Forensic multi-mode agent system oriented to cause-of-cause analysis and cause-of-cause analysis method

Abstract

The invention discloses a forensic multi-mode agent system oriented to cause-of-death analysis and a cause-of-death analysis method. The forensic multi-mode agent system comprises a decision chain agent, a routing chain agent, an execution chain agent, an reflection chain agent, an integration chain agent and an update chain agent. The decision chain agent decomposes the cause-of-death analysis task into a plurality of sub-tasks, and generates an optimized sub-task set according to the cue word; the routing chain agent judges whether an external tool library needs to be called or not; the execution chain agent is used for generating an execution result of the current subtask; the reflection chain agent generates defect and improvement suggestions of each subtask; the integration chain intelligent agent retrieves a reference female parent for the updated execution result of each subtask and writes a preliminary cause analysis description; the updating chain intelligent agent updates a writing expression mode of the cause analysis description; and fine-tuning the large language model agent to generate a final cause analysis description. According to the method, the death reason can be automatically analyzed to generate a clear death reason analysis description.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and specifically relates to a forensic multimodal intelligent agent system for cause of death analysis, and also relates to a cause of death analysis method of the forensic multimodal intelligent agent system for cause of death analysis. Background Art

[0002] Forensic death analysis uses a variety of scientific methods such as autopsy, toxicology, pathology, forensic imaging, and DNA analysis to determine the cause and manner of death, helping forensic doctors and judicial institutions to clarify the truth of the case. Autopsy is one of the most commonly used methods in forensic analysis. Through external and internal examination of the body, it can find signs of trauma, disease or poisoning that may have caused death; toxicology analysis is used to detect poisons or chemicals in the body of the deceased, especially in cases of drug overdose, suicide or poisoning; pathology analysis focuses on the disease state of the deceased, and through the examination of tissue samples, it can find fatal health conditions such as chronic diseases or acute conditions. Forensic imaging technology (such as X-rays, CT scans, MRI, etc.) obtains internal images without damaging the body, showing fractures, foreign objects or bleeding, especially in cases such as traffic accidents, fires or explosions. DNA analysis is not only used for identity recognition, but also can be used to detect biological traces on the body to confirm the identity of the deceased or whether there are third-party DNA traces, which is crucial in homicide or sexual assault cases. These methods work together to target different analytical needs in order to reach accurate conclusions on the cause of death, ultimately providing strong support for judicial justice and public safety through scientific and comprehensive analysis.

[0003] However, it is often very difficult to give a clear and definite diagnosis of the cause of death. Forensic cause of death analysis requires comprehensive consideration of the above-mentioned multiple modalities, such as autopsy, toxicology analysis, pathology analysis, forensic imaging, and DNA analysis, so it is a complex and time-consuming task. Each analysis method involves the processing of a large amount of data and meticulous judgment, which places extremely high demands on the professional knowledge and patience of forensic doctors. With the development of technology, we hope to use artificial intelligence (AI) to assist forensic doctors to complete this work more efficiently, help practitioners process complex data more quickly, reduce the burden of manual work, and improve accuracy and efficiency.

[0004] An AI agent is an autonomous computing entity that perceives its environment through sensors, performs operations on it using actuators, and aims to achieve specific goals by making decisions without direct human intervention. Key features include perception (collecting information), action (influencing the environment), autonomy (operating independently), goal-directed behavior (guided by goals), and adaptability (learning and improving over time). AI agents can be classified into simple reflex agents, model-based agents, goal-based agents, utility-based agents, or learning agents, each with varying levels of complexity and functionality. In a multi-agent system, multiple AI agents interact in the same environment to collaborate or compete, enhancing problem-solving capabilities and simulating complex behaviors. Currently, AI agents are an integral part of applications such as virtual assistants, autonomous vehicles, robots, game AI, and recommendation systems, but they have not been effectively applied in forensic pathological cause-of-death analysis tasks.

[0005] Therefore, the present invention proposes a brand-new method of introducing AI agent technology into forensic pathological cause-of-death analysis tasks to achieve fully automated processing of cause-of-death analysis. Summary of the Invention

[0006] The object of the present invention is to provide a cause-of-death analysis method for a forensic multi-modal agent system for cause-of-death analysis, which can automatically analyze the cause of death and efficiently and accurately generate clear analysis reports and expert opinions.

[0007] Another object of the present invention is to provide a forensic multi-modal agent system for cause-of-death analysis.

[0008] The technical solution adopted by the present invention for the cause-of-death analysis method of the forensic multi-modal agent system for cause-of-death analysis is as follows: Step 1: Build a large language model agent architecture, including a decision chain agent, a routing chain agent, an execution chain agent, a reflection chain agent, an integration chain agent, and an update chain agent; Step 2: The decision chain agent decomposes the cause-of-death analysis task into multiple subtasks according to the background information of the cause-of-death analysis task, and generates an optimized subtask set through prompt learning; Step 3: The routing chain agent generates a judgment score based on the subtasks and background information in each optimized subtask set to determine whether to call an external tool library; Step 4: The execution chain agent is used to generate the execution result of the current subtask; Step 5: The reflection chain agent analyzes the execution result of each subtask, generates the defects and improvement suggestions of each subtask, and re-enters the defects and improvement suggestions into Step 4 to generate the updated execution result of the subtask; Step 6, the integration chain agent matches the updated execution results of each subtask with the case database, selects the case with the highest similarity as the reference master copy, and writes a preliminary cause-of-death analysis description; Step 7, the update chain agent updates the writing expression of the cause-of-death analysis description through the method of chain of thought; Step 8, fine-tune the large language model agent locally to generate the final cause-of-death analysis description.

[0009] The features of the present invention also lie in that, The specific process of Step 2 is as follows: Input the background information D of the cause-of-death analysis task into the decision chain agent. The decision chain agent decomposes the cause-of-death analysis task into multiple subtasks , and forms a subtask set with the multiple subtasks . Generate the optimal subtask combination under the current conditions according to the prompt word A , , and then generate an optimized subtask set according to the prompt word B .

[0010] The specific process of Step 3 is as follows: Input each subtask in the optimized subtask set obtained in Step 2 and the background information D into the routing chain agent at the same time. The routing chain agent generates a judgment score ; If , it means that the current information cannot complete this subtask , and an external tool library needs to be called; If , it means that the current information is sufficient to complete this subtask , and there is no need to call the external tool library.

