the cleaned report text. The cleaned report text was semantically encoded using a medical pre-trained language model, and the text semantic feature vector was extracted. The medical pre-trained language model is a pre-trained model based on the Transformer architecture, and the training corpus includes medical literature, drug instructions, clinical guidelines and adverse reaction report texts.

[0006] Furthermore, in the aforementioned adverse reaction report analysis method based on a deep learning model, the step of identifying key entities in the report using a BiLSTM-CRF model based on text semantic feature vectors, and then extracting relationships between entities through an attention graph convolutional network to construct structured heterogeneous graph data includes: Based on the text semantic feature vector, the BiLSTM-CRF bidirectional long short-term memory network-conditional random field model is used to perform sequence labeling on the adverse reaction report text and identify key entities in the report. The key entities include drug name, adverse reaction symptoms, patient demographic information, drug dosage and course of treatment, information on concomitant medications and past medical history. Attention-based graph convolutional networks extract inter-entity relationships from identified key entities. These inter-entity relationships include drug-adverse reaction causal relationships, drug-drug interaction relationships, dose-adverse reaction intensity relationships, and time-adverse reaction occurrence relationships. The identified key entities and their relationships are constructed into a structured adverse reaction event representation, generating heterogeneous graph data containing entity nodes and relationship edges.

[0007] Furthermore, in the aforementioned adverse reaction report analysis method based on a deep learning model, the extraction of drug molecular structure features, adverse reaction semantic features, and patient individualization features from heterogeneous graph data, and the adaptive weighted fusion of these three types of features through a cross-modal attention mechanism to generate a multimodal fusion feature vector, includes: The SMILES sequences of drugs in the heterogeneous graph data are obtained, and the drug molecular graph is encoded by a graph attention network to extract the molecular structure feature vector of the drug. The molecular graph is constructed with atoms as nodes and chemical bonds as edges. Semantic embedding vectors of adverse reaction symptoms are obtained from heterogeneous graph data based on medical knowledge graphs, which include various biomedical entities such as diseases, symptoms, drugs, genes, and pathways and their relationships. The patient's age, gender, weight, liver and kidney function indicators, and genotype information are encoded to generate a personalized feature vector for the patient; A cross-modal attention fusion mechanism is used to adaptively weight and fuse the drug molecular structure feature vector, semantic embedding vector, and patient individualized feature vector to generate a multimodal fusion feature vector.

[0008] Furthermore, in the aforementioned adverse reaction report analysis method based on a deep learning model, the construction of a heterogeneous graph with multiple types of nodes and edges, and the input of the multimodal fusion feature vector into the heterogeneous graph for representation learning using a meta-path-based heterogeneous graph neural network to obtain a risk analysis model, includes: A heterogeneous graph neural network model is constructed. The heterogeneous graph neural network model is based on a heterogeneous graph data structure and includes various types of nodes such as drug nodes, adverse reaction nodes, patient nodes, and disease nodes, as well as various types of relationship edges such as drug-adverse reaction relationship edges, drug-drug interaction edges, patient-drug use relationship edges, and disease-adverse reaction association edges. The multimodal fusion feature vector is input into the heterogeneous graph, and the representation learning of the heterogeneous graph is performed using a meta-path-based heterogeneous graph neural network. Neighborhood sampling and feature aggregation are performed in the heterogeneous graph through a preset meta-path template to generate enhanced feature representations for each node. A multi-task learning framework is constructed to jointly train the heterogeneous graph neural network model, and a contrastive learning strategy is adopted to optimize the representation learning of the heterogeneous graph neural network model to obtain a risk analysis model.

[0009] Furthermore, in the aforementioned adverse reaction report analysis method based on a deep learning model, the construction of a multi-task learning framework for joint training of the heterogeneous graph neural network model includes: The adverse reaction classification task is to predict the probability of various adverse reactions of a target drug to a target patient based on enhanced feature representation; The adverse reaction severity prediction task is based on enhanced feature representation to predict the severity level of an adverse reaction that has occurred. The adverse reaction time prediction task is to predict the expected time window for the occurrence of adverse reactions based on enhanced feature representation.

[0010] Furthermore, in the aforementioned adverse reaction report analysis method based on a deep learning model, the step of inputting the report to be analyzed into a risk analysis model, outputting risk prediction results, and using attention mechanisms and gradient attribution methods to trace and locate key features to generate an interpretable analysis report includes: Input the adverse reaction report data to be analyzed into the risk analysis model, and output the adverse reaction risk prediction results; Based on the risk prediction results, the attention mechanism is used to trace and locate the feature with the highest contribution to the model prediction, and an interpretable analysis report is generated. The interpretable analysis report includes the key entities, key relationships and their contribution weights that lead to the prediction results. Adverse reactions are classified and warned according to preset risk thresholds, and a structured analysis report containing risk level, risk factors, and recommended measures is generated.

