Method and device for automatic generation of cause-of-death chain and determination of underlying cause of death based on intelligent inference of fusion multi-model
By employing a four-level screening process and multi-model fusion of electronic medical record data, an interpretable cause-of-death chain is generated, solving the accuracy and efficiency issues in cause-of-death determination in existing technologies and achieving high-quality automated generation and determination of cause-of-death chains.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI MUNICIPAL CENT FOR DISEASE CONTROL & PREVENTION
- Filing Date
- 2026-05-06
- Publication Date
- 2026-06-05
AI Technical Summary
Existing technologies for determining the cause of death suffer from problems such as unstable data quality, inaccurate coding selection, black-box reasoning process, and mismatched probability models, resulting in low accuracy and efficiency in determining the underlying cause of death.
A four-level sequential screening mechanism is used to preprocess electronic medical record data. Combined with a retrieval-enhanced generation framework and a hybrid probability model with time-series-causal dual constraints, candidate cause-of-death chains are generated. These chains are then evaluated by a hard rule engine and a controllable reasoning engine to ultimately achieve an interpretable determination of the root cause of death.
It significantly improves the accuracy of ICD encoding mapping and the accuracy of fundamental cause of death determination, automates the cause of death chain generation process, improves work efficiency, and continuously optimizes model performance through human-machine collaborative closed-loop learning.
Smart Images
Figure CN122158094A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the intersection of medical information technology and artificial intelligence. Specifically, it relates to a method, electronic device, and computer-readable storage medium for automated and interpretable generation of cause-of-death chains and determination of the underlying cause of death using large language models, retrieval-enhanced generation, graph neural networks, and multi-source data fusion technologies. Background Technology
[0002] The standardized completion of medical certificates for death and the accurate determination of the underlying cause of death are the cornerstones of public health decision-making, disease burden assessment, and optimal allocation of health resources. Currently, this work heavily relies on the professional knowledge and experience of clinicians and faces numerous challenges.
[0003] On the one hand, there are limitations in data sources and application models. In many regions, information on the cause of death still relies on post-mortem manual investigations, which not only consumes a lot of human and material resources, but also often results in data bias due to recall, leading to unstable data quality. Existing auxiliary technical solutions, such as the method disclosed in prior art 1 (CN120450052A), are mostly used as tools for post-mortem data quality control, failing to intervene at the source of death certificate issuance and thus unable to fundamentally solve the data quality problem.
[0004] On the other hand, existing technologies have significant shortcomings at the algorithmic level. In the mapping stage of the International Classification of Diseases (ICD) coding, existing technologies largely rely on simple semantic retrieval, lacking a multi-dimensional comprehensive evaluation of clinical context, temporal logic, and historical statistical data, making it difficult to guarantee the accuracy of code selection. The reasoning process suffers from a "black box" problem. For example, existing technology 1 uses a single-round, comprehensive prompt word to guide a large language model to directly generate conclusions. This approach is prone to logical jumps or "illusions" in complex cases involving multiple etiologies, and the decision-making process is untraceable and difficult to audit, making it difficult for doctors to fully trust its output. Probabilistic modeling methods suffer from mismatch problems. Existing technology 1 uses the traditional N-Gram model to calculate the probability of the cause-of-death chain, but this model struggles to effectively handle long-distance temporal dependencies, data sparsity, and complex causal relationships in the disease development process, leading to inaccurate probability estimates and consequently affecting the reliability of the underlying cause of death inference.
[0005] Therefore, there is an urgent need for a fully automated cause-of-death chain generation technology that can deeply integrate multi-source objective clinical data, achieve precise intelligent mapping of ICD codes, perform controllable and interpretable reasoning, and have continuous learning capabilities, thereby improving the objectivity, accuracy, and efficiency of cause-of-death reporting at the source. Summary of the Invention
[0006] The purpose of this invention is to overcome the above-mentioned defects of the prior art and provide a method and system for automatic generation of cause-of-death chains and determination of the root cause of death that integrates multi-model intelligent reasoning, so as to solve the technical problems in the prior art such as fuzzy ICD coding judgment standards, black box reasoning process, mismatch of probability models and lack of efficient human-machine collaborative closed loop.