[0011] The external tool library includes the PubMed paper database, the forensic pathology book database, and Internet search.

[0012] The specific process of Step 4 is as follows: The execution chain agent generates the execution result of the current subtask according to the content of the current subtask, the background information D, the execution result of the previous subtask, and the external knowledge called for executing the current subtask.

[0013] The specific process of Step 5 is as follows: The reflection chain agent includes an evaluation agent E and a reflection agent R; The evaluation agent E is used to judge whether the execution result of each subtask meets the requirements of the subtask and generate a reward score; The reflection agent R analyzes the differences between the execution results of each subtask and the requirements of that subtask based on the reward scores generated by the evaluation agent E, generates defects and improvement suggestions, and takes the obtained defects and improvement suggestions as new tasks to re - execute step 4, generating the updated execution results of the subtasks.

[0014] The specific process of step 6 is as follows: The integration chain agent randomly aggregates the updated execution results of the subtasks and the execution results of the subtasks that meet the expected task goals, inputs each aggregated result into the case database to retrieve the historical cases most similar to the current aggregated result through the RAG method, takes each most similar historical case as a reference master, extracts the cause - of - death analysis methods, logical reasoning paths, and related pathological conclusions of each reference master, learns from each reference master. After the learning is completed, the integration chain agent combines the updated execution results of all subtasks of the current case and the execution results of the subtasks that meet the expected task goals to write a preliminary cause - of - death analysis description; The preliminary death analysis description includes the independent conclusions of each subtask and their logical consistency with the reference master.

[0015] The specific process of step 7 is as follows: The update chain agent extracts high - quality cause - of - death analysis reports from historical reports and authoritative forensic analysis cases as expression examples, takes the key steps of the step - by - step reasoning process in the expression examples as learning examples for COT. The update chain agent can form a chain - type logic when generating the cause - of - death analysis description by learning the learning examples of COT, update the cause - of - death analysis description, and conduct a comparison test between the updated cause - of - death analysis description and the expression examples to obtain the expression deviation, and further optimize the COT reasoning chain according to the expression deviation until the updated cause - of - death analysis description has the same reasoning process and expression quality as the expression examples.

[0016] Another technical solution adopted by the present invention is a forensic multi - modal intelligent agent system for cause - of - death analysis, including: A decision - making chain agent, which is used to decompose the cause - of - death analysis task into multiple subtasks according to the background information of the cause - of - death analysis task and generate an optimized subtask set through the prompt learning method; A routing chain agent, which is used to generate a judgment score according to the subtasks and background information in each optimized subtask set to judge whether it is necessary to call an external tool library; An execution chain agent, which is used to generate the execution result of the current subtask; A reflection chain agent, which is used to analyze the execution results of each subtask, generate defects and improvement suggestions for each subtask, and re - input the defects and improvement suggestions into step 4 to generate the updated execution results of each subtask; The integration chain agent is used to match the execution results after the update of each optimized subtask with the case database, take the case with the highest similarity as the reference master copy, and write a preliminary cause of death analysis description; The update chain agent is used to update the writing expression of the cause of death analysis description.

[0017] The beneficial effects of the present invention are as follows: (1) The cause of death analysis method of the present invention automatically integrates various analysis means such as autopsy, toxicology, pathology, imaging, and DNA, effectively reducing the complexity in manual operations, and improving the data processing speed and analysis accuracy; (2) The cause of death analysis method of the present invention, through the autonomy and adaptability of the agent, assists forensic doctors in quickly processing complex data analysis tasks, reduces the burden on forensic doctors in dealing with a large amount of repetitive data, and at the same time reduces the risk of human errors; (3) The cause of death analysis method of the present invention makes it possible to fully automate the cause of death analysis through the agent, not only significantly improving the analysis efficiency, but also promoting the modernization and intelligentization process of forensic pathology, and laying a foundation for future forensic technology innovation. Brief Description of the Drawings

[0018] Figure 1 is a flowchart of the cause of death analysis method of the forensic multi-modal agent system for cause of death analysis of the present invention; Figure 2 is a schematic diagram of the cause of death analysis method of the forensic multi-modal agent system for cause of death analysis of the present invention; Figure 3 is a detailed diagram of the specific implementation method of the cause of death analysis method of the forensic multi-modal agent system for cause of death analysis of the present invention; Figure 4 is a display diagram of the web version application corresponding to the forensic multi-modal agent system for cause of death analysis of the present invention. Detailed Embodiments

[0019] The present invention will be described in detail below in conjunction with the drawings and specific embodiments.

[0020] Example 1 The cause of death analysis method of the forensic multi-modal agent system for cause of death analysis of the present invention, as Figure 1 、 Figure 2 and Figure 3 shown, the specific steps are as follows: Step 1, build a large language model agent architecture, including a decision chain agent, a routing chain agent, an execution chain agent, a reflection chain agent, an integration chain agent, and an update chain agent; Step 2: The decision chain agent decomposes the cause-of-death analysis task into multiple subtasks according to the background information of the cause-of-death analysis task, and generates an optimized subtask set through the prompt learning method; Step 3: The routing chain agent generates a judgment score based on the subtasks and background information in each optimized subtask set to determine whether to call the external tool library; Step 4: The execution chain agent is used to generate the execution result of the current subtask; Step 5: The reflection chain agent analyzes the execution results of each subtask, generates the defects and improvement suggestions of each subtask, and re-enters the defects and improvement suggestions into Step 4 to generate the updated execution result of the subtask; Step 6: The integration chain agent matches the updated execution result of each subtask with the case database, takes the case with the highest similarity as the reference master copy, and writes a preliminary cause-of-death analysis description; Step 7: The update chain agent updates the writing expression of the cause-of-death analysis description through the chain-of-thought method; Step 8: Fine-tune the large language model agent locally to generate the final cause-of-death analysis description.