[0011] Furthermore, in the adverse reaction report analysis system based on a deep learning model, the evaluation system includes the following modules: The multi-source data acquisition module is used to acquire multi-source adverse reaction reports. After preprocessing the multi-source adverse reaction reports, the module uses a medical pre-trained language model to perform semantic encoding on the report text and extract the text semantic feature vector. The heterogeneous graph data acquisition module is used to identify key entities in the report based on text semantic feature vectors, using the BiLSTM-CRF model, and then extracting the relationships between entities through an attention graph convolutional network to construct structured heterogeneous graph data. The feature vector fusion module is used to extract drug molecular structure features, adverse reaction semantic features, and patient individualization features from heterogeneous graph data. It adaptively weights and fuses the three types of features through a cross-modal attention mechanism to generate a multimodal fusion feature vector. The analysis model building module is used to construct a heterogeneous graph with multiple types of nodes and multiple types of edges. The multimodal fusion feature vector is input into the heterogeneous graph and representation learning is performed using a heterogeneous graph neural network based on meta-path to obtain a risk analysis model. The reaction report analysis module is used to input the report to be analyzed into the risk analysis model, output the risk prediction results, and use the attention mechanism and gradient attribution method to trace the source of key features and generate an interpretable analysis report.

[0012] Furthermore, in the adverse reaction report analysis system based on a deep learning model, the heterogeneous graph data acquisition module includes the following sub-modules: The annotation submodule is used to perform sequence annotation on the adverse reaction report text based on the text semantic feature vector and adopt the BiLSTM-CRF bidirectional long short-term memory network-conditional random field model to identify key entities in the report. The key entities include drug name, adverse reaction symptoms, patient demographic information, drug dosage and course of treatment, information on combined medications and past medical history. The extraction submodule is used to extract inter-entity relationships from the identified key entities using a graph convolutional network based on an attention mechanism. These inter-entity relationships include drug-adverse reaction causal relationships, drug-drug interaction relationships, dose-adverse reaction intensity relationships, and time-adverse reaction occurrence relationships. The construction submodule is used to build a structured adverse reaction event representation of the identified key entities and the relationships between entities, and generate heterogeneous graph data containing entity nodes and relationship edges.

[0013] Furthermore, in the adverse reaction report analysis system based on a deep learning model, the reaction report analysis module includes the following sub-modules: The input submodule is used to input the adverse reaction report data to be analyzed into the risk analysis model and output the adverse reaction risk prediction results. The positioning submodule is used to locate the feature with the highest contribution to the model prediction based on the risk prediction result using an attention mechanism, and generate an interpretable analysis report. The interpretable analysis report includes the key entities, key relationships and their contribution weights that lead to the prediction result. The early warning submodule is used to classify and issue early warnings for adverse reactions based on preset risk thresholds, and generate a structured analysis report that includes risk level, risk factors, and recommended measures.

[0014] Its beneficial effects lie in the fact that, through semantic encoding of a medical pre-trained language model and refined sequence annotation of a BiLSTM-CRF model, it can accurately identify key medical entities such as drugs, symptoms, dosages, and medical history from free text. Furthermore, it extracts complex semantic relationships between entities through an attention graph convolutional network, effectively transforming unstructured text into a structured knowledge graph, significantly improving the quality and efficiency of information extraction from adverse reaction reports. Adaptive weighted fusion is achieved through a cross-modal attention mechanism. This multimodal fusion strategy enables the model to comprehensively assess adverse reaction risks simultaneously at the molecular, semantic, and individual levels, significantly improving the accuracy and comprehensiveness of predictions. A joint analysis framework based on heterogeneous graph neural networks is constructed. By constructing a heterogeneous graph containing multiple node types and relational edges such as drugs, adverse reactions, patients, and diseases, and employing a meta-path-based heterogeneous graph neural network for representation learning, this invention can fully utilize the rich and diverse entity relationships in the pharmacovigilance domain. Simultaneously, through a multi-task learning framework, it completes three related tasks: adverse reaction classification, severity prediction, and time prediction. Furthermore, by optimizing representation learning through a contrastive learning strategy, it significantly improves the model's generalization ability and prediction accuracy. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 This is a schematic diagram of the first embodiment of the adverse reaction report analysis method based on a deep learning model in this invention. Figure 2 This is a schematic diagram of the second embodiment of the adverse reaction report analysis method based on a deep learning model in this invention. Figure 3 This is a schematic diagram of the first embodiment of the adverse reaction report analysis system based on a deep learning model in this invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms "one," "an," and "this" used herein may also include the plural forms. It should be further understood that the terminology used in this specification includes the presence of features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0019] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 As shown, the adverse reaction report analysis method based on a deep learning model includes the following steps: Step 101: Obtain multi-source adverse reaction reports. After preprocessing the multi-source adverse reaction reports, use a medical pre-trained language model to perform semantic encoding on the report text and extract the text semantic feature vector. Specifically, this embodiment obtains adverse reaction report data from multiple data sources. These data sources include, but are not limited to: Individual Case Safety Reports (ICSRs), which originate from proactive reporting systems of drug marketing authorization holders, medical institutions, and regulatory authorities, containing structured basic patient information and unstructured adverse event description text; adverse reaction records in electronic health records, originating from medical records, discharge summaries, and outpatient medical records in hospital information systems; adverse reaction descriptions in medical literature, originating from case reports and clinical studies in biomedical journals; and self-reported adverse drug reaction texts on social media, originating from internet platforms such as patient forums and health communities. Due to the wide range of data sources and varying formats, this invention first performs text cleaning and standardization on the obtained report data, specifically including removing irrelevant characters and noise information, standardizing medical terminology (e.g., unifying "myocardial infarction," "myocardial infarction," and "MI" as the standard term "myocardial infarction"), and performing Chinese word segmentation and part-of-speech tagging. Based on this, this invention utilizes a medical pre-trained language model to semantically encode the cleaned report text and extract text semantic feature vectors. This medical pre-trained language model is built on the Transformer architecture. Its training corpus includes large-scale medical domain corpora such as medical literature, drug instructions, clinical guidelines and adverse reaction reports. After domain-adaptive pre-training, it can better understand and encode professional semantic information in medical texts.