[0007] To achieve the above objectives, the present invention provides a method for automatically generating a cause-of-death chain and determining the underlying cause of death, comprising: The system obtains patients' electronic medical record data from the hospital information system and performs structured preprocessing on the electronic medical record data. Based on the preprocessed electronic medical record data, a four-level sequential screening mechanism is adopted to map clinical texts into a set of candidate ICD codes. The four-level sequential screening mechanism includes initial screening of semantic matching degree, fine screening of clinical context fit degree, verification of time series rationality and optimization of historical statistical support degree. A retrieval-enhanced generation framework is employed to retrieve domain knowledge related to the electronic medical record data from a locally deployed medical knowledge base; a dynamic multi-round controllable inference engine is driven to perform causal relationship inference on the candidate ICD codes in combination with the domain knowledge, generating at least one candidate cause of death chain; A hybrid probabilistic model with time-series and causal dual constraints is used to evaluate the at least one candidate cause-of-death chain. The hybrid probabilistic model includes a Transformer branch for learning the time-series dependence of the disease and a graph neural network branch for modeling the causal association of the disease. Combined with a hard rule engine, the comprehensive probability and corresponding confidence of each candidate cause-of-death chain are calculated to determine the underlying cause of death. The candidate cause-of-death chain is sorted based on the comprehensive probability and the confidence level, and the sorting results are presented through a visualization interface; user feedback on the sorting results is received, and the feedback is converted into structured data for incremental updates to the hybrid probability model and the medical knowledge base.
[0008] In some embodiments, the four-level sequential screening mechanism specifically includes: performing the initial semantic matching screening, calculating the semantic similarity score between the disease description in the clinical text and the standard interpretation of the ICD code using a biomedical pre-trained language model, and retaining ICD codes with scores higher than a first preset threshold; performing the fine screening of clinical context fit based on the initial screening results, using a context-aware model to evaluate the logical conformity of each ICD code with the patient's age, gender, past medical history, comorbidities, and department of visit, and removing ICD codes with conformity lower than a second preset threshold; performing the time series rationality verification, mapping the remaining ICD codes and their corresponding onset times to a disease evolution time series diagram, and verifying whether there are logical conflicts in the time series diagram according to a preset medical causal logic rule base, retaining conflict-free ICD codes; and performing the historical statistical support optimization, querying the historical case database, calculating the co-occurrence frequency of the remaining ICD codes in similar cases, and reordering the candidate ICD codes.
[0009] In some embodiments, the time series rationality verification further includes: constructing the disease evolution time series graph as a directed acyclic graph, where nodes represent disease events corresponding to the carefully selected ICD codes, and directed edges represent the chronological order of disease events; loading a set of time series constraint rules from the medical causal logic rule base, each of the time series constraint rules defining a prohibited chronological relationship between a pair of disease events; traversing all directed edges in the directed acyclic graph to check for edges that violate the time series constraint rules; and calculating a time series conflict score for each ICD code. The timing conflict scoring The number of non-compliant edges formed by the ICD encoding is proportional to the number of non-compliant edges; the timing conflict score is... Below the preset timing conflict threshold The ICD code is used as the verified code and passed to the historical statistical support optimization step.
[0010] In some embodiments, the step of evaluating the candidate cause-of-death chains using the hybrid probability model includes: inputting the ICD codes and their corresponding timestamps in the candidate cause-of-death chains into the Transformer branch, and having the Transformer branch output a temporal probability characterizing the rationality of the temporal evolution of the cause-of-death chain. The ICD codes in the candidate cause-of-death chains are used as nodes, and causal relationship edges between these nodes are extracted based on a pre-constructed global disease association knowledge graph to form a candidate causal subgraph. The candidate causal subgraph is then input into the graph neural network branch, which outputs a causal probability representing the causal logic strength of the cause-of-death chain. The candidate cause-of-death chain is submitted to the hard rule engine, which performs matching and verification according to the cause-of-death determination rules published by the World Health Organization, and outputs a rule verification score. .
[0011] In some embodiments, the comprehensive probability is calculated by using a preset gated multiplication formula. The formula is: in, A score used to characterize the overall plausibility of the cause-of-death chain; The rule gating coefficients output by the hard rule engine; the rule gating coefficients It is a binary variable: it takes the value of 1 when the candidate cause of death chain fully conforms to the cause of death determination rule, and takes the value of 0 when there is any violation of the rule; the fundamental cause of death is determined by the comprehensive score of the fundamental cause of death of each node in the chain, and the comprehensive score is calculated by weighting the prediction probability of the node, time weight and rule conformity of the model.
[0012] In some embodiments, the step of the retrieval enhancement generation framework retrieving relevant domain knowledge from the medical knowledge base includes: using a sparse retrieval engine to perform a retrieval based on keywords in the electronic medical record data to obtain a first set of relevant knowledge fragments; using a dense vector retrieval engine to encode the electronic medical record data into query vectors, and performing a similarity retrieval in a vector database to obtain a second set of relevant knowledge fragments; and using a cross-attention reordering model to fuse and reorder the first and second sets of relevant knowledge fragments, and selecting the knowledge fragments most relevant to the current case as the domain knowledge.