[0021] Embodiment 2 Based on Embodiment 1, the specific process of Step 2 is as follows: Input the background information D of the cause-of-death analysis task into the decision chain agent. The decision chain agent understands the background information D of the cause-of-death analysis task, decomposes the cause-of-death analysis task into multiple subtasks, and forms a subtask set with the multiple subtasks, generates the optimal subtask combination under the current conditions according to the prompt A, and then generates an optimized subtask set according to the prompt B; The decision chain agent uses the large language model GPT-4o-mini to execute the above process; wherein, represents the function for specifically generating each subtask, which is autonomously completed by GPT-4o-mini; ; The decision chain agent uses the large language model GPT-4o-mini to execute the above process; wherein,

[0022]

[0023] In the formula, represents the function for specifically generating each subtask, which is autonomously completed by GPT-4o-mini;

[0024] In the formula, Represents a decision chain agent, is the prompt A; Specifically: Select several reasoning modules crucial for solving the given task: All reasoning module descriptions: {reasoning_modules}; Task: {task_description}; Selecting several modules is crucial for solving the above task:

[0025] In the formula, is the prompt B; Specifically: Rewrite and specify each reasoning module so that it better helps solve the task: SELECTED module description: {selected_modules} Task: {task_description} Adapt the descriptions of each reasoning module to better solve the task: Background information D includes all the details of the case, the core factors of the cause of death analysis, and the key details in the case.

[0026] Example 3 Based on Example 2, the specific process of step 3 is as follows: Input each sub-task in the optimized sub-task set obtained in step 2 and the background information D into the routing chain agent at the same time. The routing chain agent uses a large language model GPT-4o capable of structured output to process these inputs and generate a judgment score which is used to evaluate whether the current information is sufficient to complete the sub-task. Specifically: If If , it means that the current information cannot complete the sub-task, and an external tool library needs to be called to obtain supplementary information to ensure that the sub-task can be effectively supported; If , it means that the current information is sufficient to complete the sub-task, and there is no need to call external tools; Among them, the judgment formula is:

[0027] Among them, the external tool library includes PubMed paper database, forensic pathology book database, and Internet search (Tavily API); they are respectively aimed at medical literature, professional forensic pathology knowledge, and real-time Internet data needs to ensure that the system effectively responds to various information needs; PubMed paper database: PubMed is a medical literature database that contains a large number of high-quality medical, pharmaceutical and forensic research results. It is an important information source in the medical field. For some subtasks that require in-depth medical knowledge, PubMed is used as the primary source of information acquisition. Through this database, highly relevant and strictly peer-reviewed academic literature is retrieved. PubMed provides the necessary authoritative medical evidence in the analysis of the cause of death, and provides important support for complex pathological judgments and rare case analysis. Forensic pathology book database: Forensic pathology knowledge is highly specialized, involving multiple fields such as anatomy, pathology, and biochemistry. To ensure that the model can fully utilize this expertise, the core books and authoritative materials on forensic pathology are converted into a vector database through the hierarchical RAG (Retrieval-Augmented Generation) method. The RAG method can effectively extract high-value information from the books, allowing the large language model to retrieve the most relevant knowledge fragments from the vector database when faced with forensic-specific subtasks. These fragments are embedded in the model as knowledge supplements to help it form conclusions that conform to the logic of forensic analysis. In this way, even when faced with complex or special pathology cases, the large language model can make accurate judgments based on authoritative materials. Internet search (Tavily API): To meet the needs of cross-domain information or the latest news, Tavily API is introduced for real-time Internet search. When a subtask involves knowledge areas other than forensics or requires the latest dynamic information, a search request is initiated through Tavily API to obtain real-time relevant content from the Internet. The introduction of this external information source ensures that the system can update the knowledge base at any time and cross the boundaries of traditional medicine and forensic medicine to effectively support the task requirements of multimodal forensic agents in a wider range of fields. For example, for subtasks involving rare poisons, specific drug side effects, or international forensic identification guidelines, Tavily API can retrieve the latest data from trusted sources to provide the necessary real-time support for the model. The PubMed database is used to obtain the latest scientific research literature in the medical field, the forensic pathology book database provides structured forensic expertise, and the Tavily API ensures that the model can access cross-domain and real-time dynamic information; it can flexibly schedule various external resources while keeping information updated and content accurate, meet the diverse needs of different subtasks, make full use of the rich knowledge systems in the medical and forensic fields, and support the multi-modal artificial intelligence agent for cause-of-death analysis in a scientific and comprehensive manner, achieving the accuracy and professionalism of complex forensic tasks.

[0028] Example 4 Based on Example 3, the execution chain agent uses the large language model GPT-4o-mini and performs dynamic in-context learning to generate the execution result of the current subtask. The specific prompt is: where {context} is the overall background of the task , {observed} is the execution result of the previous subtask, {content} is the external knowledge invoked for the current subtask, and {plan} is the task to be executed at the current time.

[0029] The specific manifestation of dynamic in-context learning is as follows: Instruction: You are a forensic professor responsible for conducting a detailed autopsy analysis to determine the cause and manner of death. A crucial part of your work involves differentiating injuries caused by trauma from potential spontaneous medical conditions, such as subarachnoid hemorrhage, which may present similar symptoms.

[0030] Differentiating trauma from disease is crucial for the integrity of the investigation and must be done with great care.

[0031] This analysis will be carried out in structured phases, with each phase focusing on a specific aspect of the investigation.

[0032] In each phase (PLAN or TASK), the results and conclusions should be closely related to the immediate environment and the phase being examined. Pay attention to key diseases such as "myocarditis", "pneumonia", "coronary heart disease", "coronary atherosclerosis", "cerebral hemorrhage", "various types of shock (toxic shock, anaphylactic shock, traumatic shock)", "subarachnoid hemorrhage", "brainstem hemorrhage", "cerebral hemorrhage", "purulent peritonitis", and "craniocerebral injury".

[0033] Also consider cases of accidental death, such as "fall", "poisoning", "asphyxiation", "drowning", "electric shock", "burning", etc.

[0034] Each stage of the investigation will build on the previous one, but each stage must be treated independently and be directly relevant to the facts at hand.

[0035] Attention must be focused only on the current tasks of each stage and ensure that all judgments are based on the specific findings and information at that moment in the investigation.

[0036] In forensic investigations, a careful distinction must be made between the cause and the direct cause of death.

[0037] Specifically, determining whether trauma is a predisposing factor or the primary cause of death is a crucial distinction that requires in-depth analysis.