[0020] Step 102: Based on the text semantic feature vector, the BiLSTM-CRF model is used to identify key entities in the report, and then the relationships between entities are extracted through the attention graph convolutional network to construct structured heterogeneous graph data. Specifically, in this embodiment, a Bidirectional Long Short-Term Memory Network-Conditional Random Field (BiLSTM-CRF) model is used to perform sequence labeling on adverse reaction report text and identify key entities in the report. The BiLSTM-CRF model can effectively capture the contextual dependencies in the text sequence and optimize the label sequence globally through the conditional random field layer, thereby improving the accuracy of entity recognition. The key entity types identified by this invention include: drug names, such as brand names and generic names; adverse reaction symptoms, such as rash, abnormal liver function, and arrhythmia; patient demographic information, such as age, gender, dosage and course of treatment, information on concomitant medications, and past medical history. After entity recognition, this invention further uses an attention-based graph convolutional network to extract inter-entity relationships from the identified key entities. The extracted relationship types include: causal relationships between drugs and adverse reactions, interaction relationships between drugs, relationships between dosage and adverse reaction intensity, and relationships between time and the occurrence of adverse reactions. Through the above entity recognition and relationship extraction operations, this invention transforms unstructured report text into a structured representation of adverse reaction events, specifically in the form of an initial heterogeneous graph data structure containing entity nodes and relationship edges.

[0021] Step 103: Extract drug molecular structure features, adverse reaction semantic features, and patient individualization features from the heterogeneous graph data. Adaptively weightedly fuse the three types of features through a cross-modal attention mechanism to generate a multimodal fusion feature vector. Specifically, this embodiment extracts features from three different modalities and integrates them through a cross-modal attention fusion mechanism. First, it extracts the molecular structure features of the drug: obtaining the drug's SMILES sequence, a linear representation of molecular structure explicitly described by ASCII strings, converting this sequence into a molecular graph with atoms as nodes and chemical bonds as edges. Then, it encodes the molecular graph using a graph attention network to extract the drug's molecular structure feature vector. The graph attention network calculates the attention coefficients between atomic nodes through a multi-head attention mechanism, focusing on key functional groups and chemical bonds closely related to biological activity in the molecular structure. Second, it extracts the semantic features of adverse reactions: obtaining semantic embedding representations of adverse reaction symptoms based on a medical knowledge graph. This medical knowledge graph contains various biomedical entities such as diseases, symptoms, drugs, genes, and pathways, and their relationships. Through knowledge graph embedding technology, adverse reaction symptoms are mapped to a high-dimensional semantic space, allowing their semantic information to be quantified and calculated. Third, it extracts individualized patient features: encoding the patient's age, gender, weight, liver and kidney function indicators, genotype information, etc., to generate individualized patient feature vectors. After extracting features from the three modalities mentioned above, this invention employs a cross-modal attention fusion mechanism to adaptively weight and fuse the drug molecular structure feature vector, adverse reaction semantic feature vector, and patient individualized feature vector. This fusion mechanism calculates the similarity between a learnable query vector and the features of each modality, and after normalization, obtains the attention weights for each modality, ultimately generating a multimodal fused feature vector. This fused feature vector serves as the initialization feature for each node in the subsequent graph neural network, ensuring that each node carries rich information across the molecular, semantic, and individual dimensions from the initial stage.