[0013] In some embodiments, the step of the dynamic multi-round controllable inference engine generating the candidate cause-of-death chain includes: using a named entity recognition model to extract medical entities of disease, symptoms, examination, and treatment from the electronic medical record data; constructing the relationship graph based on the medical entities and the candidate ICD codes, wherein the nodes of the graph are the medical entities and ICD codes, and the edges are the association relationships determined based on timestamps and medical logic; driving a controlled large language model to perform restricted decoding with the relationship graph as a constraint condition, generating a text sequence that conforms to the graph topology and causal logic, and parsing the text sequence into the candidate cause-of-death chain.
[0014] In some embodiments, the step of converting the feedback operation into structured data for incremental updates includes: if the feedback operation is that the user has confirmed the correct one among multiple candidate cause-of-death chains, then the correct candidate cause-of-death chain is used as a positive sample, and the rest are used as negative samples for fine-tuning the hybrid probability model; if the feedback operation is that the user has modified the node order or relationship in a candidate cause-of-death chain, then the modified structure is converted into a new rule and added to the rule base of the hard rule engine.
[0015] To achieve the above objectives, the present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory, wherein the processor executes the program to implement the method described above.
[0016] To achieve the above objectives, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.
[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. Significantly improves accuracy and consistency: By directly connecting to objective electronic medical record data from hospitals and employing an innovative four-level sequential progressive screening mechanism and a hybrid probability model with time-series and causal dual constraints, this invention can significantly improve the accuracy of ICD coding mapping and the accuracy of determining the root cause of death, thereby ensuring the high quality and regional comparability of cause-of-death data from the source.
[0018] 2. Significantly improve work efficiency: The complex process of generating and determining the cause of death is automated, and the time for filling out and coding a single death certificate is greatly reduced from tens of minutes to minutes. Clinicians only need to perform quick review and confirmation, which improves work efficiency by dozens of times and effectively liberates clinical productivity.
[0019] 3. Strong interpretability and high acceptability: The dynamic multi-round controllable inference engine designed in this invention "white-boxes" the reasoning process, making each decision traceable. The system's output of multiple candidate chains, confidence scores, and complete evidence chains makes the AI decision-making process transparent and credible, and the expected acceptance rate among doctors is over 90%.
[0020] 4. Possesses continuous evolution capabilities: It has constructed a technical and business closed loop of "intelligent generation - human-machine collaboration - feedback learning". The system can transform doctors' feedback operations into structured data, which is used for incremental updates and continuous optimization of the model and knowledge base, enabling it to continuously absorb expert experience, adapt to medical advancements, and maintain optimal performance in the long term. Attached Figure Description
[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0022] Figure 1 This is a diagram illustrating the overall system architecture and workflow provided in this embodiment of the invention.
[0023] Figure 2 This is a flowchart of the four-level serial screening process for ICD encoding provided in an embodiment of the present invention.
[0024] Figure 3 This is a schematic diagram of the optimized RAG framework and dynamic multi-round controllable inference process provided in the embodiments of the present invention.
[0025] Figure 4 This is the temporal-causal dual-constraint hybrid probability model architecture and information flow graph provided in the embodiments of the present invention.
[0026] Figure 5 This is a schematic diagram of an example of a human-machine collaborative review interface provided in an embodiment of the present invention. Detailed Implementation
[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0028] Example 1 This embodiment provides a method for automatically generating a cause-of-death chain and determining the underlying cause of death. For example... Figure 1 As shown, this method constructs a closed-loop intelligent system encompassing "data fusion and preprocessing → ICD-encoded intelligent mapping → knowledge enhancement and controllable reasoning → hybrid probabilistic modeling and decision-making → human-machine collaboration and closed-loop learning." This system is used to automatically and accurately generate cause-of-death chains that conform to medical logic and international disease classification standards, directly starting from objective electronic medical record data in the hospital information system, and provides decision support for determining the underlying cause of death.
[0029] Step S201: Data Acquisition and Preprocessing The system interfaces with the hospital's information system to securely acquire the electronic medical record (EMR) data of designated patients. This EMR data is an objective clinical record, which may include outpatient and emergency history, inpatient progress notes, diagnostic information, laboratory test reports, key findings from imaging examinations, and timelines of medical orders.
[0030] After acquiring the data, the processor performs structured preprocessing on the electronic medical record data. This preprocessing process is used to form a high-quality, unified, standardized data pool, and the specific operations include: (1) Data denoising: Remove redundant, formatted or clinically irrelevant information from the data.
[0031] (2) Standardization of medical terminology: Based on a pre-set clinical thesaurus, non-standard and colloquial medical terms are standardized into standard terms.
[0032] (3) Time information extraction and normalization: Extract time-related descriptions from the text and unify them into a standardized time format or quantitative representation.