[0038] This distinction has significant implications for the legal and medical understanding of the case.

[0039] Therefore, conclusions about the cause of death must be drawn with the utmost caution, ensuring that all factors are fully considered, including potential medical conditions, pre-existing vulnerabilities, and the role of external trauma.

[0040] Rigorous evidence-based reasoning is necessary to prevent premature or biased conclusions, as such decisions can affect clinical and forensic outcomes.

[0041] Note: Context is the main element of the case; The information observed is the execution of the previous plan of the case; Forensic knowledge is information obtained from reliable internet sources and should be used to support the solution of the plan; irrelevant elements can be ignored in the implementation planning; The results must be based on what is obtained after a comprehensive analysis of all content, i.e., the context of this case; The background (key information) of this case: {context}; The information observed in this case (supplementary information): {observed}; The forensic knowledge of this case (supplementary materials): {content}; The plan of this case that needs to be completed (what needs to be done for the current task): {plan}.

[0042] The key elements of the execution chain agent include: (1) Content description of subtasks: In the execution chain agent, the description of each subtask combines the background information D and the execution result of the previous subtask, enabling the large execution chain agent to analyze and make decisions within the existing context, rather than processing each subtask in isolation. Through this explicit description, the execution chain agent can understand the logical flow of the task, ensuring the consistency and high quality of the generated results; (2) Input of background information D: As the key context in the execution chain, the background information D runs through the execution of each subtask. This background information D includes all the content of the case, the core factors of the cause-of-death analysis, and the key details in the case, enabling the large language model agent to refer to the complete case information at each step, thereby making more reasonable and scientific inferences. The input of this global information gives the large language model agent an overall perspective, thus enhancing the reasoning ability and the coherence of the results; (3) Transmission of the execution result of the previous subtask: To maintain the logic and coherence of task execution, the execution chain advances the progress of the task step by step by transmitting the execution results of the previous subtasks; the execution result of each subtask is transmitted as input to the next task, enabling the execution chain agent to effectively utilize the previous results in the new subtask and achieve layer-by-layer reasoning. This design ensures the result dependency between subtasks, avoids the gap between tasks, and makes the final result more logical and complete; (4) Embedding of the external knowledge invoked: In the execution chain agent, when specific subtasks require external knowledge, the routing chain agent will automatically call the corresponding external tool library and embed the obtained external knowledge into the input of the current task. These knowledge supplements include information obtained from PubMed, the forensic pathology database, and Internet searches, enabling the routing chain agent to select and use the most relevant external data sources according to specific needs, thereby solving the difficult problems in subtasks. The embedding of external knowledge ensures that the large language model agent has the necessary domain knowledge when dealing with complex medical and forensic tasks, and the generated results are more accurate and scientific.

[0043] This step realizes the sequentiality, coherence, and comprehensiveness of knowledge of subtasks. In the entire task process, the execution chain agent constructs a structured and coherent analysis path by integrating the global information of the case, the results of previous tasks, and external knowledge support. The final execution result is generated based on the gradually accumulated subtask results, and is logical, scientific, and highly accurate, providing strong support for complex forensic multimodal analysis tasks.

[0044] Example 5 On the basis of Example 4, the specific process of step 5 is as follows: The reflection chain agent consists of an evaluation agent E and a reflection agent R; the evaluation agent is used to analyze each subtask Whether the execution result meets the expected task goal to generate a reward score , and the reward score reflects the performance of the generated result in the current subtask content, the execution result of the current subtask and the performance in background information D. When the reward score is lower than the preset value, the reward score, the execution result of each subtask and the expected task goal of each subtask are input into the reflection agent R. When the reward score is not lower than the preset value, directly execute step 6; The reflection agent R analyzes the difference between the execution result of each subtask and the requirements of the subtask according to the reward score generated by the evaluation agent E, generates defects and improvement suggestions, and uses the obtained defects and improvement suggestions as new tasks to re-execute step 4 to generate the updated execution result of the subtask; When the evaluation agent analyzes that the execution result of each subtask meets the expected task goal, directly execute step 6 with the execution result of the subtask; When the evaluation agent analyzes that the execution result of each subtask does not meet the expected task goal, it needs to be input into the reflection agent R for analysis, generate defects and improvement suggestions, and use the defects and improvement suggestions as new tasks to re-execute step 4 to generate the updated execution result; Among them,

[0045] The evaluation agent judges through the self-ability of the large language model (GPT-4o-mini) and the given context content ( ) under the given prompt (Prompt) the relevant relationship between, let {input} be , {prediction} be .

[0046] Prompt: Please act as an impartial judge and evaluate the quality of the AI assistant's answer to the user question shown below. Your evaluation should consider factors such as the usefulness, relevance, accuracy, depth, creativity, and detail of the response. Start the evaluation by providing a brief explanation. Be as objective as possible. After providing the explanation, you must rate the answer on a scale of 1 to 10; for example: "Rating: 5 points".

[0047] Question: { input} Start of assistant's answer: {prediction} End of assistant's answer.

[0048] When evaluating the agent If the score output by the agent is less than 5, the reflection agent R will conduct a reflection; through the capabilities of the large language model (GPT-4o-mini) itself and the given context content ( ), conduct a reflection under the given prompt, and generate defect analysis and improvement suggestions. Analyze under the given prompt the gap between, and generate the content of further queries to be made. Let {task} be {context} be .

[0049] Prompt: You are an experienced forensic expert responsible for critically evaluating the results of complex case investigations. Criticism and reflection: Based on the task and the specific context of this case, strictly review the findings and analysis results. Identify any gaps, inconsistencies, or areas where the analysis can be improved. Be strictly critical to ensure the maximum enhancement and reliability of the results. Recommended search queries: Suggest targeted search queries that can be used to gather additional information. These queries should be aimed at filling in gaps, validating results, and strengthening the overall analysis. Ensure that the suggestions are specific, actionable, and directly relevant to the case context; Task: {task}, Context of this case: {context}.