[0022] Step 104: Construct a heterogeneous graph with multiple types of nodes and multiple types of edges. Input the multimodal fusion feature vector into the heterogeneous graph and use a heterogeneous graph neural network based on meta-paths to perform representation learning to obtain a risk analysis model. Specifically, this embodiment constructs a heterogeneous graph neural network model. Based on the initial heterogeneous graph data structure described above, this model includes various types of nodes such as drug nodes, adverse reaction nodes, patient nodes, and disease nodes, as well as various types of relationship edges such as drug-adverse reaction relationship edges, drug-drug interaction edges, patient-drug use relationship edges, and disease-adverse reaction association edges. In model construction, this invention employs a meta-path-based heterogeneous graph neural network to learn the representation of the heterogeneous graph. A meta-path is a path template connecting different types of nodes in the heterogeneous graph. For example, a drug-adverse reaction-drug meta-path can capture the potential association between two drugs causing the same adverse reaction, and a patient-drug-adverse reaction meta-path can capture the complete event chain of a specific patient experiencing a specific adverse reaction after taking a specific drug. Through various preset meta-path templates, the model performs neighborhood sampling and feature aggregation in the heterogeneous graph, generating enhanced feature representations for each node. In terms of multi-task learning, this invention constructs a joint learning framework comprising three sub-tasks: the first sub-task is an adverse reaction classification task, which predicts the probability of various adverse reactions to a target drug in a target patient based on node-enhanced feature representations; the second sub-task is an adverse reaction severity prediction task, which predicts the severity level of an already occurring adverse reaction, such as mild, moderate, severe, or life-threatening; and the third sub-task is an adverse reaction timing prediction task, which predicts the expected time window for the occurrence of an adverse reaction, such as immediately after medication, short-term, or long-term. The three sub-tasks share the underlying heterogeneous graph neural network representation learning module, and joint optimization improves the prediction performance of each sub-task. Furthermore, this invention employs a contrastive learning strategy to optimize the model's representation learning capability. This is achieved by constructing positive and negative sample pairs—positive sample pairs being enhanced sample pairs generated from different adverse reaction reports of the same drug through random masking or feature perturbation, and negative sample pairs being sample pairs of adverse reaction reports from different drugs—bringing similar entity representations closer together and pushing dissimilar entity representations apart, thereby enhancing the feature discrimination capability of the risk analysis model.

[0023] Step 105: Input the report to be analyzed into the risk analysis model, output the risk prediction results, and use the attention mechanism and gradient attribution method to trace and locate the key features to generate an interpretable analysis report.

[0024] Specifically, in this embodiment, after the risk analysis model is trained, the adverse reaction report data to be analyzed is input into the trained heterogeneous graph neural network model, and the model outputs the adverse reaction risk prediction results. To enhance the credibility and usability of the model's prediction results, this invention utilizes an attention mechanism to trace and locate the features with the highest contribution to the model's predictions—by calculating the contribution of each input feature to the prediction results using a gradient-based attribution method, and ranking them by contribution to determine the key entities and key relationships that have the greatest impact on the prediction results, thereby generating an interpretable analysis report. This interpretable analysis report includes the key entities, key relationships, and their contribution weights that lead to the prediction results, enabling users to understand why the model makes a certain risk judgment. Finally, this invention performs graded warnings for adverse reactions based on preset risk thresholds, generating a structured analysis report containing risk levels, risk factors, and recommended measures, providing direct reference for clinical decision-making and pharmacovigilance work.

[0025] Its beneficial effects lie in the fact that, through semantic encoding of a medical pre-trained language model and refined sequence annotation of a BiLSTM-CRF model, it can accurately identify key medical entities such as drugs, symptoms, dosages, and medical history from free text. Furthermore, it extracts complex semantic relationships between entities through an attention graph convolutional network, effectively transforming unstructured text into a structured knowledge graph, significantly improving the quality and efficiency of information extraction from adverse reaction reports. Adaptive weighted fusion is achieved through a cross-modal attention mechanism. This multimodal fusion strategy enables the model to comprehensively assess adverse reaction risks simultaneously at the molecular, semantic, and individual levels, significantly improving the accuracy and comprehensiveness of predictions. A joint analysis framework based on heterogeneous graph neural networks is constructed. By constructing a heterogeneous graph containing multiple node types and relational edges such as drugs, adverse reactions, patients, and diseases, and employing a meta-path-based heterogeneous graph neural network for representation learning, this invention can fully utilize the rich and diverse entity relationships in the pharmacovigilance domain. Simultaneously, through a multi-task learning framework, it completes three related tasks: adverse reaction classification, severity prediction, and time prediction. Furthermore, by optimizing representation learning through a contrastive learning strategy, it significantly improves the model's generalization ability and prediction accuracy.