[0033] (4) Intelligent imputation of missing values: For missing key information, machine learning models can be used to imput it reasonably based on the common development patterns of similar diseases and historical case patterns.
[0034] To ensure data security, the system strictly adheres to data privacy protection methods throughout the entire processing, such as employing data anonymization techniques, implementing role-based access control, and recording full operation logs. When model training is involved, technologies such as national cryptographic algorithms are used to ensure that training data does not leave the original medical institution.
[0035] Step S202: ICD encoding mapping based on a four-level cascaded progressive filtering mechanism After obtaining the preprocessed electronic medical record data, the system employs an innovative four-level sequential filtering mechanism to precisely map clinical text descriptions to a set of candidate ICD codes. For example... Figure 2 As shown, this mechanism simulates the decision-making process of medical experts from broad screening to precise judgment, specifically including: Level 1: Initial Semantic Matching Screening. The goal of this stage is to quickly identify all semantically relevant codes from the complete ICD encoding library. The processor uses a pre-trained biomedical language model to calculate the semantic similarity score between disease descriptions in clinical texts and standard ICD encoding interpretations. Subsequently, the system retains all ICD codes with scores higher than a first preset threshold as candidates. Through this step, the number of codes to be processed can be significantly reduced from tens of thousands to hundreds.
[0036] Level Two: Clinical Context Adaptability Screening. This stage aims to eliminate codes that, while semantically relevant, are inconsistent with the specific circumstances of the current patient. The processor utilizes a context-aware model, combined with the patient's age, gender, past medical history, comorbidities, and the department visited, to validate candidate codes using a hard rule base. The system will discard ICD codes with a logical conformity below a second preset threshold.
[0037] Level 3: Time Series Rationality Verification. This stage ensures that the selected encoding matches the disease's developmental stage and temporal logic. The processor analyzes the temporal clues in the medical record and verifies whether these clues are consistent with the temporal attributes of the ICD encoding itself.
[0038] In one specific implementation, this verification step is further refined. The system constructs a directed acyclic graph (DAG) using the carefully selected ICD codes and their corresponding onset times, where nodes represent disease events and directed edges represent the chronological order of events. The processor loads a set of temporal constraint rules from a medical causal logic rule library. Subsequently, it traverses all directed edges in the graph, checking for edges that violate the temporal constraint rules. The processor calculates a temporal conflict score for each ICD code. The score is proportional to the number of non-compliant edges formed by the encoding. Finally, only those with temporal conflict scores below a preset threshold are retained. The ICD code is passed to the next level.
[0039] Level 4: Historical Statistical Support Optimization. This stage leverages expert consensus from historical cases to select the best among the best. The processor queries a large-scale, high-quality historical case coding database to calculate the co-occurrence frequency or usage frequency of the remaining candidate ICD codes in previous cases similar to the current case. Based on this statistical support, the candidate codes are reordered, prioritizing the codes with the highest historical support.
[0040] Through the above four-level sequential screening, the system finally outputs a set of highly relevant candidate ICD codes that have undergone multi-dimensional verification.
[0041] It is worth mentioning that, in the above four-level screening process, the system introduces adaptive backoff and dynamic threshold mechanisms to ensure the robustness of the system: Dynamic threshold setting: In the fourth level of historical statistical support optimization, the threshold is not a fixed value, but is dynamically calculated based on the type of candidate ICD codes (such as common diseases or rare diseases) and the amount of historical data samples. For rare disease codes, the system will automatically lower the support threshold to avoid incorrect filtering due to sparse historical samples.
[0042] Fallback Strategy: If the candidate set is empty after any level of screening (e.g., level N), the system will trigger a fallback mechanism. Specific strategies include: automatically falling back to the candidate set of the previous level (level N-1) and selecting the highest-scoring code from that level; or automatically lowering the screening threshold of the current level (level N) (e.g., reducing the threshold by 10%) and re-performing the screening. This mechanism ensures that when faced with atypical or vaguely described clinical text, the system can still output the closest valid code, rather than directly reporting an error or outputting an empty result.
[0043] In the initial semantic matching screening step, to improve the recall rate, the system first expands the extracted medical terms with synonyms, mapping clinical colloquialisms to a set of standard medical terms. Then, semantic similarity is calculated. In this embodiment, the first preset threshold is set to 0.7, meaning that ICD codes with a similarity score greater than or equal to 0.7 are retained. In the time series reasonableness verification step, the system calculates the time reasonableness score and retains codes with scores higher than 0.75 (i.e., the normalized complement of the time series conflict threshold) to ensure the rigor of the time logic.