[0050] This closed-loop design enables the model to continuously optimize based on the results of the previous execution, gradually make up for the defects of the initial results, and improve the accuracy and consistency of the results. Under this feedback mechanism, the system can make full use of the output results of the reflection agent R, making the execution results of each subtask more precise after multiple rounds of improvement. The reflection chain agent plays an important role in self-optimization in the entire cause-of-death analysis task; through the combination of the evaluation agent and the reflection agent , a dynamic self-adjustment mechanism is formed to gradually improve the analysis task. The reflection chain agent not only improves the scientificity and rigor of the task results, but also enables the system to have stronger adaptability and intelligent analysis capabilities, and can improve the reliability of the results in multiple feedbacks and loops.

[0051] Example 6 Based on Example 5, the specific process of Step 6 is as follows: The integrated chain agent randomly aggregates the updated execution results obtained in Step 5 and the execution results of the subtasks that meet the expected task objectives, and inputs each aggregated result into the case database to retrieve the historical cases most similar to the current case through the RAG (Retrieval-Augmented Generation) method. Each most similar historical case is used as a reference template, and the cause-of-death analysis method, logical reasoning path, and relevant pathological conclusions of each reference template are extracted. After learning from each reference template, the integrated chain agent combines the updated execution results of all subtasks of the current case and the execution results of the subtasks that meet the expected task objectives to write a preliminary cause-of-death analysis description. The integrated chain agent uses the large language model GPT-4o-mini and dynamic context learning to perform the above process; The specific prompt is as follows: where {context} is the background information D of the task, {obs_result} is the summary information of all execution results, {supplements} is the execution result of the reflection chain, and {fewshot} is the similar cases obtained through hierarchical RAG search.

[0052] Introduction: Write a cause-of-death analysis description based on the detailed case information provided.

[0053] In forensic investigations, it is crucial to carefully distinguish between the causative and direct causes of death.

[0054] Specifically, determining whether a trauma is a predisposing factor or the main cause of death is a key distinction that requires in-depth analysis.

[0055] This distinction has significant implications for the legal and medical understanding of the case. Therefore, conclusions about the cause of death must be drawn with extreme caution, ensuring that all factors are fully considered, including potential medical conditions, pre-existing vulnerabilities, and the role of external trauma.

[0056] Rigorous evidence-based reasoning is necessary to prevent premature or biased conclusions, as such decisions can affect clinical and forensic outcomes.

[0057] Guidelines for writing the analysis description: 1. Autopsy results Summarize the key pathological features or injuries observed during the autopsy.

[0058] Briefly describe the location and nature of these findings (e.g., fractures, tissue damage, organ changes) in 3 - 4 sentences.

[0059] 2. Cause-of-death analysis Analyze the direct or indirect connections between these pathological features or injuries and the cause of death.

[0060] Discuss the relationship between external factors (such as surgery, accidents, violence) and these injuries or pathological conditions.

[0061] 3. Pathological background and death correlation Consider the patient's medical history and lifestyle to discuss how pre-existing health conditions affected the death.

[0062] Evaluate the role of chronic diseases or existing health problems in the death.

[0063] 4. Supplementary test results Mention any forensic examinations (such as toxicology, histology) conducted.

[0064] Explain how these results support or challenge the preliminary cause-of-death analysis.

[0065] 5. Conclusions and reasoning Clearly summarize the relationship between the autopsy findings and the cause of death.

[0066] Based on a comprehensive analysis of the autopsy and supplementary test results, provide the final forensic opinion.

[0067] Ensure that your analysis is thorough, specific, and clearly outlines the scientific principles and logical reasoning behind each step.

[0068] 6. Additional notes The output should be in continuous paragraphs, without subheadings or bullet points.

[0069] Ensure the use of precise medical terminology for professional accuracy.

[0070] Maintain objectivity and scientific rigor, avoid using any speculative language (e.g., "might"), and ensure that the structure, format, and expressions are as similar as possible to the given examples. The output should be in Chinese.

[0071] Background of this case: {context}; Observed findings: {obs_result}; Supplementary information: {supplements}; Examples: {fewshot}; Analysis description of this case (refer to the example expressions as much as possible).

[0072] Example 7 Based on Example 6, the specific process of Step 7 is as follows: The update chain extracts high-quality cause-of-death analysis reports from historical reports and authoritative forensic analysis cases as expression examples, and takes the key steps of the step-by-step reasoning process in these expression examples as the learning examples for COT. By learning these COT learning examples, the update chain can form a chain of logic when generating the cause-of-death analysis description, update the cause-of-death analysis description generated in Step 6 through the update chain, compare the updated cause-of-death analysis description with the expression examples for testing, obtain the expression deviation, and further optimize the COT reasoning chain based on the expression deviation until the updated cause-of-death analysis description has the same reasoning process and expression quality as the expression examples.

[0073] The key steps of the step-by-step reasoning process include pathological analysis, logical reasoning, causal association, and forensic evidence support.

[0074] The update chain agent uses the large language model GPT-4o-mini and dynamic context learning to execute the above process.

[0075] The specific prompt is: where {example_o} and {example_r} are examples of few-shot thinking chains, and {context} is the initial result output in Example 6; Instruction: Your task is to enhance the following statement for cause-of-death analysis to make it more scientific, rigorous, and detailed.

[0076] First, review the provided examples, which illustrate how to create well-structured cause-of-death statements (revised "good examples" vs. original "bad examples").

[0077] Then, modify the given statement with the same level of rigor, paying particular attention to the methods of analyzing injury relationships, as shown in the examples.

[0078] The revised statement should be in Chinese and within the specified word limit.

[0079] The revised statement must not contain words such as "may" or "might".

[0080] The revised statement must be strictly affirmative.

[0081] In forensic investigations, it is crucial to carefully distinguish between the pathogenic cause and the direct cause of death.

[0082] Specifically, determining whether a trauma is a predisposing factor or the main cause of death is a key distinction that requires in-depth analysis.

[0083] This distinction has a significant impact on the legal and medical understanding of the case.

[0084] Therefore, conclusions about the cause of death must be drawn with the utmost caution, ensuring that all factors are fully considered, including potential medical conditions, pre-existing vulnerabilities, and the role of external trauma.

[0085] Rigorous evidence-based reasoning is necessary to prevent premature or biased conclusions, as such decisions can impact clinical and forensic outcomes.