[0026] Please see Figure 2 In the adverse reaction report analysis method based on deep learning models, the report to be analyzed is input into a risk analysis model, which outputs risk prediction results. Attention mechanisms and gradient attribution methods are then used to trace and locate key features, generating an interpretable analysis report. This process includes the following steps: Step 201: Input the adverse reaction report data to be analyzed into the risk analysis model and output the adverse reaction risk prediction results; Step 202: Based on the risk prediction results, use the attention mechanism to trace and locate the features that contribute the most to the model prediction, and generate an interpretability analysis report. The interpretability analysis report includes the key entities, key relationships and their contribution weights that lead to the prediction results. Step 203: Classify and issue early warnings for adverse reactions based on preset risk thresholds, and generate a structured analysis report containing risk level, risk factors, and recommended measures.

[0027] The above describes embodiments of the adverse reaction report analysis method based on a deep learning model of the present invention. Please refer to [link / reference]. Figure 3 In the adverse reaction report analysis system based on a deep learning model, the evaluation system includes the following modules: The multi-source data acquisition module is used to acquire multi-source adverse reaction reports. After preprocessing the multi-source adverse reaction reports, the module uses a medical pre-trained language model to perform semantic encoding on the report text and extract the text semantic feature vector. The heterogeneous graph data acquisition module is used to identify key entities in the report based on text semantic feature vectors, using the BiLSTM-CRF model, and then extracting the relationships between entities through an attention graph convolutional network to construct structured heterogeneous graph data. The feature vector fusion module is used to extract drug molecular structure features, adverse reaction semantic features, and patient individualization features from heterogeneous graph data. It adaptively weights and fuses the three types of features through a cross-modal attention mechanism to generate a multimodal fusion feature vector. The analysis model building module is used to construct heterogeneous graphs with multiple types of nodes and multiple types of edges. The multimodal fusion feature vector is input into the heterogeneous graph and representation learning is performed using a heterogeneous graph neural network based on meta-path to obtain the risk analysis model. The reaction report analysis module is used to input the report to be analyzed into the risk analysis model, output the risk prediction results, and use the attention mechanism and gradient attribution method to trace the source of key features and generate an interpretable analysis report.

[0028] In this embodiment, individual safety report data is retrieved from the pharmacovigilance database of a drug marketing authorization holder. These reports are stored in a combination of structured tables and unstructured narrative text, where the unstructured narrative text records key information such as the occurrence of adverse events, the patient's clinical manifestations, and medication usage. Simultaneously, anonymized outpatient and inpatient medical record data is retrieved from the electronic health record system of a partner hospital, from which relevant records containing descriptions of adverse reactions are extracted. Furthermore, case report literature related to the target drug is retrieved and downloaded from medical literature databases, and patient-reported medication experience texts are collected from publicly available health forums.

[0029] In the data preprocessing stage, the multi-source data was first cleaned by removing HTML tags, special symbols, repeated spaces, and other irrelevant characters; Traditional Chinese was converted to Simplified Chinese; and English medical terms were standardized to standard Chinese translations. Then, medical terminology was standardized, mapping renal failure, renal insufficiency, and other conditions to corresponding levels in the standard terminology system according to their severity. After cleaning and standardization, a pre-trained medical language model based on the Transformer architecture was used to semantically encode the report text. This model underwent further domain-adaptive pre-training on a large-scale medical corpus, including 20 million medical literature abstracts, 100,000 drug instructions, 5 million clinical guideline paragraphs, and 1 million adverse reaction report texts, enabling it to encode each report text paragraph into a 768-dimensional semantic feature vector.

[0030] Based on the obtained text semantic feature vectors, a BiLSTM-CRF model is used for sequence labeling. The model's input is the word sequence of the report text after word segmentation, and the output is the entity label corresponding to each word. The entity label adopts the BIOES labeling system, where B represents the start of an entity, I represents the inside of an entity, O represents a non-entity, E represents the end of an entity, and S represents an entity consisting of a single word. For example, for the text segment "A patient developed a rash after taking amoxicillin," the model outputs the labeled sequence: "Patient / O took / O amoxicillin / B-Drug after / O developed / O rash / S-Symptom," thus identifying amoxicillin as the drug entity and the rash as the symptom entity.

[0031] After entity recognition is completed, a graph convolutional network based on an attention mechanism is used for relation extraction. Using the identified entities as nodes, the model predicts the relationship type between entity pairs. For example, for the entity pair "amoxicillin" and "rash," the model predicts its relationship type as a drug-adverse reaction causal relationship with a confidence level of 0.92; for the entity pair "amoxicillin" and "patient age 65 years," the model extracts the patient-drug use relationship. Through these operations, each adverse reaction report is converted into a structured representation containing several entity nodes and relation edges.