[0044] Step S203: Generation of candidate cause-of-death chains based on optimized RAG and controllable reasoning After obtaining the candidate ICD codes, the system enters the knowledge enhancement and reasoning stage to generate a candidate cause-of-death chain that conforms to medical logic. For example... Figure 3 As shown, the retrieval process is initiated by the query builder. The query builder not only generates vectors for dense retrieval but also constructs structured query statements containing Boolean logic. The system employs a hybrid retrieval strategy, performing vector semantic retrieval and precise keyword retrieval in parallel to ensure that crucial information is not lost when dealing with rare diseases or specific terms.
[0045] This process consists of two parts: 1. Optimized retrieval enhancement generation framework: To overcome the potential issues of knowledge lag or "illusion" inherent in large language models, the system first employs an optimized retrieval-enhanced generation framework to retrieve the most relevant domain knowledge for the current case from a locally deployed, specialized medical knowledge base. This knowledge base includes ICD coding rules, clinical guidelines, and a large number of expert-reviewed cases of cause-of-death determination.
[0046] The retrieval process employs a hybrid retrieval method, executing two retrieval paths in parallel: A precise retrieval engine based on structured queries: Unlike traditional full-text keyword matching, this step first uses named entity recognition technology to extract disease entities and logical relation words from electronic medical records and maps them to ICD schema symbols. Then, the system constructs structured query statements containing Boolean logic, such as a query in the format (entity:A OR ICD:CodeA) AND (entity:B OR ICD:CodeB) AND (relation:cause). The system uses this structured query to perform precise matching in a relational database storing ICD rule tables and historical case libraries, thereby accurately retrieving rule fragments with strong business relevance.
[0047] Dense Vector Retrieval Engine: It uses a sentence embedding model to encode electronic medical record data into query vectors, and then performs efficient similarity retrieval in a vector database to obtain a second set of semantically related knowledge fragments.
[0048] The results from both searches are fed into a cross-attention reordering model. This model, specifically trained on death determination scenario data, can intelligently fuse and reorder all retrieved knowledge fragments based on the complete context of the current case, ultimately selecting the Top-K knowledge fragments most relevant to the current case as the domain knowledge required for subsequent reasoning.
[0049] 2. Dynamic multi-turn controllable inference engine: The system drives a dynamic, multi-round controllable inference engine, decomposing the complex cause-of-death chain generation task into three sequential steps with strict input-output specifications, thereby achieving a "white-box" and auditable inference process: Step 1: Entity and Context Extraction. The system uses a named entity recognition model to automatically extract key medical entities related to diseases, symptoms, examinations, and treatments from electronic medical record data.
[0050] Step 2: Relationship Graph Construction. Based on the extracted medical entities, candidate ICD codes, and their timestamp information, a local relationship graph is constructed. The nodes of the graph represent entities and codes, while the edges represent the associations determined according to chronological order and medical logic.
[0051] Step 3: Causal chain generation under rule constraints. The system drives a controlled large language model, using the relational graph generated in the first two steps as strong constraints, and combines it with precise domain knowledge obtained from the retrieval-enhanced generation framework for constrained decoding. This forces the model to generate text sequences that conform to the graph topology and medical causal logic, and finally parses the sequence into one or more candidate cause-of-death chains.
[0052] Step S204: Assessment and determination of the underlying cause of death based on a mixed probability model After generating candidate cause-of-death chains, the system uses a hybrid probabilistic model with both temporal and causal constraints to evaluate and score them, in order to determine the most likely cause-of-death chain and the underlying cause of death. Figure 4 As shown, this model abandons the limitations of the traditional N-Gram model, and its architecture includes three core branches: Transformer Branch (Temporal Modeling): This branch is used to learn long-range temporal dependencies in disease sequences. The processor inputs the ICD-coded sequences and their corresponding timestamps from candidate cause-of-death chains into this branch. This Transformer model employs a causal masking mechanism to ensure that when processing a disease node in the sequence, only its historical information is considered, thus accurately simulating the natural evolutionary path of the disease. This branch ultimately outputs a temporal probability characterizing the rationality of the temporal evolution of the cause-of-death chain. .
[0053] Graph Neural Network (GNN) Branch (Causal Modeling): This branch is used to explicitly model the strength of causal associations between disease nodes. The processor uses ICD codes in the candidate cause-of-death chains as nodes and extracts causal relationship edges between these nodes based on a pre-built global disease association knowledge graph, forming a candidate causal subgraph. This subgraph is input to the GNN branch. The GNN calculates the influence of each node as a "causal source" through a message-passing mechanism. This branch ultimately outputs the causal probability characterizing the causal logic strength of the cause-of-death chain. .
[0054] Hard Rule Engine: This branch is used to inject the hard rules for determining the cause of death published by the World Health Organization (WHO), as strong constraints that cannot be violated. These rules are transformed into logical judgment functions, such as: R code exclusion rule: If the disease code begins with 'R', it is prohibited as the underlying cause of death.