[0086] For example: Source language: {example_o}; Revision: {example_r}; In this case: Source language: {context}; Revision: Example 8 Based on Example 7, the specific process of Step 8 is as follows: Using the method of knowledge instruction tuning, the large language model agent is fine-tuned by combining the dataset and specific prompts using the LORA (Low-Rank Adaptation) method. Here, we selected three open-source large language models as the base models, including: LLAMA3.1, GLM4, and QWEN2; 9508 autopsy reports were selected as the dataset. During the fine-tuning process, the analysis data and conclusions in each autopsy report were extracted, and a standardized prompt template was created through the extracted analysis data and conclusions. The prompt template and the extracted analysis conclusions were input into the large language model agent for fine-tuning to obtain the fine-tuned large language model agent. The task of cause-of-death analysis to be analyzed was input into the fine-tuned large language model to obtain the final cause-of-death analysis description.

[0087] The dataset comes from autopsy reports of different institutions in China. These reports are distributed in six centers, covering various causes of death. The six centers are: 1517 reports from Xi'an Jiaotong University, 3354 reports from Sun Yat-sen University, 1996 reports from Hebei Medical University, 997 reports from Jining Medical University in Shandong, 852 reports from Xinxiang Medical University in Henan, and 792 reports from Bemei Judicial Appraisal Center; The above autopsy reports not only cover the detailed pathological analysis results of each deceased, but also contain the final identification conclusions, providing rich and multi-dimensional data support for cause-of-death judgment.

[0088] Example 9 The forensic multi-modal intelligent agent system of the present invention for cause-of-death analysis includes: A decision chain intelligent agent, which is used to decompose the cause-of-death analysis task into multiple sub-tasks according to the background information of the cause-of-death analysis task, and generate an optimized set of sub-tasks through prompt learning methods; The routing chain agent is used to generate a judgment score based on the subtasks and background information in each optimized subtask set to determine whether to call the external tool library; The execution chain agent is used to generate the execution result of the current subtask; The reflection chain agent is used to analyze the execution result of each subtask, generate the defects and improvement suggestions of each subtask, and re-enter the defects and improvement suggestions into step 4 to generate the updated execution result of each subtask; The integration chain agent is used to match the updated execution result of each optimized subtask with the case database, use the case with the highest similarity as the reference master copy, and write a preliminary cause-of-death analysis description; The update chain agent is used to update the writing expression of the cause-of-death analysis description.

[0089] Embodiment 10 Figure 4 This is a display diagram of the web version application (APP) of the forensic multi-modal intelligent agent system for cause-of-death analysis according to the present invention; (1) Specifically, the task decomposition results of the decision chain agent in this web version application are shown as follows: File name: test_ 2024.docx Using the self-discovery method to decompose your task! (This step is relatively slow and takes 40s!) Steps to be executed for this case: 1. Review the autopsy report and case background records: Carefully review the autopsy report and medical records of Jiao XX, focusing on injuries caused by the car accident, such as head and hip injuries, and other organ injuries that may affect death. At the same time, analyze the role of his long-term diseases (such as bronchial asthma, heart disease) in the accident and evaluate whether these chronic diseases exacerbated the injuries caused by the car accident.

[0090] 2. Deeply analyze the disease course and acute onset: Carefully analyze the interaction between Jiao XX's chronic diseases (such as heart disease, lung disease) and acute injuries during the car accident. Explore whether these chronic diseases were exacerbated by the car accident, especially the injuries to the heart and lungs, and how these interactions may have led to a sharp deterioration of his condition.

[0091] 3. Evaluate the impact of trauma on his own diseases: Specifically analyze whether the trauma caused by the car accident, such as head injuries and hip injuries, triggered or exacerbated Jiao XX's chronic diseases (such as heart disease and asthma), and evaluate how these traumas led to unstable vital signs and ultimately death.

[0092] 4. Comprehensive Analysis and Conclusion Writing: Based on the autopsy results, medical records, and details of the car accident, comprehensively analyze the causal relationship between the accident and the disease, identify the main cause of death and the predisposing factors, and write a detailed analysis report to provide a scientific basis for forensic identification.

[0093] (2) Specifically, the result display of the routing chain agent in the web version application is as follows: This task is being executed: 4. Comprehensive Analysis and Conclusion Writing: Based on the autopsy results, medical records, and details of the car accident, comprehensively analyze the causal relationship between the accident and the disease. Identify the main cause of death and the predisposing factors, and write a detailed analysis report to provide a scientific basis for forensic identification.

[0094] Searching... Looking for corresponding websites and the corresponding PubMed database! The query information is: Causes of death in elderly patients with multiple injuries and pre-existing conditions; The website query results and the PubMed database are: The causes of death in elderly patients with multiple injuries and pre-existing conditions are multifaceted and influenced by multiple factors. The following are some key points based on recent research and literature: Anemia: Anemia is an important predictor of mortality in severely injured patients. Studies have shown that moderate to severe anemia increases the risk of death and prolongs hospital stay. Specifically, severe anemia (Hb < 7 g / dl) is associated with a higher mortality rate.

[0095] The content obtained from forensic pathology textbooks is: When writing a comprehensive analysis and conclusion, first, based on the autopsy results, medical records, and details of the car accident, comprehensively analyze the causal relationship between the accident and the disease. Forensic experts should understand in detail information such as the time and place of death, the deceased's past medical history, and the manifestations before death to form hypotheses about the cause and manner of death. Through systematic autopsy, histological examination, and extraction and examination of relevant specimens, determine the lethal lesions and injuries and analyze their mechanisms of death. Identify the main cause of death and the predisposing factors, and distinguish different situations such as the underlying cause of death, the immediate cause of death, the contributory cause of death, and the predisposing factors. Finally, write a detailed analysis report to provide a scientific basis for forensic identification.

[0096] (3) Specifically, the execution result of the execution chain agent in the web version application is as follows: The conclusion drawn from the analysis of this case at this step: According to the autopsy results and medical records, the death of Jiao XX was the result of the combined effects of the trauma caused by the car accident and his chronic diseases. The head and hip injuries caused by the car accident, combined with his chronic diseases and the stress response after the car accident, ultimately led to his death.

[0097] Jiao XX suffered severe trauma to the head and hip, especially multiple fractures of the right hip and lacerations of the head. Although no obvious bleeding was seen on the cranial CT, the scalp laceration was accompanied by active bleeding, indicating that the head had been significantly impacted.