[0032] For the target drug amoxicillin, its SMILES sequence CC1(C)SC2C(NC(=O)C(C3=CC=CC=C3)N)C(=O)N2C1C(=O)O is first obtained and converted into a molecular graph structure. This molecular graph contains 24 atomic nodes and 26 chemical bond edges. A graph attention network is used to encode this molecular graph. The network contains 3 graph attention layers, each with 4 attention heads, and finally outputs a 256-dimensional feature vector of the drug molecule structure.

[0033] For adverse reaction rashes, their semantic embedding representations are obtained from a medical knowledge graph. This knowledge graph contains approximately 5 million biomedical entities and 20 million relationships between entities. The rash nodes are mapped to 256-dimensional semantic feature vectors using graph embedding techniques.

[0034] For individualized patient characteristics, the patient's age (65 years), gender (male), liver function indicators (ALT 45 U / L), and kidney function indicators (eGFR 72 mL / min / 1.73 mcg) were considered. 2 The information is encoded into a 128-dimensional individualized feature vector.

[0035] A cross-modal attention fusion mechanism is employed to fuse the features of the three modalities. The learnable query vector is used to calculate attention scores with the features of each of the three modalities, and the scores are then normalized using Softmax to obtain the weights for each modality. In this embodiment, assuming the weight of the molecular structure feature is 0.35, the weight of the semantic feature is 0.40, and the weight of the individualization feature is 0.25, the final multimodal fused feature vector is the weighted sum of the three, with a dimension of 256.

[0036] A heterogeneous graph was constructed based on all training data. This graph contains approximately 100,000 nodes, including about 5,000 drug nodes, 8,000 adverse reaction nodes, 80,000 patient nodes, and 7,000 disease nodes; it also contains approximately 500,000 relational edges, covering various types such as drug-adverse reaction causal edges, drug-drug interaction edges, patient-drug use edges, and disease-adverse reaction association edges. A meta-path-based heterogeneous graph neural network was used for representation learning. The preset meta-paths include drug-adverse reaction-drug (for discovering drug groups with similar adverse reaction profiles), patient-drug-adverse reaction (for establishing a complete event chain of patient-drug use-adverse reaction), and drug-disease-adverse reaction (for associating drug indications with adverse reactions), totaling five path patterns. The model's representation learning module consists of two heterogeneous graph convolutional layers, each outputting 256-dimensional node features.

[0037] In the multi-task learning framework, the three sub-tasks share the output of the representation learning module described above. The adverse reaction classification task is a multi-class classification problem covering 200 common adverse reaction types, employing a cross-entropy loss function. The severity prediction task is a four-class classification problem (mild, moderate, severe, life-threatening), employing a weighted cross-entropy loss function to address class imbalance. The time prediction task is a three-class classification problem (within 24 hours after medication, 2 to 7 days, and more than 7 days after medication), employing a cross-entropy loss function. The loss functions of the three tasks are summed according to their weight coefficients to form the total loss function, which is determined through grid search. In the contrastive learning strategy, for each drug node, two positive enhancement samples are generated from different adverse reaction reports of that drug using random feature masks, while five negative samples are randomly sampled from adverse reaction reports of other drugs. The contrastive learning loss accounts for 10% of the total loss function.

[0038] A new suspected adverse reaction report was input into the trained model. The report described a 72-year-old female patient who developed diarrhea and fever three days after taking an antibiotic for a lung infection. The model, through entity recognition and relation extraction, identified the drug entity, symptom entity, diarrhea, fever, and patient entities: 72 years old, female. Multimodal fusion feature vectors were generated and then input into a heterogeneous graph neural network for inference. The model's predicted output was: an 87% probability of antibiotic-associated diarrhea, moderate severity, and an expected onset time of 3 to 7 days after medication.

[0039] In the interpretability analysis phase, the contribution of each input feature to the prediction result was calculated using the gradient attribution method. The results showed that the top three features with the highest contributions were: patient age (28%), known association strength between the drug and diarrhea symptoms (35%), and baseline gut microbiota characteristics (18%). Based on this, the model generated an interpretable report indicating that the patient was older and had poor gut microbiota stability. Considering the known gastrointestinal adverse reaction risks of the drug, the report recommended close monitoring of diarrhea symptoms and, if necessary, routine stool examination and microbiota analysis. Simultaneously, based on a preset risk threshold (a probability greater than 80% is considered high-risk), the system marked the report as high-risk and generated a structured analysis report, which was then sent to the pharmacovigilance department.

[0040] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended technical solutions and equivalents.