[0055] Timing rule: If disease A occurs later than disease B, then A cannot be the underlying cause of death of B, unless there is a stronger modifying rule.
[0056] External cause priority rule: If there is an external cause in the cause of death chain that is coded between V01 and Y98, it must be taken as the starting point of the fundamental cause of death chain unless it is caused by a specific disease such as epilepsy or osteoporosis.
[0057] The processor submits candidate death cause chains to the engine for matching and verification, and outputs a rule verification score. .
[0058] Subsequently, the system uses the dual prediction head output by the hybrid probability model to calculate the overall rationality score of the cause-of-death chain and the location probability of the underlying cause of death, respectively.
[0059] Specifically, the hybrid probability model comprises two independent output layers: (1) Fundamental Cause of Death Identification Head: This branch is based on the output of the multimodal feature fusion layer, and calculates each disease node in the cause of death chain through a fully connected layer and a Softmax function. Probability distribution of the underlying cause of death .
[0060] (2) Cause of death chain score head: This branch is based on the global pooling representation of the entire sequence and outputs a scalar score through the Sigmoid function. This characterizes the overall rationality of the cause-of-death chain in terms of its temporal and causal logic.
[0061] In calculating the final overall probability, this embodiment employs gated multiplication logic to ensure the veto power of the hard rules. Overall Probability The calculation formula is as follows: in, The probability value (ranging from 0 to 1) output by the cause of death chain scoring head. This refers to the rule gating coefficients output by the hard rule engine. When the candidate cause-of-death chain completely conforms to all cause-of-death determination rules, The value is 1; when any rule is violated (such as violating the timing rule, R-code as the root cause of death, etc.), The value is forced to be 0.
[0062] Furthermore, when determining the underlying cause of death, the system calculates the comprehensive score of the underlying cause of death for each node in the chain. : in, w 1 、w 2 、w 3 represents the weight coefficients corresponding to node probability, time, and rule compliance, respectively; The node probabilities predicted by the model; Time-weighted, with higher weight given to earlier onset of illness; The system selects the rule conformity degree of the node as the fundamental cause of death (e.g., whether it belongs to an external cause, whether it is a non-R code). The highest node was identified as the root cause of death.
[0063] Step S205: Intelligent Decision Support and Human-Machine Collaborative Closed-Loop Learning Finally, the system integrates the evaluation results, provides intelligent decision support, and establishes a closed-loop learning mechanism for human-machine collaboration.
[0064] The system sorts all candidate cause-of-death chains in descending order based on the calculated overall probability and confidence level, and presents the Top-N results to clinicians or coders through a visual interface. Figure 5 As shown, the interface clearly displays the recommended primary cause of death chain, alternative cause of death chains, their respective confidence scores, key evidence chains, and decision logs.
[0065] The system automatically captures all feedback actions taken by doctors on the visualization interface (such as modifying codes, adjusting the order of cause-of-death chains, specifying the underlying cause of death, etc.) and transforms them into standardized feedback triplets (AI recommendation value, human correction value, and reason for correction). Based on this feedback data, the system executes a hierarchical closed-loop learning strategy: High-frequency corrections are transformed into business rules (real-time learning): If the system detects that a specific correction pattern (e.g., changing "Disease A" as the underlying cause of death to "Disease B") occurs frequently within a short period of time and exceeds a preset frequency threshold, the system automatically extracts the pattern and generates a new candidate hard rule. After manual review and approval, the rule is immediately injected into the hard rule engine to achieve rapid response to business logic.
[0066] High-confidence corrections are used for model fine-tuning (incremental learning): For infrequent but high-impact corrections made by experts, the system adds them to the training dataset as difficult examples. The system periodically (e.g., weekly or monthly) uses this accumulated high-quality feedback data to update the parameter weights in the mixture probability model using few-shot learning or fine-tuning techniques, thereby continuously improving the model's generalization ability on long-tailed and difficult cases.
[0067] like Figure 5 As shown, taking a male patient (ID: 425**423059) diagnosed with "thalamic hemorrhage" as an example, the system automatically extracted key evidence from the electronic medical record, including "sudden onset of headache and vomiting," "loss of consciousness," and "CT scan confirming basal ganglia hemorrhage rupture into the ventricular system." The system generated three candidate cause-of-death chains through reasoning and calculated their confidence levels: Primary recommendation chain (confidence 96%): Direct cause of death "brainstem compression / brain herniation" (G93.5) The intermediate cause of death was "intracranial hemorrhage ruptured into the ventricular system" (I61.9). The underlying cause of death was thalamic hemorrhage (I61.802). The system determined that this chain of events was complete and consistent with the pathophysiological timeline. Alternative chain (94% confidence): Logically similar, but focusing on ventricular casts. Figure 5On the right-hand side, the expert review interface allows doctors to make interactive modifications. For example, if a doctor believes that "brainstem compression" is not accurate enough as a direct cause of death, they can revise it to "brain herniation (G93.5)" and explain the reason in the remarks. The system will immediately capture this correction (as shown in the "Expert Correction Record" in the figure) and incorporate it into the closed-loop learning system to optimize the terminology recommendation logic for similar neurological cases in the future.