[0098] Jiao XX suffered from chronic diseases such as bronchial asthma, emphysema, and aortic sclerosis. These diseases exacerbated his body's stress response after the car accident, especially leading to difficulty breathing.

[0099] The myocardial edema and brain edema mentioned in the autopsy report may be due to the systemic stress response caused by the car accident.

[0100] The toxicology test results showed that there were multiple drug components in Jiao XX's body. These drugs may have been used during the treatment of his chronic diseases. Although these drugs themselves may not have directly caused death, they may have had an adverse impact on his body under the stress state after the car accident.

[0101] In summary, Jiao XX's death was the result of the combined effects of the trauma caused by the car accident and his chronic diseases. The head and hip injuries caused by the car accident, combined with his chronic diseases and the stress response after the car accident, ultimately led to his death.

[0102] (4) Specifically, the display of the matching results between the integrated chain agent and the case database in the web version application is as follows: 3 cases similar to this case were found in the known case database! Analysis and explanation: 1. No common drug components such as amphetamine, methamphetamine, 3,4-methylenedioxymethamphetamine, ketamine, dolantin, cathinone, methcathinone, cocaine, benzoylecgonine, morphine, monoacetylmorphine, codeine, heroin, methadone, 2-ethylidene-1,5-dimethyl-3,3-diphenylpyrrolidine, and tetrahydrocannabinolic acid were detected in the blood of the deceased Zhang Xuanxiang during the toxicological test of the submitted materials, so death due to poisoning by the above drugs can be excluded. 2. The ethanol content in the blood during the toxicological test of the submitted materials was 320.37 mg / 100 mL, indicating that the deceased Zhang Xuanxiang drank alcohol before his death. 3. During the autopsy, organ examination, and histopathological examination, subcutaneous hemorrhage in the right occipital scalp, fracture of the left temporal bone, fractures of the occipital top and the right side of the occipital bone, fracture of the skull base, local subarachnoid hemorrhage, and focal cerebral contusion were found; abrasions and lacerations on the head and face, fractures of the nasal bone, left zygomatic arch, upper and left mandible; abrasions on the chest and abdomen, multiple rib fractures on both sides, fractures of the sternum and thoracic vertebrae; laceration of the mediastinal soft tissue, hemothorax on both sides, contusion and bleeding of the lower lobes of both lungs; rupture of the pericardium; rupture of the left diaphragm, herniation of abdominal organs into the thoracic cavity; rupture of the stomach, rupture of the esophagus; rupture of the liver, rupture of the right kidney, hemorrhage around the adrenal gland and local mesentery; multiple fractures of the extremities accompanied by abrasions and lacerations; anemia of multiple organs such as the lungs, liver, and spleen. Combining the case situation for comprehensive analysis, Zhang Xuanxiang died due to multiple injuries to the cranium, chest and abdomen, and extremities caused by a traffic accident.

[0103] Analysis and Explanation: Based on the basic theories and knowledge in textbooks of various medical disciplines and forensic medicine, the cause of death was identified. (I) Injury Analysis 1. According to the forensic autopsy examination of the body system and clinical case records analysis, Huang Jiongzeng suffered injuries such as a scalp hematoma on the left frontal part, an acute subdural hematoma on the left frontotemporal parietal part, a contusion and focal hemorrhage in the left frontotemporal lobe of the brain, soft tissue contusions on the right hand and wrist, and a skin abrasion on the anterior side of the left knee due to a car accident on July 14, 2021; combined with the analysis of injury characteristics, the above injuries were caused by blunt external forces and had a direct causal relationship with this traffic accident. 2. According to the clinical case records, when the deceased was admitted to the hospital, he was conscious, with an apathetic expression and slow reaction. The GCS score was 14 points. The bilateral pupils were equal in size and round, with a diameter of about 3 mm, sensitive to light reflex. The muscle strength of the four limbs was grade V, the muscle tone was normal, the physiological reflexes were present, and the pathological reflexes were not elicited. Auxiliary examinations indicated that the subdural hematoma and cerebral hemorrhage of the deceased were gradually absorbed and recovered. Forensic examinations could also confirm this. The above indicated that the isolated effect of the deceased's craniocerebral injury was not sufficient to directly cause death. (II) Disease Analysis 1. Autopsy and histological examinations showed that the deceased had the following pathological...; Specifically, the result of the writing expression of the update chain agent in the web version application for updating the cause of death analysis and explanation is shown as follows: The final analysis and explanation obtained by AI: The main findings from the forensic examination of Jiao XX's body were: severe trauma to the head and hip, especially multiple fractures of the right hip and lacerations of the head. Although no obvious bleeding was seen on the head CT, the scalp laceration was accompanied by active bleeding, indicating that the head had been significantly impacted. Pathological examinations showed that Jiao XX suffered from chronic diseases such as bronchial asthma, emphysema, and aortic sclerosis, which exacerbated his body's stress response after the car accident, especially leading to difficulty in breathing. The toxicology test results showed that there were multiple drug components in Jiao XX's body, which might have been used during the treatment of his chronic diseases. Although these drugs themselves might not directly cause death, they might have had an adverse effect on his body under the stress state after the car accident. Through comprehensive analysis, Jiao XX's death was the result of the combined effect of the trauma caused by the car accident and his chronic diseases. The head and hip injuries caused by the car accident, combined with his chronic diseases and the stress response after the car accident, ultimately led to his death.

[0104] Specifically, the display of the final and only possible cause of death analysis and explanation generated after local fine-tuning the large language model agent in the web version application is as follows: Possible expert opinion: The cause of Jiao XX's death was multiple injuries caused by a traffic accident on the basis of suffering from chronic diseases such as bronchial asthma, emphysema, and aortic sclerosis.

[0105] Jiao XX died of multiple injuries caused by a car accident.

[0106] The death of Jiao XX was caused by the combined effect of the trauma resulting from a car accident and his chronic diseases.

[0107] The cause of Jiao XX's death was the combined result of head and hip injuries caused by the car accident, along with his chronic diseases and the stress reaction after the accident. Jiao XX had chronic diseases such as bronchial asthma and emphysema, and due to the traffic accident, he suffered multiple injuries to the head, hip, etc., leading to acute respiratory failure and death.