Claims

1. A method for analyzing adverse reaction reports based on a deep learning model, characterized in that, The adverse reaction report analysis method includes the following steps: Obtain multi-source adverse reaction reports, preprocess the multi-source adverse reaction reports, and then use a medical pre-trained language model to perform semantic encoding on the report text and extract the text semantic feature vector; Based on text semantic feature vectors, the BiLSTM-CRF model is used to identify key entities in the report, and then the relationships between entities are extracted through attention graph convolutional network to construct structured heterogeneous graph data. Drug molecular structure features, adverse reaction semantic features, and patient individualization features are extracted from heterogeneous graph data. The three types of features are adaptively weighted and fused through a cross-modal attention mechanism to generate a multimodal fusion feature vector. A heterogeneous graph with multiple types of nodes and multiple types of edges is constructed. The multimodal fusion feature vector is input into the heterogeneous graph and representation learning is performed using a heterogeneous graph neural network based on meta-paths to obtain a risk analysis model. The report to be analyzed is input into the risk analysis model, which outputs risk prediction results. Attention mechanisms and gradient attribution methods are used to trace and locate key features, generating an interpretable analysis report.

2. The adverse reaction report analysis method based on a deep learning model as described in claim 1, characterized in that, The process of obtaining multi-source adverse reaction reports involves preprocessing the reports and then using a medical pre-trained language model to semantically encode the report text, extracting semantic feature vectors, including: Acquire multi-source adverse reaction report data, which includes pharmacovigilance case safety reports, adverse reaction records in electronic health records, adverse reaction descriptions in medical literature, and self-reported adverse drug reactions in social media. The adverse reaction report data is cleaned and standardized, including removing irrelevant characters and noise information, standardizing medical terminology, and performing Chinese word segmentation and part-of-speech tagging to obtain the cleaned report text. The cleaned report text was semantically encoded using a medical pre-trained language model, and the text semantic feature vector was extracted. The medical pre-trained language model is a pre-trained model based on the Transformer architecture, and the training corpus includes medical literature, drug instructions, clinical guidelines and adverse reaction report texts.

3. The adverse reaction report analysis method based on a deep learning model as described in claim 1, characterized in that, The process involves identifying key entities in the report using a BiLSTM-CRF model based on textual semantic feature vectors, then extracting relationships between entities through an attention graph convolutional network to construct structured heterogeneous graph data, including: Based on the text semantic feature vector, the BiLSTM-CRF bidirectional long short-term memory network-conditional random field model is used to perform sequence labeling on the adverse reaction report text and identify key entities in the report. The key entities include drug name, adverse reaction symptoms, patient demographic information, drug dosage and course of treatment, information on concomitant medications and past medical history. Attention-based graph convolutional networks extract inter-entity relationships from identified key entities. These inter-entity relationships include drug-adverse reaction causal relationships, drug-drug interaction relationships, dose-adverse reaction intensity relationships, and time-adverse reaction occurrence relationships. The identified key entities and their relationships are constructed into a structured adverse reaction event representation, generating heterogeneous graph data containing entity nodes and relationship edges.

4. The adverse reaction report analysis method based on a deep learning model as described in claim 1, characterized in that, The extracted drug molecular structure features, adverse reaction semantic features, and patient individualization features from the heterogeneous graph data are adaptively weighted and fused using a cross-modal attention mechanism to generate a multimodal fused feature vector, including: The SMILES sequences of drugs in the heterogeneous graph data are obtained, and the drug molecular graph is encoded by a graph attention network to extract the molecular structure feature vector of the drug. The molecular graph is constructed with atoms as nodes and chemical bonds as edges. Semantic embedding vectors of adverse reaction symptoms are obtained from heterogeneous graph data based on medical knowledge graphs, which include various biomedical entities such as diseases, symptoms, drugs, genes, and pathways and their relationships. The patient's age, gender, weight, liver and kidney function indicators, and genotype information are encoded to generate a personalized feature vector for the patient; A cross-modal attention fusion mechanism is used to adaptively weight and fuse the drug molecular structure feature vector, semantic embedding vector, and patient individualized feature vector to generate a multimodal fusion feature vector.

5. The adverse reaction report analysis method based on a deep learning model as described in claim 1, characterized in that, The process of constructing a heterogeneous graph with multiple node and edge types, inputting the multimodal fusion feature vector into the heterogeneous graph, and performing representation learning using a meta-path-based heterogeneous graph neural network to obtain a risk analysis model includes: A heterogeneous graph neural network model is constructed. The heterogeneous graph neural network model is based on a heterogeneous graph data structure and includes various types of nodes such as drug nodes, adverse reaction nodes, patient nodes, and disease nodes, as well as various types of relationship edges such as drug-adverse reaction relationship edges, drug-drug interaction edges, patient-drug medication relationship edges, and disease-adverse reaction association edges. The multimodal fusion feature vector is input into the heterogeneous graph, and the representation learning of the heterogeneous graph is performed using a meta-path-based heterogeneous graph neural network. Neighborhood sampling and feature aggregation are performed in the heterogeneous graph through a preset meta-path template to generate enhanced feature representations for each node. A multi-task learning framework is constructed to jointly train the heterogeneous graph neural network model, and a contrastive learning strategy is adopted to optimize the representation learning of the heterogeneous graph neural network model to obtain a risk analysis model.