[0068] Through this closed-loop learning and continuous optimization mechanism, the system can continuously absorb the tacit knowledge of experts and achieve continuous performance evolution.
[0069] Example 2 This embodiment provides an electronic device for performing the method described in Embodiment 1. Figure 1 As shown, the electronic device can be a server, workstation, or dedicated medical computing device. Internally, it includes hardware components such as a processor, memory, and communication interfaces.
[0070] The memory is a computer-readable storage medium, such as a hard disk, solid-state drive, or memory, which stores computer programs and various types of data, including but not limited to: electronic medical record databases, ICD encoding libraries, locally deployed medical knowledge bases, and various trained model parameters.
[0071] The processor may be a central processing unit, a graphics processing unit, or a dedicated artificial intelligence chip. Its core function is to execute computer programs stored in memory to perform all the calculation, judgment, reasoning, and control tasks in the method of Embodiment 1.
[0072] The communication interface, such as an Ethernet interface or a wireless communication module, is used to exchange data with external devices, particularly to connect with a hospital information system to obtain patients' electronic medical record data.
[0073] During operation, the processor executes a program, first retrieving electronic medical record data from the hospital information system via a communication interface and storing it in memory. Next, the processor loads and runs a preprocessing module to clean and standardize the data. Subsequently, the processor sequentially calls a four-level cascaded progressive screening module, a retrieval enhancement generation module, a dynamic multi-round controllable inference module, and a hybrid probability model evaluation module to complete the entire process from ICD encoding mapping to candidate cause-of-death chain generation and final evaluation. The calculation results, including the sorted cause-of-death chains, confidence levels, and evidence chains, are formatted by the processor and output to a visualization interface for review. When feedback is received, the processor calls a closed-loop learning module to transform the feedback into training data or new rules and update the corresponding model parameters or knowledge base in memory, thereby achieving incremental learning and evolution of the system.
[0074] Example 3 This embodiment provides a computer-readable storage medium on which a computer program is stored. The storage medium can be non-volatile, such as a read-only memory, flash memory, hard disk, solid-state drive, or optical disk. When the computer program is loaded and executed by the processor of an electronic device, it can implement all the steps of the method for automatically generating the cause-of-death chain and determining the underlying cause of death described in Embodiment 1.
[0075] It should be noted that the above embodiments are merely preferred embodiments of the present invention, intended to provide a more detailed description of the technical solution of the present invention, but the scope of protection of the present invention should not be construed as limited thereto. Any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for automatically generating cause-of-death chains and supporting the determination of the root cause of death by integrating multi-model intelligent reasoning, characterized in that, include: The system obtains patients' electronic medical record data from the hospital information system and performs structured preprocessing on the electronic medical record data. Based on the preprocessed electronic medical record data, a four-level sequential screening mechanism is adopted to map clinical texts into a set of candidate ICD codes. The four-level sequential screening mechanism includes initial screening of semantic matching degree, fine screening of clinical context fit degree, verification of time series rationality and optimization of historical statistical support degree. A retrieval-enhanced generation framework is employed to retrieve domain knowledge related to the electronic medical record data from a locally deployed medical knowledge base. A dynamic, multi-round controllable inference engine is driven to perform causal reasoning on the candidate ICD codes, based on the domain knowledge, to generate at least one candidate cause-of-death chain. A hybrid probabilistic model with time-series and causal dual constraints is used to evaluate the at least one candidate cause-of-death chain. The hybrid probabilistic model includes a Transformer branch for learning the time-series dependence of the disease and a graph neural network branch for modeling the causal association of the disease. Combined with a hard rule engine, the comprehensive probability and corresponding confidence of each candidate cause-of-death chain are calculated to determine the underlying cause of death. The candidate cause-of-death chain is sorted based on the comprehensive probability and the confidence level, and the sorting results are presented through a visualization interface; user feedback on the sorting results is received, and the feedback is converted into structured data for incremental updates to the hybrid probability model and the medical knowledge base.