[0108] The cause of Jiao XX's death was consistent with head and hip trauma caused by the car accident, resulting in acute respiratory and circulatory failure and death.

[0109] Jiao XX died due to multiple injuries caused by a traffic accident on the basis of having chronic diseases such as bronchial asthma and emphysema.

[0110] Jiao XX died due to acute respiratory dysfunction caused by severe craniocerebral injury and multiple fractures of the right hip caused by the car accident.

[0111] The death of Jiao XX was the result of the combined effect of the trauma caused by the car accident and his chronic diseases.

[0112] The final expert opinion obtained by AI was: The cause of Jiao XX's death was the combined result of head and hip injuries caused by the car accident, along with his chronic diseases and the stress reaction after the accident.

Claims

1. A method for cause of death analysis of a forensic multimodal agent system for cause of death analysis, characterized in that: The specific steps are as follows: Step 1: Build a large language model agent architecture, including decision chain agent, routing chain agent, execution chain agent, reflection chain agent, integration chain agent, and update chain agent; Step 2: The decision chain agent decomposes the cause of death analysis task into multiple subtasks according to the background information of the cause of death analysis task, and generates an optimized subtask set through a prompt learning method; Step 3: The routing chain agent generates a judgment score based on the subtasks and background information in each optimized subtask set to determine whether an external tool library needs to be called; Step 4: The execution chain agent is used to generate the execution result of the current subtask; Step 5: The reflection chain agent analyzes the execution results of each subtask, generates defects and improvement suggestions for each subtask, and re-enters the defects and improvement suggestions into step 4 to generate updated execution results of the subtask; Step 6: The integration chain agent matches the updated execution results of each subtask with the case database, takes the case with the highest similarity as the reference master, and writes a preliminary cause of death analysis statement; Step 7, the update chain agent updates the written expression of the cause of death analysis statement through the chain thinking method; Step 8: Fine-tune the large language model agent locally to generate the final cause of death analysis description.

2. The method for cause of death analysis of a forensic multimodal intelligent agent system for cause of death analysis according to claim 1, characterized in that: The specific process of step 2 is: The background information D is input into the decision chain agent, and the decision chain agent understands the background information D and completes the task of cause of death analysis. Break it down into subtasks , and multiple subtasks Constitute a subtask set , based on the prompt word A, generate the optimal subtask combination under the current conditions , , and then generate the optimized subtask set based on prompt word B , .

3. The method for cause of death analysis of a forensic multimodal intelligent agent system for cause of death analysis according to claim 2, characterized in that: The specific process of step 3 is: the optimized subtask set obtained in step 2 Each subtask in The routing chain agent generates a judgment score ; like , indicating that the current information cannot complete the subtask , external tool library needs to be called; like , indicating that the current information is sufficient to complete the subtask , no need to call external tool libraries.

4. The method for cause of death analysis of a forensic multimodal intelligent agent system for cause of death analysis according to claim 3, characterized in that: External tool libraries include PubMed paper database, forensic pathology book database, and Internet search.

5. The method for cause of death analysis of a forensic multimodal intelligent agent system for cause of death analysis according to claim 1, characterized in that: The specific process of step 4 is: the execution chain agent generates the execution result of the current subtask based on the content of the current subtask, background information D, the execution result of the previous subtask, and the external knowledge called by the execution of the current subtask.

6. The method for cause of death analysis of a forensic multimodal intelligent agent system for cause of death analysis according to claim 1, characterized in that: The specific process of step 5 is as follows: the reflection chain agent includes the evaluation agent E and the reflection agent R; The evaluation agent E is used to determine whether the execution result of each subtask meets the requirements of the subtask and generate a reward score; The reflective agent R analyzes the difference between the execution results of each subtask and the requirements of the subtask based on the reward scores generated by the evaluation agent E, and generates defects and improvement suggestions; The obtained defect and improvement suggestion are used as a new task to re-execute step 4 to generate an updated execution result of the subtask.

7. The method for cause of death analysis of a forensic multimodal intelligent agent system for cause of death analysis according to claim 1, characterized in that: The specific process of step 6 is as follows: the integrated chain agent randomly aggregates the updated execution results of the subtasks and the execution results of the subtasks that meet the expected task goals, inputs each aggregated result into the case database, retrieves the historical case most similar to the current aggregated result through the RAG method, takes each most similar historical case as a reference parent, extracts the cause of death analysis method, logical reasoning path and related pathological conclusions of each reference parent, and learns each reference parent. After the learning is completed, the integrated chain agent combines the updated execution results of all subtasks of the current case and the execution results of the subtasks that meet the expected task goals to write a preliminary cause of death analysis description; The preliminary mortality analysis description includes the independent conclusions of each subtask and its logical consistency with the reference parent.

8. The method for cause of death analysis of a forensic multimodal intelligent agent system for cause of death analysis according to claim 1, characterized in that: The specific process of step 7 is as follows: the update chain agent extracts high-quality cause of death analysis reports from historical reports and authoritative forensic analysis cases as expression examples, and uses the key steps of the step-by-step reasoning process in the expression example as the learning examples of COT. By learning the learning examples of COT, the update chain agent can form a chain logic when generating cause of death analysis instructions, update the cause of death analysis instructions, and compare the updated cause of death analysis instructions with the expression examples to obtain expression deviations, and further optimize the COT reasoning chain based on the expression deviations until the updated cause of death analysis instructions have consistent reasoning processes and expression quality with the expression examples.

9. A forensic multimodal agent system for cause of death analysis, characterized by: include: A decision chain agent is used to decompose the cause of death analysis task into multiple subtasks according to the background information of the cause of death analysis task, and generate an optimized subtask set through a prompt learning method; The routing chain agent is used to generate a judgment score based on the subtasks and background information in each optimized subtask set to determine whether an external tool library needs to be called; Execution chain agent, used to generate the execution result of the current subtask; The reflection chain agent is used to analyze the execution results of each subtask, generate defects and improvement suggestions for each subtask, and re-input the defects and improvement suggestions into step 4 to generate updated execution results for each subtask; The integrated chain agent is used to match the updated execution results of each optimized subtask with the case database, taking the case with the highest similarity as the reference master to write a preliminary cause of death analysis statement; Update chain agent for updating the written representation of the cause of death analysis statement.

Images