6. The adverse reaction report analysis method based on a deep learning model as described in claim 5, characterized in that, The construction of a multi-task learning framework for joint training of the heterogeneous graph neural network model includes: The adverse reaction classification task is to predict the probability of various adverse reactions of a target drug to a target patient based on enhanced feature representation; The adverse reaction severity prediction task is based on enhanced feature representation to predict the severity level of an adverse reaction that has occurred. The adverse reaction time prediction task is to predict the expected time window for the occurrence of adverse reactions based on enhanced feature representation.

7. The adverse reaction report analysis method based on a deep learning model as described in claim 5, characterized in that, The process of inputting the report to be analyzed into the risk analysis model, outputting risk prediction results, and using attention mechanisms and gradient attribution methods to trace and locate key features to generate an interpretable analysis report includes: Input the adverse reaction report data to be analyzed into the risk analysis model, and output the adverse reaction risk prediction results; Based on the risk prediction results, the attention mechanism is used to trace and locate the feature with the highest contribution to the model prediction, and an interpretable analysis report is generated. The interpretable analysis report includes the key entities, key relationships and their contribution weights that lead to the prediction results. Adverse reactions are classified and warned according to preset risk thresholds, and a structured analysis report containing risk level, risk factors, and recommended measures is generated.

8. An adverse reaction report analysis system based on a deep learning model, characterized in that, The adverse reaction report analysis system includes the following modules: The multi-source data acquisition module is used to acquire multi-source adverse reaction reports. After preprocessing the multi-source adverse reaction reports, the report text is semantically encoded using a medical pre-trained language model to extract the text semantic feature vector. The heterogeneous graph data acquisition module is used to identify key entities in the report based on text semantic feature vectors, using the BiLSTM-CRF model, and then extracting the relationships between entities through an attention graph convolutional network to construct structured heterogeneous graph data. The feature vector fusion module is used to extract drug molecular structure features, adverse reaction semantic features, and patient individualization features from heterogeneous graph data. It adaptively weights and fuses the three types of features through a cross-modal attention mechanism to generate a multimodal fusion feature vector. The analysis model building module is used to construct a heterogeneous graph with multiple types of nodes and multiple types of edges. The multimodal fusion feature vector is input into the heterogeneous graph and representation learning is performed using a heterogeneous graph neural network based on meta-path to obtain a risk analysis model. The reaction report analysis module is used to input the report to be analyzed into the risk analysis model, output the risk prediction results, and use the attention mechanism and gradient attribution method to trace the source of key features and generate an interpretable analysis report.

9. The adverse reaction report analysis system based on a deep learning model as described in claim 8, characterized in that, The heterogeneous graph data acquisition module includes the following sub-modules: The annotation submodule is used to perform sequence annotation on the adverse reaction report text based on the text semantic feature vector and adopt the BiLSTM-CRF bidirectional long short-term memory network-conditional random field model to identify key entities in the report. The key entities include drug name, adverse reaction symptoms, patient demographic information, drug dosage and course of treatment, information on combined medications and past medical history. The extraction submodule is used to extract inter-entity relationships from the identified key entities using a graph convolutional network based on an attention mechanism. These inter-entity relationships include drug-adverse reaction causal relationships, drug-drug interaction relationships, dose-adverse reaction intensity relationships, and time-adverse reaction occurrence relationships. The construction submodule is used to build a structured adverse reaction event representation of the identified key entities and the relationships between entities, and generate heterogeneous graph data containing entity nodes and relationship edges.

10. The adverse reaction report analysis system based on a deep learning model as described in claim 8, characterized in that, The reaction report analysis module includes the following sub-modules: The input submodule is used to input the adverse reaction report data to be analyzed into the risk analysis model and output the adverse reaction risk prediction results. The positioning submodule is used to locate the feature with the highest contribution to the model prediction based on the risk prediction result using an attention mechanism, and generate an interpretable analysis report. The interpretable analysis report includes the key entities, key relationships and their contribution weights that lead to the prediction result. The early warning submodule is used to classify and issue early warnings for adverse reactions based on preset risk thresholds, and generate a structured analysis report that includes risk level, risk factors, and recommended measures.

Images