2. The method according to claim 1, characterized in that, The four-level sequential screening mechanism specifically includes: The semantic matching degree initial screening is performed. Through a biomedical pre-trained language model, the semantic similarity score between the disease description in the clinical text and the ICD encoding standard interpretation is calculated, and ICD codes with scores higher than a first preset threshold are retained. Based on the initial screening results, the clinical context fit fine screening is performed. Using a context-aware model, the logical conformity of each ICD code with the patient's age, gender, past medical history, comorbidities, and department of visit is evaluated, and ICD codes with a conformity lower than the second preset threshold are removed. Perform the time series rationality verification, map the remaining ICD codes and their corresponding onset times to a disease evolution time series diagram, and verify whether there are logical conflicts in the time series diagram according to the preset medical causal logic rule base, and retain the conflict-free ICD codes; Perform the historical statistical support optimization, query the historical case database, calculate the co-occurrence frequency of the remaining ICD codes in similar cases, and reorder the candidate ICD codes.
3. The method according to claim 2, characterized in that, The time series reasonableness verification further includes: The disease evolution time sequence graph is constructed as a directed acyclic graph, where nodes represent disease events corresponding to the refined ICD codes, and directed edges represent the chronological order of disease events. A set of temporal constraint rules are loaded from the medical causal logic rule base. Each of the temporal constraint rules defines the prohibited chronological relationship between a pair of disease events. Traverse all directed edges in the directed acyclic graph and check if there are any edges that violate the temporal constraint rules; Calculate timing conflict score for each ICD code. The timing conflict scoring It is directly proportional to the number of illegal edges formed by the ICD encoding; The timing conflict score Below the preset timing conflict threshold The ICD code is used as the verified code and passed to the historical statistical support optimization step.
4. The method according to claim 1, characterized in that, The steps for evaluating the candidate cause-of-death chains using the hybrid probability model include: The ICD codes and their corresponding timestamps in the candidate cause-of-death chains are input into the Transformer branch, which outputs a temporal probability characterizing the rationality of the temporal evolution of the cause-of-death chain. ; The ICD codes in the candidate cause-of-death chains are used as nodes, and causal relationship edges between these nodes are extracted based on a pre-constructed global disease association knowledge graph to form a candidate causal subgraph. The candidate causal subgraph is then input into the graph neural network branch, which outputs a causal probability representing the causal logic strength of the cause-of-death chain. ; The candidate cause-of-death chains are submitted to the hard rule engine, which performs matching and verification according to the cause-of-death determination rules published by the World Health Organization and outputs a rule verification score. .
5. The method according to claim 4, characterized in that, The calculation method for the comprehensive probability is as follows: The comprehensive probability is calculated using a preset gated multiplication formula. The formula is: in, A score used to characterize the overall plausibility of the cause-of-death chain; The rule gating coefficients output by the hard rule engine; The rule gating coefficient It is a binary variable: it takes the value of 1 when the candidate cause of death chain completely conforms to the cause of death determination rule, and takes the value of 0 when there is any violation of the rule; The root cause of death is determined by the comprehensive score of the root cause of death of each node in the chain. The comprehensive score is calculated by weighting the predicted probability of the node, time weight, and rule compliance of the model.
6. The method according to claim 1, characterized in that, The steps of the retrieval enhancement generation framework to retrieve relevant domain knowledge from the medical knowledge base include: Using a sparse search engine, a search is conducted based on keywords in the electronic medical record data to obtain the first set of relevant knowledge fragments; Using a dense vector retrieval engine, the electronic medical record data is encoded into query vectors, and a similarity retrieval is performed in the vector database to obtain a second set of relevant knowledge fragments; A cross-attention rearrangement model is used to fuse and rearrange the relevant knowledge fragments in the first and second groups, and the knowledge fragments most relevant to the current case are selected as the domain knowledge.
7. The method according to claim 1, characterized in that, The steps of generating the candidate cause-of-death chain by the dynamic multi-round controllable inference engine include: A named entity recognition model is used to extract medical entities related to diseases, symptoms, examinations, and treatments from the electronic medical record data. A relationship graph is constructed based on the medical entity and the candidate ICD code, wherein the nodes of the graph are the medical entity and the ICD code, and the edges are the association relationships determined based on timestamps and medical logic. The controlled large language model is driven to perform restricted decoding with the relation graph as a constraint, generate a text sequence that conforms to the graph topology and causal logic, and parse the text sequence into the candidate cause-of-death chain.
8. The method according to claim 1, characterized in that, The step of converting feedback operations into structured data for incremental updates includes: If the feedback operation is that the user confirms the correct one among the multiple candidate cause of death chains, then the correct candidate cause of death chain is used as a positive sample, and the rest are used as negative samples to fine-tune the mixed probability model. If the feedback operation involves the user modifying the node order or relationship in a candidate cause-of-death chain, the modified structure is transformed into a new rule and added to the rule base of the hard rule engine.
9. An electronic device, comprising: A memory, a processor, and a computer program stored on the memory, characterized in that the processor, when executing the program, implements the method as described in any one of claims 1 to 8.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 8.